A novel NLP-based method and algorithm to discover RNA-binding protein (RBP) motifs, contexts, binding preferences, and interactions.

A novel NLP-based method and algorithm to discover RNA-binding protein (RBP) motifs, contexts, binding preferences, and interactions - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice RNA . 2026 Jun;32(6):885–914. doi: 10.1261/rna.080892.125 A novel NLP-based method and algorithm to discover RNA-binding protein (RBP) motifs, contexts, binding preferences, and interactions Shaimae I Elhajjajy 1, ✉ , Zhiping Weng 1, ✉ Author information Article notes Copyright and License information 1 Department of Genomics and Computational Biology, University of Massachusetts Chan Medical School, Worcester, Massachusetts 01605, USA ✉ Corresponding authors: sielhajjajy@gmail.com ; zhipingweng@gmail.com Handling editor: Benjamin Blencowe ✉ Corresponding author. Received 2025 Dec 4; Accepted 2026 Feb 1. © 2026 Elhajjajy and Weng; Published by Cold Spring Harbor Laboratory Press for the RNA Society This article, published in RNA , is available under a Creative Commons License (Attribution-NonCommercial 4.0 International), as described at http://creativecommons.org/licenses/by-nc/4.0/ . PMC Copyright notice PMCID: PMC13182611  PMID: 41714110 Previous version available: This article is based on a previously available preprint posted on bioRxiv on January 22, 2025: " A novel NLP-based method and algorithm to discover RNA-binding protein (RBP) motifs, contexts, binding preferences, and interactions ". Abstract RNA-binding proteins (RBPs) are essential modulators in the regulation of mRNA processing. The binding patterns, interactions, and functions of most RBPs are not well-characterized. Previous studies have shown that motif context is an important contributor to RBP binding specificity, but its precise role remains unclear. Despite recent computational advances to predict RBP binding, existing methods are challenging to interpret and largely lack a categorical focus on RBP motif contexts and RBP–RBP interactions. There remains a need for interpretable predictive models to disambiguate the contextual determinants of RBP binding specificity in vivo. Here, we present a novel and comprehensive pipeline to address these knowledge gaps. We devise a natural language processing–based method to deconstruct sequences into entities comprising a target k -mer and its flanking regions, then use this representation to formulate RBP binding prediction as a weakly supervised multiple instance learning problem. To interpret our predictions, we introduce a deterministic motif discovery algorithm to leverage our data structure, recapitulating the established motifs of numerous RBPs as validation. Importantly, we characterize the binding motifs and binding contexts for 71 RBPs in HepG2 and 74 RBPs in K562, with many of them being novel. Finally, through feature integration, transitive inference, and a new cross-prediction approach, we propose novel cooperative and competitive RBP–RBP interaction partners and hypothesize their potential regulatory functions. In summary, we present a complete framework for investigating the contextual determinants of specific RBP binding, and we demonstrate the significance of our findings in delineating RBP binding patterns, interactions, and functions. Keywords: RNA-binding proteins, RBP interactions, NLP INTRODUCTION RNA-binding proteins (RBPs) extensively govern the biogenesis and maturation of nascent messenger RNAs (mRNAs) in co- and post-transcriptional manners ( Bentley 2014 ; Gerstberger et al. 2014 ). The human genome encodes ∼1500 RBPs, each of which has diverse roles in regulating nearly every step of mRNA processing, including transcription, canonical and alternative splicing, 3′ end cleavage and polyadenylation, stability, localization and transport, degradation, and translation ( Gerstberger et al. 2014 ). RBPs accomplish these tasks by binding their mRNA targets to form messenger ribonucleoprotein (mRNP) complexes whose constituents are dynamically remodeled based on the functional requirements at each stage of mRNA processing ( Müller-McNicoll and Neugebauer 2013 ; Gerstberger et al. 2014 ; Gehring et al. 2017 ). Thus, the activities and functions of RBPs are major determinants of the homeostatic expression and heterogeneity of the steady-state transcriptome. The dysregulation of RBPs can manifest into a variety of Mendelian ( Castello et al. 2013 ) and somatic genetic diseases ( Gebauer et al. 2021 ), including, most notably, neurological disorders ( Nussbacher et al. 2015 ; Conlon and Manley 2017 ) and cancers ( Pereira et al. 2017 ). Because the majority of RBPs catalogued to-date are not well-characterized, an improved understanding of RBP binding patterns, activities, functions, and interactions is needed to delineate their constitutive roles in mediating mRNA metabolism and to discern their causal roles in disease. RBPs interact with their target mRNAs through RNA-binding domains (RBDs), which confer upon the RBPs their sequence and structural binding preferences ( Müller-McNicoll and Neugebauer 2013 ; Gerstberger et al. 2014 ; Gehring et al. 2017 ). The complex and enigmatic binding patterns of RBPs can be attributed in part to their structural composition: They can possess single or multiple RBDs that belong to the same or different domain classes ( Gerstberger et al. 2014 ), and these RBDs may be either well-characterized or noncanonical ( Ray et al. 2023 ), together enabling a sizable potential interactome. Due to their short motif length (between 3 and 8 nt) and low nucleotide complexity (often containing only a subset of the four canonical nucleotides), RBPs with structurally diverse RBDs can recognize highly similar motifs, but they still assert a remarkable degree of transcriptomic binding specificity ( Dominguez et al. 2018 ). Based on large-scale in vitro experiments, previous studies have posited that the flanking sequences of RBP binding sites may play a critical role in enabling RBP binding specificity by fostering conformations conducive to stable binding; however, these findings were based on short, synthetic oligonucleotide sequences in isolated, exogenous conditions lacking a standard cellular environment ( Dominguez et al. 2018 ). Further investigation is necessary to corroborate the contribution of these flanking regions to RBP binding specificity in vivo. Adding another layer of complexity is the high propensity for RBPs to interact with one another, working in tandem to jointly regulate common targets ( Dassi 2017 ). To account for the vastness of the transcriptome and the diversity of associated regulatory functions, RBPs are further believed to interact combinatorially, forming complexes of varying sizes and composed of varying components ( Khoroshkin et al. 2024 ); this is yet another attribute that may enable RBPs to achieve their observed degree of binding specificity ( Lang et al. 2021 ; Street et al. 2024 ). These RBP–RBP interactions may occur via direct (i.e., requiring RNA binding and physical interaction between RBPs) or indirect (i.e., requiring RNA binding but not physical interaction between RBPs) binding modes ( Street et al. 2024 ), and they may operate in cooperative, competitive, or mutual manners ( Dassi 2017 ). Such interactions, which may involve as few as two RBPs or as many as required for large macromolecular complexes like the spliceosome, are vital to the fidelity of mRNA regulatory processes ( Lang et al. 2021 ). Previous work has provided experimental evidence of an extensive RBP–RBP interactive network ( Dassi 2017 ; Quattrone and Dassi 2019 ; Lang et al. 2021 ; Street et al. 2024 ), but the biological mechanisms for the majority of these interactions have not yet been determined. In short, the exponentially large landscape and functional implications of the RBP–RBP interactome remain underexplored ( Khoroshkin et al. 2024 ). Machine learning and deep learning have become increasingly prevalent in genomics applications, especially for the prediction of RBP binding ( Alipanahi et al. 2015 ; Ghanbari and Ohler 2020 ; Sharma et al. 2021 ; Sun et al. 2021 ). While such applications at large have contributed to our current understanding of RBP binding, there remain significant areas that have been under- or never before investigated, such as (1) the structure, composition, and role of motifs to RBP binding and interactions, (2) the salient features within the regions flanking RBP binding motifs, (3) the in vivo preferences of RBP motifs and flanking regions, (4) the role of flanking regions in facilitating RBP binding and interactions, and (5) the similarities in flanking regions across RBPs. Furthermore, current motif discovery methods rely on enumerative or probabilistic approaches (D'haeseleer 2006a) but do not consider the contributions of k -mers in the surrounding flanking regions when constructing motifs, which may cause noise or inaccuracies to be introduced into the discovered motif. Therefore, we elected to focus on two areas of great import to mRNA regulation: the contribution of motifs and their flanking regions to RBP binding and RBP–RBP interactions, as well as the combinatorial capacity of RBP–RBP interactions. We defined domain requirements to guide our formulation and construction of a comprehensive pipeline to address these RBP knowledge gaps. Namely, we aimed to develop an architecture-independent solution—to maximize flexibility and generalizability—that balances the performance-interpretability tradeoff intrinsic to neural networks, whose “black box” nature makes their decision-making rationales ambiguous to the user ( Hwang et al. 2024 ). We strove to design an interpretable strategy that achieves high accuracy in RBP binding prediction and motif discovery, both of which are integral to the investigation of motif flanking regions and RBP–RBP interactions. As opposed to commonly practiced top-down approaches that search for significant regions within a full-length, intact sequence, we instead adopt a bottom-up approach in which our goal is to deconstruct a sequence into its constituent k -mers and regions, discover the role and importance of each constituent, and then reconstruct the sequence based on the significant elements that were discovered. Given that RBPs only bind motif instances that are overrepresented and that are located in a favorable environment with significant flanking regions, we sought to understand the composition, structure, and contribution of these flanking regions to sites of RBP binding events. To this end, we required a method by which to identify the overrepresented k -mers that constitute the motif and its proximal flanking regions, the similarities between k -mer instances of the binding motif, the relationships between k -mers comprising the motif and k -mers comprising the motif's flanking regions, and the relationships between k -mers within the flanking regions themselves. The genomics and computational biology fields have speculated that DNA and RNA possess an underlying language and linguistic structure, whose components and inner workings remain elusive ( Durbin et al. 1998 ). In our endeavor to understand the k -mers contributing to RBP binding, we regarded RNA as possessing features and structures akin to those found in natural languages and were thus motivated to model RNA using concepts from linguistics and natural language processing (NLP). While natural language has a defined lexicon, syntax, and semantics, as well as established delimiters which enable text parsing, analogous structures and representations of genomic sequences are either missing, difficult to formalize, or not yet discovered. For our work, we decided to model sentences as sequences, phrases as significant regions, and words as k -mers. Furthermore, a crucial similitude between natural language and genomic language is that context confers meaning, which makes NLP particularly apt for this study. Here, we present a novel pipeline that integrates state-of-the-art concepts from NLP, weak supervision, and deep learning, as well as a novel motif discovery algorithm, to bridge the knowledge gaps pertaining to RBP binding patterns. To this end, we devised an NLP-based sequence decomposition strategy that establishes a new representation for candidate RBP binding sites and their flanking regions, which we then used to train neural networks to predict RBP binding under a multiple instance learning (MIL) framework. We also developed a consensus-based motif discovery algorithm to interpret our predictions, demonstrating its ability to accurately recapitulate the motifs of several well-characterized RBPs. Consequently, and importantly, we systematically identified and profiled the flanking regions and binding preferences for 71 RBPs in the HepG2 cell line and 74 RBPs in the K562 cell line, the majority of which are novel, and we conducted a thorough analysis of the similarities and differences in binding patterns between RBPs. Finally, we integrated our discovered motifs and flanking regions with several analysis metrics and a new cross-prediction approach to propose novel cooperative and competitive RBP–RBP interactions, for which we additionally hypothesize underlying mechanisms and functions. Taken together, we present a comprehensive computational approach that enables characterization of the features contributing to RBP binding specificity, as well as investigation into the participants and functions of RBP interactions. RESULTS A novel computational framework to model RBP binding We drew on theoretical concepts from linguistics and NLP to design a novel computational framework to discover the motifs and sequence contexts of RBP binding sites. We utilized merged replicate enhanced cross-linking and immunoprecipitation (eCLIP) sequencing data ( Supplemental Table S1 ; Van Nostrand et al. 2016 ) in the HepG2 and K562 cell lines from the ENCODE consortium due to its coverage of a large array of RBPs, its standardized, uniformly processed nature, and its use of input controls to ensure high-confident, high-quality peaks ( Van Nostrand et al. 2020a ). To construct our training data set, we defined positive sequences as centered eCLIP peaks resized to a uniform length of 101 bp; for our negative set, we performed random sampling of sequences with equivalent length from matched genomic regions ( Supplemental Fig. S1A ). We first introduced a linguistics-inspired method of sequence decomposition to deconstruct each sequence in our data set into a series of regions called “contexts,” each of which consists of a central target k -mer and its flanking subsequences ( Supplemental Fig. S1B,C ; Supplemental Note S1 ). We used this sequence representation to formulate the RBP binding prediction task as a multiple instance learning (MIL) learning problem ( Dietterich et al. 1997 ; Herrera et al. 2016 ; Carbonneau et al. 2018 ), in which we modeled each sequence in our data set as a “bag” and the sequence's constituent contexts as the bag's “instances” ( Supplemental Fig. S1A ; Supplemental Note S1 ). We then trained weakly supervised neural networks to perform (1) bag-level classification, to predict whether a sequence is bound or unbound, and (2) instance-level classification, to predict which context(s) within a sequence are responsible for the sequence's label ( Supplemental Fig. S1D ; Supplemental Note S1 ). To remove labeling noise that is inherent to MIL, we also introduced a strategy called iterative relabeling, which helps to filter out “weak” contexts that have little to no contribution to an RBP binding event while simultaneously retaining “strong” contexts that have the most significant contribution to the binding event ( Supplemental Note S1 ). To evaluate how well our models can successfully distinguish between true binding sites and the matched random control sequences, we utilized the area under the precision-recall (AUPR) curve, a performance metric that assesses how well a classifier can achieve good precision (i.e., the positive predictive value, or the proportion of predicted positives that are in fact true positives) and good recall (i.e., the true positive rate, or the proportion of true positives that are correctly predicted positive) ( Boyd et al. 2013 ); a perfect classifier would achieve an AUPR of 1.0. Looking at the distribution of AUPR values across an array of RBPs in each cell line, our models generally performed well, with a final median bag AUPR of 0.865 in HepG2 ( Fig. 1 A; Supplemental Table S2A,B ) and 0.860 in K562 ( Fig. 1 B; Supplemental Table S2C,D ). The consistency of performance across both cell lines demonstrates the robustness of our method. Notably, the bag performance was largely preserved before and after relabeling, indicating that our trained models could successfully achieve the removal of labeling noise from positive sequences while simultaneously maintaining bag-level prediction accuracy ( Supplemental Note S1 ; Supplemental Table S3 ). FIGURE 1. Open in a new tab A novel computational framework and algorithm for RBP binding motif and context discovery. ( A ) Distribution of instance-level and bag-level model performances (in AUPR) across 71 RBPs in HepG2, before and after iterative relabeling. ( B ) Distribution of instance-level and bag-level model performances (in AUPR) across 74 RBPs in K562, before and after iterative relabeling. ( C ) Exemplification of motif discovery validation, with respect to the literature, for four representative RBPs in HepG2 with motifs of diverse nucleotide content. Literature motifs were obtained from Ray et al. (2013) , Dominguez et al. (2018) , and Van Nostrand et al. (2020a) . For each RBP, one in vitro and one in vivo motif logo is displayed, with the exception of QKI, for which only in vitro logos are shown; however, QKI's motif has been well-characterized in vivo by multiple studies, such as in Hall et al. (2013) . ( D ) Unique instances comprising the RBFOX2 canonical GCAUG motif; six motif instances recapitulate known primary and secondary k -mers bound by RBFOX2 in vitro ( Begg et al. 2020 ). ( E ) Context logo for RBFOX2; the target motif GCAUG is preceded by a U and enriched in a G-rich sequence context, as previously demonstrated in the literature ( Dominguez et al. 2018 ). ( F ) Nucleotide preference of RBFOX2 binding, visualized as a line plot. ( G ) Nested violin-boxplots of the distribution of log 2 (eCLIP read enrichment over size-matched input [SMInput]) for discovered motifs ( left ) and contexts ( right ), relative to matched controls. Wilcoxon rank-sum test P -values test whether the motifs or contexts have significantly higher eCLIP enrichment than their corresponding control regions. A novel motif discovery algorithm for accurate discovery of RBP binding motifs and contexts To derive RBP motifs and, importantly, RBP binding contexts directly from the MIL predictions, we required a method of interpretation for our models that could accommodate the context data structure. To this end, we developed a deterministic, consensus-based motif discovery algorithm that relies on empirically derived measurements of three theoretically significant k -mer properties to discover motifs ( Supplemental Note S2 ; Supplemental Fig. S2 ). As a proof-of-principle, we first inspected RBFOX2 (HepG2), a well-characterized RBP with established regulatory functions in mRNA splicing and mRNA stability that is implicated in neurological developmental processes and disease ( Begg et al. 2020 ). Our algorithm was able to reproduce RBFOX2's canonical GCAUG motif, as supported by motifs derived from both in vitro and in vivo assays ( Fig. 1 C, bottom row; Van Nostrand et al. 2020a ). Previous studies have also demonstrated that approximately half of RBFOX2's binding sites contain alternative, non-GCAUG motifs, which have since been identified in vitro after successive rounds of iterative enrichment calculations from a synthetic RNA library modeled solely after 3′UTRs ( Begg et al. 2020 ). Crucially, our discovered motif contains six of the eight k -mers identified in the in vitro study: two primary (GCAUG and GCACG) and four secondary (GCUUG, GUAUG, GUGUG, and GUUUG) ( Fig. 1 D). To the best of our knowledge, this marks the first documentation of an in vivo–derived motif that successfully captures the majority of RBFOX2's alternative lower-affinity k -mers. Our RBFOX2 motif also contained four additional k -mers that have not been previously identified in Begg et al. (2020) . GACUG and GCGUG maintain the critical G1, U4, and G5 positions of the canonical motif; although these two k- mers were not included in the set of GHNUG k -mers characterized in Begg et al. (2020) , they can still be denoted by this IUPAC nucleotide representation ( Cornish-Bowden 1985 ). In conjunction with their relatively frequent occurrence in our discovered motif, especially in comparison to some of the established primary and secondary k -mers ( Supplemental Table S4 ), these properties suggest that GACUG and GCGUG may serve as novel candidates for lower-affinity RBFOX2 secondary k -mers; in fact, GACUG has been previously reported in another in vitro study ( Zhou et al. 2007 ). The remaining two k -mers, GGUAG and GUUAG, which do not share the GHNUG representation and which occur less frequently, may also present novel, lower-affinity secondary motifs with a distinct nucleotide content and lower similarity to the GCAUG consensus, or they may be attributed to noise. We reason to propose that, because they satisfy the same conditions and constraints as the known established motif instances, it is probable that these k -mers may be novel motif instances, and we further comment on this later in our work. Our context logo for RBFOX2 revealed a strong U directly preceding the GCAUG motif, a finding that coincides with its well-profiled canonical 6 -mer motif, (U)GCAUG ( Begg et al. 2020 ). The nucleotide preference of RBFOX2's motif context also corroborated in vitro findings ( Dominguez et al. 2018 ), with RBFOX2 demonstrating a clear predisposition for binding in G-rich regions ( Fig. 1 E,F). To further validate our discovered motifs and contexts for RBFOX2, we performed a comparative analysis of eCLIP read enrichments relative to matched control regions (Materials and Methods). We found that motif instances located within the discovered contexts were very significantly and more strongly enriched in eCLIP read coverage compared to their matched controls containing motif instances not located within the proper sequence context ( Fig. 1 G), thus providing further experimental support for our discovered motifs and contexts for RBFOX2 in HepG2. Robust validation reveals consistency of motif and context discovery In total, we report the discovered motifs, contexts, and binding preferences for 71 RBPs in the HepG2 cell line and 74 RBPs in the K562 cell line ( Supplemental Tables S5, S6 ). For all RBPs in our data set, we further demonstrate that our discovered motifs and contexts are significantly more enriched in eCLIP read coverage compared to their corresponding matched random control regions (Materials and Methods; Supplemental Table S7 ), which provides additional support that our method discovers true, high-confidence binding sites located within meaningful sequence contexts that exhibit preferred nucleotide preferences. To quantitatively validate the algorithm, we demonstrate our ability to recapitulate known motifs from the literature for a series of diverse RBPs. To curate this ground-truth set, we compiled known motifs for 14 well-characterized RBPs in HepG2 and K562 independently (Materials and Methods). While the size of these validation sets was the same for both cell lines, two RBPs from HepG2's validation set (PABPN1, PCBP2) were unavailable in K562, and were thus replaced with two other RBPs (PUM2, TARDBP). Our algorithm correctly discovered the motifs for 13 of 14 RBPs in the respective validation sets, yielding a consistent accuracy of ∼92.9% in both cell lines ( Supplemental Table S8 ). We highlight representative examples of discovered motifs from HepG2 in Figure 1 C. The sole RBP from the HepG2 ground-truth set for which our discovered motif did not correspond with that from the literature was PABPN1, a nuclear poly(A) binding protein ( Huang et al. 2023 ) for which we discovered a motif consensus of GCUGG and for which the poly(A) 5 -mer was not enriched in our data set ( Supplemental Note S3 ; Supplemental Fig. S3 ; Supplemental Table S9A,B ). Similarly, the sole RBP from the K562 ground-truth set for which our discovered motif did not correspond with that from the literature was RBFOX2, and we demonstrate that the canonical GCAUG is quite lowly enriched in the data set, which helps to explain why our algorithm did not rank the GCAUG motif as the primary motif ( Supplemental Note S3 ; Supplemental Fig. S4 ; Supplemental Table S9C–F ). For a more widespread comparison of context discovery, we compared our discovered contexts with in vitro–derived contexts for eight RBPs that were present in our collection and deemed to have significant nucleotide preferences in Dominguez et al. (2018) , showing that we could recapitulate the putative nucleotide preferences for five of them ( Supplemental Table S10A ); for the remaining three, we closely examined the contexts to explain any discrepancies observed in comparison to the in vitro study. For HNRNPL, we discovered a symmetrically oscillating CA-rich context, while an average preference for Cs was reported in vitro; however, the in vitro context also had a relatively strong underlying presence of As—especially at the positions adjacent to the motif—and displayed a depletion of Gs similar to that observed in our context. These differences were observed despite the fact that our discovered HNRNPL motif consensus (ACACA) was concordant with the in vitro–derived motif. Second, while our discovered context for IGF2BP1 did not exhibit a strong preference for any specific nucleotide, it did demonstrate a slight, albeit asymmetric, preference for Cs; on the other hand, the in vitro context was strongly C-rich on average. Finally, SFPQ's discovered context was strongly G-rich, while its in vitro average nucleotide preference was reported to be C-rich; interestingly, the more detailed in vitro context for SFPQ showed that the context was actually C-rich on the left-hand side of the motif but G-rich on the right-hand side of the motif. The discordant context preferences for these three RBPs could be attributed to a number of reasons, including possible limitations of our approach, although we support the following explanations: (1) an averaged representation of in vitro nucleotide preference may obscure position-specific preferences, and (2) in vitro experiments are inherently limited in their ability to simulate representative RNA sequences, native contexts, endogenous trans -factors, and cellular signals. However, we believe that the strong accuracy of our motif discovery algorithm, as demonstrated, provides support for the discovered motifs and contexts for these RBPs based on the available in vivo eCLIP data, especially given that (1) accurate motif discovery suggests that their corresponding contexts are valid, (2) accurate motif discovery suggests that the k -mer properties used to derive motifs are valid, and (3) two of those properties—namely, k -mer enrichment and k -mer co-occurrence—use information from the context of the motif instances to construct the motif itself. To further demonstrate the robustness of our framework, we applied our method to RBP binding data sets from other CLIP-based assays, namely, iCLIP ( König et al. 2010 ), HITS-CLIP ( Licatalosi et al. 2008 ), and PAR-CLIP ( Hafner et al. 2010 ; Supplemental Fig. S5 ). For RBPs that were present in our eCLIP data set, we demonstrate in vivo motif consistency with other CLIP assays as additional validation of the high accuracy of our model. Indeed, the discovered motifs for RBFOX2 (GCAUG), QKI (CUAAC), and SRSF1 (GAGGA) corresponded with those from our eCLIP data sets ( Supplemental Fig. S5A,B ). Interestingly, we noticed slight differences in the contexts of these motifs, and nucleotide preferences were generally less striking. For example, for RBFOX2 iCLIP, while a strong U preceded the canonical GCAUG motif as expected, the discovered context was not G-rich as in our HepG2 results ( Supplemental Fig. S5A , left). Similarly, for QKI HITS-CLIP, while an A preceded the canonical CUAAC as expected, the discovered context was not as strongly AU-rich as in our HepG2 and K562 results ( Supplemental Fig. S5B , left). We attribute these observations to differences in the CLIP variants: While eCLIP ensures that peaks are highly enriched over a size-matched input, which serves as a filtering mechanism to retain only high-confidence binding sites that are most likely to be functional, other CLIP assays may use less stringent filtering that might result in the capture of sequence-specific but transient interactions that are not necessarily functional ( Ramanathan et al. 2019 ), thus influencing the observed nucleotide preference. Application of our framework to other CLIP-based assays also exemplifies the broad generalizability of our method, which allows for the discovery of RBP binding motifs from more diverse data sources than solely eCLIP. To this end, we also demonstrated our discovery of motifs for RBPs that were not present in our eCLIP data set. Three of these RBPs had U-rich motifs: LSM6 is known to bind oligo(U)s ( Supplemental Fig. S5A , middle; Achsel et al. 1999 , 2001 ), EIF4G1 has been shown to bind oligo(U)s in yeast ( Supplemental Fig. S5A , right; Zinshteyn et al. 2017 ), and ELAVL1 recognizes U-rich motifs in in vitro assays ( Supplemental Fig. S5C , left; Ray et al. 2013 ). HNRNPA1's motif contains the core AG(G) motif ( Supplemental Fig. S5C , middle; Abdul-Manan and Williams 1996 ; Beusch et al. 2017 ), while IGF2BP1 contains AC-repeats present in previously published motifs ( Supplemental Fig. S5C , right; Ray et al. 2013 ; Van Nostrand et al. 2020a ). Lastly, we present a novel GC-rich motif for MSI2, for which previous studies typically report UAG but that was not enriched in our data set ( Supplemental Fig. S5B , right; Supplemental Table S9G ; Duggimpudi et al. 2018 ). We further note that these CLIP data sets originated from cell types other than HepG2 and K562, which demonstrates that our framework is also generalizable to diverse cellular contexts. Finally, as a negative control, we used our algorithm to perform motif discovery on a set of random genomic and random synthetic sequences (Materials and Methods; Supplemental Note S4 ). Despite the fact that both our algorithm and STREME ( Bailey 2021 ) returned motif hits for this set of negative controls ( Supplemental Figs. S6, S7 ), we noted several distinctive differences in sequence composition and feature constitution that clearly distinguish the motifs and contexts from RBP data sets from those discovered from random sequences. First, MIL binary classification on the random sequences performed only as well as a random classifier (i.e., an AUPR of ∼0.5 when 50% of observations in a data set are positive), indicating that these models are unable to differentiate between the foreground and background distributions present within the random data set. This performance was significantly worse than the performance for both a subset of sequence-specific RBPs and a set of non-sequence-specific RBPs ( Supplemental Fig. S8A,B ). We also found that (1) the enrichment values for the motif consensuses of random motifs were significantly lower than those from sequence-specific and non-sequence-specific RBP motifs ( Supplemental Fig. S8C ) and (2) the P -values of motif consensuses from random motifs were higher than those from sequence-specific and non-sequence-specific RBP motifs ( Supplemental Fig. S8D ). We observed similar findings for the enrichment and P -values of all k -mers in each of these data sets, with the random data sets also displaying much fewer numbers of significantly enriched k -mers overall ( Supplemental Fig. S9 ). Across all of these features, sequence-specific and non-sequence-specific RBPs displayed comparable properties with one another. Taken together, these analyses demonstrate that it is the properties of the discovered motifs and contexts, rather than the presence or absence of a hit by a motif discovery algorithm, that are informative of a good negative control ( Supplemental Note S4 ). Further, these analyses demonstrate that our algorithm discovers meaningful RBP motifs and contexts that are clearly distinct from those discovered from random genomic and random synthetic sequences. We find that RBP data sets contain k -mers with enrichments greatly exceeding 1 and P -values considerably less than 0.05, whereas random sequences contain k -mers with enrichments and P -values bordering 1 and 0.05, respectively ( Supplemental Note S4 ); these properties are highly informative and can be leveraged to determine whether a given set of sequences is more likely to originate from an RBP or a random data set. This is of particular importance for interpreting the results of non-sequence-specific RBPs, which are typically known to bind to the RNA phosphate backbone rather than to nucleotide bases. We hypothesize that the discovered motifs and contexts for non-sequence-specific RBPs, which display preferences distinct from random negative controls, may be reflective of their frequent association with other sequence-specific RBPs, thus causing the CLIP cross-linking and pulldown to capture regions of RNA that demonstrate the sequence preferences of the interacting partner; we propose such a case later in this work. Importantly, these analyses definitively demonstrate that non-sequence-specific RBPs do not constitute a valid negative control for motif discovery, as their motifs and contexts are significantly different from those discovered from random sequences; thus, although these RBPs are classified as non-sequence-specific, the sequences associated with their binding sites are nonrandom. Motif and context discovery across cell lines reveals similarities and differences in binding preferences We sought to comprehensively characterize the discovered binding contexts for the set of RBPs covered in our study. In the HepG2 cell line, we discovered novel in vivo contexts for a total of 71 RBPs. For the contexts of the five RBPs for which we corroborated with in vitro data ( Supplemental Table S10A ), we provide additional detail by reporting a larger context region, but also by delineating nucleotide preferences for each component of the context (below). For other RBPs, such as those known to bind polypyrimidine tracts (e.g., PTBP1, U2AF2) ( Sayers et al. 2022 ), we recapitulated the nucleotide preferences of their contexts, but also provided a more detailed report of these preferences per position. For the remaining RBPs, which comprise the majority of the set, we present for the first time, to the best of our knowledge, a comprehensive profile of their preferred contexts for binding in vivo. First, we comprehensively determined the nucleotide preferences for each RBP's motif and context ( Supplemental Table S10B ). Notably, we observed a strong degree of symmetry for the majority of our discovered contexts ( N = 59; 83.1%), where we define symmetry to occur if the left and right contexts flanking the motif share the same nucleotide preference (Materials and Methods). We noticed that the symmetric preference was for either a single nucleotide or for two frequently occurring nucleotides. The remaining 12 RBPs had contexts that were asymmetric in nucleotide preference, which displayed a bit more variance. In the case of ILF3, for example, each side of the context displayed preference for a single, albeit different, nucleotide, where its left context showed preference for C but its right for G. In other cases, each flank of the context displayed preference for 2 nucleotides of different identity (e.g., IGF2BP1), or the left and right contexts each displayed preference for a different number of nucleotides (e.g., PABPN1). By visualizing the position probability matrices (PPMs), we noted that some RBPs had stronger, more pronounced nucleotide preferences than other RBPs. Furthermore, we observed that the majority of RBPs ( N = 46; 64.8%) had preference for the same nucleotide in both their motif and context (Materials and Methods; Supplemental Table S10B ); for example, G3BP1 is AG-rich in both its motif and contexts. For such RBPs in which the context's nucleotide preferences appear to be extensions of the motif's nucleotide preferences, we also considered that this pattern may represent a known phenomenon in which multiple motif sites occur together to form clusters ( Supplemental Note S5 ; Supplemental Fig. S10 ). While the remaining 25 RBPs had different nucleotide preferences in their motif and contexts, the difference in preference was generally not very striking. For example, while HLTF's motif is AG-rich, its contexts are overall A-rich, indicating that there is still some degree of similarity between the nucleotide preference of each element. One exception to this phenomenon is QKI, which demonstrates a marked difference in the nucleotide preferences of its elements, having an AC-rich motif but a U-rich context ( Supplemental Tables S5, S10B ). We next performed the same analyses for the 74 RBPs in K562. Similarly to HepG2, the vast majority of RBPs ( N = 60; 81.1%) in K562 exhibited symmetric nucleotide preferences in their contexts, while only 14 (18.9%) had asymmetric contexts ( Supplemental Table S10C ). The majority of RBPs also had the same nucleotide preference in the motif and context ( N = 49; 66.2%) ( Supplemental Table S10C ). We next compared the discovered binding patterns for RBPs with data sets in both cell lines, for which there were 40 RBPs. Broadly, we found that 60% of these RBPs shared preferences for the same nucleotide in their motifs, while 67.5% of RBPs shared preferences for the same nucleotide in both the left context and right context, respectively, and 65% of RBPs shared the same preference across both sides of the context ( Fig. 2 A). An even greater majority, 82.5%, displayed the same pattern of (a)symmetry across cell lines, while 70% exhibited the same nucleotide preference in both the motif and context. FIGURE 2. Open in a new tab Motif and context nucleotide preferences in HepG2 and K562. ( A ) Comparison of nucleotide preferences between common RBPs in both cell lines. ( B ) Clustering of motif ( left ) and context ( right ) position probability matrices (PPMs) for common RBPs in both cell lines, with representative examples shown in the middle . We investigated the degree of similarity in the motifs and contexts between the two cell lines for a given RBP by assessing whether the RBP from both HepG2 and K562 were assigned to the same cluster upon hierarchical clustering of their PPMs ( Fig. 2 B). Only 45% of RBPs ( N = 18) from one cell line clustered together by motif PPM with their counterpart from the other cell line ( Fig. 2 B, left), while 57.5% ( N = 23) clustered together by context PPM ( Fig. 2 B, right), suggesting that both motifs and contexts may differ to a sizable degree in a cell type–specific manner. Some RBPs, such as HNRNPK and PTBP1, had strikingly similar motifs and contexts in both cell lines ( Fig. 2 B, middle). For some RBPs that did not cluster together with their counterpart in the other cell line, the differences were visually apparent, such as with MATR3, whose motifs differed strongly by a single nucleotide but whose contexts were otherwise similar; for others, the difference could be explained by the fact that one motif was a “shadow” of the other (e.g., with QKI, whose motif in HepG2 was identified as [A]CUAAC but whose motif in K562 was identified as ACUAA). On the other hand, more substantial discrepancies were present for some RBPs that had similar motifs but different contexts in the two cell lines (e.g., EXOSC5) or vastly different motifs and contexts between the two cell lines (e.g., UPF1, DDX55). While these analyses can be considered preliminary, this work demonstrates the importance of studying RBP binding patterns in a cell type–specific manner, and future work will be important to investigate the degree of cell type–specificity's effects on RBP motifs and contexts as well as their functions and implications in RNA regulatory mechanisms. A comprehensive characterization of the binding motifs and contexts for 71 RBPs in HepG2 For a closer, more detailed, and higher resolution inspection of RBP binding preferences, we next decided to focus on our results from the HepG2 cell line for the remainder of our work. To gain an improved understanding of our discovered motifs for the full set of 71 RBPs in HepG2, we classified each into one of five categories ( Supplemental Table S11 ): Consistently reported ( N = 14): The RBP has a well-characterized binding motif that has been consistently reported across multiple sources in the literature. Multiple reported ( N = 22): Multiple different motifs have previously been reported for the RBP; we either present a motif that supports at least one of these reported motifs or a novel motif that does not correspond with any of those previously reported. Weakly characterized ( N = 25): Limited information is available regarding the binding motif of the RBP (e.g., only one source available, only nucleotide preferences [but not a concrete motif] available); we present a more specific motif. Non-sequence-specific ( N = 6): The RBP is believed to recognize features other than sequence, such as RNA secondary structure. No existing characterization ( N = 4): To the best of our knowledge, no motif has been previously reported for the RBP, and we present a novel binding motif. We have addressed the “Consistently reported” classification in the preceding section. In our “Multiple reported” category, our discovered motif for a given RBP often corresponded with at least one of the reported motifs. For example, FUBP3, an RBP for which we discovered a motif with a UGUGU consensus, corroborated the UG-repeat motif discovered in an in vivo study ( Gau et al. 2011 ) but was distinct from the UA-rich motifs discovered in vitro ( Dominguez et al. 2018 ); this discrepancy could be attributed to technical reasons underlying the environments of the respective assays used to generate the binding data. On the other hand, discrepancies between multiple reported motifs could also be attributed to biological reasons based on the RBP's binding mode. For example, KHSRP, an RBP belonging to the same protein family as FUBP3 ( Gau et al. 2011 ) and for which we also discovered a UGUGU motif, has been shown to bind both GU-rich and AU-rich elements in vivo ( Cook et al. 2011 ; Briata et al. 2016 ). KHSRP has four KH RBDs, each having varying degrees of specificity, with the third KH domain in particular demonstrating a strong preference for G-containing sequences ( García-Mayoral et al. 2008 ; Palzer et al. 2022 ). We thus hypothesize that our UGUGU motif may reflect that KHSRP's utilization of KH-3 for RNA binding was more prevalent during eCLIP capture. Our KHSRP motif also exhibited a characteristic depletion in Cs that has previously been observed ( Supplemental Table S5 ; García-Mayoral et al. 2008 ). We also examined the additional motifs that were discovered by our algorithm but not ranked first by our scoring strategy, which we refer to collectively as secondary motifs. In the case of LIN28B, we discovered a GGAGA secondary motif that supported the GGAG motif discovered for LIN28B in a previous study of its transcriptomic targets ( Graf et al. 2013 ) as well as a GGAGA motif discovered for LIN28B's paralog, LIN28A, with which it shares a similar structure ( Wilbert et al. 2012 ; Maklad et al. 2023 ). This demonstrates that our method still has the power to discover previously reported motifs, even if not ranked first by our scoring method. For our “Weakly characterized” category, we propose new, specific motifs for 25 RBPs that we considered to be weakly characterized. For 10 of these RBPs, a previous study had reported their candidacy for binding to rG4s ( Lee et al. 2020 ; Busa et al. 2021 ; He et al. 2021 ), and our discovered motifs (and contexts)—which were all G-rich—supported this theory. For some RBPs that only had a reported motif from a single source, our discovered motifs were concordant, thus providing additional support; for example, PPIG and PRPF8, which both have GUGAG motifs ( Supplemental Table S5 ; Feng et al. 2019 ). For other RBPs, we propose novel motifs (e.g., GGGUG for BUD13, CACAC for CDC40, GAAGA for FXR2; Supplemental Table S11 ). Our “Non-sequence-specific” category contains a number of RBPs that have previously been hypothesized to bind in a non-sequence-specific manner. Many of these RBPs are DEAD-box helicases, which recognize the RNA sugar-phosphate backbone as opposed to its nucleotides ( Linder and Jankowsky 2011 ), while the other two have either been reported to have double-stranded RBDs (ILF3) ( Li et al. 2020 ) or to bind highly structured RNAs (XPO5) ( Wang et al. 2020 ). We nevertheless present novel binding motifs for these RBPs, hypothesizing that the motifs may instead reveal the sequence preferences of other trans -factors that interact with the RBP in question, as we elaborate on in the following sections. Finally, to the best of our knowledge, the motifs for four of the 71 total RBPs have not been previously characterized, comprising our “No existing characterization” category. This set includes EXOSC5, HLTF, RPS3, and ZNF800. Investigating the relationships of binding patterns between RBPs Equipped with our characterization of RBP binding preferences, we next sought to understand the similarities and differences in these preferences between RBPs. First, we used HDBSCAN ( McInnes et al. 2017 ) to cluster the flattened PPMs of the motifs and contexts, respectively, of all RBPs, then visualized the dimension-reduced clusters using a two-dimensional UMAP embedding ( Fig. 3 A,B; McInnes et al. 2018 ). Using HDBSCAN and UMAP with the Euclidean metric to compute distances between points—with the former technique seeking to assign clusters based on point proximity and the latter technique aiming to preserve the relative distances observed in high-dimensional space in its lower-dimensional mapping—we can use the results as a method of motif comparison, where a small Euclidean distance and same cluster membership suggest that two RBPs in question may possess similar motif and context patterns, as well as similar nucleotide binding preferences. FIGURE 3. Open in a new tab RBPs cluster based on similarities between their respective motifs and contexts. Clustering of motif ( A ) and context ( B ) PPMs with HDBSCAN and visualization with UMAP. Clusters are assigned number labels, with cluster colors roughly mapping between the two embeddings. ( C ) Motif and context logos for representative RBPs from selected clusters; cluster labels and nucleotide preferences underlying the clustering assignments from ( A ) and ( B ) for target ( left ) and context ( right ) motifs are labeled in the periphery. RBPs whose names are labeled in black are those that migrated clusters between the target and context clustering, with the cause for migration evident from the corresponding logos. Indeed, we observed that RBPs clustered together by nucleotide preference. Clustering by motif yielded nine clusters in total ( Fig. 3 A), with distinct groups forming for RBPs with generally UG-rich (Cluster 1), U-rich (Cluster 2), C-rich (Cluster 3), CG-rich (Cluster 4), and A-rich (Cluster 9) motifs, each with a moderate degree of variation ( Fig. 3 C). Cluster 6 contained three RBPs with relatively different motifs from one another but are characterized by their strong G1 and U5 positions. The last three clusters consisted of RBPs with largely G-rich motifs, but their division into subgroups can be attributed to the presence of a distinct C5 position (Cluster 8), strong G's at the odd-indexed or solely peripheral positions and U's at either or both of the even-indexed positions (Cluster 5), and the remaining, predominantly G-rich motifs (Cluster 7). The latter cluster is fairly large ( N = 17), with a considerable degree of variation present, but with less well-defined patterns, which may help to explain why it was not further divided into subgroups. In our context-based clustering results, we observed that the majority of RBPs remained clustered with the same members as in the motif-based clustering, with 16 of the 71 (22.5%) migrating to a new cluster ( Fig. 3 B, labeled in black). We sought to understand the underlying causes that fostered these migrations, and we did so by carefully inspecting the motif and context logos of the associated RBPs ( Fig. 3 C). For example, due to its strong CUUUC consensus, PTBP1's motif was distinct enough from the set of RBPs having U-rich motifs (Motif Cluster 2, orange) and instead clustered with the set of RBPs having C-rich motifs (Motif Cluster 3, green). However, upon the addition of context, PTBP1's overwhelmingly strong preference for U's in its binding context led to its subsequent clustering with the set of RBPs having U-rich contexts (Context Cluster 3, orange), as opposed to remaining in the same cluster that it previously belonged to, whose members ultimately have C-rich contexts (Context Cluster 5, green). A similar phenomenon was observed for CDC40, HNRNPL, and QKI, which, despite demonstrating a binding preference for C's and A's, have motifs with strong enough C's that they were also clustered with the set of RBPs having C-rich motifs (Motif Cluster 3, green). When contexts were incorporated, these three RBPs detached to form their own cluster (Context Cluster 4, blue). CDC40 and HNRNPL both had CA-rich contexts, although the latter had strong U undertones. Interestingly, in spite of its context preference for Us, QKI was still included in this cluster, likely due to its nucleotide preference similarities in motif with CDC40 and HNRNPL. This demonstrates a case in which the motif is more influential than the context in the clustering outcome. In short, the migration of certain RBPs between the motif and context clustering demonstrates that context plays an important role in differentiating between binding outcomes, particularly for RBPs sharing similar target motifs. We next sought to examine the similarities and differences between RBP binding patterns using a probabilistic approach, reasoning that the PPM representation of motifs and contexts fundamentally comprises a series of position-specific probability distributions across the four canonical nucleotides. To this end, we used the Kullback–Leibler divergence (KLD) to quantify the degree to which one probability distribution Q (i.e., the approximating distribution) can approximate another probability distribution P (i.e., the reference distribution), where smaller values of KLD represent a closer approximation ( Bishop and Nasrabadi 2006 ). Unlike the Euclidean distance metric, the KLD does not qualify as a bona fide metric in the formal, mathematical sense due to its asymmetric nature, and thus we instead refer to KLD as a measure of comparison. By computing the cumulative KLD across each position of a given pair of PPMs, we can obtain a quantitative method to compare PPMs between RBPs (Materials and Methods). For this reason, we used the KLD as a quantitative measure of the distributional similarities and differences between RBP motifs and contexts for all future analyses ( Supplemental Note S6 ; Supplemental Fig. S11 ). We also computed the cumulative KLD between RBPs in a pairwise manner, performed hierarchical clustering on the global KLD profiles across all RBPs, and annotated each branch in the resulting dendrogram based on the corresponding nucleotide preferences of the motif ( Fig. 4 A) and context ( Fig. 4 B). Groups of RBPs with mutually low KLDs are visible as patches of blue in the heat map. Notably, the hierarchy of the dendrogram allowed for ease of visualization and interpretation of the comparison between RBP binding patterns. For example, for motif KLDs ( Fig. 4 A), the first main branch of the dendrogram splits based on whether the motif is rich in G's or rich in C's, U's, or A's; because this is a major divide, and based on the number of RBPs in each of the two sets, we observe that (1) the majority of RBPs in this collection have predominantly G-rich motifs (albeit with some variation) and (2) the preference for G's is a major differentiator between the two main groups. On the other hand, when we performed the same analyses for the context KLDs, we observed deviations from the patterns observed in the motif KLD clusters, in that the nucleotide preferences at the context level did not always heavily influence the clustering of RBPs; some clusters exhibited a strong, consistent nucleotide preference, while other clusters did not. Such behavior may be attributed to (1) the increased length and complexity upon addition of the contexts, (2) the combined influences of both the motif and the contexts, and/or (3) the overall nucleotide preference of each RBP not being the decisive factor in clustering, as RBPs may have contexts with sufficient similarity to be assigned to the same cluster while also having different nucleotide preferences ( Fig. 4 B). FIGURE 4. Open in a new tab Clustering of cumulative Kullback–Leibler divergence (KLD) values. ( A ) Heat map and dendrogram of cumulative motif KLDs between RBP pairs. ( B ) Heat map and dendrogram of cumulative context KLDs between RBP pairs. Enrichment analyses in differential gene regulation demonstrate the functionality of discovered contexts To demonstrate the functionality of our discovered motifs and contexts, we leveraged the RBP knockdown (KD) shRNA-seq data sets from the ENCODE Consortium, for which data were available for 60 RBPs in our HepG2 data set ( Supplemental Table S1D ; Luo et al. 2020 ; Van Nostrand et al. 2020a ). We first used DESeq2 ( Love et al. 2014 ) to identify differentially expressed genes upon KD of each RBP and then computed the degree to which our discovered contexts were enriched in differentially expressed (DE) genes relative to a background expressed gene set (Materials and Methods). To closely examine the functional effects of each RBP's KD, we stratified these context enrichment calculations based on (1) the direction in which a DE gene was over- or underexpressed (i.e., up- or downregulated) and (2) the genomic region in which the RBP's context was located, with a specific focus on regions known to most strongly influence gene expression (i.e., 5′UTR, CDS, 3′UTR) (Materials and Methods). In total, we found that 45 of the 60 RBPs with KD shRNA-seq data (75%) had enriched contexts in DE genes in at least one set of conditions (i.e., in any group defined by a direction–region pair; Fig. 5 A). Furthermore, 23 RBPs demonstrated enriched contexts in upregulated genes, while 39 RBPs demonstrated enriched contexts in downregulated genes, suggesting that most of the RBPs in this data set positively regulate gene expression under normal conditions. Some RBPs had enriched contexts in genes that were differentially expressed in both directions, while other RBPs had enriched contexts in more than one genomic region of differentially expressed genes. FIGURE 5. Open in a new tab Significant enrichment of discovered contexts in RBP KD experiments reveals functional insights. ( A ) Volcano plot of RBP context enrichment in up- ( top row) and down- ( bottom row) regulated genes upon RBP KD, and stratified by whether the context enrichment is observed in 5′UTRs ( left column), CDSs ( middle column), and 3′UTRs ( right column). For each panel, the top 12 most significant points are labeled with the RBP name. ( B ) Schematic demonstrating the four regions of interest for differential splicing analyses. ( C ) Volcano plot of RBP context enrichment in up- ( top row) and down- ( bottom row) differentially spliced (DS) skipped exon (SE) events upon RBP KD, and stratified by the region of interest in which the context enrichment is observed. Upregulation of events indicates increased exon inclusion, while downregulation of events indicates increased exon skipping. Twenty-three of the 45 RBPs with enriched contexts in DE genes have previous annotation of regulatory roles in gene expression–related functions such as 3′end processing, RNA stability, or translation, thus representing 88.5% of the RBPs in our data set that had annotations for these functions ( Van Nostrand et al. 2020a ). For the remaining 22 RBPs, 11 have annotations for roles in splicing. The final 11 are primarily composed of RBPs that have been previously annotated as “Novel RBPs” ( Van Nostrand et al. 2020a ); therefore, based on their significant context enrichment in differentially expressed genes upon RBP KD, we hypothesize that these RBPs (DDX55, DDX59, FAM120A, FUBP3, GTF2F1, HLTF, NKRF, SUB1, UCHL5) have functions in the regulation of gene expression. We hypothesized that RBPs with enriched binding contexts located in specific genomic regions of up- or downregulated genes would be indicative of their regulatory functions ( Fig. 5 A). Because 3′UTRs are hotspots for binding by RBPs that regulate mRNA stability, we first examined context enrichments in 3′UTRs. UPF1, whose discovered contexts were primarily located in 3′UTRs ( Supplemental Fig. S12A ), demonstrated significantly enriched contexts in the 3′UTRs of target genes that were upregulated upon UPF1 KD (enrichment = 1.95; P -value = 4.71 × 10 −23 ; Fig. 5 A, top right), which supports its known function as an RNA helicase that plays important roles in diverse RNA decay pathways ( Kim and Maquat 2019 ). On the other hand, CSTF2T, a component of the cleavage stimulation factor (CSTF) complex with regulatory roles in 3′end processing and alternative polyadenylation (APA) ( Yao et al. 2013 ), demonstrated significant context enrichment in the 3′UTRs of target genes that were downregulated upon CSTF2T KD (enrichment = 7.93; P -value = 1.15 × 10 −63 ; Fig. 5 A, bottom right), suggesting that the absence of CSTF2T may lead to widespread alterations in 3′end processing that ultimately destabilize its RNA targets. CSTF2T's functions are generally not well-characterized, and its additional context enrichment observed in other genomic regions of its downregulated targets upon CSTF2T KD ( Fig. 5 A, bottom left and middle) may indicate additional regulatory roles, particularly given that the majority of its discovered contexts (∼78%) reside in proximal and distal introns ( Supplemental Fig. S12B ). As a final example in 3′UTRs, IGF2BP3 exhibits context enrichment in both upregulated (enrichment = 1.96, P -value = 0.00757) and downregulated (enrichment = 5.26, P -value = 9.52 × 10 −32 ) genes, which aligns with its known functions in regulating RNA degradation ( Liu et al. 2024 ) and RNA stability ( Huang et al. 2018 ), respectively ( Fig. 5 A, right; Supplemental Fig. S12C ). We next sought to validate our discovered motifs and contexts in relation to splicing regulation. We used rMATS-turbo ( Wang et al. 2024 ) to detect and quantify differential splicing events for each RBP, focusing primarily on the most prominent class of event—skipped exon (SE) (Materials and Methods; Supplemental Figs. S13, S14 ). Because splicing events occur in localized regions, we first defined four regions of interest (adapted from Yee et al. 2019 ; Fig. 5 B), then computed the degree to which our discovered contexts were enriched in differentially spliced (DS) SE events relative to a background set, stratified by direction of regulation and region of interest ( Fig. 5 C; Materials and Methods). In total, we found that 34 of the 60 RBPs with KD shRNA-seq data (56.7%) had enriched contexts in DS events in at least one set of conditions. Twenty-three RBPs had enriched contexts in upregulated events, while 18 RBPs had enriched contexts in downregulated events. Nineteen of the 34 RBPs (55.9%) with enriched contexts in DS events have previously annotated roles in splicing regulation and the spliceosome, thus representing 67.9% of the RBPs in our data set that had annotations for these functions ( Van Nostrand et al. 2020a ). For the nine RBPs in our data set with previous annotation of splicing but without enriched contexts in DS events, we hypothesize that some may have more predominant functions in other splicing event classes, such as BUD13, which is known to suppress intron retention ( Frankiw et al. 2019 ), while others have small data sets that ultimately impact the detection of context enrichment in the identified DS events, such as SRSF9, which has only ∼3000 sequences in its data set. For the remaining 15 RBPs of the 34 with enriched contexts in DS events, 11 also had enriched contexts in DE genes, while four did not; we hypothesize novel functions in splicing regulation for these RBPs. In total, across our differential expression and splicing analyses, we detected context enrichment for 52 of the 60 RBPs (86.7%) with KD shRNA-seq data. We observed direction- and region-specific context enrichment for RBPs that corroborate their known splicing functions. For example, RBFOX2 demonstrated significantly enriched contexts at the 5′ splice site (SS) of cassette exons downregulated upon RBFOX2 KD (enrichment = 4.85, P -value = 9.67 × 10 −21 ; Fig. 5 C). Furthermore, its motif consensus GCAUG was significantly enriched in the 5′SS of downregulated events for both event-centric (enrichment = 1.67, P -value = 1.77 × 10 −16 , Supplemental Fig. S15A ) and context-centric k -mer enrichments (enrichment = 1.87, P -value = 1.62 × 10 −4 , Supplemental Fig. S15B ; Materials and Methods). Indeed, RBFOX2 is known to promote exon inclusion when bound to sites proximally downstream from the cassette exon ( Yeo et al. 2009 ). RBFOX2 is also known to promote exon skipping when bound proximally upstream of the cassette exon, albeit with a weaker signal relative to its downstream binding sites ( Yeo et al. 2009 ). We observe similar findings, with RBFOX2's contexts being ∼2.3-fold more enriched in the cassette 5′SS of downregulated events than in the cassette 3′SS of upregulated events (enrichment = 2.13, P -value = 1.40 × 10 −2 , Fig. 5 C, top row, second column). Finally, we leveraged our KD analyses to study each of our proposed novel motif instances for RBFOX2 ( Supplemental Note S7 ). Notably, QKI demonstrated the most highly significantly enriched contexts in a position-specific manner. First, QKI had significantly enriched contexts at the 3′SS of upregulated cassette exons—signifying increased exon inclusion upon KD (enrichment = 9.18, P -value = 3.88 × 10 −44 ; Fig. 5 C)—and QKI's consensus CUAAC was significantly enriched in these 3′SSs by both event-centric (enrichment = 1.65, P -value = 7.81 × 10 −8 , Supplemental Fig. S16A ) and context-centric (enrichment = 2.28, P -value = 1.29 × 10 −4 , Supplemental Fig. S16B ) k -mer enrichment methods (Materials and Methods). Indeed, under normal conditions, QKI binds upstream of cassette exons to promote exon skipping ( Hall et al. 2013 ; Hinkle et al. 2022 ; Montañés-Agudo et al. 2023 ). Second, QKI had significantly enriched contexts at the 5′SS of downregulated cassette exons, signifying increased exon skipping upon KD (enrichment = 14.07, P -value = 1.62 × 10 −60 ; Fig. 5 C), and the QKI consensus CUAAC was enriched in these 5′SSs (event-centric: enrichment = 2.24, P -value = 7.65 × 10 −16 , Supplemental Fig. S16A ; context-centric: enrichment = 1.79, FDR = 1.94 × 10 −3 , Supplemental Fig. S16B ). These findings corroborate QKI's known functions of binding downstream from cassette exons to promote exon inclusion ( Hall et al. 2013 ; Hinkle et al. 2022 ; Montañés-Agudo et al. 2023 ). In closing, we demonstrate that our discovered contexts are not only enriched in regions implicated in splicing regulation, but also that the directional and positional context enrichment patterns are indicative of RBP functions in regulating expression and splicing, thus providing further validation for the results of our framework. Cross-prediction: a novel approach to investigate RBP binding specificity The determinants governing RBP binding specificity, particularly for those RBPs sharing similar motifs and contexts, remain an important outstanding question. RBPs recognizing similar sequence patterns must be able to differentiate between candidate sites in vivo in order to maintain the degree of specificity necessary for efficient and accurate mRNA regulation. We asked whether we could simulate the biological process of binding site selection in silico, especially to determine whether our models are capable of capturing the binding specificity exhibited by RBPs that recognize similar motifs. To this end, we developed and implemented a novel strategy that we refer to as cross-prediction, whereby the model trained on one RBP's contexts is presented with a distinct set of contexts belonging to another RBP and predicts the binding outcome for each context in that set ( Fig. 6 A). The cross-prediction performance, an aggregated measure across all contexts in the data set, serves as a proxy for how successfully an RBP, by way of its model, can distinguish between its cognate binding sites versus those belonging to another RBP. Because this approach implicitly compares the salient features learned by each RBP's model, it provides a measure of comparison between the binding patterns of different RBPs. FIGURE 6. Open in a new tab Cross-prediction as a novel approach to reveal similarities in global binding patterns between RBPs. ( A ) Schematic demonstrating the operations involved in the cross-prediction approach. ( B ) Clustered heat map of the pairwise cross-prediction AUPRs between all RBPs. Regions highlighted in D and E are labeled in the heat map. ( C ) Enlarged dendrogram from the clustered heat map in B , revealing hierarchical clustering of RBPs based on similarity of their respective cross-prediction profiles across all other RBPs. ( D , E ) Highlighted regions of high cross-prediction performance from the heat map in B ; the arrow symbol in the bottom left corner indicates that the inset was rotated 90° clockwise relative to the original heat map in B . We hypothesized that a strong cross-prediction performance between a given pair of RBPs may suggest that the RBPs share certain salient features that are significant to their respective binding activities, including similarity in motifs, contexts, nucleotide preferences, and targets; it may also signify that the motif of one RBP is strongly enriched in the contexts of another RBP's motif. At the global level, a high cross-prediction performance may indicate structural, behavioral, and functional similarity between the two involved RBPs. In cases where the cross-prediction between two RBPs is strongly asymmetric, it suggests that the RBP possessing the higher cross-prediction performance recognizes a larger percentage of the salient features present in the second RBP's contexts than the other way around. We performed cross-prediction in a pairwise manner to comprehensively compare the binding patterns and activities of all RBPs. Only 316 pairs, or 6.36% of the total 4970 possible pairs (i.e., 71 choose 2 × 2), had a cross-prediction AUPR >86.5% (i.e., the median cognate performance from above), indicating that the vast majority of RBP pairs demonstrate binding patterns that are distinct from one another ( Supplemental Table S12 ). We subsequently generated a heat map of cross-prediction performances, which provides an effective medium for visualizing the pairs of RBPs with strong cross-predictions ( Fig. 6 B). We also conducted hierarchical clustering on the RBPs based on their respective cross-prediction profiles across the entire set of RBPs ( Fig. 6 C), where assignment to the same cluster is suggestive of RBPs exhibiting similar global binding patterns to one another. Of note, certain collections of RBPs demonstrated striking cross-prediction patterns ( Supplemental Note S8 ), with the largest group ( Fig. 6 B, top left; Fig. 6 D, inset) being most prominent: More than 50% (40 of 71 RBPs, excluding HNRNPM and DDX55) predicted relatively well on a set of 14 RBPs. This set of 14 comprised RBPs with strongly G-rich motifs and contexts, and different combinations of these RBPs clustered together by motif and context similarity measures ( Figs. 3 , 4 ). Upon further investigation, we noticed that each of the 14 RBPs has either exhibited binding of RNA G-quadruplexes (rG4s) (CSTF2T, DDX3X, FAM120A, GRSF1, GTF2F1, PRPF4), or showed binding enrichment at putative rG4s (EFTUD2, FASTKD2, FKBP4, NCBP2, NKRF, RBFOX2, SLTM, XRN2) in previous studies ( Herdy et al. 2018 ; Lee et al. 2020 ; Busa et al. 2021 ; Dumas et al. 2021 ; Bourdon et al. 2023 ). The considerable degree of universality exhibited by cross-predictions on these 14 RBPs may indicate that their contexts contain highly similar and prevalent nucleotide patterns and/or structures; furthermore, the commonality between target genes for certain subsets of these RBPs ( Lee et al. 2020 ) may suggest that these regions undergo coordinated regulation by a large number of factors, which might in turn reveal the importance of their target genes. A group of RBPs (DDX59, HNRNPK, PCBP1, and PCBP2) with predominantly C-rich motifs and contexts ( Fig. 3 ) showed reciprocated, strong cross-prediction ( Fig. 6 E; Supplemental Table S12 ). The lower cross-prediction by IGF2BP1/3 could be attributed to the more variable content of their motifs and contexts that did not exhibit a very strong preference for any given nucleotide. Importantly, PTBP1 also performed relatively well on the four strongly C-preferring RBPs (and vice versa) ( Supplemental Table S12 ), suggesting that they together possess similar global binding patterns, despite PTBP1's distinct, highly dissimilar U-rich motifs and contexts ( Fig 3 ); PTBP1 did not cluster with any of the four C-preferring RBPs by motif and context KLD ( Fig. 4 ). This suggests that our cross-prediction results may reveal novel potential relationships between RBPs. Taken together, the novel cross-prediction strategy we introduce proves informative of the degree of similarity between the global recognition and binding patterns of RBPs. Importantly, we posit that a strong cross-prediction performance and/or membership to the same cross-prediction cluster may be suggestive of potential RBP–RBP interactions. While the clustering assignments observed by cross-prediction may differ from those produced by HDBSCAN ( Fig. 3 ) or KLD ( Fig. 4 ), each of the three measures provides an alternative viewpoint of the underlying data (i.e., pattern-based, Euclidean, and probabilistic, respectively). As we demonstrate below, integration of these features and the information they carry not only helps to provide a more complete depiction of the relationship between RBPs, but also helps to increase our confidence in interaction discovery—especially if the corresponding RBPs consistently cluster together across methods. A novel method to discover proximal interactions between RBP–RBP pairs Given that RBPs frequently interact cooperatively or competitively to jointly regulate common targets ( Dassi 2017 ), we developed a method to investigate RBP–RBP interactions. We first devised a conceptual delineation of the potential RBP–RBP interaction modes into concrete categories based on properties of the involved partners, including whether the RBPs bind proximally or distally to one another, whether one or both RBPs bind RNA, and whether the RBPs physically interact ( Fig. 7 A; Supplemental Note S9 ). We used this classification to outline a scope within which we performed our RBP–RBP interaction discovery. Because our analyses are context-based, do not yet incorporate RNA secondary structure data, and do not yet account for long-range interactions across genomic features, we limited our method to consider only proximal interactions between pairs of RBPs ( Fig. 7 A, left-hand major branch). We define four proximal interaction binding modes ( Supplemental Note S9 ). FIGURE 7. Open in a new tab A novel integrative method to infer RBP–RBP interactions. ( A ) Conceptual representation for the classification of RBP–RBP interaction categories, which depend on properties of the involved components. Distal interactions, while demonstrated here, are not considered in this analysis. ( B ) Interaction discovery method, represented as a decision tree based on features of the candidate RBP–RBP pair. Each path (I through IV) along the tree represents a set of possible interaction modes based upon the fulfilled conditions along the path. A candidate pair may be involved in a cooperative interaction, competitive interaction, or, potentially, both. ( C ) Feature summary and ( D ) GO analysis (top 50 terms with FDR <0.05 shown) for the proposed interaction between AKAP1 and UPF1. ( E ) Feature summary and ( F ) GO analysis (top 50 terms with FDR <0.05 shown) for the proposed interaction between DDX59, HNRNPK, and PCBP2. Within this scope, we formulated an RBP–RBP interaction discovery method based on empirical observations and guided by reported interactions in the literature, which allowed us to define a set of characteristic features that we represent as a set of conditions in the form of a decision tree ( Fig. 7 B). Each branch of the tree in turn represents a set of conditions that must be satisfied for a given candidate RBP–RBP pair to qualify as belonging to the corresponding interaction category ( Fig. 7 B; Supplemental Note S9 ). The set of conditions used in our interaction discovery method comprise motif similarity, context similarity, motif co-occurrence, and motif overlap. As a strategy to reduce the search space of candidate interacting pairs, we began our investigation by referencing the clustering results from KLD and cross-prediction analyses, reasoning that RBPs assigned to the same cluster(s) may share similar motifs, contexts, and/or global binding profiles, and thus have a stronger likelihood of being potential interacting pairs, especially if the RBPs cluster together repeatedly across multiple measures. To validate the conditions of this method, we first examined known interacting RBP pairs established in the literature as positive controls ( Supplemental Table S13 ). We started by examining pairs of RBPs with strong, reciprocated cross-prediction scores from Figure 6 E. We noticed that HNRNPK and PCBP2 consistently demonstrated high similarity to one another, as witnessed by their respectably low motif and context KLDs ( Supplemental Table S13A ), membership to the same cluster in global KLD comparisons ( Fig. 4 ), and remarkably strong and symmetric cross-prediction performances ( Fig. 6 E; Supplemental Table S13A ). We discovered C-rich motifs and contexts for both of these RBPs ( Fig. 3 C), which is supported by their belonging to the poly(rC) binding protein family ( Choi et al. 2009 ). As heterogeneous ribonucleoproteins, their functional range is extensive and diverse, shuttling between the nucleus and cytoplasm to regulate pre-mRNA splicing, mRNA transport, mRNA stability, and mRNA translation ( Kim et al. 2000 ; Lang et al. 2021 ). Experimental evidence has revealed an interaction between HNRNPK and PCBP2, with hypotheses that they may form an hnRNP complex to subsequently facilitate secondary interactions with regulatory effectors ( Kim et al. 2000 ), but it remains unclear whether this interaction is cooperative or competitive, and whether one or both partners bind target mRNAs. Based on our finding that their motif co-occurrence substantially exceeds their motif overlap ( Supplemental Table S13A ), we believe that HNRNPK and PCBP2 have a higher propensity to form cooperative interactions in which both partners bind RNA ( Fig. 7 B, Path I). For the less frequent cases in which their motifs overlap, this could be attributed to a cooperative mechanism in which the RBPs physically interact but only one partner binds RNA, or a competitive mechanism in which the RBPs compete for a binding site ( Fig. 7 B, Path II). This case study serves as a validation of our method's ability to detect an established RBP–RBP interaction with existing experimental evidence, and we additionally propose a novel hypothesis for their potential mode(s) of interaction. With this positive control and others ( Supplemental Note S10 ), we demonstrate that our proposed method successfully identifies known RBP–RBP interactions. These examples also demonstrate the combinatorics of RBP–RBP interactions ( Khoroshkin et al. 2024 ), with certain RBPs (e.g., PCBP2, PTBP1, MATR3) being involved in multiple different interaction pairs, thus pointing to the exponential scale at which the full range of human RBPs interact. In total, we discovered 796 interactions with our method, which accounts for ∼32% of the total 2485 possible RBP–RBP pairs in our collection ( Supplemental Table S14 ; Supplemental Notes S9, S11 ; Supplemental Fig. S17 ). As further validation, we cross-referenced our method with known, documented interactions in the BioGRID database ( Oughtred et al. 2021 ; Supplemental Notes S9, S11 ). Of the 436 documented interacting RBP–RBP pairs that were present both in our collection and in BioGRID, 145 of these interactions (33.36%) were detected by our method, demonstrating our method's ability and efficacy. We next sought to further investigate novel interacting pairs detected by our method and to hypothesize their potential functions and mechanisms. NMD factor UPF1's strong association with AKAP1 suggests its surveillant role in local translation at organelle membranes We noted that AKAP1 and UPF1, despite visually appearing to have dissimilar motifs and nucleotide preferences ( Supplemental Table S5 ), occurred in the same cross-prediction cluster ( Fig. 6 C), which may be suggestive of a potential interaction. The relatively large divergence between their target motif logos ( Fig. 7 C, “Motif KLD”), with AKAP1's motif logo being predominantly C-rich (consensus: CUCCC) and UPF1's motif logo being predominantly G-rich (consensus: GGGGG), indicated dissimilarity between their motifs. Conversely, their contexts exhibited some degree of similarity: AKAP1's context logo appears C-rich but has G-rich undertones, while UPF1's context logo is GC-rich. This context similarity is supported by the context KLD increasing only slightly compared to the motif KLD (from 31.36 for motif to 33.32 for context, AKAP1 vs. UPF1; from 33.64 for motif to 35.59 for context, UPF1 vs. AKAP1; Figure 7 C, “Context KLD”), despite being a cumulative sum across a considerably larger region of 55 nt compared to the target length of 5 nt ( Supplemental Notes S9, S12 ). Their relatively high and reciprocated cross-prediction scores ( Fig. 7 C), as well as their membership in the same cross-prediction cluster ( Fig. 6 C), indicate that the two RBPs likely share similar global binding patterns. Finally, as expected from their highly divergent motif PPMs, AKAP1's ( N = 3775) and UPF1's ( N = 6225) motif instances shared very little overlap ( N = 14; Fig. 7 C, “Percent of Motif Instance Overlaps”); however, their respective co-occurrence rates were relatively high, with 30.65% of AKAP1's motif instances overlapping UPF1's motif instances and 18.57% of UPF1's motif instances overlapping AKAP1's motif instances ( Fig. 7 C). Taken together, these features correspond with our conditions for cooperative interactions along Path III, and we therefore hypothesized that AKAP1 and UPF1 may work together in a synergistic manner to regulate common targets. To further support our hypothesis, we demonstrate that overlapping AKAP1–UPF1 contexts have significantly higher eCLIP enrichment from both AKAP1 ( Supplemental Fig. S18A ) and UPF1 ( Supplemental Fig. S18B ) data sets than contexts belonging to either RBP that are not implicated in the interaction (Materials and Methods). We also demonstrate that overlapping AKAP1–UPF1 contexts were significantly enriched in the 3′UTRs of differentially expressed genes upon knockdown of both AKAP1 and UPF1 independently ( Supplemental Note S12 ). We sought to understand how this pair may cooperate. We observed that the vast majority (93.3%) of overlapping AKAP1–UPF1 contexts were located in 3′UTRs, which are known to have roles in mRNA translation regulation ( Mayr 2019 ), a process for which both RBPs have documented functions: AKAP1 spatially regulates the local translation of essential proteins at the mitochondrial outer membrane ( Gabrovsek et al. 2020 ; Sharma and Fazal 2024 ), while UPF1 is an essential RBP in the nonsense-mediated decay (NMD) pathway ( Supplemental Note S12 ; Kim and Maquat 2019 ). Upon examination of the gene set containing overlapping contexts, we not only found a number of genes encoding important mitochondrial proteins, but also significant enrichment for localization (e.g., protein localization), membrane (e.g., endomembrane system organization), and transport (e.g., vesicle-mediate transport) gene ontology (GO) terms ( Fig. 7 D; Supplemental Note S12 ). Based on these findings, we propose a cooperative mechanism between AKAP1 and UPF1 in which they work jointly, by way of Path III ( Fig. 7 B), to ensure accurate local translation of specialized mRNA transcripts into full-length protein products at certain organelles. Our results may suggest that UPF1 maintains a sustained interaction with AKAP1, whereby UPF1 intervenes in circumstances of local translation where AKAP1 has anchored mRNA transcripts containing premature termination codons to organelle membranes. This could present a quality control mechanism for the accurate translation and efficient use of proteins that are essential to organelle function, especially at the mitochondria and the endomembrane system. The DEAD-box RNA helicase DDX59 facilitates the mRNA regulatory roles of HNRNPK and PCBP2 Interestingly, DDX59 demonstrated highly similar features with both HNRNPK and PCBP2 individually ( Fig. 7 E), despite being a distinct RBP belonging to a large family of highly conserved DEAD-box RNA helicases ( Linder and Jankowsky 2011 ; Shamseldin et al. 2013 ). Of note, DDX59's model demonstrated superior performance on the contexts of HNRNPK (AUPR = 97.44%) and PCBP2 (AUPR = 94.75%) compared to its own contexts (AUPR = 92.44%), suggesting a very strong degree of similarity in the underlying patterns of their respective binding regions. DDX59 also exhibited a higher degree of motif co-occurrence and motif overlap with HNRNPK and PCBP2, compared to those observed between the known interactors HNRNPK and PCBP2 ( Fig. 7 E), which provides additional support for a potential interaction with DDX59. Although DEAD-box helicases interface with the RNA backbone of their targets in a non-sequence-specific manner ( Jarmoskaite and Russell 2011 ; Linder and Jankowsky 2011 ), we discovered a C-rich nucleotide binding preference ( Fig. 3 C) for DDX59, which may be a byproduct of frequent associations with poly(rC) binding proteins like HNRNPK and PCBP2 and subsequent co-immunoprecipitation. While the exact functions of DDX59 have not yet been deduced, its family members have annotated roles in ATP-dependent local RNA duplex unwinding, RNA clamping, and RNP assembly, stabilization, remodeling, and disassembly ( Linder and Jankowsky 2011 ; Bohnsack et al. 2023 ). Based on our empirical evidence, we hypothesize that DDX59 interacts cooperatively with both HNRNPK and PCBP2 by any or all of the following possible mechanisms: (1) unwinding short regions of double-stranded RNA to enable binding by the single-stranded RNA-preferring KH domains of HNRNPK and PCBP2, (2) serving as an anchor point at which HNRNPK and PCBP2 can interact to form an hnRNP complex, (3) remodeling hnRNPs, and/or (4) disassembling hnRNPs. This hypothesis is further supported by the fact that HNRNPK has previously been shown to interact with a number of DEAD-box helicases to enact various regulatory processes ( Chen et al. 2002 , 2021 ; Good et al. 2019 ; Li et al. 2019 ). Because each pair of RBPs within this trio exhibited strong features indicative of cooperative interactions, we leveraged transitive relations to hypothesize that DDX59, HNRNPK, and PCBP2 may also interact in unison to form a complex ( Supplemental Note S13 ). To investigate this possibility, we performed a three-way intersection between their respective contexts (Materials and Methods) and found that 21.79% of DDX59's contexts overlapped with both HNRNPK's and PCBP2's, with most of these events located in introns or coding sequences. From GO analysis of the genes containing context overlaps between all three RBPs, we discovered enrichment for anatomical structure development and its child term anatomical structure morphogenesis ( Fig. 7 F); further inspection of additional child terms for both revealed many related to the development of the face (face, head, and craniofacial suture morphogenesis), oral cavity (hard and soft palate morphogenesis, roof of mouth development, tooth eruption), and digits (limb joint and appendage morphogenesis) ( Binns et al. 2009 ). Mutations in DDX59 have been associated with orofaciodigital syndrome, a disease which leads to varied disfigurements in features of the mouth, face, and digits ( Shamseldin et al. 2013 ; Faily et al. 2017 ), which may help to explain the discovered GO terms. In conclusion, we not only hypothesize novel and specific roles for DDX59 in regulating mRNA processing via its strong association with either or both of the hnRNPs HNRNPK and PCBP2, but we also utilize combinatorics and transitive inference to propose an interaction between DDX59, HNRNPK, and PCBP2—either as individual pairs, a complex of three proteins, or both—that holds mechanistic, functional, and disease significance, thereby demonstrating the utility of our approach. DISCUSSION In this paper, we presented a novel, comprehensive computational pipeline to study the binding preferences, patterns, activities, functions, and interactions of RBPs. Inspired by state-of-the-art concepts from linguistics and NLP, we developed a sequence representation strategy that enabled us to clearly delineate RBP binding motifs and their contexts, and we subsequently leveraged multiple instance learning to predict RBP binding of these contexts in a weakly supervised fashion. We also developed an accurate, consensus-based motif discovery algorithm to identify the binding and context motifs for 71 RBPs in the HepG2 cell line and 74 RBPs in the K562 cell line. Importantly, we discovered novel in vivo contexts for these RBPs, which is an important contribution to understanding the inherent nucleotide binding preferences exhibited by RBPs in vivo as well as their role in facilitating RBP binding specificity. With feature integration and transitive inference, we were able to leverage these findings and introduce a method for identifying known and novel RBP–RBP interactions, illustrating examples of each. Our approach to RBP–RBP interaction discovery allowed us to identify many novel interacting pairs, and we were additionally able to hypothesize functional mechanisms pertaining to mRNA regulation for a subset of those pairs. In total, we characterized 2485 RBP–RBP pairs ( Supplemental Table S14 ). Based on our findings, the majority of detected interactions are likely cooperative, and this appears to align with previous work that has remarked on the biological and technical challenges involved in detecting competitive interactions ( Lang et al. 2021 ). In addition, the mode of interaction between RBPs ( Street et al. 2024 ) greatly influences its ability to be detected by our method, which for the time being addresses proximal, RNA-mediated interactions between pairs of RBPs. Long-range cooperative interactions in which RBPs bind distally from one another and are subsequently brought into closer proximity by virtue of RNA secondary structures may be more difficult to detect ( Dassi 2017 ), as are long-range competitive interactions in which RBPs do not directly contend for the same binding site but instead bind distally from one another to promote divergent outcomes, such as in the proposed regulatory mechanism of alternative splicing in muscle differentiation between PTBP1 and QKI ( Fig. 7 A; Hall et al. 2013 ). For other binding modes, such as the cooperative and competitive modes in Path II ( Fig. 7 B), we do not yet have the resolution to distinguish whether or not the feature set indicates a cooperative or competitive interaction, for which other types of data will be required. Though not explored in detail here, BCLAF1 and TRA2A present a strong example for the cooperative and/or competitive modes of Path II ( Supplemental Table S14 ), and it would be of great interest to understand by which mode and for which functions they interact. We also do not consider homomeric interactions here. We additionally note that, due to its reliance on the percentage of motif instance and context overlaps, our approach is more attuned to detect RBP–RBP interactions that occur more frequently, which in turn implies that these interactions have more widespread regulatory roles on a larger set of mRNA targets. Indeed, the case studies we highlighted here ( Fig. 7 ) are those with very prominent features. On the other hand, less frequently occurring interactions that may govern the regulation of a small, specialized subset of mRNAs—which are still detected but have much lower overlap percentages—will require more detailed investigation in order to elucidate their precise functions. By using the theory of transitive relation, we were also able to propose a novel n -way interaction ( n > 2), as exemplified by our DDX59 case study. Importantly, this potential cooperativity between DDX59, HNRNPK, and PCBP2 has significant physiological and disease implications, thereby demonstrating the utility of our discovery. As additional support for this approach, we were also able to detect a putative interaction between RBPs recognizing G-rich motifs (namely, CSTF2T, FAM120A, GTF2F1, and PRPF4; Supplemental Table S15A ), which has previously been identified ( Lee et al. 2020 ) and therefore serves as a form of validation for our use of transitivity. We further propose that interactions may occur between other members of the 14 rG4-binding candidates isolated from our cross-prediction results ( Fig. 6 D; Supplemental Table S15B ); for example, the highly similar G-rich motif, preceded by a U and followed by a C, within a strongly G-rich context, belonging to ETUD2, FAM120A, GRSF1, and NKRF may be indicative of a potential interaction between the four RBPs. It is logical, therefore, to also extend this method in the future to infer larger macromolecular complexes. In fact, we were able to obtain preliminary results in this direction, identifying pairwise interactions between multiple components of the (pre)catalytic spliceosome ( Supplemental Table S16 ), with some RBPs being associated with either the U2, U5, or U4/U6 snRNPs, or the U4/U6 × U5 tri-snRNP complex ( Sayers et al. 2022 ) and many having been captured together in existing cryo-EM structures of the spliceosome ( Szklarczyk et al. 2023 ). Two RBPs in this set, AQR and BUD13, have not, to the best of our knowledge, been reported in the literature, and we present them as a potential novel interacting pair within the spliceosome that is also supported by their discovered interactions with other spliceosomal proteins ( Supplemental Tables S14, S16 ). We may even be able to obtain a finer resolution by using our feature integration and transitive inference to build up such complexes piecewise based on progressive degrees of similarity. To facilitate detections of these large-scale interactions, the cross-prediction and KLD dendrograms may prove helpful, as their inherent hierarchical nature may provide clues regarding the combinatorics, hierarchy, and order in which RBP–RBP interactions are formed. We also note that, while we used the KLD and cross-prediction clusters to guide our identification of candidate RBP pairs, this does not preclude the possibility that interactions also occur between RBPs from different clusters; such cases, which will also contribute to the discovery of larger complexes, require further examination. Overall, our results provide a promising advancement and outlook for charting the vast landscape of RBP–RBP interactions. Finally, future avenues could entail analyses from numerous angles. Firstly, additional studies will be needed to investigate the proposed mechanisms of novel interactions discussed in this paper. The many interactions we identified but did not elaborate on here ( Supplemental Table S14 ) will also require additional exploration to understand their underlying mechanisms, functions, and scale. We are also interested in extending our analyses to encompass mutual interactions, which we were able to preview in our AKAP1–UPF1 case study: The gene set containing context overlaps between the two RBPs contains the UPF1 gene itself, indicating that (1) UPF1 regulates the translation of its own cognate mRNA, a process referred to as an autogenous interaction, and (2) AKAP1 is involved in the regulation of UPF1 mRNA, a process referred to as a heterogeneous interaction ( Dassi 2017 ). Solely within our collection of RBPs, our preliminary results reveal a strong prevalence of mutual interactions: ∼97% (69 of the 71) are involved in any type of mutual interaction, while 100% of those detected mutual interactions demonstrated heterogeneous interactions and ∼30% ( N = 21) also demonstrated autogenous interactions. The full extent of the mutual interaction landscape will become clearer as these analyses are expanded to encompass the entire set of human RBPs, as will their functional consequences and implications. Other biological questions of interest include (1) to what extent RBP binding patterns and RBP–RBP interactions exhibit cell type specificity, (2) how RNA secondary structures contribute to enable specific RBP binding and distal RBP–RBP interactions, (3) how RBP binding preferences are influenced by cooperative and competitive interactions with microRNAs ( Ciafrè and Galardi 2013 ; Kim et al. 2021 ), (4) how RBP activities, functions, and interactions can be further elucidated via RBP regulatory network modeling, and (5) how the discovered RBP binding patterns and interactions play a role in disease mechanisms. In conclusion, the method we have presented herein serves as a valuable contribution to the field's current knowledge of RBP binding patterns, context preferences, and interactions, all enabled by the advent of state-of-the-art machine learning, deep learning, and modeling techniques; we believe our work sets an important foundation for future studies in these areas. MATERIALS AND METHODS Comparative analyses of eCLIP read enrichment over discovered motifs and contexts As further validation, we performed comparative analyses to demonstrate that our discovered motifs and contexts are significantly more enriched in eCLIP read coverage compared to a set of matched control regions. To define a set of comparable control regions, for every motif instance in our discovered motif, we randomly sampled a genomic instance of the same k -mer that met the following conditions: (1) belongs to a gene expressed in the same cell line (i.e., HepG2), (2) belongs to the same chromosome as the motif instance, (3) belongs to the same genomic region as the motif instance, (4) when extended symmetrically to the same length as the discovered contexts (i.e., 55 nt), does not overlap with any discovered contexts. We next downloaded the eCLIP read alignment .bam files from the ENCODE portal for all RBPs and their corresponding size-matched input experiments ( Supplemental Table S1C ). For each RBP, we performed the following procedure for both (1) the discovered motifs and contexts and (2) the matching control “motifs” and “contexts,” based on previous eCLIP studies ( Van Nostrand et al. 2016 , 2020a , b ). We first used BEDTools coverage ( Quinlan and Hall 2010 ) to determine the number of eCLIP reads overlapping each region (normalized by read depth of the sequencing experiment) and averaged across both replicates; we also computed normalized eCLIP read coverage overlapping each region for the single corresponding size-matched input (SMInput) experiment. Next, we computed each region's eCLIP enrichment by: log 2 ( fold change ) = log 2 FC = log 2 ( normalized eCLIP read coverage normalized SMInput read coverage ) To compute the significance of each region's enrichment, we ran a χ 2 test if the raw read count overlapping the region was >5 reads for both the eCLIP and input experiments, and ran Fisher's exact test otherwise. Because each eCLIP replicate had a different raw read count, we used Fisher's method to combine the resulting P -values from both replicates ( Fisher 1973 ). Finally, we performed multiple test correction across all regions for a given RBP by computing the FDR. We designate a region as significantly enriched in eCLIP if the log 2 FC is >0 and the FDR is <0.05. To compare whether discovered motifs had a significantly greater eCLIP enrichment than their matched controls, we used a Wilcoxon rank-sum test, then visualized the corresponding distributions as nested violin-box plots. We performed the same for the comparison between discovered contexts and their matched controls. We also leveraged the eCLIP enrichments to support our discovered RBP–RBP interactions, following a similar procedure to the one above. For a given RBP–RBP pair, we computed the eCLIP enrichment independently for both RBPs implicated in the interaction, over four sets of regions: (1) contexts belonging to RBP #1 that are not implicated in the interaction (i.e., do not overlap with contexts belonging to RBP#2), (2) contexts belonging to RBP#1 that are implicated in the interaction (i.e., overlap with contexts belonging to RBP#2), (3) contexts belonging to RBP#2 that are implicated in the interaction, and (4) contexts belonging to RBP#2 that are not implicated in the interaction. We again used the Wilcoxon rank-sum test to determine whether the predicted interaction's binding sites (containing overlapping or co-occurring motifs) demonstrated significantly stronger eCLIP enrichment than each RBP's noninteracting binding sites. Quantification of motif discovery accuracy We quantified the accuracy of our motif discovery algorithm by curating a set of ground-truth motifs for RBPs with strongly characterized binding patterns in the literature, which we define as having specified logos, consensus sequences, or nucleotide preferences that are concordant across multiple sources. This resulted in a set of 14 RBPs each for HepG2 and K562. For this evaluation, we opted against performing a cross-reference analysis using tools that compare user-provided motifs with those from existing databases in order to determine the identity of the proteins to which the motifs belong ( Bailey et al. 2015 ); while such an approach is certainly useful in specific circumstances, we deemed it unsuitable for our purposes because (1) most available databases are transcription factor–centric, (2) the available RBP databases are primarily derived from in vitro data only, (3) the available RBP databases are limited in the number of RBPs and motifs they contain with respect to the full scope of literature available, and (4) the outcomes of such analyses may be misleading, as they do not account for RBPs that may recognize different binding motifs depending on which RBD is interacting with its target mRNA at the time of experimental capture. Calculation of sequence and context k -mer enrichments To compute k -mer enrichments, we performed the following procedure. For sequence-level k -mer enrichments, for each unique k -mer, we computed the ratio between the proportion of positively predicted and negatively predicted sequences (from the baseline model) containing the k -mer. For context-level k -mer enrichments, for each unique k -mer, we computed the ratio between the frequency of the k -mer as the target of positively predicted contexts and the frequency of the k -mer as the target of negatively predicted contexts (from the contexts model). For each, we used the associated counts to perform χ 2 testing and obtain a P -value for each k -mer indicating the significance of its enrichment. We considered a k -mer enriched if it had an enrichment >1 and a P -value <0.05. Random motif discovery To generate a set of random genomic “peaks,” we implemented the following procedure. For each sampled sequence, we (1) randomly selected a chromosome from {chr1-chr22, chrX}, (2) randomly selected a genomic region from {CDS, 3′UTR, 5′UTR, 5′SS, 3′SS, proximal intron, distal intron} from the chromosome selected in step 1, (3) randomly selected a position within the region selected in step 2 to serve as the central position, and (4) extended ±50 nt around the central position selected in step 3. We repeated this procedure 10,000 times, making sure that each sampled sequence did not overlap eCLIP peaks from any RBP from either the HepG2 or K562 data sets. The final set, containing 10,000 random sequences of length 101 nt, served as our set of random “positive” sequences. To generate the set of corresponding “negative” sequences, we performed the same procedure we used to derive the set of negative sequences for training our models ( Supplemental Note S1 ). Briefly, for each “peak” in our positive set, we randomly sampled a sequence of the same length from a matched chromosome and genomic region and resampled as needed to ensure that no sequence overlapped with the “positive” set or with any eCLIP peaks. The final set served as our set of random “negative” sequences, leading to a total random data set size of 20,000 sequences. For comparison, we used the same procedure to generate data sets of size 2000 and 10,000. To generate a set of random synthetic “peaks,” we generated random sequences of length 101 nt by sampling from the alphabet {A, C, G, U} under a uniform distribution. We similarly generated sets of 2000, 10,000 and 20,000 sequences, and performed the same analyses on each (i.e., motif discovery with our algorithm and with STREME). For each of the six random data sets we generated, we first deconstructed the set of random sequences into contexts using our sequence decomposition method, then ran our motif discovery algorithm on the resulting contexts. For comparison, we also submitted the contexts to STREME. To contrast the properties of the results from random sequences (i.e., prediction AUPR, consensus/ k -mer enrichments, and P -values, etc.) with those of RBPs, we selected a subset of seven sequence-specific RBPs (HNRNPC, HNRNPK, MATR3, PTBP1, QKI, RBFOX2, SRSF1) as well as six non-sequence-specific RBPs (DDX3X, DDX55, DDX59, DDX6, ILF3, XPO5) from our work in the HepG2 cell line for comparison. Determination of motif and context nucleotide preference and symmetry We determined the nucleotide preference of each RBP's motif and context by examining its PPM representation. Starting with the left context, we identified the most prevalently occurring nucleotide for every position, then ranked each of the 4 nucleotides in descending order based on the number of positions at which each occurred most prevalently. We also computed the ratio of the number of positions for each nucleotide relative to the maximum of these values. To account for cases in which RBPs demonstrate a clear preference for more than 1 nt (e.g., HNRNPL), we deemed any nucleotide for which its position count was at least half that of the top-ranking nucleotide as also demonstrating a strong preference. We performed these operations for the left context, motif, and right context individually to obtain nucleotide preferences at a finer resolution, and finally performed it for the full context (excluding the motif) to determine the overall nucleotide preference of the context. We deemed a context symmetric if its preferred nucleotide(s) in the left and right contexts were the same. Similarly, we compared the preferred nucleotide(s) in the motif and the full context to determine whether they shared the same preferences. Computation of cumulative motif and context Kullback–Leibler divergence (KLD) The Kullback–Leibler divergence (KLD) is a statistical measure that quantifies the degree to which one probability distribution Q (i.e., the approximating distribution) is different from another probability distribution P (i.e., the reference distribution). In information theory, the KLD is also known as the relative entropy, where smaller values indicate that Q is a closer approximation of P , and the minimum value of 0 indicates that the two probability distributions are identical ( Bishop and Nasrabadi 2006 ). The KLD does not qualify as a metric in the formal mathematical sense because it does not satisfy two main properties that are required of a metric. First, the KLD is asymmetric, such that KLD( P , Q ) does not necessarily equal KLD( Q , P ). Second, the KLD does not satisfy the triangle inequality: If two approximating distributions Q and S are each independently similar to the true distribution P , it does not necessarily mean that Q and S are similar to one another. Thus, we use KLD as a statistical measure of distribution similarity. In our application, we reasoned that the position probability matrix (PPM), which we use to represent both motifs and contexts, fundamentally comprises a series of position-specific probability distributions across the four canonical nucleotides, where each position has its own distribution. For a given position in a PPM, we can use the KLD to approximate the similarity between the corresponding nucleotide probability distributions at that position for any given pair of RBPs. For a single position within any given sequence, the KLD is computed as follows: KLD ( P , Q ) = ∑ α ∈ { A , C , G , U } ⁡ P α , j × log 2 P α , j Q α , j Note that the KLD has some finite maximum value that is dependent on each distribution, provided that no values in the distribution are equal to 0. Because we sought to compare the distributions across the full length of the motifs and contexts, we herein introduce the cumulative KLD, which sums together the position-specific KLDs for the entire motif and context, respectively: Cumulative KLD ( P , Q ) = ∑ j = 1 k ⁡ ∑ α ∈ { A , C , G , U } ⁡ P α , j × log 2 P α , j Q α , j , where k = 5 for motifs and k = 55 for the full contexts. Thus, the values that we present in this work are cumulative KLDs that represent the sum of position-specific KLDs across an entire PPM. Processing of RBP knockdown shRNA-seq data Differential expression analyses To functionally connect our motifs and contexts to the regulation of gene expression, we ran DESeq2 ( Love et al. 2014 ) on the shRNA-seq raw count data, defining any gene whose expression in the KD data set was 1.5 times greater or less than the expression in the corresponding control data set [i.e., log 2 FC > log 2 (1.5)] and whose corresponding P -value and FDR were both <0.05 as differentially expressed. We also defined the set of expressed genes in any experiment as those whose average TPM across two replicates was >0 in at least one of the control or KD conditions. For each RBP, we designated the set of differentially expressed (DE) genes as the foreground set, and the remaining expressed but not DE genes as the background (bkgd) set. We then computed the extent to which our discovered contexts were enriched in DE genes relative to background, based on the general formula: context enrichment = % DE genes containing a discovered context % bkgd genes containing a discovered context To more closely examine the functional effects of RBP KD, we reasoned that this general metric required stratification on two levels in order to prevent information loss: (1) based on whether or not the DE gene was up- or downregulated, and (2) based on the genomic region in which the RBP's context is located. Therefore, we computed a directional, region-based context enrichment as follows: For direction d , s.t. d ∈ {up, down}, and region r , s.t. r ∈ { 5 ′ U T R , C D S , 3 ′ U T R } : context enrichment d , r = % DE genes , regulated in direction d , containing a discovered context in region r % bkgd genes containing a discovered context in region r Differential splicing analyses To functionally connect our discovered motifs and contexts to the regulation of alternative splicing, we ran rMATS-turbo ( Wang et al. 2024 ) on the shRNA-seq bam files ( Supplemental Table S1D ). We focused primarily on the skipped exon (SE) class of events, as they were the most prevalent event class, representing ∼70% of all detected events for each RBP ( Supplemental Fig. S13A ). We filtered detected SE events using the following criteria: (1) events must have a replicate-averaged PSI value between 0.05 and 0.95 in at least one of the control or knockdown conditions ( Dominguez et al. 2018 ; Van Nostrand et al. 2020a ), and (2) events must have a minimum total inclusion + exclusion junction read coverage of 10 ( Dominguez et al. 2018 ). This filtering yielded ∼13,439 SE events (on average) across all RBPs, representing 32.78% (on average) of all detected SE events across all RBPs ( Supplemental Fig. S13B ). We defined an event as being differentially spliced (DS) if its deltaPSI between control and KD conditions was >0.05, and both its P -value and FDR were <0.05 ( Van Nostrand et al. 2020a ). We defined all remaining SE events as background (bkgd) events. The number of DS events demonstrated more variability across RBP KD experiments, with ∼1029 DS events (on average) per RBP, representing 7.31% (on average) of all filtered SE events per RBP ( Supplemental Fig. S14A,B ). The majority of RBPs demonstrated a preference for up- or downregulated events upon knockdown, which provides insight into their regulatory effects on exon inclusion or exon skipping under normal conditions ( Supplemental Fig. S13C ). For enrichment analyses, we defined the set of DS events as our foreground and all other filtered SE events as our background. To determine whether our discovered motifs and contexts are enriched in these events, we first defined regions of interest that have been established to contain sequence elements for splicing factors to bind. We defined four regions of interest that are typically implicated in the splicing of cassette exons (adapted from Yee et al. 2019 ) at (1) the 3′ end of the upstream exon (“upstream”), (2) the 3′SS of the cassette exon (“cassette 3′SS”), (3) the 5′SS of the cassette exon (“cassette 5′SS”), and (4) the 5′ end of the downstream exon (“downstream”) ( Fig. 5 B). Each of the four regions span 50 nt into the associated exon and 300 nt into the associated intron, except for cases that fall within the following exceptions: (1) if the upstream or downstream exons are <50 nt long, the region only extends to the exon boundary; (2) if the cassette exon is shorter than 100 nt long, the region only extends halfway into the exon; and (3) if either of the introns are shorter than 600 nt long, the region only extends halfway into the intron. Similarly to the computation of context enrichments in our DE analyses, we also computed the extent to which our discovered contexts were enriched in DS SE events relative to background, based on the general formula: context enrichment = % DS SE events containing a discovered context % bkgd SE events containing a discovered context It thus follows that the directional, region-based context enrichment is computed as follows: For direction d , s.t. d ∈ {up, down}, and region r , s.t. r ∈ { upstream , cassette 3 ′ SS , cassette 5 ′ SS , downstream } : context enrichment d , r = % DS SE events , regulated in direction d , containing a discovered context in region r % bkgd SE events containing a discovered context in region r We computed P -values and visualized our results in a similar manner to the differential expression analyses. We used bedtools getfasta ( Quinlan and Hall 2010 ) to fetch the genomic sequences corresponding to all regions of all events. We computed k -mer enrichments using one of two methods. In the first method, which is event-centric, we computed the enrichment of k -mer i in region r , s.t. r ∈ { upstream , cassette 3 ′ SS , cassette 5 ′ SS , downstream } , in DS events regulated in direction d , s.t. d ∈ {up, down}, as follows: k − m e r e n r i c h m e n t i , d , r = % total k − mers in region r of DS events regulated in direction d that are k − mer i % total k − mers in region r of bkgd events that are k − mer i In the second method, which is context-centric, we computed the enrichment of k -mer i using our original starting context data set. We first used BEDTools intersect ( Quinlan and Hall 2010 ) to isolate any contexts that overlap either DS or background events and then defined our foreground as contexts overlapping DS events and our background as contexts overlapping background events. For any context overlapping both a DS event and a background event, we considered that context as part of the background and removed it from the foreground. We then computed the context-centric enrichment for k -mer i in region r , s.t. r ∈ { upstream , cassette 3 ′ SS , cassette 5 ′ SS , downstream } , in DS events regulated in direction d , s.t. d ∈ {up, down}, as follows: k − mer enrichment i , d , r = % total k − mers in contexts overlapping region r of DS events regulated in direction d that are k − mer i % total k − mers in contexts region r of b k g d events that are k − mer i For both methods of k -mer enrichment computation, we evaluated statistical significance and visualized results as previously described. Context intersections for DDX59, HNRNPK, and PCBP2 To identify distinct interactions between contexts belonging to any pair of RBPs within this trio, we first performed BEDTools intersect ( Quinlan and Hall 2010 ) between the contexts of two RBPs (e.g., HNRNPK and PCBP2), then performed BEDTools intersect again, this time with the contexts of the third RBP (e.g., DDX59) using the -v option; this produced the set of contexts from HNRNPK and PCBP2 that overlapped with each other, but that simultaneously did not overlap with DDX59's contexts. We repeated this procedure for the other two pairs (HNRNPK and DDX59; PCBP2 and DDX59). To identify all overlaps between contexts belonging to each of the three RBPs, we sought to account for all possible combinations of orderings (with respect to genomic coordinates) in which these overlaps could potentially occur. For example, three contexts—one from each RBP—may all overlap one another simultaneously. In other cases, the three contexts may sequentially (with respect to genomic coordinates) overlap each other in a staggered manner, such that there is continuity between all three, but they do not all overlap with one another; the order in which the three can overlap in such a manner is combinatorial. We first performed intersections with BEDTools ( Quinlan and Hall 2010 ) between two RBPs (e.g., HNRNPK and PCBP2) to identify their set of overlapping contexts. Then, we intersected the third RBP's contexts (e.g., DDX59) independently with the subset of HNRNPK's contexts that overlapped PCBP2's and with the subset of PCBP2's contexts that overlapped HNRNPK's; note that both of these operations are required to capture all possible interactions with DDX59. We merged the output that resulted from these last two intersections and filtered out duplicates to arrive at the subset of DDX59's contexts that overlap both HNRNPK's and PCBP2's contexts. Finally, to account for the combinatorics of the possible order of overlaps, we repeated this multistep intersection procedure two more times: once where we first intersected HNRNPK's and DDX59's contexts and then with PCBP2's contexts; and once where we first intersected PCBP2's and DDX59's contexts and then with HNRNPK's contexts. Note that all intersections were performed in a strand-specific manner. Gene ontology (GO) analysis In accordance with the pipeline used to process eCLIP data at the ENCODE portal ( Luo et al. 2020 ), we used the GENCODE v29 annotation ( Harrow et al. 2012 ) for all gene-level analyses. For a given RBP–RBP pair, we used BEDTools ( Quinlan and Hall 2010 ) to intersect the genomic coordinates of their motif instances and co-occurrences with the GENCODE annotation to determine in which genes they were located; the resulting unique set of genes served as our test set for GO analysis. To derive a background set, we used all genes that were expressed in an ENCODE poly(A) RNA-seq data set from the same HepG2 cell line (experiment accession: ENCSR000CPE; file accession: ENCFF168QKW; TPM threshold = 1). We used the gProfiler package ( Kolberg et al. 2023 ) to perform GO analysis of biological processes, using FDR <0.05 as a cutoff to identify significantly enriched GO terms. DATA DEPOSITION Data are publicly available at the ENCODE portal. Accessions for eCLIP data used in this study are detailed in Supplemental Table S1 . The data set and machine learning clip arts in Figure 6 A, tote bag clip art in Supplemental Figure S1A , and filter clip art in Supplemental Figure S2 were obtained from Vecteezy; the sliding window clip art in Supplemental Figure S1A was obtained from Freepik; all clip arts are available under free licenses. CODE AVAILABILITY Code is publicly available at https://github.com/orbitalse/novel_RBP_discovery_methods . SUPPLEMENTAL MATERIAL Supplemental material is available for this article. COMPETING INTEREST STATEMENT Z.W. cofounded Rgenta Therapeutics, and she serves as a scientific advisor for the company and is a member of its board. ACKNOWLEDGMENTS We thank Dr. Athma A. Pai for critical reading of and feedback on the manuscript. This work was supported by the National Institutes of Health grant U24HG012343 to Z.W. Author contributions : S.I.E.: Problem formulation, conceptualization, data curation, data processing, investigation, formal analysis, methodology, visualization, writing of manuscript (original draft), review of manuscript, editing of manuscript. Z.W.: Research area, discussions, review of manuscript. All authors reviewed and approved the final version of the manuscript. Footnotes Article is online at http://www.rnajournal.org/cgi/doi/10.1261/rna.080892.125 . Freely available online through the RNA Open Access option. MEET THE FIRST AUTHOR Shaimae I. Elhajjajy. Open in a new tab Meet the First Author(s) is an editorial feature within RNA , in which the first author(s) of research-based papers in each issue have the opportunity to introduce themselves and their work to readers of RNA and the RNA research community. Shaimae I. Elhajjajy is the first author of this paper, “A novel NLP-based method and algorithm to discover RNA-binding protein (RBP) motifs, contexts, binding preferences, and interactions.” Shaimae completed her PhD in Dr. Zhiping Weng's lab in the Department of Genomics and Computational Biology at the University of Massachusetts Chan Medical School. Her research focuses on leveraging state-of-the-art machine learning techniques to decipher the activities and functions of RNA-binding proteins (RBPs) in the regulation of RNA processing, particularly through an investigative lens informed by natural language to define linguistic forms, rules, representations, and models that enable the characterization of RBP binding patterns. What are the major results described in your paper, and how do they impact this branch of the field? In this work, we present a novel linguistics-inspired framework and computational pipeline to investigate sequence-based determinants of RBP binding specificity. RBPs bind with a remarkable degree of selectivity that is imperative to ensuring accurate RNA regulation, gene expression, and transcriptomic diversity, but many open questions and challenges remain surrounding the mechanisms by which this selectivity is achieved. Despite significant advances in recent years, deciphering the precise determinants and functions of RBP binding specificity in the regulation of co- and post-transcriptional mRNA processing has still proven to be a challenging task. At its inception, our study was primarily motivated by three existing notions: First, that sequence context has been shown to facilitate specific RBP binding in vitro; second, that combinatorial interactions between RBPs are hypothesized to increase the specificity landscape; and third, that DNA and RNA possess an underlying, though enigmatic, “language.” In our study, we sought to leverage concepts from linguistics and natural language to define lexical, syntactic, and semantic rules to investigate sequence-based RBP binding patterns in vivo, including the precise contributions of sequence context and RBP–RBP interactions. Integrating state-of-the-art techniques in natural language processing and deep learning, we constructed predictive models of RBP binding under a multiple instance learning framework, designed a significant region extraction method, and demonstrated our models’ ability to achieve accurate and robust performance. We then developed a motif and context discovery algorithm that we extensively validated through multiple controls and experiments, which allowed us to comprehensively characterize novel motifs, sequence contexts, and binding preferences for an array of RBPs in the HepG2 and K562 cell lines. We further devised an integrative approach to discover novel RBP–RBP interactions, which we identified, annotated, and classified according to diverse proximal binding modes. In summary, our study has enabled the comprehensive characterization of RBP motifs, contexts, binding preferences, and interactions at large-scale in vivo, contributes to the broader understanding of contextual and interactive determinants of RBP binding specificity, and provides a novel linguistics view to model RBP binding. What led you to study RNA or this aspect of RNA science? At the start of my graduate studies, one of the major and widely investigated topics of research in the genomics and computational biology fields was the prediction of transcription factor binding to DNA and the investigation of their downstream regulatory effects in transcriptional regulation. By contrast, the prediction of RBP binding had not witnessed the same degree of widespread attention, or perhaps the field was still in its earlier stages. My broad research interests were at the intersection of bioinformatics and RNA biology, and my interest grew as I learned more about the complexity and essentiality of RBP-mediated RNA processing in health and disease through independent study, collaborations, and attendance at seminars hosted by the RNA Therapeutics Institute at UMass Chan. I therefore decided to pursue research in this area. The prominence of AI and deep learning applications in genomics also presented an exciting opportunity to extend similar concepts and techniques to transcriptomic applications. Further, my keen interest in languages and grammars presented a synergistic complement for researching a sequence-based problem like RBP binding from a linguistics-inspired lens. During the course of these experiments, were there any surprising results or particular difficulties that altered your thinking and subsequent focus? There were several findings during the course of this work that influenced our approach to and outlook of the project, which together contributed to its success. First, our conceptualization of our syntactic form, which clearly demarcates potential target k -mers from their flanking regions, proved to be highly beneficial by introducing structure into sequences, significantly facilitating downstream processing, and guiding interpretations. Second, during the development of our motif and context discovery algorithm, we found that considering the properties of candidate motif instances in isolation is generally not sufficient for good discovery, and that relationships between candidate motif instances and k -mers in their flanking regions are also important for accurate motif instance selection. Third, when formulating our RBP–RBP interaction discovery approach, we realized that we first needed to define a scope within which our method had the ability to effectively discover interactions based on the span and structure of our contexts. This led us to delineate different proximal RBP–RBP interaction modes, which helped us to perform interaction discovery with a more precise focus and at higher resolution. What are your subsequent near- or long-term career plans? My primary goal is to continue researching the fundamental mechanisms governing RBP binding specificity and their functional effects on RNA regulation, with a focus on their roles and implications in diseases such as neurological disorders and cancers. There remains much to uncover regarding the sequence-based determinants of RBP binding, and I aim to expand my linguistics-based focus to further decipher these elements. I am also interested in non-sequence-based features, such as RNA secondary structure, RNA modifications, and RBP–miRNA interactions, to name a few. I believe the integration of sequence- and non-sequence-based determinants will be crucial for developing a more complete, comprehensive model of RBP binding, which I believe will enhance our collective understanding of how RBPs dynamically, spatially, temporally, and combinatorially regulate cell type–specific mRNA isoform selection and expression under homeostatic and disease states. I look forward to leveraging cutting-edge machine learning techniques, AI methods, and mathematical modeling to address these open questions. Ultimately, I aspire to leverage the discoveries gleaned from my computational models to guide studies of disease pathogenesis and to develop therapeutic strategies for the treatment of diseases in which dysregulated RBP binding is the primary cause. REFERENCES Abdul-Manan N, Williams KR. 1996. hnRNP A1 binds promiscuously to oligoribonucleotides: utilization of random and homo-oligonucleotides to discriminate sequence from base-specific binding. Nucleic Acids Res 24: 4063–4070. 10.1093/nar/24.20.4063 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Achsel T, Brahms H, Kastner B, Bachi A, Wilm M, Lührmann R. 1999. A doughnut-shaped heteromer of human Sm-like proteins binds to the 3′-end of U6 snRNA, thereby facilitating U4/U6 duplex formation in vitro. EMBO J 18: 5789–5802. 10.1093/emboj/18.20.5789 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Achsel T, Stark H, Lührmann R. 2001. The Sm domain is an ancient RNA-binding motif with oligo(U) specificity. Proc Natl Acad Sci 98: 3685–3689. 10.1073/pnas.071033998 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Alipanahi B, Delong A, Weirauch MT, Frey BJ. 2015. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Nat Biotechnol 33: 831–838. 10.1038/nbt.3300 [ DOI ] [ PubMed ] [ Google Scholar ] Bailey TL. 2021. STREME: accurate and versatile sequence motif discovery. Bioinformatics 37: 2834–2840. 10.1093/bioinformatics/btab203 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bailey TL, Johnson J, Grant CE, Noble WS. 2015. The MEME suite. Nucleic Acids Res 43: W39–W49. 10.1093/nar/gkv416 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Begg BE, Jens M, Wang PY, Minor CM, Burge CB. 2020. Concentration-dependent splicing is enabled by Rbfox motifs of intermediate affinity. Nat Struct Mol Biol 27: 901–912. 10.1038/s41594-020-0475-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bentley DL. 2014. Coupling mRNA processing with transcription in time and space. Nat Rev Genet 15: 163–175. 10.1038/nrg3662 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Beusch I, Barraud P, Moursy A, Cléry A, Allain FH-T. 2017. Tandem hnRNP A1 RNA recognition motifs act in concert to repress the splicing of survival motor neuron exon 7. eLife 6: e25736. 10.7554/eLife.25736 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Binns D, Dimmer E, Huntley R, Barrell D, O'Donovan C, Apweiler R. 2009. QuickGO: a web-based tool for Gene Ontology searching. Bioinformatics 25: 3045–3046. 10.1093/bioinformatics/btp536 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bishop CM, Nasrabadi NM. 2006. Pattern recognition and machine learning. Springer, NY. [ Google Scholar ] Bohnsack KE, Yi S, Venus S, Jankowsky E, Bohnsack MT. 2023. Cellular functions of eukaryotic RNA helicases and their links to human diseases. Nat Rev Mol Cell Biol 24: 749–769. 10.1038/s41580-023-00628-5 [ DOI ] [ PubMed ] [ Google Scholar ] Bourdon S, Herviou P, Dumas L, Destefanis E, Zen A, Cammas A, Millevoi S, Dassi E. 2023. QUADRatlas: the RNA G-quadruplex and RG4-binding proteins database. Nucleic Acids Res 51: D240–D247. 10.1093/nar/gkac782 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Boyd K, Eng KH, Page CD. 2013. Area under the precision-recall curve: point estimates and confidence intervals. In Machine Learning and Knowledge Discovery in Databases (ed. Blockeel H, et al.), ECML PKDD 2013. Lecture Notes in Computer Science, Vol. 8190. Springer, Berlin, Heidelberg. [ Google Scholar ] Briata P, Bordo D, Puppo M, Gorlero F, Rossi M, Perrone-Bizzozero N, Gherzi R. 2016. Diverse roles of the nucleic acid-binding protein KHSRP in cell differentiation and disease. Wiley Interdiscip Rev RNA 7: 227–240. 10.1002/wrna.1327 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Busa VF, Favorov AV, Fertig EJ, Leung AKL. 2021. Spatial correlation statistics enable transcriptome-wide characterization of RNA structure binding. Cell Rep Methods 1: 100088. 10.1016/j.crmeth.2021.100088 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Carbonneau M-A, Cheplygina V, Granger E, Gagnon G. 2018. Multiple instance learning: a survey of problem characteristics and applications. Pattern Recognit 77: 329–353. 10.1016/j.patcog.2017.10.009 [ DOI ] [ Google Scholar ] Castello A, Fischer B, Hentze MW, Preiss T. 2013. RNA-binding proteins in Mendelian disease. Trends Genet 29: 318–327. 10.1016/j.tig.2013.01.004 [ DOI ] [ PubMed ] [ Google Scholar ] Chen H-C, Lin W-C, Tsay Y-G, Lee S-C, Chang C-J. 2002. An RNA helicase, DDX1, interacting with poly(A) RNA and heterogeneous nuclear ribonucleoprotein K. J Biol Chem 277: 40403–40409. 10.1074/jbc.M206981200 [ DOI ] [ PubMed ] [ Google Scholar ] Chen C-C, Yang J-H, Fu S-L, Lin W-J, Lin C-H. 2021. Arginine methylation of hnRNPK inhibits the DDX3-hnRNPK interaction to play an anti-apoptosis role in osteosarcoma cells. Int J Mol Sci 22: 9764. 10.3390/ijms22189764 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Choi HS, Hwang CK, Song KY, Law P-Y, Wei L-N, Loh HH. 2009. Poly(C)-binding proteins as transcriptional regulators of gene expression. Biochem Biophys Res Commun 380: 431–436. 10.1016/j.bbrc.2009.01.136 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ciafrè SA, Galardi S. 2013. microRNAs and RNA-binding proteins: a complex network of interactions and reciprocal regulations in cancer. RNA Biol 10: 935–942. 10.4161/rna.24641 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Conlon EG, Manley JL. 2017. RNA-binding proteins in neurodegeneration: mechanisms in aggregate. Genes Dev 31: 1509–1528. 10.1101/gad.304055.117 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Cook KB, Kazan H, Zuberi K, Morris Q, Hughes TR. 2011. RBPDB: a database of RNA-binding specificities. Nucleic Acids Res 39: D301–D308. 10.1093/nar/gkq1069 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Cornish-Bowden A. 1985. Nomenclature for incompletely specified bases in nucleic acid sequences: recommendations 1984. Nucleic Acids Res 13: 3021–3030. 10.1093/nar/13.9.3021 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Dassi E. 2017. Handshakes and fights: the regulatory interplay of RNA-binding proteins. Front Mol Biosci 4: 67. 10.3389/fmolb.2017.00067 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Dietterich TG, Lathrop RH, Lozano-Pérez T. 1997. Solving the multiple instance problem with axis-parallel rectangles. Artif Intell 89: 31–71. 10.1016/S0004-3702(96)00034-3 [ DOI ] [ Google Scholar ] Dominguez D, Freese P, Alexis MS, Su A, Hochman M, Palden T, Bazile C, Lambert NJ, Van Nostrand EL, Pratt GA, et al. 2018. Sequence, structure, and context preferences of human RNA binding proteins. Mol Cell 70: 854–867.e9. 10.1016/j.molcel.2018.05.001 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Duggimpudi S, Kloetgen A, Maney SK, Münch PC, Hezaveh K, Shaykhalishahi H, Hoyer W, McHardy AC, Lang PA, Borkhardt A, et al. 2018. Transcriptome-wide analysis uncovers the targets of the RNA-binding protein MSI2 and effects of MSI2's RNA-binding activity on IL-6 signaling. J Biol Chem 293: 15359–15369. 10.1074/jbc.RA118.002243 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Dumas L, Herviou P, Dassi E, Cammas A, Millevoi S. 2021. G-quadruplexes in RNA biology: recent advances and future directions. Trends Biochem Sci 46: 270–283. 10.1016/j.tibs.2020.11.001 [ DOI ] [ PubMed ] [ Google Scholar ] Durbin R, Eddy SR, Krogh A, Mitchison G. 1998. Biological sequence analysis. Cambridge University Press, Cambridge, UK. [ Google Scholar ] Faily S, Perveen R, Urquhart J, Chandler K, Clayton-Smith J. 2017. Confirmation that mutations in DDX59 cause an autosomal recessive form of oral-facial-digital syndrome: further delineation of the DDX59 phenotype in two new families. Eur J Med Genet 60: 527–532. 10.1016/j.ejmg.2017.07.009 [ DOI ] [ PubMed ] [ Google Scholar ] Feng H, Bao S, Rahman MA, Weyn-Vanhentenryck SM, Khan A, Wong J, Shah A, Flynn ED, Krainer AR, Zhang C. 2019. Modeling RNA-binding protein specificity in vivo by precisely registering protein-RNA crosslink sites. Mol Cell 74: 1189–1204.e6. 10.1016/j.molcel.2019.02.002 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Fisher RA. 1973. Statistical methods for research workers, 15th ed. Macmillan, London, UK. [ Google Scholar ] Frankiw L, Majumdar D, Burns C, Vlach L, Moradian A, Sweredoski MJ, Baltimore D. 2019. BUD13 promotes a type I interferon response by countering intron retention in Irf7. Mol Cell 73: 803–814.e6. 10.1016/j.molcel.2018.11.038 [ DOI ] [ PubMed ] [ Google Scholar ] Gabrovsek L, Collins KB, Aggarwal S, Saunders LM, Lau H-T, Suh D, Sancak Y, Trapnell C, Ong S-E, Smith FD, et al. 2020. A-kinase-anchoring protein 1 (dAKAP1)-based signaling complexes coordinate local protein synthesis at the mitochondrial surface. J Biol Chem 295: 10749–10765. 10.1074/jbc.RA120.013454 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] García-Mayoral MF, Díaz-Moreno I, Hollingworth D, Ramos A. 2008. The sequence selectivity of KSRP explains its flexibility in the recognition of the RNA targets. Nucleic Acids Res 36: 5290–5296. 10.1093/nar/gkn509 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gau B-H, Chen T-M, Shih Y-HJ, Sun HS. 2011. FUBP3 interacts with FGF9 3′ microsatellite and positively regulates FGF9 translation. Nucleic Acids Res 39: 3582–3593. 10.1093/nar/gkq1295 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gebauer F, Schwarzl T, Valcárcel J, Hentze MW. 2021. RNA-binding proteins in human genetic disease. Nat Rev Genet 22: 185–198. 10.1038/s41576-020-00302-y [ DOI ] [ PubMed ] [ Google Scholar ] Gehring NH, Wahle E, Fischer U. 2017. Deciphering the mRNP code: RNA-bound determinants of post-transcriptional gene regulation. Trends Biochem Sci 42: 369–382. 10.1016/j.tibs.2017.02.004 [ DOI ] [ PubMed ] [ Google Scholar ] Gerstberger S, Hafner M, Tuschl T. 2014. A census of human RNA-binding proteins. Nat Rev Genet 15: 829–845. 10.1038/nrg3813 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ghanbari M, Ohler U. 2020. Deep neural networks for interpreting RNA-binding protein target preferences. Genome Res 30: 214–226. 10.1101/gr.247494.118 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Good AL, Haemmerle MW, Oguh AU, Doliba NM, Stoffers DA. 2019. Metabolic stress activates an ERK/hnRNPK/DDX3X pathway in pancreatic β cells. Mol Metab 26: 45–56. 10.1016/j.molmet.2019.05.009 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Graf R, Munschauer M, Mastrobuoni G, Mayr F, Heinemann U, Kempa S, Rajewsky N, Landthaler M. 2013. Identification of LIN28B-bound mRNAs reveals features of target recognition and regulation. RNA Biol 10: 1146–1159. 10.4161/rna.25194 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hafner M, Landthaler M, Burger L, Khorshid M, Hausser J, Berninger P, Rothballer A, Jr AM, Jungkamp A-C, Munschauer M, et al. 2010. Transcriptome-wide identification of RNA-binding protein and microRNA target sites by PAR-CLIP. Cell 141: 129–141. 10.1016/j.cell.2010.03.009 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hall MP, Nagel RJ, Fagg WS, Shiue L, Cline MS, Perriman RJ, Donohue JP, Ares M Jr. 2013. Quaking and PTB control overlapping splicing regulatory networks during muscle cell differentiation. RNA 19: 627–638. 10.1261/rna.038422.113 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Harrow J, Frankish A, Gonzalez JM, Tapanari E, Diekhans M, Kokocinski F, Aken BL, Barrell D, Zadissa A, Searle S, et al. 2012. GENCODE: the reference human genome annotation for The ENCODE Project. Genome Res 22: 1760–1774. 10.1101/gr.135350.111 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] He X, Yuan J, Wang Y. 2021. G3BP1 binds to guanine quadruplexes in mRNAs to modulate their stabilities. Nucleic Acids Res 49: 11323–11336. 10.1093/nar/gkab873 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Herdy B, Mayer C, Varshney D, Marsico G, Murat P, Taylor C, D'Santos C, Tannahill D, Balasubramanian S. 2018. Analysis of NRAS RNA G-quadruplex binding proteins reveals DDX3X as a novel interactor of cellular G-quadruplex containing transcripts. Nucleic Acids Res 46: 11592–11604. 10.1093/nar/gky861 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Herrera F, Ventura S, Bello R, Cornelis C, Zafra A, Sanchez-Tarrago D, Vluymans S. 2016. Multiple instance learning, 1st ed. Springer, Cham, Switzerland. [ Google Scholar ] Hinkle ER, Wiedner HJ, Torres EV, Jackson M, Black AJ, Blue RE, Harris SE, Guzman BB, Gentile GM, Lee EY, et al. 2022. Alternative splicing regulation of membrane trafficking genes during myogenesis. RNA 28: 523–540. 10.1261/rna.078993.121 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Huang H, Weng H, Sun W, Qin X, Shi H, Wu H, Zhao BS, Mesquita A, Liu C, Yuan CL, et al. 2018. Recognition of RNA N6-methyladenosine by IGF2BP proteins enhances mRNA stability and translation. Nat Cell Biol 20: 285–295. 10.1038/s41556-018-0045-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Huang L, Li G, Du C, Jia Y, Yang J, Fan W, Xu Y-Z, Cheng H, Zhou Y. 2023. The polyA tail facilitates splicing of last introns with weak 3′ splice sites via PABPN1. EMBO Rep 24: e57128. 10.15252/embr.202357128 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hwang H, Jeon H, Yeo N, Baek D. 2024. Big data and deep learning for RNA biology. Exp Mol Med 56: 1293–1321. 10.1038/s12276-024-01243-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jarmoskaite I, Russell R. 2011. DEAD-box proteins as RNA helicases and chaperones. Wiley Interdiscip Rev RNA 2: 135–152. 10.1002/wrna.50 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Khoroshkin M, Buyan A, Dodel M, Navickas A, Yu J, Trejo F, Doty A, Baratam R, Zhou S, Lee SB, et al. 2024. Systematic identification of post-transcriptional regulatory modules. Nat Commun 15: 7872. 10.1038/s41467-024-52215-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kim YK, Maquat LE. 2019. UPFront and center in RNA decay: UPF1 in nonsense-mediated mRNA decay and beyond. RNA 25: 407–422. 10.1261/rna.070136.118 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kim JH, Hahm B, Kim YK, Choi M, Jang SK. 2000. Protein-protein interaction among hnRNPs shuttling between nucleus and cytoplasm. J Mol Biol 298: 395–405. 10.1006/jmbi.2000.3687 [ DOI ] [ PubMed ] [ Google Scholar ] Kim S, Kim S, Chang HR, Kim D, Park J, Son N, Park J, Yoon M, Chae G, Kim Y-K, et al. 2021. The regulatory impact of RNA-binding proteins on microRNA targeting. Nat Commun 12: 5057. 10.1038/s41467-021-25078-5 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kolberg L, Raudvere U, Kuzmin I, Adler P, Vilo J, Peterson H. 2023. g:Profiler—interoperable web service for functional enrichment analysis and gene identifier mapping (2023 update). Nucleic Acids Res 51: W207–W212. 10.1093/nar/gkad347 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] König J, Zarnack K, Rot G, Curk T, Kayikci M, Zupan B, Turner DJ, Luscombe NM, Ule J. 2010. iCLIP reveals the function of hnRNP particles in splicing at individual nucleotide resolution. Nat Struct Mol Biol 17: 909–915. 10.1038/nsmb.1838 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lang B, Yang J-S, Garriga-Canut M, Speroni S, Aschern M, Gili M, Hoffmann T, Tartaglia GG, Maurer SP. 2021. Matrix-screening reveals a vast potential for direct protein-protein interactions among RNA binding proteins. Nucleic Acids Res 49: 6702–6721. 10.1093/nar/gkab490 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lee DSM, Ghanem LR, Barash Y. 2020. Integrative analysis reveals RNA G-quadruplexes in UTRs are selectively constrained and enriched for functional associations. Nat Commun 11: 527. 10.1038/s41467-020-14404-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Li J, Chen Y, Xu X, Jones J, Tiwari M, Ling J, Wang Y, Harismendy O, Sen GL. 2019. HNRNPK maintains epidermal progenitor function through transcription of proliferation genes and degrading differentiation promoting mRNAs. Nat Commun 10: 4198. 10.1038/s41467-019-12238-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Li K, Wu J-L, Qin B, Fan Z, Tang Q, Lu W, Zhang H, Xing F, Meng M, Zou S, et al. 2020. ILF3 is a substrate of SPOP for regulating serine biosynthesis in colorectal cancer. Cell Res 30: 163–178. 10.1038/s41422-019-0257-1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Licatalosi DD, Mele A, Fak JJ, Ule J, Kayikci M, Chi SW, Clark TA, Schweitzer AC, Blume JE, Wang X, et al. 2008. HITS-CLIP yields genome-wide insights into brain alternative RNA processing. Nature 456: 464–469. 10.1038/nature07488 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Linder P, Jankowsky E. 2011. From unwinding to clamping - the DEAD box RNA helicase family. Nat Rev Mol Cell Biol 12: 505–516. 10.1038/nrm3154 [ DOI ] [ PubMed ] [ Google Scholar ] Liu C, Dou X, Zhao Y, Zhang L, Zhang L, Dai Q, Liu J, Wu T, Xiao Y, He C. 2024. IGF2BP3 promotes mRNA degradation through internal m7G modification. Nat Commun 15: 7421. 10.1038/s41467-024-51634-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Love MI, Huber W, Anders S. 2014. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 15: 550. 10.1186/s13059-014-0550-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Luo Y, Hitz BC, Gabdank I, Hilton JA, Kagda MS, Lam B, Myers Z, Sud P, Jou J, Lin K, et al. 2020. New developments on the Encyclopedia of DNA Elements (ENCODE) data portal. Nucleic Acids Res 48: D882–D889. 10.1093/nar/gkz1062 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Maklad A, Sedeeq M, Chan KM, Gueven N, Azimi I. 2023. Exploring Lin28 proteins: unravelling structure and functions with emphasis on nervous system malignancies. Life Sci 335: 122275. 10.1016/j.lfs.2023.122275 [ DOI ] [ PubMed ] [ Google Scholar ] Mayr C. 2019. What are 3′ UTRs doing? Cold Spring Harb Perspect Biol 11: a034728. 10.1101/cshperspect.a034728 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] McInnes L, Healy J, Astels S. 2017. hdbscan: Hierarchical density based clustering. J Open Source Softw 2: 205. 10.21105/joss.00205 [ DOI ] [ Google Scholar ] McInnes L, Healy J, Melville J. 2018. UMAP: uniform manifold approximation and projection for dimension reduction. arXiv [statML]. http://arxiv.org/abs/1802.03426 Montañés-Agudo P, Aufiero S, Schepers EN, van der Made I, Cócera-Ortega L, Ernault AC, Richard S, Kuster DWD, Christoffels VM, Pinto YM, et al. 2023. The RNA-binding protein QKI governs a muscle-specific alternative splicing program that shapes the contractile function of cardiomyocytes. Cardiovasc Res 119: 1161–1174. 10.1093/cvr/cvad007 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Müller-McNicoll M, Neugebauer KM. 2013. How cells get the message: dynamic assembly and function of mRNA-protein complexes. Nat Rev Genet 14: 275–287. 10.1038/nrg3434 [ DOI ] [ PubMed ] [ Google Scholar ] Nussbacher JK, Batra R, Lagier-Tourenne C, Yeo GW. 2015. RNA-binding proteins in neurodegeneration: seq and you shall receive. Trends Neurosci 38: 226–236. 10.1016/j.tins.2015.02.003 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Oughtred R, Rust J, Chang C, Breitkreutz B-J, Stark C, Willems A, Boucher L, Leung G, Kolas N, Zhang F, et al. 2021. The BioGRID database: a comprehensive biomedical resource of curated protein, genetic, and chemical interactions. Protein Sci 30: 187–200. 10.1002/pro.3978 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Palzer K-A, Bolduan V, Käfer R, Kleinert H, Bros M, Pautz A. 2022. The role of KH-type splicing regulatory protein (KSRP) for immune functions and tumorigenesis. Cells 11: 1482. 10.3390/cells11091482 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Pereira B, Billaud M, Almeida R. 2017. RNA-binding proteins in cancer: old players and new actors. Trends Cancer Res 3: 506–528. 10.1016/j.trecan.2017.05.003 [ DOI ] [ PubMed ] [ Google Scholar ] Quattrone A, Dassi E. 2019. The architecture of the human RNA-binding protein regulatory network. iScience 21: 706–719. 10.1016/j.isci.2019.10.058 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Quinlan AR, Hall IM. 2010. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26: 841–842. 10.1093/bioinformatics/btq033 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ramanathan M, Porter DF, Khavari PA. 2019. Methods to study RNA-protein interactions. Nat Methods 16: 225–234. 10.1038/s41592-019-0330-1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ray D, Kazan H, Cook KB, Weirauch MT, Najafabadi HS, Li X, Gueroussov S, Albu M, Zheng H, Yang A, et al. 2013. A compendium of RNA-binding motifs for decoding gene regulation. Nature 499: 172–177. 10.1038/nature12311 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ray D, Laverty KU, Jolma A, Nie K, Samson R, Pour SE, Tam CL, von Krosigk N, Nabeel-Shah S, Albu M, et al. 2023. RNA-binding proteins that lack canonical RNA-binding domains are rarely sequence-specific. Sci Rep 13: 5238. 10.1038/s41598-023-32245-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sayers EW, Bolton EE, Brister JR, Canese K, Chan J, Comeau DC, Connor R, Funk K, Kelly C, Kim S, et al. 2022. Database resources of the national center for biotechnology information. Nucleic Acids Res 50: D20–D26. 10.1093/nar/gkab1112 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Shamseldin HE, Rajab A, Alhashem A, Shaheen R, Al-Shidi T, Alamro R, Al Harassi S, Alkuraya FS. 2013. Mutations in DDX59 implicate RNA helicase in the pathogenesis of orofaciodigital syndrome. Am J Hum Genet 93: 555–560. 10.1016/j.ajhg.2013.07.012 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sharma S, Fazal FM. 2024. Localization of RNAs to the mitochondria-mechanisms and functions. RNA 30: 597–608. 10.1261/rna.079999.124 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sharma NK, Gupta S, Kumar A, Kumar P, Pradhan UK, Shankar R. 2021. RBPSpot: learning on appropriate contextual information for RBP binding sites discovery. Iscience 24. 10.1016/j.isci.2021.103381 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Street LA, Rothamel KL, Brannan KW, Jin W, Bokor BJ, Dong K, Rhine K, Madrigal A, Al-Azzam N, Kim JK, et al. 2024. Large-scale map of RNA-binding protein interactomes across the mRNA life cycle. Mol Cell 84: 3790–3809.e8. 10.1016/j.molcel.2024.08.030 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sun L, Xu K, Huang W, Yang YT, Li P, Tang L, Xiong T, Zhang QC. 2021. Predicting dynamic cellular protein-RNA interactions by deep learning using in vivo RNA structures. Cell Res 31: 495–516. 10.1038/s41422-021-00476-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Szklarczyk D, Kirsch R, Koutrouli M, Nastou K, Mehryary F, Hachilif R, Gable AL, Fang T, Doncheva NT, Pyysalo S, et al. 2023. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Res 51: D638–D646. 10.1093/nar/gkac1000 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Van Nostrand EL, Pratt GA, Shishkin AA, Gelboin-Burkhart C, Fang MY, Sundararaman B, Blue SM, Nguyen TB, Surka C, Elkins K, et al. 2016. Robust transcriptome-wide discovery of RNA-binding protein binding sites with enhanced CLIP (eCLIP). Nat Methods 13: 508–514. 10.1038/nmeth.3810 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Van Nostrand EL, Freese P, Pratt GA, Wang X, Wei X, Xiao R, Blue SM, Chen J-Y, Cody NAL, Dominguez D, et al. 2020a. A large-scale binding and functional map of human RNA-binding proteins. Nature 583: 711–719. 10.1038/s41586-020-2077-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Van Nostrand EL, Pratt GA, Yee BA, Wheeler EC, Blue SM, Mueller J, Park SS, Garcia KE, Gelboin-Burkhart C, Nguyen TB, et al. 2020b. Principles of RNA processing from analysis of enhanced CLIP maps for 150 RNA binding proteins. Genome Biol 21: 90. 10.1186/s13059-020-01982-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang J, Lee JE, Riemondy K, Yu Y, Marquez SM, Lai EC, Yi R. 2020. XPO5 promotes primary miRNA processing independently of RanGTP. Nat Commun 11: 1845. 10.1038/s41467-020-15598-x [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang Y, Xie Z, Kutschera E, Adams JI, Kadash-Edmondson KE, Xing Y. 2024. rMATS-turbo: an efficient and flexible computational tool for alternative splicing analysis of large-scale RNA-seq data. Nat Protoc 19: 1083–1104. 10.1038/s41596-023-00944-2 [ DOI ] [ PubMed ] [ Google Scholar ] Wilbert ML, Huelga SC, Kapeli K, Stark TJ, Liang TY, Chen SX, Yan BY, Nathanson JL, Hutt KR, Lovci MT, et al. 2012. LIN28 binds messenger RNAs at GGAGA motifs and regulates splicing factor abundance. Mol Cell 48: 195–206. 10.1016/j.molcel.2012.08.004 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yao C, Choi E-A, Weng L, Xie X, Wan J, Xing Y, Moresco JJ, Tu PG, Yates JR III, Shi Y. 2013. Overlapping and distinct functions of CstF64 and CstF64τ in mammalian mRNA 3’ processing. RNA 19: 1781–1790. 10.1261/rna.042317.113 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yee BA, Pratt GA, Graveley BR, Van Nostrand EL, Yeo GW. 2019. RBP-Maps enables robust generation of splicing regulatory maps. RNA 25: 193–204. 10.1261/rna.069237.118 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yeo GW, Coufal NG, Liang TY, Peng GE, Fu X-D, Gage FH. 2009. An RNA code for the FOX2 splicing regulator revealed by mapping RNA-protein interactions in stem cells. Nat Struct Mol Biol 16: 130–137. 10.1038/nsmb.1545 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhou H-L, Baraniak AP, Lou H. 2007. Role for Fox-1/Fox-2 in mediating the neuronal pathway of calcitonin/calcitonin gene-related peptide alternative RNA processing. Mol Cell Biol 27: 830–841. 10.1128/MCB.01015-06 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zinshteyn B, Rojas-Duran MF, Gilbert WV. 2017. Translation initiation factor eIF4G1 preferentially binds yeast transcript leaders containing conserved oligo-uridine motifs. RNA 23: 1365–1375. 10.1261/rna.062059.117 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Articles from RNA are provided here courtesy of The RNA Society ACTIONS View on publisher site PDF (5.8 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Prot2RNA (v1): Dataset, Code and Model Weights#1079147
• Knowledge-guided Pattern Discovery via Coupled Tensor Factorizations#451308
• Intronic RNA of yeast <i>RPL22</i> paralogs acts as an allosteric switch#1159494
• It Usually Begins with the Gene Ontology#732106
• Gapped BLAST and PSI-BLAST: a new generation of protein database search programs#6260