PEARL: integrative multi-omics classification and omics feature discovery via deep graph learning.

PEARL: integrative multi-omics classification and omics feature discovery via deep graph learning - 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 Bioinformatics . 2026 May 9;42(6):btag253. doi: 10.1093/bioinformatics/btag253 PEARL: integrative multi-omics classification and omics feature discovery via deep graph learning Quan Zhao 1, ‡ , Jiawen Du 2, ‡ , Muqing Zhou 3 , Xu-Wen Wang 4, ✉ , Quan Sun 5, 6, ✉ , Can Chen 7, 8, 9, 10, ✉ Editor: Jonathan Wren Author information Article notes Copyright and License information 1 Carolina Health Informatics Program, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 2 Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 3 Department of Genetics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 4 Channing Division of Network Medicine, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA 02115, United States 5 Center for Computational and Genomic Medicine, Children’s Hospital of Philadelphia, Phildaelphia, PA 19104, United States 6 Department of Pathology and Laboratory Medicine, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA 19104, United States 7 Carolina Health Informatics Program, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 8 Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 9 School of Data Science and Society, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States 10 Department of Mathematics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States ✉ Corresponding authors. Xu-Wen Wang, Channing Division of Network Medicine, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA 02115, United States. E-mail: spxuw@channing.harvard.edu ; Quan Sun, Center for Computational and Genomic Medicine, Children’s Hospital of Philadelphia, Phildaelphia, PA 19104, United States. E-mail: sunq@chop.edu ; Can Chen, School of Data Science and Society, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States. E-mail: canc@unc.edu . ‡ Equal contribution. Roles Quan Zhao : Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing - original draft Jiawen Du : Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing - original draft Muqing Zhou : Methodology, Software, Validation Xu-Wen Wang : Supervision, Writing - review & editing Quan Sun : Supervision, Writing - review & editing Can Chen : Conceptualization, Investigation, Project administration, Supervision, Writing - review & editing Jonathan Wren : Associate Editor Received 2025 Oct 19; Revised 2026 Mar 30; Accepted 2026 Apr 28; Collection date 2026 Jun. © The Author(s) 2026. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( https://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. PMC Copyright notice PMCID: PMC13224962  PMID: 42108553 Abstract Motivation Integrating multi-omics data provides valuable insights into biological processes by capturing information across multiple molecular layers, enabling a comprehensive understanding of complex diseases and driving advancements in precision medicine. However, existing computational methods for multi-omics integration face significant challenges, such as low reliability and poor generalizability, due to the high dimensionality and low sample size nature of omics data. Results To address these challenges, we present PEARL (Pearson-Enhanced spectrAl gRaph convoLutional networks), a novel deep graph learning method for biomedical classification and functional important omics features identification. PEARL leverages a simple yet effective learning architecture to achieve superior and robust performance in high-dimensional, low-sample-size multi-omics settings. Our results demonstrate that PEARL significantly outperforms existing state-of-the-art methods on both synthetic and real biomedical datasets. Furthermore, applied to Alzheimer’s disease (AD) brain multi-omics data, features prioritized by PEARL lead to functionally important genes that demonstrate significant enrichment in AD-related pathways. These findings highlight PEARL’s practical utility in biomedical research and its potential to enhance biological interpretability in multi-omics studies. Availability and implementation The source code of our computational framework is available at https://github.com/zqq121017/PEARL . 1 Introduction Advances in high-throughput sequencing technologies have revolutionized biomedical research by enabling the generation of diverse and extensive omics data, including genomic, transcriptomic, and epigenetic profiles, which are crucial for understanding complex diseases ( Huang et al. 2017 , Nicora et al. 2020 , Subramanian et al. 2020 , Wörheide et al. 2021 , Shahrajabian and Sun 2023 , Du et al. 2026 ). Integrating these multi-omics datasets provides a powerful approach to unraveling disease heterogeneity, uncovering functional important genes, and revealing dynamic molecular interactions underlying disease progression ( Hasin et al. 2017 , Karczewski and Snyder 2018 , Kreitmaier et al. 2023 , Chen et al. 2023a , Yan et al. 2024 ). Moreover, as personalized medicine advances, multi-omics integration has become indispensable for refining disease diagnosis, tailoring individualized therapies, and discovering new therapeutic targets ( Reska et al. 2021 , Mohr et al. 2024 , Molla and Bitew 2024 ). Therefore, developing robust multi-omics integration methods is salient to fully harness the potential of these diverse data types, leading to more accurate disease biology models, deeper understanding of molecular-level disease mechanisms, and more effective therapeutic strategies. Numerous computational approaches have been developed for patient stratification, disease subtype classification, and drug target discovery through the integration of multi-omics data ( Bersanelli et al. 2016 , Tini et al. 2019 , Wang et al. 2023 , Chen et al. 2023a ). In general, integration strategies can be categorized into three types: early integration, which directly concatenates features across omics; intermediate integration, which first learns omics-specific representations before combining them; and late integration, which fuses predictions from separate models. Among these, early integration methods typically combine features from multiple omics datasets into a single representation, while ensemble approaches aggregate predictions from separate classifiers, each trained individually on distinct omics data types ( Xu et al. 2021 , Vahabi and Michailidis 2022 , Spooner et al. 2025 ). However, these methods struggle to accommodate the intrinsic properties of omics data, such as low sample sizes and high dimensionality, and fail to capture the complex correlations across different types of omics data. Most recent approaches therefore fall into intermediate and late integration categories, as they provide more flexible and effective ways to capture cross-omics relationships. The classical statistical method DIABLO ( Singh et al. 2019 ) employs canonical correlation analysis for multi-view integration but still faces difficulties in effectively modeling the nonlinear relationships across omics platforms with low sample sizes. With advanced computational capabilities, deep learning methods have emerged as powerful tools for multi-omics analysis. For instance, DeepKEGG ( Lan et al. 2024 ) introduces biologically-informed hierarchical frameworks that leverage pathway–gene relationships for cancer recurrence prediction and functional genes discovery. DeepMo ( Lin et al. 2020 ) employs unsupervised learning approaches to stratify disease subtypes, while MOMA ( Moon and Lee 2022 ) utilizes attention mechanisms to identify prognostic subtypes from multi-omics profiles. More recently, DeepCentroid ( Xie et al. 2024 ) was introduced as a deep cascaded ensemble classifier that utilizes feature scanning and centroid-based learning to maintain robust performance in high-dimensional, small-sample-size regimes. Despite their improved performance over traditional methods, these deep learning approaches often require substantial training data and computational resources, presenting barriers to clinical adoption. Deep graph learning methodologies, such as graph convolutional networks (GCNs), have proven to be promising solutions for enhancing multi-omics integration. These methods excel at capturing both intra- and inter-omics relationships by representing samples as nodes in similarity networks. For example, MOGONET ( Wang et al. 2021 ) is a patient classification method that employs omics-specific GCNs constructed from sample similarity networks, integrating their predictions through a view correlation discovery network. Similarly, MOGDx ( Ryan et al. 2024 ) utilizes a similarity network fusion to construct integrated patient networks, enabling robust disease classification even with heterogeneous data sources. MVGNN ( Ren et al. 2024 ) and MOGAT ( Tanvir et al. 2024 ) further advance these approaches by incorporating attention mechanisms to fuze multi-omics features and weigh the significance of molecular interactions. Recent surveys highlight the growing use of GCNs in multi-omics cancer research ( Zohari and Chehreghani 2025 ). Large-scale comparative studies across 31 cancer types have benchmarked GCN, GAT, and GTN architectures, underscoring both their potential and the tradeoffs in scalability and robustness ( Alharbi et al. 2025b ). Other GCN-based approaches integrate embedded feature selection, such as differential expression and LASSO-based strategies, to enhance interpretability and biomarker discovery ( Alharbi et al. 2025a ). However, despite their potential, these methods often rely on highly complex architectures (e.g. graph attention networks), making them prone to overfitting and resulting in poor predictive performance when sample size is small. To address these challenges, we propose PEARL (Pearson-Enhanced spectrAl gRaph convoLutional networks), a novel deep graph learning method for biomedical classification and functional important genes identification. PEARL leverages a simple but effective learning architecture, including weighted Pearson correlation, simple spectral GCNs, and a multi-layer perceptron (MLP), offering robust performance against sample size variations and noise commonly present in omics data. We systematically evaluate PEARL and other existing state-of-the-art multi-omics integration methods using both simulated and real biomedical datasets. Our results demonstrate that PEARL significantly outperforms existing methods, particularly under conditions of small sample sizes, while effectively identifying critical disease-associated functional genes, highlighting its potential for advancing precision medicine. 2 Materials and methods 2.1 Overview of PEARL PEARL is a supervised intermediate multi-omics integration approach designed for biomedical classification and functional important omics features identification, effectively tackling the challenges of high dimensionality and limited sample sizes inherent in omics data. The framework consists of three key components: similarity network construction, feature refinement with GCNs, and feature integration ( Fig. 1 ). Figure 1. Open in a new tab Overview of PEARL. PEARL is a novel multi-omics integration method based on deep graph learning for biomedical classification and functional important genes identification. Leveraging a simple yet effective architecture, PEARL first defines the graph structure for omics data using a weighted Pearson correlation similarity matrix. Omics-specific simple spectral graph convolutional networks are then employed for robust feature extraction from each view. For multi-view feature integration, PEARL offers two alternative strategies: a concatenated MLP or a combined pooling MLP, allowing for flexible fusion before the final prediction. The efficacy of PEARL has been validated across both simulated and real biomedical datasets. First, PEARL constructs sample similarity networks for different omics data using weighted Pearson correlation ( Langfelder and Horvath 2008 ), where features with higher mean values have greater influence. This weighting strategy is based on the assumption that features with lower mean values are more susceptible to noise, thus offering greater robustness than other network construction techniques, such as cosine similarity. Importantly, the similarity networks are built solely from omics feature values and do not incorporate any class label information. This ensures that no information from the testing data labels leaks into the training process, preventing overfitting. Next, the similarity networks are represented as graphs for feature refinement using simple spectral graph convolutional networks (SSGConvs) ( Zhu and Koniusz 2021 ). SSGConv is a simple yet effective graph learning architecture that enhances feature smoothing, reduces overfitting, and improves computational efficiency compared to traditional GCNs, making it particularly well-suited for high-dimensional, low-sample-size omics data ( Zhu and Koniusz 2021 ). The refinement process involves an initial SSGConv layer followed by ReLU activation and dropout for regularization, as well as a second SSGConv layer with activation. This dual-layer design allows the model to capture both local and global patterns in the data while minimizing overfitting. Finally, the refined features from each omics dataset are integrated using combined pooling and/or concatenation, and fed into a MLP for the final prediction. The combined pooling method unifies various operations such as maximum, minimum, mean, and max–min pooling, which enables the model to capture inter-omics patterns, such as dominant signals from particular modalities and variability across omics layers, without requiring the construction of a complex heterogeneous graph. Meanwhile, concatenation effectively integrates multiple feature sets by preserving all available information from diverse data sources. Detailed descriptions of each component can be found in the following. 2.2 Network construction PEARL employs weighted Pearson correlation ( Langfelder and Horvath 2008 ) to assign feature-specific weights and compute a correlation network for each view. Unlike the traditional Pearson correlation coefficient, the weighted version incorporates a tailored weighting scheme for each feature, effectively addressing a key challenge in multi-omics data analysis, namely, that not all features contribute equally to underlying biological processes. Our approach leverages feature means as biologically-informed weights, based on the observation that features with higher average expression or activity levels often reflect greater biological relevance. Specifically, given two samples with feature vectors x and y , and a feature weight vector w derived from the mean value of each feature across all samples, we first compute the weighted feature representations as x weighted = w ⊙ x and y weighted = w ⊙ y , respectively, where ⊙ denotes element-wise multiplication. We then center the weighted features by subtracting their respective means, i.e. x centered = x weighted − mean ( x weighted ) and y centered = y weighted − mean ( y weighted ) . Finally, the weighted Pearson correlation between the two samples is computed as: corr = x centered ⊤ · y centered ‖ x centered ‖ · ‖ y centered ‖ . (1) Using this simple yet biologically meaningful formulation, we effectively construct a sample similarity network for each modality that captures latent relationships between samples. 2.3 Feature refinement PEARL utilizes SSGConv layers ( Zhu and Koniusz 2021 ) to refine sample features over the constructed similarity networks. SSGConv is a simple but sophisticated spectral-based deep learning architecture that is particularly well-suited for multi-omics data analysis scenarios characterized by high dimensionality and limited sample sizes. For each omics-specific similarity network, the SSGConv layer is defined as H ( l + 1 ) = ( α · A ^ H ( l ) + ( 1 − α ) · H ( l ) ) Θ ( l ) , (2) where A ^ is the normalized adjacency matrix of the network with added self-loops, H ( l ) is the feature matrix at layer l ( H ( 0 ) represents the original omics data), Θ ( l ) is the learned parameter matrix at layer l , and α is a weighting factor that balances the smoothness of the neighborhood aggregation with the retention of the previous features. SSGConv can be implemented efficiently using sparse matrix operations, crucial for handling high-dimensional omics data. In PEARL, we employ a two-layer SSGConv architecture to refine each omics feature representation over its corresponding similarity network, incorporating ReLU as the nonlinear activation function and dropout for regularization. The use of SSGConv equips PEARL with a robust framework for handling high-dimensional, low-sample-size multi-omics data, enabling effective feature learning while ensuring computational efficiency. 2.4 Feature integration PEARL offers two feature integration strategies: combined pooling and concatenation. The combined pooling approach integrates features from multiple omics views using a variety of pooling operations, such as maximum, minimum, average, and difference pooling, followed by a fully connected MLP for classification. Given the refined features from M views { h ( 1 ) , h ( 2 ) , … , h ( M ) } for each sample, the pooling operations can be mathematically defined as: MaxPool ( { h ( 1 ) , h ( 2 ) , … , h ( M ) } ) i = max m = 1 , 2 , … , M h i ( m ) , MinPool ( { h ( 1 ) , h ( 2 ) , … , h ( M ) } ) i = min m = 1 , 2 , … , M h i ( m ) , AvePool ( { h ( 1 ) , h ( 2 ) , … , h ( M ) } ) i = 1 M ∑ m = 1 M h i ( m ) , DiffPool ( { h ( 1 ) , h ( 2 ) , … , h ( M ) } ) i = max m = 1 , 2 , … , M h i ( m ) − min m = 1 , 2 , … , M h i ( m ) The pooled features are concatenated to form a unified feature vector for each sample, denoted by h . This unified feature vector is then passed through an MLP with a final output dimension of two for binary classification, i.e. h ( l + 1 ) = σ ( W ( l ) h ( l ) + b ( l ) ) , (3) where W ( l ) and b ( l ) represent the weights and biases of the l th layer, respectively, and σ denotes the activation function (e.g. Softmax). The second pooling approach, i.e. concatenation, is relatively simpler, where it directly concatenates the refined features from different omics views and feeds them into an MLP for final classification. In our experiment, we used combined pooling for multi-class datasets and concatenated MLP for binary classification tasks, while providing both integration strategies as user-definable input options in our published implementation. These two pooling strategies in PEARL provide a comprehensive yet flexible framework for integrating features across diverse omics types. 2.5 Feature identification Beyond its prediction power, PEARL also has the ability to identify the important features of omics data that contribute the most to prediction. Specifically, for a given feature, we first set its value vector to zero to simulate the absence of this feature as a perturbed dataset. This dataset is then processed as usual and is used to train a perturbed prediction model. We define the importance score of this feature by calculating the difference in model performance metrics ( F 1 score or F 1 macro) between the perturbed model and the baseline model that includes this feature. We note that each omics data may contribute to the prediction unequally. Thus, we normalize the importance scores across all the features within each omics layer. In our experiments, we performed five repeats of each dataset and calculated the average of feature importance scores to account for randomness of training process. 2.6 GO and KEGG enrichment analysis To explore the biological functions and pathways associated with the functional important genes identified by PEARL, we first used MethylationEPIC v2.0 annotation from Illumina and TargetScan to map the corresponding genes from epigeneticmarkers and miRNA. After obtaining the gene set, GO and KEGG pathway enrichment analysis were conducted with the enrichGO and enrichKEGG functions, respectively, from the clusterProfiler R package ( Yu et al. 2012 ). To account for multiple testing, FDR was calculated by adjusting P -values using the Benjamini–Hochberg method. GO terms and pathways with adjusted P -values <.05 are considered to be significant. 3 Results 3.1 PEARL demonstrates consistent superiority over existing methods across simulated scenarios We first evaluated PEARL on synthetic data generated by InterSIM ( Chalise et al. 2016 ), a method for generating synthetic multi-omics datasets that simulate real biological data. We considered three types of omics data, including methylation, gene expression, and protein data, and examined various cluster mean shifts and noise levels (i.e. difficulty levels) across sample sizes of 300, 500, and 800, resulting in a total of nine synthetic datasets ( Note 1 , available as supplementary data at Bioinformatics online). PEARL was compared with state-of-the-art multi-omics integration methods MOGONET ( Wang et al. 2021 ) (late integration), MOMA ( Moon and Lee 2022 ) (late integration), MOGDx ( Ryan et al. 2024 ) (intermediate integration) ( Note 2 , available as supplementary data at Bioinformatics online), as well as baseline classifiers, including K -nearest neighbor (KNN, early integration) and random forest (RF, early integration). We used accuracy, weighted average F 1 score ( F 1 weighted), and macro-averaged F 1 score ( F 1 macro) as evaluation metrics and reported the mean and SD of those metrics from 30 random train-test splits with train-test ratio 8:2. As illustrated in Fig. 2 , PEARL consistently outperforms all competing methods across varying sample sizes and difficulty levels, excelling in all evaluated metrics (accuracy, F 1 weighted, and F 1 macro). Notably, PEARL demonstrates exceptional robustness in low-sample-size settings ( n = 300 ). For example, at the intermediate difficulty level ( Fig. 2b ), PEARL achieves an average of 25% higher relative accuracy compared to the second-best method MOGONET ( P -value = 4.69E−10, one-sided paired Wilcoxon signed-rank). This performance advantage persists as sample sizes increase, with 30% and 32% higher accuracy compared to MOGONET in n = 500 and 800 scenarios ( P -values = 4.18E−11 and 9.97E−14) under the same intermediate difficulty level. Additionally, PEARL effectively adapts to tasks of varying complexity, handling both noise-heavy scenarios with subtle patterns and simpler cases with clear class separations. Across all conditions, PEARL significantly surpasses alternatives like MOGONET, MOMA, and MOGDx with an average of 27%, 28%, and 27% higher accuracy. These results on synthetic data highlight the superiority of PEARL in various real-world data scenarios, especially given that multi-omics datasets are often limited by small sample sizes and high noise. Figure 2. Open in a new tab Performance comparison on synthetic datasets. Three sets of parameters were used with InterSIM to generate synthetic datasets. Hard refers to the cases when the cluster mean shift (delta) is equal to 0.1 and the noise level is equal to 0.4; intermediate refers to the cases when the cluster mean shift is equal to 0.15 and the noise level is equal to 0.45; and then easy refers to the cases when the cluster mean shift is equal to 0.2, and the noise level is equal to 0.5. For each case, we used three sample sizes ( n ) to generate synthetic datasets, which are 300, 500, and 800. We utilized accuracy, F 1 weighted, and F 1 macro to measure the performance of PEARL, MOGONET, MOGDx, MOMA, KNN, and RF on the synthetic datasets. To assess statistical significance, we conducted paired t -tests comparing PEARL with the second-best method in each case. The resulting P -values are indicated with asterisks in the plots, where three asterisks denote a P -value less than .01 across all sample sizes. 3.2 PEARL outperforms existing methods in various real datasets We applied our method and other competing algorithms to three real datasets, including the Religious Orders Study/Memory and Aging Project (ROSMAP) for Alzheimer’s disease (AD) patients versus normal control (NC) classification ( Bennett et al. 2012 , De Jager et al. 2018 ), BReast CAncer (BRCA) ( Colaprico et al. 2016 ) for breast invasive carcinoma PAM50 subtype classification, and Lung Adenocarcinoma Collection (LUAD) from The Cancer Genome Atlas (TCGA) ( Albertina et al. 2016 ) for lung cancer pathological stage classification. ROSMAP and BRCA contain three types of omics data: mRNA expression data (mRNA), DNA methylation data (meth), and miRNA expression data (miRNA), and LUAD only contains mRNA and methylation data. The numbers of features used for training are included in Table 1 . Table 1. Summary of datasets. a Dataset Categories Number of features mRNA, meth, miRNA Number of features training mRNA, meth, miRNA ROSMAP NC: 169, AD: 182 55 889, 23 788, 309 200, 200, 200 BRCA Normal-like: 115, Basal-like: 131, HER2-enriched: 46, Luminal A: 436, Luminal B: 147 20 531, 20 106, 503 1000, 1000, 503 LUAD Stage I: 278, Stage II: 124, Stage III/IV: 110 20 531, 22 601, – 1000, 1000, – METABRIC Basal: 209, Her2: 224, Normal: 148, LumA: 700, LumB: 475 20 385, 13 184, – 782, 468, – Open in a new tab a mRNA refers to mRNA expression data. meth refers to DNA methylation data. miRNA refers to miRNA expression data. The ROSMAP dataset is for the classification of Alzheimer’s disease (AD) patients versus normal control (NC). The BRCA dataset is for breast invasive carcinoma (BRCA) PAM50 subtype classification with normal-like, basal-like, human epidermal growth factor receptor 2 (HER2)-enriched, Luminal A, and Luminal B subtypes. The LUAD dataset is for lung adenocarcinoma pathological stage classification with Stage I, Stage II, and Stage III/IV. METABRIC is an independent breast cancer cohort from the Molecular Taxonomy of Breast Cancer International Consortium, used for external validation of models trained on BRCA, with the same PAM50 subtype classification task using mRNA and DNA methylation data. To ensure fairness and comparability, we followed the exact same preprocessing and feature selection strategy as used in MOGONET. Specifically, features were pre-filtered by applying variance thresholds followed by ANOVA F -values to retain those with the greatest discriminative potential. This procedure is a standard practice in multi-omics studies to reduce dimensionality while preserving biologically relevant signals ( St and World 1989 , Golugula et al. 2011 , Chen et al. 2023b ). Each dataset was randomly split into 80% training and 20% testing sets, repeated 30 times to ensure robust performance evaluation. For the binary ROSMAP dataset, we used accuracy (ACC), F 1 score ( F 1), and area under the receiver operating characteristic curve (AUC) as evaluation metrics. Due to the multi-class characteristic of the BRCA and LUAD, we calculated ACC, average F 1 score weighted by support (F1_weighted), and macro-averaged F 1 score (F1_macro). We reported mean and SD of each evaluation metrics across the 30 realizations. To further ensure fairness and to address integration strategies across categories, we additionally included XGBoost ( Chen and Guestrin 2016 ), two most recent late integration methods [DeepCentroid ( Xie et al. 2024 ), 2024; fuseMLR ( Fouodo et al. 2025 ), 2025], and one intermediate integration method [iDeepViewLearn ( Wang et al. 2024 ), 2024] in the real data experiment. Our results show that PEARL consistently outperforms all other methods in all datasets ( Fig. 3 ), with notable improvements in all metrics. In the BRCA dataset with a challenging situation of classifying five different subtypes of breast cancer ( Fig. 3a–c ), PEARL outperforms the second best by 2% in accuracy, 3% in F 1 weighted, and 1% in F 1 macro ( P -value = .14, .018, .05, respectively). In LUAD dataset ( Fig. 3d–f ) with situation of classifying three pathological states, PEARL demonstrates remarkable improvements compared to the second best by 5% in accuracy, 12% in F 1 weighted and 13% in F 1 macro ( P -value = 1E−3, 1.6E−7, 1E−5). Figure 3. Open in a new tab Performance comparison on real datasets. We compared the performance of PEARL with nine other methods, MOGONET, MOGDx, MOMA, KNN, RF, fuseMLR, iDeepViewLearn, XGBoost, and DeepCentroid on the BRCA, LUAD, and ROSMAP datasets. We used accuracy, F 1 weighted, and F 1 macro as metrics on the BRCA and LUAD multi-view datasets. For the ROSMAP binary dataset, we calculated accuracy, F 1 score, and AUROC as evaluation metrics. To assess statistical significance, we conducted Wilcoxon signed-rank tests comparing PEARL with the second-best method in each case. The resulting P -values are indicated with asterisks in the plots [ P -value < .1(*)/.05(**)/.01(***)]. In the ROSMAP dataset ( Fig. 3g–i ), which presents the challenging task of distinguishing AD patients from NCs, PEARL demonstrates improvements compared to the second-best model, XGBoost, achieving gains of 2% in accuracy and AUC, and 3% in F 1 score ( P -value = .013, .05, .018, respectively). These improvements are particularly meaningful in the context of AD diagnosis, where even small increases in accuracy can potentially translate to earlier and more reliable disease detection, potentially enabling earlier intervention strategies. All methods were run with their default parameter configurations to ensure consistency in comparative evaluation in above analysis. To evaluate the impact of parameter tuning on existing methods, we adopted a grid search strategy, and the results confirm the superiority of PEARL ( Fig. 1 , available as supplementary data at Bioinformatics online). We further performed an ablation study to examine the impact of our key innovations in PEARL ( Note 3 , available as supplementary data at Bioinformatics online). The results demonstrate that removing any one of our key components, including weighted Pearson correlation, SSGConv, or feature integration, causes a stepwise decline in all performance metrics, confirming that each innovation is essential for PEARL’s best results ( Fig. 2 , available as supplementary data at Bioinformatics online). To further demonstrate the real-world applicability of our framework, we conducted an independent cross-study validation. We trained PEARL on the BRCA dataset and evaluated its performance on METABRIC dataset ( Curtis et al. 2012 , Pereira et al. 2016 ) ( Table 1 ), which serves as an independent resource for breast cancer multi-omics. In this independent validation, PEARL outperformed the second-best model (i.e. MOGONET) by 42.4% in accuracy, 97.3% in F 1 weighted and 33.3% in F 1 macro ( Fig. 4 ). This independent validation underscores PEARL’s ability to transcend study specific batch effects, confirming the reliability for clinical applications across different data sources. Figure 4. Open in a new tab Performance on independent dataset. We trained PEARL on BRCA and tested on METABRIC cohort. We compared the performance of PEARL with nine other methods, MOGONET, MOGDx, MOMA, KNN, RF, fuseMLR, iDeepViewLearn, XGBoost, and DeepCentroid. Accuracy, F 1 weighted, and F1 macro were used as evaluation metrics. 3.3 Identification of disease-associated genes and pathways Finally, we investigated the key omics features that contributed the most to the prediction ( Tables 2–4 ). Details of the identification process are provided in Section 2. Table 2. Feature identification results from ROSMAP by PEARL. Omics data type Features mRNA expression ENSG00000117266.11, ENSG00000155980.6, ENSG00000105643.3, ENSG00000204219.5, ENSG00000182310.7, ENSG00000168743.8, ENSG00000071054.10, ENSG00000102032.8, ENSG00000166682.6 DNA methylation cg07232688, cg22518733, cg27091787, cg25736482, cg14387505, cg12447832, cg08418111, cg02008416, cg06946880, cg10157098, cg24765079, cg04983977, cg14847483 miRNA expression hsa-miR-517c_hsa-miR-519a, hsa-miR-424, hsa-miR-216a, hsa-miR-651, hsa-miR-127-3p, hsa-miR-640, hsa-miR-199b-5p, hsa-miR-199a-5p Open in a new tab Table 3. Feature identification results from BRCA by PEARL. Omics data type Features mRNA expression SOX11|6664, HPDL|84842, CNGA1|1259, KIF2C|11004, AR|367, CAPN13|92291, CDC6|990, KIF14|9928, PTCHD1|139411, C1orf64|149563, PHYHD1|254295 DNA methylation C14orf72, OR1J4, TM4SF18, LEMD1, ROPN1, ATP8B1, C21orf82, GPR37L1, KDM2B, ATP10B, LGALS2 miRNA expression hsa-mir-187, hsa-mir-9–1, hsa-mir-9–2, hsa-mir-203, hsa-mir-21, hsa-mir-184, hsa-mir-944, hsa-mir-551b Open in a new tab Table 4. Feature identification results from LUAD by PEARL. Omics data type Features mRNA expression TMEM8B, VCL, MRPL19, PHF1, PLK1, NAPB, SERPINI1, RPS6KA5, CPT1B, C2CD4D-AS1, ZC3H15, TRIM13, PGM1, NPHP3, GPR55, MAT2A, PGM5, ORMDL3, CLK4, KLHDC1, ZC3HAV1L, PYGL, FOXM1, SRGAP3, AGPS, LRFN4 DNA methylation cg21770145, cg07015079, cg05322019, cg08554114 Open in a new tab In the ROSMAP dataset, we chose the top 50 features with highest importance scores, across mRNA, DNA methylation, and miRNA data. We annotated the cpg sites to their nearest gene using Illumina MethylationEPIC v2.0 manifest file. For miRNA, they were filtered to those present in the TargetScan human database ( McGeary et al. 2019 ) and further mapped to corresponding miRNA families. Further, the target genes were extracted from the TargetScan’s target predictions. All gene symbols were subsequently converted to ENSEMBL gene IDs using the org . Hs.eg.db ( Carlson et al. 2019 ) annotation database. Overall, these 50 features correspond to 2861 genes. Then, we performed enrichment analysis for the functional genes identified in ROSMAP, using both Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) databases ( Kanehisa and Goto 2000 ). GO terms within three ontology categories are considered: biological process (BP), cellular components (CCs), and molecular function (MF). To account for multiple testing, false discovery rate (FDR) correction was applied and the adjusted P -values were reported. We found these genes were enriched for several GO terms that have been reported in other studies to be associated with AD. Notable examples include small GTPase-mediated signal transduction (biological process, GO:0007264, adjusted P -value = 1.07E−16, Fig. 5 ) ( Hu et al. 2015 ), axongenesis (biological process, GO:0007409, adjusted P -value = 5.11E−13, Fig. 5 ) ( Huang et al. 2018 ), axon development (biological process, GO:0061564, adjusted P -value = 2.56E−12, Fig. 5 ) ( Santa-Maria et al. 2015 ), glutamatergic synapse (cellular composition, GO:0098978, adjusted P -value = 1.36E−23, Fig. 3 , available as supplementary data at Bioinformatics online) ( Ma et al. 2024 ), and protein serine/threonine kinase activity (molecular function, GO:0004674, adjusted P -value = 2.98E−9, Fig. 4 , available as supplementary data at Bioinformatics online) ( Zhang et al. 2022 ). The full GO enrichment results are provided in Tables 3–5 , available as supplementary data at Bioinformatics online. Figure 5. Open in a new tab Top 20 enriched biological processes from pathway analysis ordered by adjusted P -value. Each dot represents a biological process, with the x -axis showing the gene ratio (the ratio of genes associated with a given term to the total number of genes analyzed), and the y -axis listing the description of each biological process. Color corresponds to the adjusted P -value, with more intense red indicating higher significance. Dot size reflects the gene count. Several notable significantly enriched pathways include processes related to GTPase-mediated signaling, axongenesis, and axon development. The KEGG pathway enrichment analysis also identified several important pathways that are known to be critically related to AD ( Table 6and Worldsupplementary data at Bioinformatics online). The two most notable pathways are axon guidance (adjusted P -value = 2.06E−13, Fig. 5 , available as supplementary data at Bioinformatics online) and MAPK signaling pathway (adjusted P -value = 1.40E−10, Fig. 5 , available as supplementary data at Bioinformatics online). Axon-guidance molecules are playing import roles in occurrence and development of AD by influencing in connectivity and repair mechanism ( Lin et al. 2009 ), and the activate MAPK signaling pathway has been considered to contribute to AD via multiple mechanism like neuronal apoptosis and transcriptional and enzymatic activation of β - and γ -secretases ( Kim and Choi 2010 ). These findings demonstrate that PEARL effectively identifies trait-associated functional genes from multi-omics data, supporting its utility in uncovering biologically meaningful insights. 4 Discussion Multi-omics integration plays crucial roles in understanding disease mechanisms, providing pivotal perspectives in disease diagnosis and prognosis. In this study, we present PEARL, a novel framework for multi-omics data integration and analysis that advances the field through several key innovations. Our comprehensive evaluation demonstrates PEARL’s superior performance across both synthetic and real-world datasets, while also providing biologically meaningful insights through its feature identification capabilities. PEARL’s performance can be attributed to several technical design choices. First, the weighted Pearson correlation approach in the preprocessing stage provides an alternative to traditional similarity measures for constructing patient similarity network. By incorporating a position-based weighting scheme, PEARL captures the varying importance of different molecular features, reflecting the biological reality that not all features contribute equally to disease mechanisms. This builds upon insights from previous studies showing that representing data as patient similarity networks can retain information and have better performance compared to standard Euclidean methods ( Li et al. 2015 , Gliozzo et al. 2020 ). Our sensitivity analysis further validates this choice, as the weighted Pearson approach consistently outperforms standard Pearson correlation, Spearman correlation, and cosine similarity ( Fig. 6 , available as supplementary data at Bioinformatics online). Second, the use of simple but effective SSGConv layers in PEARL’s GCN architecture provides another crucial advantage. Unlike traditional approaches that rely on simple feature concatenation ( Sun et al. 2007 ) or majority voting for network integration ( Gliozzo et al. 2022 ), SSGConv’s ability to balance between feature preservation and neighborhood aggregation is especially valuable in handling the noise and heterogeneity inherent in multi-omics data. This is reflected in the consistently smaller SDs observed in our results, particularly in the ROSMAP dataset. Third, the dual integration strategies, combined pooling and concatenation MLP, demonstrate PEARL’s robustness to different classification scenarios. This flexibility addresses limitations identified in existing methods ( Kim et al. 2015 ) where integrative analysis often struggles with different types of classification tasks. Each component of PEARL contributes to its superiority over competing methods, validated through our results with enhanced performance for both synthetic and real datasets. PEARL addresses several limitations of existing approaches. Unlike MOGONET, which relies on a fixed integration framework, PEARL offers more flexible handling of different data types and classification scenarios. Compared to MOMA, which faces challenges with complex optimization in multi-task learning settings that can create negative transfer effects, PEARL’s architecture avoids these conflicts while maintaining computational efficiency. Similarly, while MOGDx effectively handles missing data through network fusion, it lacks the sophisticated feature weighting mechanisms present in PEARL. The weighted Pearson correlation and SSGConv architecture in PEARL provide a more robust framework for high-dimensional, low-sample-size data compared to these methods. Furthermore, PEARL’s adaptive integration approach delivers superior performance across varying classification scenarios without extensive parameter tuning, offering better scalability and robustness as demonstrated in our synthetic data experiments across different sample sizes. Computationally, PEARL is designed to balance predictive performance with resource efficiency. By utilizing a SSGConv, the framework avoids the expensive matrix operations often required in attention-based architectures. The overall time complexity is approximately O ( M · N 2 · d + K · | E | · d ) , where M is the number of omics modalities, N is the sample size, d is the number of omics features, K is the number of propagation steps, and | E | is the number of edges in the similarity graph. Since feature propagation is handled as a one-time preprocessing step, PEARL requires fewer trainable parameters. Empirically, using a standard workstation (13th Gen Intel Core i7-13700K, NVIDIA GeForce RTX 3080 Ti with 12 GB VRAM), PEARL trained on ROSMAP dataset in 928 s, with a peak memory utilization of 2.48 GB. PEARL not only achieves higher accuracy in classification tasks, but also has the ability to identify biologically meaningful important features from various omics data, providing valuable insights into disease mechanisms. Like recent successful methods ( Serra et al. 2015 ), PEARL effectively integrates multiple omics data types while maintaining interpretability. In the ROSMAP data analysis, the underlying genes of prioritized omics features by PEARL show significant enrichment in AD-related pathways, particularly axon guidance and MAPK signaling, consistent with current understanding of AD pathophysiology. Of particular note is PEARL’s ability to identify functionally important pathway associations. Like recent network-based approaches ( Wang et al. 2014 ), PEARL excels at capturing strong latent relationships between samples. However, PEARL goes beyond simple network fusion by incorporating weighted correlations and advanced graph convolution techniques, allowing for more nuanced feature importance estimation. To further enhance model parsimony, we embedded a LASSO (L1) penalty into the PEARL loss function. The primary advantage of LASSO is to produce a sparser, more interpretable model by forcing the weights of noncontributory features to exactly zero. In addition, it is more computational efficient for feature selection than the original method, e.g. setting a feature to zero and re-training the entire model. Our analysis shows that the penalized model yielded comparable classification performance to the original framework ( Tables 3 and 4 , available as supplementary data at Bioinformatics online). Furthermore, the feature importance scores between two models were highly correlated ( ρ = 0.95 and 0.99 for the BRCA and ROSMAP datasets, respectively), underscoring the robustness of PEARL’s native structure. However, the use of a LASSO penalty has some limitations, particularly in biological data ( O’Brien 2016 ). It can arbitrarily select one feature from a group of highly correlated variables, such as genes within the same pathway, while discarding the rest. This can lead to the loss of potentially important co-linear signals. Therefore, a tradeoff exists that while the original PEARL framework may be preferable for retaining a comprehensive feature set, the LASSO option offers a valuable alternative for studies where model simplicity and the identification of a minimal set of biomarkers are the primary goals. While PEARL demonstrates significant advantages, several areas warrant further investigation. First, the current implementation focuses on classification tasks; extending the framework to handle regression problems could broaden its clinical applications. This could potentially build upon approaches used in previous studies for survival prediction ( Chaudhary et al. 2018 ). Second, while PEARL’s complexity is highly optimized for current high-dimensional, small-sample-size datasets, the O ( N 2 ) space complexity required for the dense similarity matrix may present memory challenges with ultra-large-scale multi-omics studies (e.g. biobank-scale cohorts). This reflects a common problem in the field ( Poirion et al. 2018 ), and future work utilizing subnetwork extraction or alternative sparse-graph methods could improve scalability. Third, although PEARL effectively uses combined pooling MLP to capture inter-omics information, explicitly modeling these interactions through cross-view graph connections remains an important objective. Developing robust methods to construct unified heterogeneous networks without exacerbating overfitting in low-sample-size cohorts could be the future direction. Fourth, we recognize that the success of feature identification may be influenced by the use of well-characterized datasets such as BRCA and ROSMAP. While population-scale biobanks could offer an alternative for validation, we could not find external datasets that exactly match the features used in our current study. Additionally, the reliance on ICD codes for case identification in large biobanks may also introduce diagnostic biases, which may yield features that reflect clinical practice patterns rather than the molecular driven signals captured in curated research cohorts. Last, another promising direction would be to integrate PEARL with emerging approaches in multi-omics integration, such as canonical correlation analysis methods ( Singh et al. 2019 ) or advanced group-regularized techniques ( Van De Wiel et al. 2016 ), which may provide additional insights into cellular heterogeneity and tissue-specific patterns in disease progression. In summary, PEARL represents a significant advancement in multi-omics data integration and analysis. Its superior performance across various datasets, coupled with its ability to identify biologically meaningful features, positions it as a valuable tool for precision medicine applications. Building on insights from previous unsupervised integration methods ( Shen et al. 2009 ), PEARL extends these capabilities to supervised learning while maintaining robust performance in the presence of noise. As multi-omics studies continue to grow in scale and complexity, we anticipate PEARL will become increasingly valuable for understanding disease mechanisms and advancing personalized medicine approaches. Supplementary Material btag253_Supplementary_Data btag253_supplementary_data.pdf (592.9KB, pdf) Acknowledgements The authors thank the anonymous reviewers for their valuable suggestions. Contributor Information Quan Zhao, Carolina Health Informatics Program, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States. Jiawen Du, Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States. Muqing Zhou, Department of Genetics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States. Xu-Wen Wang, Channing Division of Network Medicine, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA 02115, United States. Quan Sun, Center for Computational and Genomic Medicine, Children’s Hospital of Philadelphia, Phildaelphia, PA 19104, United States; Department of Pathology and Laboratory Medicine, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA 19104, United States. Can Chen, Carolina Health Informatics Program, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States; Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States; School of Data Science and Society, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States; Department of Mathematics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, United States. Author contributions Quan Zhao (Data curation [equal], Formal analysis [equal], Methodology [lead], Software [lead], Validation [equal], Visualization [equal], Writing—original draft [equal]), Jiawen Du (Data curation [equal], Formal analysis [equal], Methodology [supporting], Software [supporting], Validation [equal], Visualization [equal], Writing—original draft [equal]), Muqing Zhou (Methodology [supporting], Software [supporting], Validation [equal]), Xu-Wen Wang (Supervision [equal], Writing—review & editing [equal]), Quan Sun (Supervision [equal], Writing—review & editing [equal]), and Can Chen (Conceptualization [lead], Investigation [lead], Project administration [lead], Supervision [equal], Writing—review & editing [equal]) Supplementary material Supplementary material is available at Bioinformatics online. Conflict of interests The authors declare that they have no competing interests. Funding None declared. Data availability The ROSMAP dataset was obtained from AMP-AD Knowledge Portal ( https://adknowledgeportal.synapse.org/ ). Omics data of BRCA and LUAD was obtained from The Cancer Genome Atlas Program (TCGA) through Broad GDAC Firehose ( https://gdac.broadinstitute.org/ ). PAM50 breast cancer subtypes of TCGA BRCA patients were obtained through the TCGAbiolinks R package v2.12.6. The source data and code of our computational framework is available at https://github.com/zqq121017/PEARL . The archival version of the code is preserved on Zenodo at https://doi.org/10.5281/zenodo.19875866 ( zqq121017 2026 ). References Albertina B, Watson M, Holback C  et al.  The cancer genome atlas lung adenocarcinoma collection (TCGA-LUAD). Cancer Imaging Arch  2016. [ Google Scholar ] Alharbi F, Budhiraja N, Vakanski A  et al.  Interpretable graph kolmogorov-arnold networks for multi-cancer classification and biomarker identification using multi-omics data. Sci Rep  2025a;15:27607. 10.1038/s41598-025-13337-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Alharbi F, Vakanski A, Zhang B  et al.  Comparative analysis of multi-omics integration using graph neural networks for cancer classification. IEEE Access  2025b;13:37724–36. 10.1109/ACCESS.2025.3540769 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bennett DA, Schneider JA, Arvanitakis Z  et al.  Overview and findings from the religious orders study. Curr Alzheimer Res  2012;9:628–45. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bersanelli M, Mosca E, Remondini D  et al.  Methods for the integration of multi-omics data: mathematical aspects. BMC Bioinformatics  2016;17 Suppl 2:15–177. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Carlson M, Falcon S, Pages H  et al.  org. hs. eg. db: genome wide annotation for human. R Package Version  2019;3:3. [ Google Scholar ] Chalise P, Raghavan R, Fridley BL.  Intersim: simulation tool for multiple integrative ‘omic datasets’. Comput Methods Programs Biomed  2016;128:69–74. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chaudhary K, Poirion OB, Lu L  et al.  Deep learning-based multi-omics integration robustly predicts survival in liver cancer. Clin Cancer Res  2018;24:1248–59. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chen C, Wang J, Pan D  et al.  Applications of multi-omics analysis in human diseases. MedComm (2020)  2023a;4:e315. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chen C, Weiss ST, Liu Y-Y.  Graph convolutional network-based feature selection for high-dimensional and low-sample size data. Bioinformatics  2023b;39:btad135. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . New York, USA: Association for Computing Machinery, 2016, 785–94. Colaprico A, Silva TC, Olsen C  et al.  TCGAbiolinks: an R/Bioconductor package for integrative analysis of TCGA data. Nucleic Acids Res  2016;44:e71. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Curtis C, Shah SP, Chin S-F  et al.  The genomic and transcriptomic architecture of 2,000 breast tumours reveals novel subgroups. Nature  2012;486:346–52. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] De Jager PL, Ma Y, McCabe C  et al.  A multi-omic atlas of the human frontal cortex for aging and Alzheimer’s disease research. Sci Data  2018;5:180142. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Du J, Zhou M, Wang H  et al. Multi-omics integration predicts the incidence of 17 diseases in the UK Biobank. Nat Comm 2026. (in press). [ DOI ] [ PMC free article ] [ PubMed ] Fouodo CJ, Bleskina M, Szymczak S.  fuseMLR: an R package for integrative prediction modeling of multi-omics data. BMC Bioinformatics  2025;26:221. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gliozzo J, Perlasca P, Mesiti M  et al.  Network modeling of patients’ biomolecular profiles for clinical phenotype/outcome prediction. Sci Rep  2020;10:3612. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gliozzo J, Mesiti M, Notaro M  et al.  Heterogeneous data integration methods for patient similarity networks. Brief Bioinform  2022;23:bbac207. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Golugula A, Lee G, Madabhushi A.  Evaluating feature selection strategies for high dimensional, small sample size datasets. In: 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society. New York, USA: IEEE, 2011, 949–52. [ DOI ] [ PubMed ] [ Google Scholar ] Hasin Y, Seldin M, Lusis A.  Multi-omics approaches to disease. Genome Biol  2017;18:83. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hu W, Lin X, Chen K.  Integrated analysis of differential gene expression profiles in hippocampi to identify candidate genes involved in Alzheimer’s disease. Mol Med Rep  2015;12:6679–87. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Huang J-L, Xu Z-H, Yang S-M  et al.  Identification of differentially expressed profiles of Alzheimer’s disease associated circular RNAs in a Panax notoginseng saponins-treated Alzheimer’s disease mouse model. Comput Struct Biotechnol J  2018;16:523–31. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Huang S, Chaudhary K, Garmire LX.  More is better: recent progress in multi-omics data integration methods. Front Genet  2017;8:84. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kanehisa M, Goto S.  Kegg: Kyoto encyclopedia of genes and genomes. Nucleic Acids Res  2000;28:27–30. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Karczewski KJ, Snyder MP.  Integrative omics for health and disease. Nat Rev Genet  2018;19:299–310. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kim D, Joung J-G, Sohn K-A  et al.  Knowledge boosting: a graph-based integration approach with multi-omics data and genomic knowledge for cancer clinical outcome prediction. J Am Med Inform Assoc  2015;22:109–20. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kim EK, Choi E-J.  Pathological roles of MAPK signaling pathways in human diseases. Biochim Biophys Acta  2010;1802:396–405. [ DOI ] [ PubMed ] [ Google Scholar ] Kreitmaier P, Katsoula G, Zeggini E.  Insights from multi-omics integration in complex disease primary tissues. Trends Genet  2023;39:46–58. [ DOI ] [ PubMed ] [ Google Scholar ] Lan W, Liao H, Chen Q  et al.  DeepKEGG: a multi-omics data integration framework with biological insights for cancer recurrence prediction and biomarker discovery. Brief Bioinform  2024;25:bbae185. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Langfelder P, Horvath S.  WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics  2008;9:559. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Li L, Cheng W-Y, Glicksberg BS  et al.  Identification of type 2 diabetes subgroups through topological analysis of patient similarity. Sci Transl Med  2015;7:311ra174. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lin L, Lesnick TG, Maraganore DM  et al.  Axon guidance and synaptic maintenance: preclinical markers for neurodegenerative disease and therapeutics. Trends Neurosci  2009;32:142–9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lin Y, Zhang W, Cao H  et al.  Classifying breast cancer subtypes using deep neural networks based on multi-omics data. Genes (Basel)  2020;11:888. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ma Y, Reyes-Dumeyer D, Piriz A  et al.  Epigenetic and genetic risk of Alzheimer disease from autopsied brains in two ethnic groups. Acta Neuropathol  2024;148:27. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] McGeary SE, Lin KS, Shi CY  et al.  The biochemical basis of microRNA targeting efficacy. Science  2019;366:eaav1741. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Mohr AE, Ortega-Santos CP, Whisner CM  et al.  Navigating challenges and opportunities in multi-omics integration for personalized healthcare. Biomedicines  2024;12:1496. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Molla G, Bitew M.  Revolutionizing personalized medicine: synergy with multi-omics data generation, main hurdles, and future perspectives. Biomedicines  2024;12:2750. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Moon S, Lee H.  Moma: a multi-task attention learning algorithm for multi-omics data interpretation and classification. Bioinformatics  2022;38:2287–96. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nicora G, Vitali F, Dagliati A  et al.  Integrated multi-omics analyses in oncology: a review of machine learning methods and tools. Front Oncol  2020;10:1030. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] O’brien CM.  Statistical learning with sparsity: The lasso and generalizations. Int Stat Rev  2016;84:156–7. [ Google Scholar ] Pereira B, Chin S-F, Rueda OM  et al.  The somatic mutation profiles of 2,433 breast cancers refine their genomic and transcriptomic landscapes. Nat Commun  2016;7:11908. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Poirion OB, Chaudhary K, Garmire LX.  Deep learning data integration for better risk stratification models of bladder cancer. AMIA Summits Transl Sci Proc  2018;2018:197. [ PMC free article ] [ PubMed ] [ Google Scholar ] Ren Y, Gao Y, Du W  et al.  Classifying breast cancer using multi-view graph neural network based on multi-omics data. Front Genet  2024;15:1363896. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Reska D, Czajkowski M, Jurczuk K  et al.  Integration of solutions and services for multi-omics data analysis towards personalized medicine. Biocybern Biomed Eng  2021;41:1646–63. [ Google Scholar ] Ryan B, Marioni RE, Simpson TI.  Multi-omic graph diagnosis (MOGDx): a data integration tool to perform classification tasks for heterogeneous diseases. Bioinformatics  2024;40:btae523. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Santa-Maria I, Alaniz ME, Renwick N  et al.  Dysregulation of microrna-219 promotes neurodegeneration through post-transcriptional regulation of tau. J Clin Invest  2015;125:681–6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Serra A, Fratello M, Fortino V  et al.  MVDA: a multi-view genomic data integration methodology. BMC Bioinformatics  2015;16:261. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Shahrajabian MH, Sun W.  Survey on multi-omics, and multi-omics data analysis, integration and application. CPA  2023;19:267–81. [ Google Scholar ] Shen R, Olshen AB, Ladanyi M.  Integrative clustering of multiple genomic data types using a joint latent variable model with application to breast and lung cancer subtype analysis. Bioinformatics  2009;25:2906–12. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Singh A, Shannon CP, Gautier B  et al.  DIABLO: an integrative approach for identifying key molecular drivers from multi-omics assays. Bioinformatics  2019;35:3055–62. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Spooner A, Moridani MK, Toplis B  et al.  Benchmarking ensemble machine learning algorithms for multi-class, multi-omics data integration in clinical outcome prediction. Brief Bioinform  2025;26:bbaf116. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] St L, Wold S.  Analysis of variance (ANOVA). Chemom Intell Lab Syst  1989;6:259–72. 10.1016/0169-7439(89)80095-4 [ DOI ] [ Google Scholar ] Subramanian I, Verma S, Kumar S  et al.  Multi-omics data integration, interpretation, and its application. Bioinform Biol Insights  2020;14:1177932219899051. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sun Y, Goodison S, Li J  et al.  Improved breast cancer prognosis through the combination of clinical and genetic markers. Bioinformatics  2007;23:30–7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tanvir RB, Islam MM, Sobhan M  et al.  MOGAT: a multi-omics integration framework using graph attention networks for cancer subtype prediction. Int J Mol Sci  2024;25:2788. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tini G, Marchetti L, Priami C  et al.  Multi-omics integration—a comparison of unsupervised clustering methodologies. Brief Bioinform  2019;20:1269–79. [ DOI ] [ PubMed ] [ Google Scholar ] Vahabi N, Michailidis G.  Unsupervised multi-omics data integration methods: a comprehensive review. Front Genet  2022;13:854752. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Van De Wiel MA, Lien TG, Verlaat W  et al.  Better prediction by use of co-data: adaptive group-regularized ridge regression. Stat Med  2016;35:368–81. [ DOI ] [ PubMed ] [ Google Scholar ] Wang B, Mezlini AM, Demir F  et al.  Similarity network fusion for aggregating data types on a genomic scale. Nat Methods  2014;11:333–7. [ DOI ] [ PubMed ] [ Google Scholar ] Wang H, Lu H, Sun J  et al.  Interpretable deep learning methods for multiview learning. BMC Bioinformatics  2024;25:69. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang T, Shao W, Huang Z  et al.  MOGONET integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification. Nat Commun  2021;12:3445. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang X-W, Wang T, Schaub DP  et al.  Benchmarking omics-based prediction of asthma development in children. Respir Res  2023;24:63. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wörheide MA, Krumsiek J, Kastenmüller G  et al.  Multi-omics integration in biomedical research—a metabolomics-centric review. Anal Chim Acta  2021;1141:144–62. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Xie K, Hou Y, Zhou X.  Deep centroid: a general deep cascade classifier for biomedical omics data classification. Bioinformatics  2024;40:btae039. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Xu F, Wang S, Dai X  et al.  Ensemble learning models that predict surface protein abundance from single-cell multimodal omics data. Methods  2021;189:65–73. [ DOI ] [ PubMed ] [ Google Scholar ] Yan H, Weng D, Li D  et al.  Prior knowledge-guided multilevel graph neural network for tumor risk prediction and interpretation via multi-omics data integration. Brief Bioinform  2024;25:bbae184. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Yu G, Wang L-G, Han Y  et al.  clusterProfiler: an r package for comparing biological themes among gene clusters. OMICS  2012;16:284–7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhang Q, Yang P, Pang X  et al.  Preliminary exploration of the co-regulation of Alzheimer’s disease pathogenic genes by microRNAs and transcription factors. Front Aging Neurosci  2022;14:1069606. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhu H, Koniusz P. Simple spectral graph convolution. In: International Conference on Learning Representations. 2021. Zohari P, Chehreghani MH. Graph neural networks in multi-omics cancer research: a structured survey. arXiv preprint, arXiv:2506.17234, 2025, preprint: not peer reviewed. zqq121017. zqq121017/pearl: Pearl-1.0, April 2026. 10.5281/zenodo.19875866 [ DOI ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Citations zqq121017. zqq121017/pearl: Pearl-1.0, April 2026. 10.5281/zenodo.19875866 [ DOI ] Supplementary Materials btag253_Supplementary_Data btag253_supplementary_data.pdf (592.9KB, pdf) Data Availability Statement The ROSMAP dataset was obtained from AMP-AD Knowledge Portal ( https://adknowledgeportal.synapse.org/ ). Omics data of BRCA and LUAD was obtained from The Cancer Genome Atlas Program (TCGA) through Broad GDAC Firehose ( https://gdac.broadinstitute.org/ ). PAM50 breast cancer subtypes of TCGA BRCA patients were obtained through the TCGAbiolinks R package v2.12.6. The source data and code of our computational framework is available at https://github.com/zqq121017/PEARL . The archival version of the code is preserved on Zenodo at https://doi.org/10.5281/zenodo.19875866 ( zqq121017 2026 ). Articles from Bioinformatics are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (4.7 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

• Graph-based machine learning and its application on multi-omics data analysis#186639
• Noesis: Bidirectional Graph-RAG with Adaptive Parallelism and Cross-Knowledge-Base Semantic Discovery#617041
• Noesis: Bidirectional Graph-RAG with Adaptive Parallelism and Cross-Knowledge-Base Semantic Discovery#616541
• Empowering Local Agriculture: A Deep Learning-Powered Web System for Identifying Bangladeshi Mango Varieties#784915
• Empowering Local Agriculture: A Deep Learning-Powered Web System for Identifying Bangladeshi Mango Varieties#783325