An adaptive weight self-distillation deep learning framework for phenotype prediction from longitudinal gut microbiome data - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy 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 BMC Microbiol . 2026 Mar 9;26:355. doi: 10.1186/s12866-026-04922-y Search in PMC Search in PubMed View in NLM Catalog Add to search An adaptive weight self-distillation deep learning framework for phenotype prediction from longitudinal gut microbiome data Kai Shi Kai Shi 1 College of Computer Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541004 China 2 Guangxi Academy of Artificial Intelligence, Nanning, Guangxi 530028 China Find articles by Kai Shi 1, 2, ✉ , Qisheng He Qisheng He 1 College of Computer Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541004 China Find articles by Qisheng He 1 , Shichuang Wang Shichuang Wang 1 College of Computer Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541004 China Find articles by Shichuang Wang 1 , Junjun Guo Junjun Guo 3 The Organ Transplantation Department of 924th Hospital of Joint Logistic Support Force of PLA, Guilin, Guangxi 541002 China Find articles by Junjun Guo 3, ✉ Author information Article notes Copyright and License information 1 College of Computer Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541004 China 2 Guangxi Academy of Artificial Intelligence, Nanning, Guangxi 530028 China 3 The Organ Transplantation Department of 924th Hospital of Joint Logistic Support Force of PLA, Guilin, Guangxi 541002 China ✉ Corresponding author. Received 2025 Oct 24; Accepted 2026 Mar 3; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13081238 PMID: 41796297 Abstract Background The gut microbiota plays a vital role in maintaining human health. In recent years, extensive researches has focused on phenotype prediction in relation to various diseases, with the gut microbiota as a key predictor. Nevertheless, most existing studies rely on single-time-point analyses, which are insufficient to capture the dynamic patterns of host states and temporal variations inherent in longitudinal data. Results In this study, we propose a deep learning framework, AWSD-CNN-LSTM, designed to classify host phenotypes using longitudinal metagenomic data. Unlike conventional approaches that treat each time point as an independent sample, our method models the sequential samples of each individual as a whole, integrating convolutional neural network (CNN) and long short-term memory network (LSTM) to effectively capture temporal dependencies in longitudinal microbiome sequencing data. In addition, the model incorporates an adaptive point-wise self-distillation mechanism to more accurately characterize host-specific patterns. Compared with state-of-the-art methods, AWSD-CNN-LSTM demonstrates superior performance on the PROTECT, DIABIMMUNE, and Infants datasets, achieving area under the receiver operating characteristic curve (AUC) values of 0.896, 0.813, and 0.894, respectively. Conclusions For the task of disease phenotype classification based on temporal data, we propose a novel framework that effectively captures the characteristics of time-series data and achieves high accuracy across multiple datasets. Our approach holds promise as a potential new tool for microbial knowledge discovery. Supplementary Information The online version contains supplementary material available at 10.1186/s12866-026-04922-y. Keywords: Disease prediction, Gut Microbiome, Self-Distillation, Inflammatory bowel disease Introduction The human body harbors trillions of microorganisms, collectively referred to as the microbiota or microbiome, forming a complex ecosystem that includes bacteria, archaea, fungi, viruses, and other microbial communities [ 1 – 3 ]. Increasing evidence has demonstrated that the human microbiome plays a crucial role in the development and maintenance of the immune system, the regulation of host metabolism, and the protection of overall health [ 4 – 7 ]. Comparative studies between healthy individuals and patients with inflammatory bowel disease (IBD), obesity, and type 2 diabetes (T2D) have revealed its potential value in disease diagnosis and risk prediction [ 8 – 10 ]. Research focusing on diseases from a microbial perspective has attracted increasing attention. Understanding the potential associations between microbes and diseases is essential for uncovering the underlying mechanisms and advancing the development of novel therapeutic strategies [ 11 ]. Advances in sequencing technologies have enabled researchers to investigate microbial communities in unprecedented ways. Analytical tools such as QIIME2 [ 12 ] and MetaPhlAn2 [ 13 ] facilitate taxonomic profiling of metagenomic data and provide a foundation for downstream disease phenotype prediction, disease association analyses, and biomarker identification [ 14 ]. With the rapid accumulation of human metagenomic data, machine learning (ML) [ 15 – 17 ] and deep learning (DL) [ 18 – 25 ] techniques have been increasingly applied to investigate microbial interactions and their associations with host diseases. Early studies relied mainly on single-time-point data, which limited their ability to capture the dynamic nature of the microbiome [ 26 ]. For example, support vector machines (SVMs) [ 27 , 28 ] and random forests (RFs) [ 29 , 30 ] can identify key microbial features but are unable to model long-term dependencies in longitudinal time series. The introduction of long short-term memory (LSTM) networks opened new opportunities for microbiome dynamics prediction. Metwally et al. [ 31 ] were the first to apply LSTM to longitudinal human microbiome data for predicting food allergy risk, demonstrating its capability in capturing temporal dynamics and highlighting the importance of time-dependent features in longitudinal microbiome analysis, thereby laying the groundwork for subsequent methodological developments. Building on this foundation, Sharma et al. [ 32 ] developed the phyLoLSTM framework, which integrates convolutional neural networks (CNN) with LSTM networks to simultaneously capture spatial features and temporal dependencies of the microbiome. To address data heterogeneity and class imbalance, they further employed time-point alignment preprocessing and weight-balancing strategies. Subsequently, Fung et al. [ 33 ] proposed an efficient hybrid deep learning framework that also integrates CNN and LSTM, but treats the sequential samples of each individual as a whole for modeling. In addition, they introduced a self-distillation mechanism to enhance the generalizability and robustness of the model. However, this approach still shows limitations in capturing inter-individual specificity. Inspired by previous studies, this work proposes a novel framework—Adaptive Weight Self-Distillation Convolutional Neural Network/Long Short-Term Memory (AWSD-CNN-LSTM)—designed to enhance the ability to capture individual-specific patterns in longitudinal metagenomic data, thereby improving the accuracy of host phenotype prediction. Specifically, we treat the sequential samples of each individual as a whole input to construct a spatiotemporal feature learning module. CNN is employed to extract the local dependency features of the microbiome, while LSTM captures the temporal dynamics among samples, thereby collaboratively capturing the complete dynamic features of each individual’s longitudinal data. Building on this, we incorporate an adaptive self-distillation mechanism [ 34 ] to integrate spatial and temporal features, further enhancing the ability to represent host-specific characteristics in cross-individual learning. Experimental results demonstrate that AWSD-CNN-LSTM consistently outperforms state-of-the-art methods on multiple publicly available longitudinal microbiome datasets, achieving superior prediction accuracy and generalization performance. Overall, our method demonstrates potential applications in phenotype prediction studies based on longitudinal microbiome data and provides a feasible approach for modeling individual-specific patterns in high-dimensional time-series data, offering valuable insights for precision medicine and disease risk prediction. Materials and methods Longitudinal collection of microbiome sequencing data Three longitudinal cohort datasets—PROTECT [ 35 ], DIABIMMUNE [ 36 ], and Infants [ 37 ]—were utilized to evaluate the proposed model. These datasets serve as critical resources for investigating pediatric inflammatory bowel disease, autoimmune disorders, and the development of the early-life gut microbiome. The PROTECT study was a multicenter prospective cohort of 428 newly diagnosed pediatric ulcerative colitis patients from the United States and Canada, with a 52-week longitudinal follow-up to investigate treatment response. They adopted a stepped treatment regimen: all patients did not receive any treatment at the baseline of week 0 and were subsequently randomly assigned to the 5-aminosalicylic acid monotherapy group or the oral/intravenous corticosteroid (CS) combined with mesalamine treatment group. Microbiome samples (fecal/rectal) were collected from 405 participants at weeks 0, 4, 12, and 52. All samples were sequenced on the Illumina Miseq platform using the 16 S rRNA gene amplicon sequencing technology [ 38 ], and 1015 operational taxonomic units (OTUs) were generated through the 16 S bioBakery standardized [ 39 ] analysis process constructed by AnADAMA2. The DIABIMMUNE study was a multinational cohort that examined the hygiene hypothesis in type 1 diabetes and other autoimmune disorders. The authors recruited 222 infants ( n = 74 per country: Finland, Estonia, and Russia) and conducted a longitudinal follow-up from birth to 36 months. Monthly stool samples were collected alongside detailed clinical and laboratory data, including breastfeeding status, dietary composition, allergy history, family history, infection history, clinical findings, and medication use. All 1,584 samples were processed using Illumina HiSeq 2500 16 S rRNA gene sequencing (V4 region). QIIME v1.8.0 [ 40 ] generated 282 OTUs after quality filtering at 97% similarity. The Swedish Infants cohort investigated how delivery mode (vaginal vs. cesarean) influenced early-life gut microbiome assembly. Researchers collected 1,679 fecal samples from 471 neonates at five timepoints: birth, 4 months, 12 months, 36 months, and 60 months. All samples were sequenced (V3-V4 region of 16 S rRNA gene) on the Illumina HiSeq platform, followed by QIIME-based quality control, chimera filtering, and OTU clustering, yielding 592 representative OTUs. For consistent and comparable analysis, we used normalized OTUs level relative abundance data from the original studies. The corresponding phenotype prediction tasks for each cohort are summarized in Supplementary Table S1. Operational taxonomic unit data processing To address missing time points in microbiome sequencing data, multiple padding strategies were applied to ensure compatibility with LSTM modeling requirements. The LSTM architecture effectively refined initially padded values through its inherent bidirectional propagation capabilities. Two primary approaches were implemented. First, for longitudinal data from the PROTECT study, when subjects had complete records at baseline (week 0) and endpoint (week 52) but lacked observations at intermediate points (weeks 4 and 12), sequential zero-value imputation (Table 1 ) was used to preserve temporal continuity. Second, for the alternative strategy presented in Table 2 , we adopt endpoint-tail zero-padding, in which zero values are appended after the observed time points to fill the missing intermediate points in the full sequence. This ensures that all input sequences maintain consistent dimensionality while preserving the integrity of the data structure. Table 1. Padding for Missing Values in Sequences Ruminococcaceae Peptostreptococcaceae Alcaligenaceae … Porphyromonadaceae Week 0 0.01541 0.13982 0.01848 … 0.00065 Week 4 0 0 0 0 0 Week 12 0 0 0 0 0 Week 52 0.00761 0.00039 0.00016 … 0.14241 Open in a new tab Table 2. Padding for Missing Values at the End Ruminococcaceae Peptostreptococcaceae Alcaligenaceae … Porphyromonadaceae Week 0 0.01541 0.13982 0.01848 … 0.00065 Week 52 0.00761 0.00039 0.00016 … 0.14241 Week 4 0 0 0 0 0 Week 12 0 0 0 0 0 Open in a new tab To further optimize data quality, dimensionality reduction was performed on the processed microbiome datasets. Principal Component Analysis (PCA) was applied, retaining the top 300 principal components to evaluate potential improvements in model performance. To avoid potential information leakage, PCA was fitted using only the training data within each cross-validation fold and then applied to the corresponding validation data. This procedure reduced computational complexity while preserving the dominant features of the microbiome profiles. Architectures of the AWSD-CNN-LSTM We propose a spatiotemporal feature learning framework based on adaptive weight self-distillation —AWSD-CNN-LSTM—designed to enhance the modeling of individual-specific patterns in longitudinal metagenomic data. As shown in Figure 1 , the framework consists of two main components: a preprocessing module and a prediction module. Fig. 1. Open in a new tab Overall framework of the AWSD-CNN-LSTM model. Longitudinal metagenomic abundance data first passes through a data preprocessing module. This module standardizes the sequential samples of each individual into complete time series and fills in missing time points via zero-padding. Next, the routing network (CNN + LSTM) extracts local dependency features and temporal dynamic features from these time series. Subsequently, the policy network employs Gumbel-Softmax to generate time-point-specific weights (W₁, W₂, W₃), enabling adaptive point-wise knowledge distillation to emphasize key individual temporal features. Finally, the learned features are aggregated at the output layer to complete the phenotype classification task During preprocessing, raw sequencing data are first converted into a species abundance matrix via taxonomic analysis. Then, sequential samples from each individual are treated as a whole input sequence to preserve the temporal continuity of longitudinal trajectories. Missing time points are filled using zero-padding to construct a complete longitudinal input matrix, enabling holistic modeling of individual temporal dependencies rather than treating samples independently. The prediction module includes a routing network [ 41 ] and a policy network. The routing network, composed of CNN and LSTM, utilizes CNN to extract local dependencies, while LSTM captures longitudinal temporal dynamics, ensuring complete preservation of individual time-series features. To further enhance individual-specific representation, an adaptive point-wise knowledge distillation mechanism is introduced. Through a teacher–student structure, the student network learns from the teacher network at different layers using knowledge distillation loss (KD Loss) [ 42 ] and Kullback–Leibler divergence loss (KL Loss) [ 43 ]. The policy network employs Gumbel-Softmax [ 44 ] for differentiable sampling of distillation strength, dynamically generating weights (W₁, W₂, W₃) and optimizing end-to-end via backpropagation. This mechanism adaptively emphasizes the contribution of key internal temporal features to the prediction task, thereby improving the model’s ability to capture individual dynamic patterns. Finally, the model outputs host phenotype predictions through a sigmoid layer, effectively capturing individual longitudinal trajectory features and significantly enhancing both prediction accuracy and model stability. Results Experimental setting In this study, we implemented a repeated 10-fold cross-validation strategy (10 repetitions) to ensure robust internal model validation, providing comprehensive performance evaluation while effectively controlling overfitting. In each cross-validation fold, the dataset was randomly partitioned into training (90%) and validation (10%) subsets, yielding 10 area under the receiver operating characteristic curve (AUC-ROC) values from the initial cross-validation. To enhance statistical reliability, the procedure was repeated 10 times, producing a total of 100 AUC-ROC values for robust evaluation. Detailed model hyperparameter settings are provided in Supplementary Table S7. Comprehensive model evaluation was performed using multiple metrics, including the mean AUC-ROC and the F1-score, the latter representing the harmonic mean of precision and recall and accounting for both false positives and false negatives. Performance comparisons were conducted between LSTM and CNN-LSTM architectures, and the effects of self-distillation and adaptive point self-distillation strategies were assessed. Sequential concatenation improves time-series microbiome models We evaluated the effects of alternative sequence concatenation strategies using two established time-series modeling approaches (LSTM and CNN-LSTM). In addition, alternative missing-data imputation methods, including Generative Adversarial Imputation Network (GAIN) and mean imputation, were evaluated for comparison (see Supplementary Tables S2–S3). Table 3 demonstrates that sequential concatenation consistently outperformed end concatenation, with particularly significant improvements on the PROTECT and DIABIMMUNE datasets. For instance, on PROTECT, the LSTM model with PCA achieved its highest AUC (0.863) and F1-score (0.837) under sequential concatenation, outperforming end concatenation. Similarly, CNN-LSTM models generally exhibited superior performance with sequential concatenation, suggesting that this strategy better captures dynamic temporal patterns in microbiome data. Across most evaluated modeling settings, sequence concatenation strategies demonstrated more stable and favorable performance than alternative constructions. Finally, we selected sequential concatenation as the standard input strategy for all subsequent experiments. Table 3. The influence of PCA and sequence Processing strategies on the Performance of LSTM/CNN-LSTM models Model performance Model PCA Processs PROTECT DIABIMMUNE Infants AUC F1 AUC F1 AUC F1 LSTM TRUE Sequence 0.863 0.837 0.810 0.849 0.890 0.857 End 0.801 0.799 0.771 0.838 0.889 0.855 FALSE Sequence 0.864 0.832 0.805 0.846 0.869 0.849 End 0.790 0.796 0.737 0.817 0.873 0.844 CNN-LSTM TRUE Sequence 0.888 0.849 0.755 0.801 0.789 0.796 End 0.817 0.812 0.728 0.838 0.780 0.774 FALSE Sequence 0.881 0.852 0.796 0.825 0.768 0.747 End 0.806 0.817 0.767 0.835 0.726 0.745 Open in a new tab The best results are marked in bold and the second-best results are underlined Prediction performance based on the PROTECT, DIABIMMUNE and Infants study To evaluate the adaptability and robustness of our proposed AWSD model across multiple publicly available longitudinal microbiome datasets, we conducted evaluations using three representative datasets: PROTECT, DIABIMMUNE, and Swedish Infants cohort. To evaluate the effectiveness of the model, we compared it against two representative methods: FD (first self-distillation), a collaborative architecture combining multiple sub-classifiers with a main classifier for label prediction, and SD (second self-distillation), a temporal distillation approach using previous training round outputs as regularization terms [ 33 ]. The AWSD model demonstrates superior overall performance compared to both distillation methods across multiple datasets, as shown in Figure 2 . Specifically, AWSD-CNN-LSTM with PCA achieved optimal performance on PROTECT (AUC = 0.896, F1-score = 0.866) and DIABIMMUNE (AUC = 0.809, F1-score = 0.847), demonstrating superior modeling capability and robustness. In contrast, FD and SD models exhibited greater performance variability across datasets, suggesting limited generalizability. Notably, AWSD-LSTM achieved state-of-the-art performance on the Swedish Infants cohort (AUC = 0.894, F1-score = 0.855), confirming the architecture’s adaptability. These comprehensive results validate the advantages of AWSD’s knowledge-guided framework, which leverages an adaptive weight self-distillation mechanism to capture individual-specific differences, unlike other methods that rely on fixed or stage-wise distillation strategies and thus lack adaptability to personalized temporal dynamics. Fig. 2. Open in a new tab Performance Comparison of PCA vs. Non-PCA Models Across Three Longitudinal Microbiome Datasets Ablation study and parameter analysis To systematically evaluate individual component contributions, we conducted ablation studies across three datasets examining: (1) PCA-based dimensionality reduction effects, (2) AWSD knowledge distillation framework incorporation, and (3) architectural variations (LSTM and CNN-LSTM) on predictive performance. As demonstrated in Tables 4 , 5 and 6 , the AWSD framework consistently achieved high performance, with AWSD-CNN-LSTM (PCA) attaining peak metrics on PROTECT (AUC = 0.896, F1-score = 0.863) and DIABIMMUNE (AUC = 0.813, F1-score = 0.840), outperforming the corresponding baseline models. These results validate the distillation mechanism’s efficacy in improving generalization. Table 4. The predictive performance of 10 times cross-validation in the PROTECT study Model PCA Model performance AUC mean AUC stdev F1 mean F1 stdev LSTM TRUE 0.863 0.040 0.837 0.032 FALSE 0.864 0.053 0.832 0.040 CNN-LSTM TRUE 0.888 0.039 0.849 0.023 FALSE 0.881 0.025 0.852 0.035 AWSD- LSTM TRUE 0.861 0.039 0.840 0.050 FALSE 0.850 0.053 0.831 0.048 AWSD- CNN-LSTM TRUE 0.896 0.045 0.866 0.018 FALSE 0.895 0.035 0.863 0.024 Open in a new tab The best results are marked in bold and the second-best results are underlined Table 5. The predictive performance of 10 times cross-validation in the DIABIMMUNE study Model PCA Model performance AUC mean AUC stdev F1 mean F1 stdev LSTM TRUE 0.810 0.056 0.849 0.045 FALSE 0.805 0.086 0.846 0.057 CNN-LSTM TRUE 0.755 0.092 0.801 0.068 FALSE 0.796 0.079 0.825 0.051 AWSD- LSTM TRUE 0.787 0.056 0.831 0.059 FALSE 0.799 0.064 0.820 0.056 AWSD- CNN-LSTM TRUE 0.809 0.093 0.847 0.066 FALSE 0.813 0.069 0.840 0.055 Open in a new tab The best results are marked in bold and the second-best results are underlined Table 6. Predictive performance of 10 times cross-validation in the Infants study Model PCA Model performance AUC mean AUC stdev F1 mean F1 stdev LSTM TRUE 0.890 0.058 0.857 0.037 FALSE 0.869 0.072 0.849 0.049 CNN-LSTM TRUE 0.789 0.061 0.796 0.027 FALSE 0.768 0.053 0.747 0.036 AWSD- LSTM TRUE 0.894 0.055 0.855 0.063 FALSE 0.875 0.082 0.849 0.060 AWSD- CNN-LSTM TRUE 0.818 0.054 0.804 0.039 FALSE 0.788 0.045 0.768 0.054 Open in a new tab The best results are marked in bold and the second-best results are underlined In comparing models with and without PCA dimensionality reduction, we observed consistent patterns: PCA provided limited or slightly reduced performance for LSTM-based models, while demonstrating variable effects on CNN-LSTM architectures. Specifically, PCA improved CNN-LSTM performance on certain datasets (e.g., PROTECT), but showed instability on others (e.g., DIABIMMUNE). These findings suggest that PCA may benefit CNN structures that are sensitive to feature dimensions, though its effectiveness is dataset-dependent. To evaluate our adaptive distillation strategy, we compared three teacher knowledge fusion approaches: random selection, average fusion, and our proposed adaptive fusion. Specifically, the random selection strategy generates fusion weights randomly during training; the average fusion strategy obtains fixed weights by simply averaging the outputs of all teacher layers; and the adaptive fusion strategy employs a policy network to dynamically generate fusion weights based on individual’s input features, enabling the model to adaptively emphasize the most informative knowledge for different subject’s temporal trajectories. As shown in Table 7 , the adaptive strategy consistently achieved the best performance across all datasets. Notably, it attained the highest scores on PROTECT (AUC = 0.896, F1-score = 0.866), with significant improvements over alternative strategies. The adaptive approach also demonstrated superior performance on DIABIMMUNE (AUC = 0.809, F1-score = 0.847) and Swedish infant cohorts (AUC = 0.818, F1-score = 0.804), outperforming both random and average fusion methods. These results confirm that dynamic weight adjustment captures complementary information among teacher models, enhancing both prediction accuracy and model robustness. Table 7. Performance Comparison of random, average and adaptive Fusion strategies Model Model performance PROTECT DIABIMMUNE Infants AUC F1 AUC F1 AUC F1 Random 0.876 0.849 0.791 0.807 0.809 0.802 Average 0.875 0.838 0.799 0.837 0.816 0.802 Adaptive 0.896 0.866 0.809 0.847 0.818 0.804 Open in a new tab The best results are marked in bold and the second-best results are underlined Sensitivity analysis To further assess the temporal dynamics of the microbiome-disease association, we performed a sensitivity analysis to investigate the impact of historical time spans on the model’s predictive performance. Using the AWSD-CNN-LSTM framework, we systematically examined the influence of different historical observation windows on model performance in the PROTECT dataset. As shown in Table 8 , when data spanning more than two months were included, the model achieved stable predictive results, with an average AUC of approximately 0.870 and an F1-score consistently above 0.830. These findings suggest that more complete time series enhance the model’s ability to capture comprehensive disease progression patterns. Table 8. Sensitivity Analysis of the Impact of Time on the Performance of the Optimal Model PROTECT Months AUC mean AUC stdev F1 mean F1 stdev 5 0.871 0.053 0.839 0.041 4 0.866 0.047 0.846 0.031 3 0.872 0.047 0.849 0.025 2 0.876 0.046 0.836 0.035 Open in a new tab We further applied the AWSD-CNN-LSTM model to the DIABIMMUNE cohort to evaluate the effects of different temporal sampling configurations. This cohort provides longitudinal follow-up data from birth to 36 months of age. The experimental results demonstrate that the model maintains robust performance across varying combinations of time points (2 to 13 time points), with no significant fluctuations observed in either AUC or F1-score. This stability indicates that the DIABIMMUNE dataset is relatively less sensitive to changes in temporal span and further shows that the model can retain strong predictive accuracy even when some time points are missing (Table 9 ). Table 9. Sensitivity Analysis of the Impact of Time on the Performance of the Optimal Model DIABIMMUNE Months AUC mean AUC stdev F1 mean F1 stdev 13 0.774 0.077 0.835 0.070 12 0.781 0.073 0.825 0.075 11 0.775 0.093 0.830 0.063 10 0.772 0.073 0.841 0.049 9 0.796 0.085 0.822 0.075 8 0.783 0.087 0.836 0.076 7 0.775 0.080 0.846 0.078 6 0.771 0.057 0.827 0.068 5 0.774 0.068 0.828 0.068 4 0.790 0.084 0.825 0.064 3 0.751 0.101 0.832 0.061 2 0.755 0.076 0.828 0.057 Open in a new tab Interpretable analysis of important features To further enhance interpretability, we employed the AWSD-CNN-LSTM framework to quantify the contribution of individual microbial features and identify the most discriminative taxa associated with disease outcomes. By ranking microbial importance using the Feature Importance Value (FIV), we identified the key bacterial species within each dataset that exert a predominant influence on phenotype prediction. Table 10 presents the top 15 microbes with the highest FIV scores for the PROTECT dataset. Table 10. Prediction Results for the Top 15 Microbes Identified in the PROTECT Study Microbe FIV LDA score Evidence Formicigenerans 0.0133 3.4543 PMID: 34,831,464 Dorea 0.0118 3.7169 PMID: 34,831,464 Clostridiaceae 0.0112 1.1721 PMID: 25,986,361 Clostridium 0.0098 3.5207 PMID: 25,986,361 Biforme 0.0095 0.4055 Unconfirmed Lachnospiraceae 0.0087 2.0208 PMID: 38,117,560 Obeum 0.0087 4.1250 PMID: 34,831,464 Gnavus 0.0086 0.3051 PMID: 26,423,113 Prausnitzii 0.0083 4.8434 PMID: 30,308,161 Coprococcus 0.0082 1.2035 PMID: 33,604,319 Bifidobacterium 0.0081 3.0358 PMID: 25,986,361 Christensenella 0.0080 1.9617 PMID: 39,294,159 Copri 0.0076 3.8132 PMID: 38,596,448 Ruminococcus 0.0075 2.0048 PMID: 38,117,560 Oribacterium 0.0072 1.6714 Unconfirmed Open in a new tab In addition, we applied Linear Discriminant Analysis Effect Size (LEfSe) to further characterize disease-associated microbial signatures. LEfSe generates Linear Discriminant Analysis (LDA) scores to quantify the magnitude of intergroup differences, with higher scores indicating more substantial microbial shifts. By integrating FIV-based rankings with LDA-derived effect sizes, we obtained a more robust and biologically interpretable set of microbial biomarkers. The PROTECT study focuses on pediatric ulcerative colitis (UC), a chronic inflammatory disease of the colon in which gut microbial dysbiosis plays a central role. Evidence from PROTECT and related UC studies has consistently demonstrated marked alterations in microbial composition. For example, Formicigenerans, Dorea, Obeum, Prausnitzii, and Coprococcus are frequently reported as depleted in UC patients, whereas Gnavus, Clostridium species, Bifidobacterium, and members of Clostridiaceae and Lachnospiraceae exhibit disease-associated fluctuations that reflect inflammation-driven shifts in the gut ecosystem. These microorganisms appear among the top features identified by our model. Notably, taxa such as Formicigenerans, Obeum, Prausnitzii, and Dorea have been repeatedly reported as reduced in the PROTECT UC cohort, consistent with their high feature importance values (FIV) in our results. In contrast, taxa such as Biforme and Oribacterium, although lacking direct evidence in the PROTECT study or broader UC literature, may still be biologically relevant. In summary, the microbial features prioritized by our framework show strong agreement with disease-associated taxa identified in the PROTECT UC cohort while also revealing potentially novel organisms. These findings demonstrate the model’s ability to identify clinically relevant microbial signatures, thereby supporting biomarker discovery and mechanistic insights in UC. Discussion Recent studies have demonstrated a close association between the human gut microbiome and various diseases. For example, a meta-analysis of large-scale metagenomic samples revealed a significant reduction in the complexity of the Firmicutes phylum in patients with IBD. Further research indicates that many species within Firmicutes possess anti-inflammatory properties, suggesting their potential role in maintaining gut health. However, most existing studies treat each time point as an independent sample, failing to capture the dynamic temporal changes of the microbiome and lacking effective models to reflect individual-specific variations. To address the limitations of existing methods in utilizing individual-specific information and temporal features, we proposed AWSD-CNN-LSTM, an innovative hybrid framework that integrates CNN, LSTM, and an adaptive point self-distillation mechanism for longitudinal microbiome analysis. The experimental results suggest that cross-individual learning combined with adaptive self-distillation can better leverage individual-specific patterns compared with conventional approaches. The AWSD-CNN-LSTM framework outperformed existing distillation methods across PROTECT, DIABIMMUNE, and Infants datasets, with ablation confirming the adaptive self-distillation mechanism and adaptive fusion further enhancing performance, while sensitivity analysis showed DIABIMMUNE was more affected by time window length than PROTECT. In addition, the sequence concatenation strategy helps preserve temporal continuity, which may lead to highly sparse input OTU vectors; however, the dimensionality reduction strategy and the adaptive point-wise self-distillation mechanism can effectively filter redundant information, thereby improving predictive performance. Despite these advances, the current framework has some limitations. First, AWSD-CNN-LSTM primarily relies on microbial features and does not fully incorporate host metadata such as age, sex, BMI, lifestyle, diet, or medication, which may influence disease outcomes. Additionally, potential batch effects arising from technical or biological sources may introduce confounding variability. Explicit batch-effect correction was not performed due to the lack of consistently annotated batch metadata in the publicly available datasets, and this should be considered when interpreting the results. Finally, although the framework has shown promising performance across the evaluated datasets, its applicability to more complex disease settings or substantially larger cohorts remains to be further investigated. Future work will focus on integrating host metadata and environmental factors to provide a more comprehensive view of disease mechanisms. Advanced deep learning approaches, such as graph neural networks or attention-based models, may further enhance the capture of temporal dynamics and inter-individual heterogeneity. Moreover, applying the framework to larger, multi-cohort, and more diverse datasets, including mental health conditions, could expand its utility and contribute to personalized disease prediction and intervention strategies. Conclusions The AWSD-CNN-LSTM framework demonstrated the potential to provide an efficient framework for longitudinal microbiome analysis and disease outcome prediction. Future research should focus on optimizing data processing pipelines, developing refined feature extraction techniques, and extending applications to additional disease domains, which could further unlock the clinical potential of microbiome data and advance the integration of bioinformatics with precision medicine. Supplementary Information Supplementary Material 1. (39.9KB, docx) Acknowledgements Not Applicable. Authors’ contributions This study was designed by K.S. and J.J.G. Data pre-processing, statistical analysis and network analysis were performed by Q.S.H. and S.C.W. The study was supervised by K.S. and J.J.G. Q.S.H. and K.S. wrote the original draft, and suggestions were made by the other authors. K.S. and J.J.G. provided funding. All authors read and approved the final manuscript prior to submission. Funding This work was supported by the National Natural Science Foundation of China (62562022), Guangxi Natural Science Foundation (2025JJA170175), Guangxi Health Commission Self-Funded Research Project (Z-C20241570), the Special Funds for Guiding Local Scientific and Technological Development by the Central Government (No. Guike ZY22096025). Data availability This study utilizes datasets obtained from publicly available sources and complies with their data usage policies. Additional approval by the local Ethics Committee was not required. All datasets can be accessed at the following URL: PRJNA436359 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJNA436359 ) , PRJNA290380 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJNA290380 ) and PRJEB38986 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJEB38986 ). To ensure full reproducibility, the complete Python implementation code and processed datasets have been made publicly available at ( https:/github.com/GUT-ShiLab/AWSD ). Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Kai Shi, Email: [email protected]. Junjun Guo, Email: [email protected]. References 1. Aggarwal N, Kitano S, Puah GRY, Kittelmann S, Hwang IY, Chang MW. Microbiome and Human Health: Current Understanding, Engineering, and Enabling Technologies. Chem Rev. 2023;123(1):31–72. 10.1021/acs.chemrev.2c00431. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Joos R, Boucher K, Lavelle A, Arumugam M, Blaser MJ, Claesson MJ, et al. Examining the healthy human microbiome concept. Nat Rev Microbiol. 2025;23(3):192–205. 10.1038/s41579-024-01107-0. [ DOI ] [ PubMed ] [ Google Scholar ] 3. Wassan JT, Wang H, Zheng HJ. Recent Advances in the Phylogenetic Analysis to Study Rumen Microbiome. Curr Bioinform. 2024;19(3):250–63. 10.2174/1574893618666230605120615. [ Google Scholar ] 4. Sharp C, Foster KR. Host control and the evolution of cooperation in host microbiomes. Nat Commun. 2022;13(1):3567. 10.1038/s41467-022-30971-8. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Wilde J, Slack E, Foster KR. Host control of the microbiome: Mechanisms, evolution, and disease. Science. 2024;385(6706):eadi3338. 10.1126/science.adi3338. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Schriefer AE, Kumar B, Zolty A, Didier A, M.G N GTG, et al. M-CAMPTM: A Cloud-based Web Platform with a Novel Approach for Species-level Classification of 16S rRNA Microbiome Sequences. Curr Bioinform. 2023;18(1):21–39. 10.2174/1574893617666220520100535. [ Google Scholar ] 7. Wu S, Feng J, Liu C, Wu H, Qiu Z, Ge J, et al. Machine learning aided construction of the quorum sensing communication network for human gut microbiota. Nat Commun. 2022;13(1):3079. 10.1038/s41467-022-30741-6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Priya S, Burns MB, Ward T, Mars RAT, Adamowicz B, Lock EF, et al. Identification of shared and disease-specific host gene-microbiome associations across human diseases using multi-omic integration. Nat Microbiol. 2022;7(6):780–95. 10.1038/s41564-022-01121-z. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Zou Q, Qu K, Luo Y, Yin D, Ju Y, Tang H. Predicting Diabetes Mellitus With Machine Learning Techniques. Front Genet. 2018;9:515. 10.3389/fgene.2018.00515. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Liu XW, Li HL, Ma CY, Shi TY, Wang TY, Yan D, et al. Predicting the role of the human gut microbiome in type 1 diabetes using machine-learning methods. Brief Funct Genomics. 2024;23(4):464–74. 10.1093/bfgp/elae004. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Chen Y, Wang Z, Wang J, Chu Y, Zhang Q, Li ZA, et al. Self-supervised learning in drug discovery. Sci China Inform Sci. 2025;68(7):170103. 10.1007/s11432-024-4453-4. [ Google Scholar ] 12. Bolyen E, Rideout JR, Dillon MR, Bokulich NA, Abnet CC, Al-Ghalith GA, et al. Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2. Nat Biotechnol. 2019;37(8):852–7. 10.1038/s41587-019-0209-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Truong DT, Franzosa EA, Tickle TL, Scholz M, Weingart G, Pasolli E, et al. MetaPhlAn2 for enhanced metagenomic taxonomic profiling. Nat Methods. 2015;12(10):902–3. 10.1038/nmeth.3589. [ DOI ] [ PubMed ] [ Google Scholar ] 14. Qin J, Li R, Raes J, Arumugam M, Burgdorf KS, Manichanh C, et al. A human gut microbial gene catalogue established by metagenomic sequencing. Nature. 2010;464(7285):59–65. 10.1038/nature08821. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Fukui H, Nishida A, Matsuda S, Kira F, Watanabe S, Kuriyama M, et al. Usefulness of Machine Learning-Based Gut Microbiome Analysis for Identifying Patients with Irritable Bowels Syndrome. J Clin Med. 2020;9(8). 10.3390/jcm9082403. [ DOI ] [ PMC free article ] [ PubMed ] 16. Li M, Liu J, Zhu J, Wang H, Sun C, Gao NL, et al. Performance of Gut Microbiome as an Independent Diagnostic Tool for 20 Diseases: Cross-Cohort Validation of Machine-Learning Classifiers. Gut Microbes. 2023;15(1):2205386. 10.1080/19490976.2023.2205386. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Li P, Li M, Chen WH. Best practices for developing microbiome-based disease diagnostic classifiers through machine learning. Gut Microbes. 2025;17(1):2489074. 10.1080/19490976.2025.2489074. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Lin CH, Lichtarge O. Using interpretable deep learning to model cancer dependencies. Bioinformatics. 2021;37(17):2675–81. 10.1093/bioinformatics/btab137. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Lo C, Marculescu R. MetaNN: accurate classification of host phenotypes from metagenomic data using neural networks. BMC Bioinformatics. 2019;20(Suppl 12):314. 10.1186/s12859-019-2833-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Shi K, Huang K, Li L, Liu Q, Zhang Y, Zheng H. Predicting microbe-disease association based on graph autoencoder and inductive matrix completion with multi-similarities fusion. Front Microbiol. 2024;15:1438942. 10.3389/fmicb.2024.1438942. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Shi K, Liu Q, Ji Q, He Q, Zhao XM. MicroHDF: predicting host phenotypes with metagenomic data using a deep forest-based framework. Brief Bioinform. 2024;25(6). 10.1093/bib/bbae530. [ DOI ] [ PMC free article ] [ PubMed ] 22. Wang N, Wu M, Gu W, Dai C, Shao Z, Subbalakshmi KP. MSFT-transformer: a multistage fusion tabular transformer for disease prediction using metagenomic data. Brief Bioinform. 2025;26(3). 10.1093/bib/bbaf217. [ DOI ] [ PMC free article ] [ PubMed ] 23. Zhu H, Hao H, Yu L. Identification of microbe-disease signed associations via multi-scale variational graph autoencoder based on signed message propagation. BMC Biol. 2024;22(1):172. 10.1186/s12915-024-01968-0. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Wang R, Jiang Y, Jin J, Yin C, Yu H, Wang F, et al. DeepBIO: an automated and interpretable deep-learning platform for high-throughput biological sequence prediction, functional annotation and visualization analysis. Nucleic Acids Res. 2023;51(7):3017–29. 10.1093/nar/gkad055. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Li H, Liu B. BioSeq-Diabolo: Biological sequence similarity analysis using Diabolo. PLoS Comput Biol. 2023;19(6):e1011214. 10.1371/journal.pcbi.1011214. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Stewart CJ, Embleton ND, Marrs ECL, Smith DP, Fofanova T, Nelson A, et al. Longitudinal development of the gut microbiome and metabolome in preterm neonates with late onset sepsis and healthy controls. Microbiome. 2017;5(1):75. 10.1186/s40168-017-0295-1. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Suykens JAK, Vandewalle J. Least Squares Support Vector Machine Classifiers. Neural Process Lett. 1999;9(3):293–300. 10.1023/A:1018628609742. [ Google Scholar ] 28. Wang Y, Zhai Y, Ding Y, Zou Q. SBSM-Pro: Support Bio-sequence Machine for Proteins. 2023. 10.48550/arXiv.2308.10275. 29. Breiman L, Random, Forests. Mach Learn. 2001;45(1):5–32. 10.1023/A:1010933404324. [ Google Scholar ] 30. Ru X, Li L, Zou Q. Incorporating Distance-Based Top-n-gram and Random Forest To Identify Electron Transport Proteins. J Proteome Res. 2019;18(7):2931–9. 10.1021/acs.jproteome.9b00250. [ DOI ] [ PubMed ] [ Google Scholar ] 31. Metwally AA, Yu PS, Reiman D, Dai Y, Finn PW, Perkins DL. Utilizing longitudinal microbiome taxonomic profiles to predict food allergy via Long Short-Term Memory networks. PLoS Comput Biol. 2019;15(2):e1006693. 10.1371/journal.pcbi.1006693. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Sharma D, Xu W. phyLoSTM: a novel deep learning model on disease prediction from longitudinal microbiome data. Bioinformatics. 2021;37(21):3707–14. 10.1093/bioinformatics/btab482. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Fung DLX, Li X, Leung CK, Hu P. A self-knowledge distillation-driven CNN-LSTM model for predicting disease outcomes using longitudinal microbiome data. Bioinform Adv. 2023;3(1):vbad059. 10.1093/bioadv/vbad059. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 34. Liu W, Zhou P, Zhao Z, Wang Z, Deng H, Ju Q. FastBERT: a Self-distilling BERT with Adaptive Inference Time. 2020. 10.48550/arXiv.2004.02178 35. Hyams JS, Davis S, Mack DR, Boyle B, Griffiths AM, LeLeiko NS, et al. Factors associated with early outcomes following standardised therapy in children with ulcerative colitis (PROTECT): a multicentre inception cohort study. Lancet Gastroenterol Hepatol. 2017;2(12):855–68. 10.1016/s2468-1253(17)30252-2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Vatanen T, Kostic AD, d’Hennezel E, Siljander H, Franzosa EA, Yassour M, et al. Variation in Microbiome LPS Immunogenicity Contributes to Autoimmunity in Humans. Cell. 2016;165(4):842–53. 10.1016/j.cell.2016.04.007. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Roswall J, Olsson LM, Kovatcheva-Datchary P, Nilsson S, Tremaroli V, Simon MC, et al. Developmental trajectory of the healthy human gut microbiota during the first 5 years of life. Cell Host Microbe. 2021;29(5):765–76.e3. 10.1016/j.chom.2021.02.021. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Schirmer M, Denson L, Vlamakis H, Franzosa EA, Thomas S, Gotman NM, et al. Compositional and Temporal Changes in the Gut Microbiome of Pediatric Ulcerative Colitis Patients Are Linked to Disease Course. Cell Host Microbe. 2018;24(4):600–10.e4. 10.1016/.chom.2018.09.009. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. McIver LJ, Abu-Ali G, Franzosa EA, Schwager R, Morgan XC, Waldron L, et al. bioBakery: a meta’omic analysis environment. Bioinformatics. 2018;34(7):1235–7. 10.1093/bioinformatics/btx754. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 40. Caporaso JG, Kuczynski J, Stombaugh J, Bittinger K, Bushman FD, Costello EK, et al. QIIME allows analysis of high-throughput community sequencing data. Nat Methods. 2010;7(5):335–6. 10.1038/nmeth.f.303. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Rosenbaum C, Klinger T, Riemer M, Routing Networks. Adaptive Selection of Non-linear Functions for Multi-Task Learning. 2017. 10.48550/arXiv.1711.01239 42. de Boer P-T, Kroese DP, Mannor S, Rubinstein RY. A Tutorial on the Cross-Entropy Method. Ann Oper Res. 2005;134(1):19–67. 10.1007/s10479-005-5724-z. [ Google Scholar ] 43. Hershey JR, Olsen PA. Approximating the Kullback–Leibler divergence between Gaussian mixture models. 2007 IEEE Int Conf Acoust Speech Signal Process (ICASSP) IEEE. 2007;IV–317. 10.1109/ICASSP.2007.366913. 44. Jang E, Gu S, Poole B. Categorical Reparameterization with Gumbel-Softmax. 2016. 10.48550/arXiv.1611.01144. Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material 1. (39.9KB, docx) Data Availability Statement This study utilizes datasets obtained from publicly available sources and complies with their data usage policies. Additional approval by the local Ethics Committee was not required. All datasets can be accessed at the following URL: PRJNA436359 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJNA436359 ) , PRJNA290380 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJNA290380 ) and PRJEB38986 ( https:/www.ncbi.nlm.nih.gov/bioproject/PRJEB38986 ). To ensure full reproducibility, the complete Python implementation code and processed datasets have been made publicly available at ( https:/github.com/GUT-ShiLab/AWSD ). Articles from BMC Microbiology are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.3 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel 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