SEARCH · PubMed Health

Results for “representation learning”

Explore indexed PubMed citations for clinical trials, systematic reviews and public health research. Read source abstracts and follow each citation to its original PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 recordsLinked to original sources

PubMed2025

An image-based protein-ligand binding representation learning framework via multi-level flexible dynamics trajectory pre-training.

MOTIVATION: Accurate prediction of protein-ligand binding (PLB) relationships plays a crucial role in drug discovery, which helps identify drugs that modulate the activity of specific targets. Traditional biological assays for measuring PLB relationships are time consuming and costly. In addition, models for predicting PLB relationships have been developed and widely used in drug discovery tasks. However, learning more accurate PLB representations is essential to meet the stringent standards required for drug discovery. RESULTS: We propose an image-based PLB representation learning framework, called ImagePLB, which equips ligand representation learner (LRL) and protein representation learner (PRL) to accept 3D multi-view ligand images and protein graphs as input, respectively, and learns rich interaction information between ligand and protein through a binding representation learner (BRL). Considering the scarcity of protein-ligand pairs, we further propose a multi-level next trajectory prediction (MLNTP) task to pre-train ImagePLB on the 4D flexible dynamics trajectory of 16 972 complexes, including ligand level, protein level, and complex level, to learn information related to trajectories. Besides, by introducing trajectory regularization (TR), we effectively alleviate the problem of high (even almost identical) feature similarity caused by adjacent trajectories. Compared with the current state-of-the-art methods, ImagePLB has achieved competitive improvements on PLB-related prediction tasks, including protein-ligand affinity and efficacy prediction tasks. This study opens the door to the image-based PLB learning paradigm. AVAILABILITY AND IMPLEMENTATION: All data and implementation details of code can be obtained from https://github.com/HongxinXiang/ImagePLB.

Ligands↗

PubMed2026

Representation learning for multi-modal spatially resolved transcriptomics data.

MOTIVATION: Spatial transcriptomics enables in-depth molecular characterization of samples on a morphology and RNA level while preserving spatial location. Integrating the resulting multi-modal data is an unsolved problem, and developing new solutions in precision medicine depends on improved methodologies. RESULTS: We introduce AESTETIK, a convolutional deep learning model that jointly integrates spatial, transcriptomics, and morphology information to learn accurate spot representations. AESTETIK yielded substantially improved cluster assignments on widely adopted technology platforms (e.g. 10x Genomics™, NanoString™) across multiple datasets. We achieved performance enhancement on structured tissues (e.g. brain) with a 21% increase in median ARI over previous state-of-the-art methods. Notably, AESTETIK also demonstrated superior performance on cancer tissues with heterogeneous cell populations, showing a 2-fold increase in breast cancer, 79% in melanoma, and 21% in liver cancer. We expect that these advances will enable a multi-modal understanding of key biological processes. AVAILABILITY AND IMPLEMENTATION: AESTETIK is implemented in Python 3 and is available as open source software at http://www.github.com/ratschlab/aestetik. The Snakemake pipeline for reproducing the results is available at http://www.github.com/ratschlab/st-rep.

Spatial Transcriptomics↗

PubMed2026

ORBIT: Oncogenic Representation Learning via Bi-Prototype Contrastive Learning in Hyperbolic Space for cancer driver gene identification.

Accurate identification of cancer driver genes is crucial for precision oncology but remains challenging due to the complexity of integrating heterogeneous data and modeling dynamic biological systems. To address these limitations, we propose ORBIT (Oncogenic Representation Learning via Bi-Prototype Contrastive Learning in Hyperbolic Space). Our framework synergistically fuses multi-omics profiles with functional network data using a context-adaptive graph reweighting mechanism to capture cancer-specific dynamics. The model employs a bi-prototype contrastive learning strategy within hyperbolic space, which aligns gene representations around distinct driver and non-driver semantic anchors while preserving the intrinsic hierarchy of biological networks. Comprehensive evaluations demonstrate that ORBIT achieves highly competitive stability in pan-cancer analysis while consistently outperforming state-of-the-art methods in cancer-specific predictions. Furthermore, functional enrichment analysis confirms that the model effectively segregates core cancer pathways, and drug sensitivity profiling validates the clinical relevance of the identified drivers. By integrating hyperbolic geometry with context-adaptive learning, ORBIT offers a robust and interpretable paradigm for precision medicine. The source codes and datasets are publicly accessible at https://github.com/spcho-dev/ORBIT.

Humans↗

PubMed2025

JASMINE: A powerful representation learning method for enhanced analysis of incomplete multi-omics data.

Integrative analysis of multi-omics data provides a more comprehensive and nuanced view of a subject's biological state. However, high-dimensionality and ubiquitous modality missingness present significant analytical challenges. Existing methods for incomplete multi-omics data are scarce, do not fully leverage both modality-specific and shared information, and produce task-biased representations. We propose JASMINE, a self-supervised representation learning method for incomplete multi-omics data that preserves both modality-specific and joint information and enhances sample similarity structure. JASMINE produces embeddings that achieve superior performance across multiple tasks for two different incomplete multi-omics datasets while requiring only a single round of training per dataset.

missing data↗

PubMed2026

A leakage-aware genomic prediction pipeline for meropenem resistance in Klebsiella pneumoniae using transformer-based resistome representation learning.

MOTIVATION: Antimicrobial resistance (AMR) in Klebsiella pneumoniae, particularly to carbapenems such as meropenem, is a major global health problem. Machine learning is increasingly used to predict resistance from genomic markers; however, many models fail to capture high-level gene-gene interactions and may exhibit inflated performance due to lineage-biased prediction. Existing genomic prediction models largely rely on flat feature representations that fail to capture epistatic gene interactions, and commonly suffer from inflated performance estimates due to phylogenetic data leakage. To address these limitations simultaneously, a leakage-aware hybrid TabTransformer-CatBoost pipeline was developed, combining self-attention-based resistome representation learning with gradient boosting classification under clade-aware data partitioning. A self-attention encoder converts sparse gene presence-absence profiles into contextualized latent embeddings, which are subsequently classified using gradient boosting to capture lineage-aware AMR patterns. RESULTS: The proposed architecture outperformed classical baselines including Logistic Regression, Random Forest, XGBoost, and optimized CatBoost models. Internal accuracy reached 92.59% for the Chained Hybrid configuration (area under the receiver operating characteristic curve, AUROC = 0.8670, F1 = 0.8537). Performance gains primarily originated from the embedding stage, as confirmed by ablation analysis. External validation across independent multinational cohorts (n = 305) demonstrated generalizability (AUROC = 0.8105; F1 = 0.7552). Permutation testing produced near-zero Matthews Correlation Coefficient (MCC) = 0.0091, indicating predictions reflect genuine biological signal rather than noise. These results establish attention-based genomic embedding with gradient boosting as a scalable, interpretable, and leakage-aware framework for clinical AMR prediction. AVAILABILITY AND IMPLEMENTATION: The source code for the TabTransformer-CatBoost framework, including preprocessing pipelines and pre-trained embeddings, is available at https://github.com/SibelKervanci/kp-meropenem-tabtransformer.

Journal Article↗

PubMed2026

Integrating histology and spatial transcriptomics via multimodal transformers and contrastive representation learning for accurate gene expression prediction.

Predicting spatial gene expression from Histological images is a fundamental task in understanding tissue organization and molecular phenotypes. However, existing methods often rely on single-model representations or lack effective alignment between image and transcriptomic features. To address these limitations, we propose a unified multimodal learning framework that integrates histological imaging and spatial transcriptomics through a shared latent representation space. Specifically, histological H&E images are encoded by a ResNet50-based convolutional stem and a MobileViT Transformer backbone to extract hierarchical visual representations. Both modalities are projected into a shared latent space via linear-GELU-dropout transformation blocks, enabling cross-modal alignment through a contrastive learning objective that maximizes agreement between the corresponding image and the spot embeddings. Experimental results on the 10x Genomics Visium dataset of human liver tissue demonstrate that MViTGene achieves significantly higher prediction accuracy than existing methods across multiple gene subsets, with improvements of 20%, 33%, and 12% in predicting marker genes, highly expressed genes, and highly variable genes, respectively. The significant improvement in relevance indicates that the model can more accurately capture the true correspondence between tissue morphology and gene expression, therefore enabling more reliable biological interpretation. It provides a computational tool for high-throughput spatial gene expression prediction that balances performance and interpretability.

Humans↗

PubMed2025

SIGEL: a context-aware genomic representation learning framework for spatial genomics analysis.

Spatial transcriptomics (ST) integrates spatial information into genomics, yet methods for generating spatially-informed gene representations are limited and computationally intensive. We present SIGEL, a cost-effective framework that derives gene manifolds from ST data by exploiting spatial genomic context. The resulting SIGEL-generated gene representations (SGRs) are context-aware, biologically meaningful, and robust across samples, making them highly effective for key downstream tasks, including imputing missing genes, detecting spatial expression patterns, identifying disease-related genes and interactions, and improving spatial clustering. Extensive experiments across diverse ST datasets validate SIGEL's effectiveness and highlight its potential in advancing spatial genomics research.

Genomics↗

PubMed2025

The signed two-space proximity model for learning representations in protein-protein interaction networks.

MOTIVATION: Accurately predicting complex protein-protein interactions (PPIs) is crucial for decoding biological processes, from cellular functioning to disease mechanisms. However, experimental methods for determining PPIs are computationally expensive. Thus, attention has been recently drawn to machine learning approaches. Furthermore, insufficient effort has been made toward analyzing signed PPI networks, which capture both activating (positive) and inhibitory (negative) interactions. To accurately represent biological relationships, we present the Signed Two-Space Proximity Model (S2-SPM) for signed PPI networks, which explicitly incorporates both types of interactions, reflecting the complex regulatory mechanisms within biological systems. This is achieved by leveraging two independent latent spaces to differentiate between positive and negative interactions while representing protein similarity through proximity in these spaces. Our approach also enables the identification of archetypes representing extreme protein profiles. RESULTS: S2-SPM's superior performance in predicting the presence and sign of interactions in SPPI networks is demonstrated in link prediction tasks against relevant baseline methods. Additionally, the biological prevalence of the identified archetypes is confirmed by an enrichment analysis of Gene Ontology (GO) terms, which reveals that distinct biological tasks are associated with archetypal groups formed by both interactions. This study is also validated regarding statistical significance and sensitivity analysis, providing insights into the functional roles of different interaction types. Finally, the robustness and consistency of the extracted archetype structures are confirmed using the Bayesian Normalized Mutual Information (BNMI) metric, proving the model's reliability in capturing meaningful SPPI patterns. AVAILABILITY: S2-SPM is implemented and freely available under the MIT license at https://github.com/Nicknakis/S2SPM.

Protein Interaction Mapping↗

PubMed2025

PharaCon: a new framework for identifying bacteriophages via conditional representation learning.

MOTIVATION: Identifying bacteriophages (phages) within metagenomic sequences is essential for understanding microbial community dynamics. Transformer-based foundation models have been successfully employed to address various biological challenges. However, these models are typically pre-trained with self-supervised tasks that do not consider label variance in the pre-training data. This presents a challenge for phage identification as pre-training on mixed bacterial and phage data may lead to information bias due to the imbalance between bacterial and phage samples. RESULTS: To overcome this limitation, we proposed a novel conditional BERT framework that incorporates label classes as special tokens during pre-training. Specifically, our conditional BERT model attaches labels directly during tokenization, introducing label constraints into the model's input. Additionally, we introduced a new fine-tuning scheme that enables the conditional BERT to be effectively utilized for classification tasks. This framework allows the BERT model to acquire label-specific contextual representations from mixed sequence data during pre-training and applies the conditional BERT as a classifier during fine-tuning, and we named the fine-tuned model as PharaCon. We evaluated PharaCon against several existing methods on both simulated sequence datasets and real metagenomic contig datasets. The results demonstrate PharaCon's effectiveness and efficiency in phage identification, highlighting the advantages of incorporating label information during both pre-training and fine-tuning. AVAILABILITY AND IMPLEMENTATION: The source code and associated data can be accessed at https://github.com/Celestial-Bai/PharaCon.

Bacteriophages↗

PubMed2025

CSGL: chemical synthesis graph learning for molecule representation.

MOTIVATION: Molecule representation learning (MRL) translates molecules into a real vector space, serving as input to downstream tasks in biology, chemistry, and computer science. This article introduces a chemical synthesis graph learning (CSGL) framework, which enhances MRL by considering both the atomic structures of molecules and their roles in chemical reactions through a hierarchical graph representation. Specifically, molecules are first modeled based on their molecular graphs, which capture atomic-level structural information. They are then further refined using a chemical synthesis graph, where nodes represent reactant and product molecule sets, and edges encode chemical transformations between reactants and products (e.g. changes in molecular structures). CSGL optimizes molecular embeddings of reactant and product nodes in a fashion that ensures the embeddings conform to a chemical balance constraint. RESULTS: Experimental results show that our method CSGL achieves strong performance on a variety of tasks, including product prediction, reaction classification, and molecular property prediction. AVAILABILITY AND IMPLEMENTATION: https://github.com/li-2023/CSGL.

Machine Learning↗

PubMed2026

Resolution-dependent self-supervised transfer in chest radiograph classification.

BACKGROUND: Self-supervised learning (SSL) has improved visual representation learning, but its value in chest radiography remains uncertain. DINOv3 extends earlier SSL models through Gram-anchored self-distillation and explicit high-resolution adaptation. Whether these changes improve transfer learning for chest radiograph classification has not been established. METHODS: We benchmarked DINOv3 against DINOv2 and supervised ImageNet initialization across seven chest radiograph datasets comprising 816,183 radiographs from pediatric and adult cohorts. ViT-B/16 and ConvNeXt-B were evaluated under full fine-tuning at 224 × 224 and 512 × 512 pixels, with targeted 1024 × 1024 experiments on three cohorts. Additional analyses examined parameter-efficient adaptation, synthetic label corruption, external validation, frozen 7B features, and computational efficiency. The primary outcome was the mean area under the receiver operating characteristic curve across labels. RESULTS: In adult cohorts, DINOv3 did not consistently outperform DINOv2 at 224 × 224 pixels, but became the strongest initialization at 512 × 512 pixels, especially with ConvNeXt-B. Gains were greatest for small focal and boundary-dependent abnormalities, whereas large-structure findings changed little. The pediatric cohort showed no significant benefit from DINOv3, higher resolution, or backbone choice. Scaling to 1024 × 1024 rarely improved performance and markedly increased computational cost. ConvNeXt-B remained superior to ViT-B/16 under both full and parameter-efficient adaptation. External validation preserved the 512 × 512 DINOv3 advantage, whereas synthetic label corruption showed that this benefit should not be interpreted simply as superior noise robustness. Frozen DINOv3-7B features underperformed relative to fully adapted 86 to 89M-parameter backbones. CONCLUSIONS: For adult chest radiograph classification, DINOv3 provides its most reliable benefit at 512 × 512 pixels, particularly with ConvNeXt-B. Fully adapted mid-sized models at 512 × 512 pixels provided the best performance-cost trade-off in our benchmark.

Journal Article↗

PubMed2026

DPAS-Graph: adaptive spatial-feature relation learning for spatial RNA-to-protein prediction and virtual protein profiling.

Paired spatial multi-omics provides a supervised basis for learning RNA-protein correspondence in situ, but predicting protein abundance from spatial transcriptomic data alone remains challenging across tissue contexts and protein panels. Here, we present DPAS-Graph, an adaptive relation-learning framework for spatial RNA-to-protein prediction. Rather than directly merging spatial proximity and transcriptomic similarity as fixed graph priors, DPAS-Graph represents them as two relation channels on a shared edge support and updates their contributions during representation learning for protein prediction. Its Niche-Coupled Field Encoder combines layer-wise edge-relation modeling, intra-branch relation refinement, and cross-branch residual correction to learn spot representations for protein abundance prediction. In a leave-one-dataset-out benchmark across seven paired spatial multi-omics datasets, DPAS-Graph achieved lower aggregate prediction errors and improved spot-level agreement of protein expression profiles, with gains mainly reflected in error-based metrics and PCC-Spot. Spatial autocorrelation and protein-derived domain agreement analyses were further used to characterize the spatial behavior of the predicted protein maps. When applied to external RNA-only spatial sections, DPAS-Graph generated qualitatively interpretable marker-level virtual protein maps, illustrating its use as a complementary tool for protein-level interpretation of transcriptomics-only spatial data.

RNA↗

PubMed2025

soFusion: facilitating tissue structure identification via spatial multi-omics data fusion.

The rapid advancement of spatial multi-omics technologies has opened new avenues for dissecting tissue architecture with unprecedented resolution. However, inherent disparities across omics modalities, such as differences in biological hierarchy and resolution, pose significant challenges for integrative analysis. To address this, we present soFusion, a method for representation learning on spatial multi-omics data that enables automated identification of tissue compartmentalization. soFusion employs a graph convolutional network (GCN) to extract latent embeddings from spatial omics profiles. To simultaneously capture both cross-modality relationships and modality-specific features, we introduce a novel strategy for intra- and inter-omics feature learning. Moreover, modality-specific decoders are designed to preserve the unique information embedded in each omics type. We evaluated soFusion on multiple datasets including gene expression, protein expression, and epigenetic features. Across all benchmarks, soFusion consistently outperformed existing methods in delineating anatomical structures and identifying spatial domains with improved continuity and reduced noise. Collectively, soFusion offers an effective solution for spatial multi-omics integration, substantially enhancing the robustness of spatial domain identification.

Humans↗

PubMed2026

Multimodal alignment improves generalizability of genomic biomarker prediction in computational pathology.

Computational pathology models that use digitized histopathology whole-slide images have the potential to become a cost-effective and scalable alternative to molecular assays for the prediction of genomic biomarkers, a key task in precision oncology. However, as new genomic biomarkers are discovered or quantified, large, labeled datasets must be prospectively collected to train new models. To address this challenge, we developed multimodal alignment for biomarker learning and generalization (MARBLE), a multimodal contrastive pretraining strategy that integrates structured biomarker knowledge into representation learning of histopathology images. MARBLE aligns histopathology-derived representations with representations of genomic biomarkers generated by a large language model (LLM) and a protein language model (PLM). This biologically informed alignment enables data-efficient generalization to novel, out-of-distribution biomarkers. Using the MSK-IMPACT cohort of over 40,000 patients across multiple biomarker panel versions, we design experiments grounded in real-world data to demonstrate the value of our proposed approach.

CP: computational biology↗

PubMed2026

GBFN: A gated bimodal fusion network leveraging foundation model embeddings for cancer drug sensitivity prediction.

Despite recent progress in deep learning for cancer drug sensitivity prediction, many existing models still rely on task-specific representation learning or relatively simple multimodal fusion, which may limit their ability to capture complex drug-cell interactions. To address this issue, we developed GBFN, a gated bimodal fusion network for continuous IC50 prediction that integrates pretrained drug and cell-line representations. Specifically, drug embeddings were obtained from SMI-TED, whereas cell-line embeddings were derived from transcriptomic profiles using BulkFormer. These two modalities were then combined through a dimension-wise gated fusion module and used to predict IC50 values in matched drug-cell line pairs. On the CCLE-based benchmark, GBFN outperformed representative neural baselines, including GraphDRP, TGSA, and TransEDRP, and achieved the best overall performance, with an R² of 0.8714 and an RMSE of 0.8938. Moreover, ablation analysis showed that the model using drug features and cell-line expression data with gated fusion performed better than the corresponding model using direct concatenation, indicating that the improvement was associated with the fusion strategy rather than with the input modalities alone. In addition, cell-line expression data were more informative than mutation data in the present setting, and adding mutation data to the model using drug features and expression data did not further improve performance. Across major cancer types, GBFN maintained generally high cell-line-level predictive performance, and perturbation-based attribution identified biologically relevant transcriptomic programs in selected drug-cell line settings. Together, these findings support GBFN as a compact and effective framework for continuous drug response prediction.

Humans↗

PubMed2025

A Graph Contrastive Learning Method for Enhancing Genome Recovery in Complex Microbial Communities.

Accurate genome binning is essential for resolving microbial community structure and functional potential from metagenomic data. However, existing approaches-primarily reliant on tetranucleotide frequency (TNF) and abundance profiles-often perform sub-optimally in the face of complex community compositions, low-abundance taxa, and long-read sequencing datasets. To address these limitations, we present MBGCCA, a novel metagenomic binning framework that synergistically integrates graph neural networks (GNNs), contrastive learning, and information-theoretic regularization to enhance binning accuracy, robustness, and biological coherence. MBGCCA operates in two stages: (1) multimodal information integration, where TNF and abundance profiles are fused via a deep neural network trained using a multi-view contrastive loss, and (2) self-supervised graph representation learning, which leverages assembly graph topology to refine contig embeddings. The contrastive learning objective follows the InfoMax principle by maximizing mutual information across augmented views and modalities, encouraging the model to extract globally consistent and high-information representations. By aligning perturbed graph views while preserving topological structure, MBGCCA effectively captures both global genomic characteristics and local contig relationships. Comprehensive evaluations using both synthetic and real-world datasets-including wastewater and soil microbiomes-demonstrate that MBGCCA consistently outperforms state-of-the-art binning methods, particularly in challenging scenarios marked by sparse data and high community complexity. These results highlight the value of entropy-aware, topology-preserving learning for advancing metagenomic genome reconstruction.

canonical correlation analysis↗

PubMed2025

PLNMFG: Pseudo-label guided non-negative matrix factorization model with graph constraint for single-cell multi-omics data clustering.

The development of single-cell multi-omics sequencing technologies has enabled the simultaneous analysis of multi-omics data within the same cell. Accurate clustering of these cells is crucial for downstream analyses of complex biological functions. Despite significant advances in multi-omics integration approaches, current methodologies exhibit two major limitations. First, they inadequately incorporate prior biological knowledge from various omic layers. Second, these methods often conduct independent dimensionality reduction on individual omic datasets, thereby failing to capture the intrinsic complementary information and potentially overlooking crucial cross-platform interactions. Motivated by these, this study investigates a non-negative matrix factorization model called PLNMFG, which integrates the unified latent representation learning that retains the features between and within omics and the cluster structure learning that retains the intrinsic structure of the data into one joint framework. Specially, PLNMFG performs adaptive imputation to handle dropout events and uses prior pseudo-labels as constraints during the process of collective non-negative matrix factorization, as a result, a more robust latent representation that preserves the double similarity information is obtained. Graph Laplacian constraint is applied during clustering which further preserves structure characteristic of multi-omics data. In addition, the weight of each omic is adaptively learned based on the omic contribution. A series of experiments on 8 benchmark datasets show that our model performs well in terms of clustering accuracy and computational efficiency.

Single-Cell Analysis↗

PubMed2026

CAGNet: a structure-aware clustering-alternated graph network for cell-cell interaction inference in spatial transcriptomics.

MOTIVATION: Understanding cell-cell interactions (CCIs) in spatial transcriptomics is crucial for uncovering the spatial organization and functional heterogeneity of tissues. However, existing graph-based models typically rely on static clustering or fixed adjacency structures, which limits their ability to capture dynamic cellular relationships. RESULTS: We propose CAGNet, a two-stage framework for CCI inference from spatial transcriptomics data. In Stage 1, a Graph Attention Network encoder with joint feature and graph reconstruction learns structure-aware node embeddings from spatial gene expression profiles. In Stage 2, an alternating optimization mechanism iteratively updates cluster centers via KL-guided soft assignment and refines node embeddings through spatial graph reconstruction, establishing a closed-loop between representation learning and clustering. Experiments on three 10x Genomics Visium datasets demonstrate that CAGNet consistently outperforms six CCI inference baselines across ACC, AUC, AP, Precision, Recall, and F1. CAGNet also achieves the highest Adjusted Rand Index on all three datasets against six spatial domain identification methods, confirming that the learned embeddings capture biologically relevant spatial organization. Information-theoretic analysis further shows that CAGNet retains the highest mutual information between input features and learned embeddings among all compared methods. Ablation studies and 5-fold cross-validation confirm the contribution of each component and the reproducibility of the results. AVAILABILITY: The proposed method is implemented in the CAGNet package available at http://github.com/mahan1233333-maker/CAGNet .

Spatial Transcriptomics↗