Drug-target affinity (DTA) prediction is a key task in computational drug discovery, yet current research is often compromised by data leakage and non-reproducible preprocessing. We present DTA-GNN, an end-to-end Python toolkit that automates the rigorous construction of target-specific datasets and streamlines the training of Graph Neural Network (GNN) based DTA predictors. To address data validity, the toolkit's dataset construction pipeline handles ChEMBL data ingestion and unit standardization, and implements scaffold- and temporal-splitting strategies to prevent overestimation of performance. Integrated leakage audits quantify split integrity prior to modeling. Following dataset construction, DTA-GNN provides a modular trainer that supports ten state-of-the-art GNN architectures and includes built-in hyperparameter optimization. In addition, DTA-GNN supports latent space analysis either by extracting learned molecular embeddings or leveraging molecular fingerprints, and provides interactive visualizations to explore chemical space and interpret model behavior. By unifying robust dataset construction with accessible model training and latent-space analysis via Python library, CLI, and Web UI, DTA-GNN enables researchers to produce standardized, reproducible, and leakage-free DTA benchmarks.

Analysis and interpretation of omics data largely benefit from the use of prior knowledge. However, this knowledge is fragmented across resources and often is not directly accessible for analytical methods. We developed OmniPath (https://omnipathdb.org/), a database combining diverse molecular knowledge from 168 resources. It covers causal protein-protein, gene regulatory, microRNA, and enzyme-post-translational modification interactions, cell-cell communication, protein complexes, and information about the function, localization, structure, and many other aspects of biomolecules. It prioritizes literature curated data, and complements it with predictions and large scale databases. To enable interactive browsing of this large corpus of knowledge, we developed OmniPath Explorer, which also includes a large language model agent that has direct access to the database. Python and R/Bioconductor client packages and a Cytoscape plugin create easy access to customized prior knowledge for omics analysis environments, such as scverse. OmniPath can be broadly used for the analysis of bulk, single-cell, and spatial multi-omics data, especially for mechanistic and causal modeling.

Abstract Biomedical discovery is hindered by fragmented, modality-specific repositories and uneven metadata, limiting integrative analysis, accessibility, and reproducibility. To address these challenges, we present CROssBARv2, a provenance-rich biomedical data-and-knowledge integration platform that unifies heterogeneous sources into a maintainable, scalable system. By consolidating diverse data types into an extensive knowledge graph enriched with standardised ontologies, rich metadata, and deep learning–based vector embeddings, CROssBARv2 alleviates the need for researchers to navigate multiple siloed databases and can facilitate downstream tasks, including predictive modelling and mechanistic reasoning, enabling applications such as drug repurposing and protein function prediction. The platform offers interactive graph exploration and embedding-based semantic search with CROssBAR-LLM, an intuitive natural language question-answering system that grounds large language model (LLM) outputs in the underlying knowledge graph to mitigate hallucinations. We assess CROssBARv2 through (i) multiple use-case analyses to test biological coherence and relational validity; (ii) knowledge-augmented biomedical question-answering benchmarks comparing CROssBAR-LLM against generalist LLMs; and (iii) a deep learning–based predictive modelling experiment for protein function prediction leveraging the heterogeneous structure of CROssBARv2. Collectively, CROssBARv2 provides a scalable, AI-ready, and user-friendly foundation that facilitates hypothesis generation, knowledge discovery, and translational research.

BACKGROUND: Learning the representation of molecules is crucial for drug discovery but is often hindered by the scarcity of labeled experimental data, which limits the performance of supervised machine learning models. While self-supervised learning (SSL) offers a solution by leveraging vast unlabeled chemical databases, many existing methods focus on learning from either local structural information or global molecular properties, but not both simultaneously. We introduce MoCL-GAT, a novel contrastive and transfer learning-based SSL framework that addresses this gap by simultaneously learning from two complementary objectives. It combines a local contrastive task on molecular subgraphs to capture fine-grained chemical environments with a global predictive task to learn holistic molecular descriptors. This dual-objective approach, powered by a Graph Attention Network, is designed to create more robust, versatile, and transferable molecular representations. RESULTS: Pre-trained on 1.9 million compounds, MoCL-GAT was fine-tuned on diverse benchmarks. It achieved state-of-the-art performance on molecular property prediction tasks, with an AUROC of 0.928 on BBBP and 0.768 on SIDER, and top-ranking RMSEs of 0.570 for ESOL and 1.818 for FreeSolv. Critically, fine-tuned models consistently and significantly outperformed models trained from scratch, confirming the value of pre-training. CONCLUSIONS: These results validate that MoCL-GAT's dual-objective approach learns highly effective and transferable representations, enabling more accurate and data-efficient predictions for key cheminformatics challenges. The source code for MoCL-GAT is publicly available on Zenodo at https://doi.org/10.5281/zenodo.16927285 .

Understanding disease-gene associations is essential for unravelling disease mechanisms and advancing diagnostics and therapeutics. Traditional approaches based on manual curation and literature review are labour-intensive and not scalable, prompting the use of machine learning on large biomedical data. In particular, graph neural networks (GNNs) have shown promise for modelling complex biological relationships. To address limitations in existing models, we propose GLaDiGAtor (Graph Learning-bAsed DIsease-Gene AssociaTiOn pRediction), a novel GNN framework with an encoder-decoder architecture for disease-gene association prediction. GLaDiGAtor constructs a heterogeneous biological graph integrating gene-gene, disease-disease, and gene-disease interactions from curated databases, and enriches each node with contextual features from well-known language models (ProtT5 for protein sequences and BioBERT for disease text). In evaluations, our model achieves superior predictive accuracy and generalisation, outperforming 14 existing methods. Literature-supported case studies confirm the biological relevance of high-confidence novel predictions, highlighting GLaDiGAtor's potential to discover candidate disease genes. These results underscore the power of graph convolutional networks in biomedical informatics and may ultimately facilitate drug discovery by revealing new gene-disease links. The source code and processed datasets are publicly available at https://github.com/HUBioDataLab/GLaDiGAtor.

MOTIVATION:Molecular representation learning is central to computational drug discovery. However, most existing models rely on single-modality inputs, such as molecular sequences or graphs, which capture only limited aspects of molecular behaviour. Yet unifying these modalities with complementary resources, such as textual descriptions and biological interaction networks, into a coherent multimodal framework remains non-trivial, hindering the development of more informative and biologically grounded representations. RESULTS:We introduce SELFormer Multimodal (SELFormerMM), a multimodal molecular representation learning framework that integrates SELFIES notations with structural graphs, textual descriptions and knowledge graph-derived biological interaction data. By aligning these heterogeneous views, SELFormerMM effectively captures complementary signals that unimodal approaches often overlook. Our performance evaluation has revealed that SELFormerMM outperforms structure-, sequence- and knowledge-based models on multiple molecular property prediction tasks. Ablation analyses further indicate that effective cross-modal alignment and modality coverage improve the model's ability to exploit complementary information. Overall, integrating SELFIES with structural, textual and biological context enables richer molecular representations and provides a promising framework for hypothesis-driven drug discovery. AVAILABILITY AND IMPLEMENTATION:SELFormerMM is available as a programmatic tool at https://github.com/HUBioDataLab/SELFormerMM. Pre-trained and fine-tuned model checkpoints, multimodal pre-training and finetuning datasets including SMILES, textual descriptions and knowledge graph, and their precomputed single-modality embedding files are available at https://huggingface.co/datasets/HUBioDataLab/SELFormerMM.

Background:The Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). Results:CAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced "partial knowledge" evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.

Foundation models have shown remarkable success across scientific domains, yet their impact in chemistry remains limited due to the absence of diverse, large-scale, high-quality datasets that reflect the field's multifaceted nature. We present the ChemPile, an open dataset containing over 75 billion tokens of curated chemical data, specifically built for training and evaluating general-purpose models in the chemical sciences. The dataset mirrors the human learning journey through chemistry – from educational foundations to specialized expertise – spanning multiple modalities and content types including structured data in diverse chemical representations (SMILES, SELFIES, IUPAC names, InChI, molecular renderings), scientific and educational text, executable code, and chemical images. ChemPile integrates foundational knowledge (textbooks, lecture notes), specialized expertise (scientific articles and language-interfaced data), visual understanding (molecular structures, diagrams), and advanced reasoning (problem-solving traces and code) – mirroring how human chemists develop expertise through diverse learning materials and experiences. Constructed through hundreds of hours of expert curation, the ChemPile captures both foundational concepts and domain-specific complexity. We provide standardized training, validation, and test splits, enabling robust benchmarking. ChemPile is openly released via HuggingFace with a consistent API, permissive license, and detailed documentation. We hope the ChemPile will serve as a catalyst for chemical AI, enabling the development of the next generation of chemical foundation models.

Discovering novel drug candidate molecules is a fundamental step in drug development. Generative deep learning models can sample new molecular structures from learned probability distributions; however, their practical use in drug discovery hinges on generating compounds tailored to a specific target molecule. Here we introduce DrugGEN, an end-to-end generative system for the de novo design of drug candidate molecules that interact with a selected protein. The proposed method represents molecules as graphs and processes them using a generative adversarial network that comprises graph transformer layers. Trained on large datasets of drug-like compounds and target-specific bioactive molecules, DrugGEN designed candidate inhibitors for AKT1, a kinase crucial in many cancers. Docking and molecular dynamics simulations suggest that the generated compounds effectively bind to AKT1, and attention maps provide insights into the model’s reasoning. Furthermore, selected de novo molecules were synthesized and shown to inhibit AKT1 at low micromolar concentrations in the context of in vitro enzymatic assays. These results demonstrate the potential of DrugGEN for designing target-specific molecules. Using the open-access DrugGEN codebase, researchers can retrain the model for other druggable proteins, provided a dataset of known bioactive molecules is available. Inhibiting AKT1 kinase can have potentially positive uses against many types of cancer. To find novel molecules targeting this protein, a graph adversarial network is trained as a generative model.

Motivation The rapid accumulation of protein sequence data, coupled with the slow pace of experimental annotations, creates a critical need for computational methods to predict protein functions. Existing models often rely on limited data types, such as sequence-based features or protein-protein interactions (PPIs), failing to capture the complex molecular relationships in biological systems. To address this, we developed ProtHGT, a heterogeneous graph transformer-based model that integrates diverse biological datasets into a unified framework using knowledge graphs for accurate and interpretable protein function prediction. Results ProtHGT achieves state-of-the-art performance on benchmark datasets, demonstrating its ability to outperform current graph-based and sequence-based approaches. By leveraging diverse biological entity types and highly representative protein language model embeddings at the input level, the model effectively learns complex biological relationships, enabling accurate predictions across all Gene Ontology (GO) sub-ontologies. Ablation analyses highlight the critical role of heterogeneous data integration in achieving robust predictions. Finally, our use-case study has indicated that it’s possible to interpret ProtHGT’s predictions via exploring the related parts of our input biological knowledge graph, offering plausible explanations to build or test new hypotheses. Availability and Implementation ProtHGT is available as a programmatic tool on [Github][1] and as a web service at [Hugging Face][2]. Contact To whom the correspondence should be addressed: Tunca Doğan (tuncadogan{at}gmail.com) ### Competing Interest Statement The authors have declared no competing interest. [1]: https://github.com/HUBioDataLab/ProtHGT [2]: https://huggingface.co/spaces/HUBioDataLab/ProtHGT

Discovering new, efficacious molecules remains slow and costly; rigorous data science-driven systems-level approaches are therefore essential to prioritise hypotheses and de-risk drug development. In this study, we present ECLIPSE, a systems-level framework for compound/ligand–protein interaction (CPI) representation and prediction, combining heterogeneous knowledge graphs (KGs), which encode large-scale entity–relation structure, with graph neural networks that exploit relational inductive biases to perform inference on graph-structured data. ECLIPSE uses our comprehensive biomedical KG-based platform, CROssBAR, incorporating genes/proteins, drugs, compounds, pathways, diseases, and phenotypes, along with their multi-layered relationships. Each entity is assigned input features derived from language or graph representation learning models and projected via type-specific neural network layers. To process these featurized biomedical KGs for bioactivity prediction, we employed the heterogeneous graph transformer (HGT) architecture. In contrast to the majority of GNN algorithms, which are restricted to homogenous graphs, HGT can handle graph heterogeneity and maintain node-and edge-type dependent representations through its attention mechanism. ECLIPSE achieves strong performance on challenging, protein-family– specific CPI benchmarks compared with baseline and state-of-the-art methods; ablations confirm performance gains from modelling graph heterogeneity and all feature sources. Use-case analyses on a druggable kinase (PIM1) and a historically undruggable receptor (HER3) illustrate generalizability across target classes and activity ranges. By leveraging direct and indirect relationships embedded in biomedical KGs, ECLIPSE provides context-aware CPI inference that is scalable to real-world settings. Code, datasets, and trained models are released to support reproducibility and reuse. ### Competing Interest Statement The authors have declared no competing interest.

Proteins play a crucial role in almost all biological processes, serving as the building blocks of life and mediating various cellular functions, from enzymatic reactions to immune responses. Accurate annotation of protein functions is essential for advancing our understanding of biological systems and developing innovative biotechnological applications and therapeutic strategies. To predict protein function, researchers primarily rely on classical homology-based methods, which use evolutionary relationships, and increasingly on machine learning (ML) approaches. Lately, protein language models (PLMs) have gained prominence; these models leverage specialized deep learning architectures to effectively capture intricate relationships between sequence, structure, and function. We recently conducted a comprehensive benchmarking study to evaluate diverse protein representations (i.e., classical approaches and PLMs) and discuss their trade-offs. The current work introduces the Protein Representation Benchmark-PROBE tool, a benchmarking framework designed to evaluate protein representations on function-related prediction tasks. Here, we provide a detailed protocol for running the framework via the GitHub repository and accessing our newly developed user-friendly web service. PROBE encompasses four core tasks: semantic similarity inference, ontology-based function prediction, drug target family classification, and protein-protein binding affinity estimation. We demonstrate PROBE's usage through a new use case evaluating ESM2 and three recent multimodal PLMs-ESM3, ProstT5, and SaProt-highlighting their ability to integrate diverse data types, including sequence and structural information. This study underscores the potential of protein language models in advancing protein function prediction and serves as a valuable tool for both PLM developers and users.

Monkeypox virus (MPXV) caused the 2022–2023 global mpox and the concurrent outbreaks in Africa, disproportionately affecting immunocompromised individuals such as people living with HIV. With no approved treatment available, we developed a robust artificial intelligence (AI) pipeline for discovering broad-spectrum poxvirus inhibitors that target the viral DNA polymerases. Among the identified leading candidates, we found that the clinically used antiretroviral drugs bictegravir and etravirine potently inhibit MPXV clade Ia, Ib and IIb infections in human intestinal and skin organoids. The broad anti-poxvirus activities of bictegravir and etravirine were further demonstrated against infections of other Orthopoxviruses such as vaccinia virus and cowpox virus. These findings support the repurposing of bictegravir and etravirine for treating mpox, especially for patients co-infected with HIV, warranting follow-up clinical investigation. The established AI pipeline and our antiviral drug discovery strategies bear major implications for responding to the ongoing mpox emergency and preparing for future poxvirus epidemics. AI-driven discovery of the antiretroviral drugs bictegravir and etravirine as potent inhibitors against monkeypox and related poxviruses infection. The effects were experimentally validated in human organoids infected with poxviruses.

Identifying new drug candidates remains a critical and complex challenge in drug development. Recent advances in deep learning have demonstrated significant potential to accelerate this process, particularly through the use of protein language models (pLMs). These models aim to effectively capture the structural and functional properties of proteins by embedding them in high-dimensional spaces, thereby providing powerful tools for predictive tasks. This review examines the application of pLMs in drug-target interaction (DTI) prediction, addressing both small-molecule and protein-based therapeutics. We explore diverse methodologies, including end-to-end learning models and those that leverage pre-trained foundational pLMs. Furthermore, we highlight the role of heterogeneous data integration-ranging from protein structures to knowledge graphs-to improve the accuracy of DTI predictions. Despite notable progress, challenges persist in accurately identifying DTIs, mainly due to data-related limitations and algorithmic constraints. Future research directions include utilising multimodal learning approaches, incorporating temporal/dynamic interaction data into training, and employing novel deep learning architectures to refine protein representations, gain a deeper understanding of biological context regarding molecular interactions, and, thus, advance the DTI prediction field.

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality worldwide, driven mainly by chronic hepatitis infections and metabolic disorders, which highlights the urgent need for novel therapeutic strategies. Sirtuins, particularly SIRT1 are crucial in HCC pathogenesis, making it a promising drug target. Indole-based molecules show potential as therapeutic agents by interacting with key proteins like sirtuins involved in cancer progression. In this study, we designed and synthesized novel indole-based small molecules and investigated their potential sirtuin inhibitory action and anticancer activity on HCC cell lines. Four of the twenty-eight tested small molecules on different cancer types were selected (4 g, 4 h, 4o, and 7j) based on their structure-activity relationship and studied on a panel of HCC cell lines. Compounds had active drug-target interactions with SIRT1 or SIRT2 based on DEEPScreen DTI predictions and docking studies which confirmed that 4o, 4 g, and 7j were most potent in their interaction with SIRT1. Compound 4 g caused the highest sirtuin activity inhibition in vitro and induced G1 arrest and apoptosis in HCC cell lines.

In this study, we propose HOPER (HOlistic ProtEin Representation), a novel multimodal learning framework designed to enhance protein function prediction (PFP) in low-data settings. The challenge of predicting protein functions is compounded by the limited availability of labeled data. Traditional machine learning models already struggle in such cases, and while deep learning models excel with abundant data, they also face difficulties when data is scarce. HOPER addresses this issue by integrating three distinct modalities - protein sequences, biomedical text, and protein-protein interaction (PPI) networks - to create a comprehensive protein representation. The model utilizes autoencoders to generate holistic embeddings, which are then employed for PFP tasks using transfer learning. HOPER outperforms existing methods on a benchmark dataset across all Gene Ontology categories, i.e., molecular function, biological process, and cellular component. Additionally, we demonstrate its practical utility by identifying new immune-escape proteins in lung adenocarcinoma, offering insights into potential therapeutic targets. Our results highlight the effectiveness of multimodal representation learning for overcoming data limitations in biological research, potentially enabling more accurate and scalable protein function prediction. HOPER source code and datasets are available at https://github.com/kansil/HOPER

Identifying unknown functional properties of proteins is essential for understanding their roles in both health and disease states. The domain composition of a protein can reveal critical information in this context, as domains are structural and functional units that dictate how the protein should act at the molecular level. The expensive and time-consuming nature of wet-lab experimental approaches prompted researchers to develop computational strategies for predicting the functions of proteins. In this study, we proposed a new method called Domain2GO that infers associations between protein domains and function-defining gene ontology (GO) terms, thus redefining the problem as domain function prediction. Domain2GO uses documented protein-level GO annotations together with proteins' domain annotations. Co-annotation patterns of domains and GO terms in the same proteins are examined using statistical resampling to obtain reliable associations. As a use-case study, we evaluated the biological relevance of examples selected from the Domain2GO-generated domain-GO term mappings via literature review. Then, we applied Domain2GO to predict unknown protein functions by propagating domain-associated GO terms to proteins annotated with these domains. For function prediction performance evaluation and comparison against other methods, we employed Critical Assessment of Function Annotation 3 (CAFA3) challenge datasets. The results demonstrated the high potential of Domain2GO, particularly for predicting molecular function and biological process terms, along with advantages such as producing interpretable results and having an exceptionally low computational cost. The approach presented here can be extended to other ontologies and biological entities to investigate unknown relationships in complex and large-scale biological data. The source code, datasets, results, and user instructions for Domain2GO are available at https://github.com/HUBioDataLab/Domain2GO. Additionally, we offer a user-friendly online tool at https://huggingface.co/spaces/HUBioDataLab/Domain2GO, which simplifies the prediction of functions of previously unannotated proteins solely using amino acid sequences.

Non-silent single nucleotide genetic variants, like nonsense changes and insertion-deletion variants, that affect protein function and length substantially are prevalent and are frequently misclassified. The low sensitivity and specificity of existing variant effect predictors for nonsense and indel variations restrict their use in clinical applications. We propose the Pathogenic Mutation Prediction (PMPred) method to predict the pathogenicity of single nucleotide variations, which impair protein function by prematurely terminating a protein's elongation during its synthesis. The prediction starts by monitoring functional effects (Gene Ontology annotation changes) of the change in sequence, using an existing ensemble machine learning model (UniGOPred). This, in turn, reveals the mutations that significantly deviate functionally from the wild-type sequence. We have identified novel harmful mutations in patient data and present them as motivating case studies. We also show that our method has increased sensitivity and specificity compared to state-of-the-art, especially in single nucleotide variations that produce large functional changes in the final protein. As further validation, we have done a comparative docking study on such a variation that is misclassified by existing methods and, using the altered binding affinities, show how PMPred can correctly predict the pathogenicity when other tools miss it. PMPred is freely accessible as a web service at https://pmpred.kansil.org/, and the related code is available at https://github.com/kansil/PMPred.

The aim of the UniProt Knowledgebase is to provide users with a comprehensive, high-quality and freely accessible set of protein sequences annotated with functional information. In this publication we describe enhancements made to our data processing pipeline and to our website to adapt to an ever-increasing information content. The number of sequences in UniProtKB has risen to over 227 million and we are working towards including a reference proteome for each taxonomic group. We continue to extract detailed annotations from the literature to update or create reviewed entries, while unreviewed entries are supplemented with annotations provided by automated systems using a variety of machine-learning techniques. In addition, the scientific community continues their contributions of publications and annotations to UniProt entries of their interest. Finally, we describe our new website (https://www.uniprot.org/), designed to enhance our users' experience and make our data easily accessible to the research community. This interface includes access to AlphaFold structures for more than 85% of all entries as well as improved visualisations for subcellular localisation of proteins.

Abstract Motivation Utilizing AI-driven approaches for drug–target interaction (DTI) prediction require large volumes of training data which are not available for the majority of target proteins. In this study, we investigate the use of deep transfer learning for the prediction of interactions between drug candidate compounds and understudied target proteins with scarce training data. The idea here is to first train a deep neural network classifier with a generalized source training dataset of large size and then to reuse this pre-trained neural network as an initial configuration for re-training/fine-tuning purposes with a small-sized specialized target training dataset. To explore this idea, we selected six protein families that have critical importance in biomedicine: kinases, G-protein-coupled receptors (GPCRs), ion channels, nuclear receptors, proteases, and transporters. In two independent experiments, the protein families of transporters and nuclear receptors were individually set as the target datasets, while the remaining five families were used as the source datasets. Several size-based target family training datasets were formed in a controlled manner to assess the benefit provided by the transfer learning approach. Results Here, we present a systematic evaluation of our approach by pre-training a feed-forward neural network with source training datasets and applying different modes of transfer learning from the pre-trained source network to a target dataset. The performance of deep transfer learning is evaluated and compared with that of training the same deep neural network from scratch. We found that when the training dataset contains fewer than 100 compounds, transfer learning outperforms the conventional strategy of training the system from scratch, suggesting that transfer learning is advantageous for predicting binders to under-studied targets. Availability and implementation The source code and datasets are available at https://github.com/cansyl/TransferLearning4DTI. Our web-based service containing the ready-to-use pre-trained models is accessible at https://tl4dti.kansil.org.