Protein complexes are fundamental to all biological processes. Public repositories have expanded to include millions of potential protein-protein interactions (PPIs) from human and diverse model organisms. Yet, large-scale structural characterization of these complexes-especially across different biological kingdoms-has lagged far behind, leaving most potential and unidentified interactions unresolved. Here, we present a comprehensive atlas of 1.1 million predicted protein-protein interaction structures generated with the AlphaFold2-based ColabFold framework. This dataset spans proteome-wide interactions from bacteria, archaea, humans, mice, plants, and human-virus pairs. Overall, we identify 181,671 high-confidence protein complex structures, especially 37,855 in the human interactome. Structural clustering revealed numerous conserved protein complex architectures shared across kingdoms, providing insights into previously uncharacterized biological functions. Supported by co-immunoprecipitation experiments, we further identify candidate viral receptors for Human mastadenovirus A and Papiine alphaherpesvirus 2. Comparative analyses integrating our complex structures with the AlphaFold monomeric structure database uncovered widespread gene fusion and fission events during evolution. Finally, we demonstrate how our dataset can enhance protein binding-surface prediction using deep learning approaches, illustrating its broad utility beyond structural modeling alone. Altogether, this atlas to our knowledge, represents one of the most extensive cross-kingdom resources and opens avenues for future discoveries in various biomedical applications.

To advance the computational simulation of cellular life, we propose a virtual yeast, an artificial intelligence (AI)-driven agent that models eukaryotic cellular behaviours by integrating multimodal biological data, mechanistic reasoning and active experimentation using Saccharomyces cerevisiae as a genetically tractable and data-rich model system. Cellular complexity is decomposed into eight function-centred modules, spanning genetic, metabolic and structural systems, each realized as a domain-specific AI tool coordinated through a large language model-based orchestration layer. Built on three data pillars, namely, mechanistic knowledge, subcellular architecture and dynamic states, the system integrates representation learning and generative modelling within a closed-loop learning pipeline that autonomously designs and executes experiments. The virtual yeast serves as both a conceptual and an operational platform to optimize biosynthetic pathways, support the generation and prioritization of hypotheses across diverse cellular processes, and accelerate target discovery. By coupling biological realism with autonomous AI reasoning, the virtual yeast establishes a generalizable blueprint for constructing virtual eukaryotic cells and advancing synthetic biology.

MOTIVATION:The plasma proteome offers a dynamic window of human health, capturing the real-time intersections between genetics and physiology. However, the application of deep learning to proteomics is currently hindered by a reliance on large-scale labeled datasets, rendering standard models ineffective for rare or novel diseases where patient samples are inherently scarce. RESULTS:Here, we present ProMeta, a meta-learning framework designed to enable robust disease modeling under extreme data restrictions. By integrating knowledge-guided pathway encoding with bi-level meta-optimization, ProMeta projects unstructured proteomic profiles into biologically interpretable functional tokens. This architecture allows the model to learn a global initialization containing transferable biological priors from biobank-scale data, facilitating rapid adaptation to novel tasks. Through comprehensive benchmark experiments, ProMeta consistently outperformed transfer learning and traditional machine learning baselines in both disease diagnosis and prediction tasks. In the most challenging 4-shot scenarios (utilizing only 2 cases and 2 controls), the model achieved robust generalization with an average AUROC of ∼0.69, representing a 24.6% relative improvement over the best-performing baseline methods. Mechanistic investigation revealed that ProMeta disentangles cases from controls in the latent space prior to task-specific adaptation, confirming the acquisition of universal biological rules rather than rote memorization. Furthermore, gradient-based interpretation identified disease-specific protein biomarkers and functional pathways consistent with known pathophysiology. Collectively, ProMeta overcomes the data-scarcity bottleneck in precision medicine, providing a scalable, interpretable framework for characterizing the full spectrum of human diseases, particularly for rare conditions lacking extensive clinical cohorts. AVAILABILITY AND IMPLEMENTATION:The source code of ProMeta is available at GitHub (https://github.com/lihan97/ProMeta).

Spatial transcriptomics reveals tissue organization but lacks in vivo lineage-tracing methods applicable to humans. We introduce SpaceTracer, a computational framework that accurately detects somatic single-nucleotide variants (SNVs) directly from spatial transcriptomics data. By leveraging naturally occurring somatic SNVs, SpaceTracer reconstructs cellular phylogenies within native tissue architecture, enabling the mapping of lineage spread, migration, lineage-coupled expression changes and lineage-aware local interactions. Applied to human cutaneous squamous cell carcinoma, it traced tumor initiation and progressions, uncovered widespread pre-invasive migration of dedifferentiated epithelial cells and characterized mutant B cells migrating from tertiary lymphoid structures (TLS) into the tumor boundary. The framework also reconstructed developmental lineages across multiple tissues and identified tissue-resident mutant immune cells. SpaceTracer thus provides a perturbation-free platform for high-resolution spatiotemporal lineage tracing, offering a transformative tool for elucidating complex biological systems--especially tumor-immune ecosystems--with direct implications for advancing cancer immunotherapy. ### Competing Interest Statement The authors have declared no competing interest. Westlake Laboratory of Life Sciences and Biomedicine (Hangzhou 310024, Zhejiang, China), under the grant "Key R&D Program of Zhejiang" () as well as the National Natural Science Foundation of China (32270682), 2024SSYS0034 The National Natural Science Foundation of China, 32270682 Research Center for Industries of the Future, Westlake University, WU2023C020

Amyotrophic lateral sclerosis (ALS) is a fatal and incurable neurodegenerative disease caused by the selective and progressive death of motor neurons (MNs). Understanding the genetic and molecular factors influencing ALS survival is crucial for disease management and therapeutics. In this study, we introduce a deep learning-powered genetic analysis framework to link rare noncoding genetic variants to ALS survival. Using data from human induced pluripotent stem cell (iPSC)-derived MNs, this method prioritizes functional noncoding variants using deep learning, links cis-regulatory elements (CREs) to target genes using epigenomics data, and integrates these data through gene-level burden tests to identify survival-modifying variants, CREs, and genes. We apply this approach to analyze 6,715 ALS genomes, and pinpoint four novel rare noncoding variants associated with survival, including chr7:76,009,472:C>T linked toCCDC146. CRISPR-Cas9 editing of this variant increasesCCDC146expression in iPSC-derived MNs and exacerbates ALS-specific phenotypes, including TDP-43 mislocalization. SuppressingCCDC146with an antisense oligonucleotide (ASO), showing no toxicity, completely rescues ALS-associated survival defects in neurons derived from sporadic ALS patients and from carriers of the ALS-associated G4C2-repeat expansion withinC9ORF72. ASO targeting ofCCDC146may be a broadly effective therapeutic approach for ALS. Our framework provides a generic and powerful approach for studying noncoding genetics of complex human diseases.

Polygenic risk score (PRS) is a widely used approach for predicting individuals' genetic risk of complex diseases, playing a pivotal role in advancing precision medicine. Traditional PRS methods, predominantly following a linear structure, often fall short in capturing the intricate relationships between genotype and phenotype. In this study, we present PRS-Net, an interpretable geometric deep learning-based framework that effectively models the nonlinearity of biological systems for enhanced disease prediction and biological discovery. PRS-Net begins by deconvoluting the genome-wide PRS at the single-gene resolution and then explicitly encapsulates gene-gene interactions leveraging a graph neural network (GNN) for genetic risk prediction, enabling a systematic characterization of molecular interplay underpinning diseases. An attentive readout module is introduced to facilitate model interpretation. Extensive tests across multiple complex traits and diseases demonstrate the superior prediction performance of PRS-Net compared with a wide range of conventional PRS methods. The interpretability of PRS-Net further enhances the identification of disease-relevant genes and gene programs. PRS-Net provides a potent tool for concurrent genetic risk prediction and biological discovery for complex diseases.

Biological tissues are composed of distinct microenvironments that spatially orchestrate gene expression and cell identity. However, the regulatory principles governing domain-specific cellular functions remain poorly understood due to the lack of effective methods for mapping gene regulatory networks (GRNs) in situ . To address this gap, we introduce STARNet, a representation learning approach that leverages heterogeneous hypergraph modeling of spatial transcriptomic and epigenomic data to resolve tissue-domain-specific regulatory interactions. By integrating graph neural networks with contrastive learning in a self-supervised framework, STARNet learns unified embeddings that preserve both multi-modal molecular features and anatomical spatial context, enabling accurate and domain-resolved GRN reconstruction within complex tissues. Benchmarking on both simulated and real datasets demonstrates that STARNet achieves state-of-the-art performance. We further demonstrate its broad applicability across diverse biological contexts, including neural development, genetic disease risk, and drug-induced developmental toxicity. In the mouse brain, it delineates region-specific regulatory networks and reconstructs spatiotemporal programs underlying neural stem cell differentiation. In human genetics, it provides a mechanistic link between genotypes and phenotypes by showing how genome-wide association study (GWAS) variants for complex diseases perturb hippocampus-specific GRNs. In developmental toxicology, STARNet reveals that drug-induced disruptions of GRNs in defined embryonic regions underlie tissue-specific vulnerability. Collectively, STARNet offers a powerful and versatile framework for resolving the spatial regulatory logic of complex tissues, providing multi-angle insights into tissue patterning, development, and disease mechanisms.

Polygenic risk scores (PRSs) predict an individual’s genetic risk for complex diseases, yet their utility in elucidating disease biology remains limited. We introduce scPRS, a graph neural network-based framework that computes single-cell-resolved PRSs by integrating reference single-cell chromatin accessibility profiles. scPRS outperforms traditional PRS approaches in genetic risk prediction, as demonstrated across multiple diseases including type 2 diabetes, hypertrophic cardiomyopathy, Alzheimer disease and severe COVID-19. Beyond risk prediction, scPRS prioritizes disease-critical cells and, when combined with a layered multiomic analysis, links risk variants to gene regulation in a cell-type-specific manner. Applied to these diseases, scPRS fine-maps causal cell types and cell-type-specific variants and genes, demonstrating its ability to bridge genetic risk with cell-specific biology. scPRS provides a unified framework for genetic risk prediction and mechanistic dissection of complex diseases, laying a methodological foundation for single-cell genetics.

In living organisms, proteins perform key functions required for life activities by interacting to form complexes. Determining the protein complex structure is crucial for understanding and mastering biological functions. Although AlphaFold2 makes a revolutionary breakthrough in predicting protein monomeric structures, accurately capturing inter-chain interaction signals and modeling the structures of protein complexes remain a formidable challenge. In this work, we report DeepSCFold, a pipeline for improving protein complex structure modeling. DeepSCFold uses sequence-based deep learning models to predict protein-protein structural similarity and interaction probability, providing a foundation for identifying interaction partners and constructing deep paired multiple-sequence alignments (MSAs) for protein complex structure prediction. Benchmark results show that DeepSCFold significantly increases the accuracy of protein complex structure prediction compared with state-of-the-art methods. For multimer targets from CASP15, DeepSCFold achieves an improvement of 11.6% and 10.3% in TM-score compared to AlphaFold-Multimer and AlphaFold3, respectively. Furthermore, when applied to antibody-antigen complexes from the SAbDab database, DeepSCFold enhances the prediction success rate for antibody-antigen binding interfaces by 24.7% and 12.4% over AlphaFold-Multimer and AlphaFold3, respectively. These results demonstrate that DeepSCFold effectively captures intrinsic and conserved protein-protein interaction patterns through sequence-derived structure-aware information, rather than relying solely on sequence-level co-evolutionary signals.

Single-cell technologies enable the dynamic analyses of cell fate mapping. However, capturing the gene regulatory relationships and identifying the driver factors that control cell fate decisions are still challenging. We present CEFCON, a network-based framework that first uses a graph neural network with attention mechanism to infer a cell-lineage-specific gene regulatory network (GRN) from single-cell RNA-sequencing data, and then models cell fate dynamics through network control theory to identify driver regulators and the associated gene modules, revealing their critical biological processes related to cell states. Extensive benchmarking tests consistently demonstrated the superiority of CEFCON in GRN construction, driver regulator identification, and gene module identification over baseline methods. When applied to the mouse hematopoietic stem cell differentiation data, CEFCON successfully identified driver regulators for three developmental lineages, which offered useful insights into their differentiation from a network control perspective. Overall, CEFCON provides a valuable tool for studying the underlying mechanisms of cell fate decisions from single-cell RNA-seq data.

>Determination of biomolecular structures is crucial for understanding biological functions and designing novel therapeutics. Numerous artificial intelligence based approaches have been developed for modeling specific types of biomolecules, aiming to improve the cost and time efficiency of wet-lab experiments. Recently, the advanced deep-learning-based methods such as Rose TTAFold All-Atom(RFAA) and AlphaFold3 have emerged for generalized biomolecular structure modeling.

The recent boom in single-cell sequencing technologies provides valuable insights into the transcriptomes of individual cells. Through single-cell data analyses, a number of biological discoveries, such as novel cell types, developmental cell lineage trajectories, and gene regulatory networks, have been uncovered. However, the massive and increasingly accumulated single-cell datasets have also posed a seriously computational and analytical challenge for researchers. To address this issue, one typically applies dimensionality reduction approaches to reduce the large-scale datasets. However, these approaches are generally computationally infeasible for tall matrices. In addition, the downstream data analysis tasks such as clustering still take a large time complexity even on the dimension-reduced datasets. We present single-cell Coreset (scCoreset), a data summarization framework that extracts a small weighted subset of cells from a huge sparse single-cell RNA-seq data to facilitate the downstream data analysis tasks. Single-cell data analyses run on the extracted subset yield similar results to those derived from the original uncompressed data. Tests on various single-cell datasets show that scCoreset outperforms the existing data summarization approaches for common downstream tasks such as visualization and clustering. We believe that scCoreset can serve as a useful plug-in tool to improve the efficiency of current single-cell RNA-seq data analyses.

Polygenic risk score (PRS) serves as a valuable tool for predicting the genetic risk of complex human diseases for individuals, playing a pivotal role in advancing precision medicine. Traditional PRS methods, predominantly following a linear structure, often fall short in capturing the intricate relationships between genotype and phenotype. We present PRS-Net, an interpretable deep learning-based framework designed to effectively model the nonlinearity of biological systems for enhanced disease prediction and biological discovery. PRS-Net begins by deconvoluting the genomewide PRS at the single-gene resolution, and then it encapsulates gene-gene interactions for genetic risk prediction leveraging a graph neural network, thereby enabling the characterization of biological nonlinearity underlying complex diseases. An attentive readout module is specifically introduced into the framework to facilitate model interpretation and biological discovery. Through extensive tests across multiple complex diseases, PRS-Net consistently outperforms baseline PRS methods, showcasing its superior performance on disease prediction. Moreover, the interpretability of PRS-Net has been demonstrated by the identification of genes and gene-gene interactions that significantly influence the risk of Alzheimer’s disease and multiple sclerosis. In summary, PRS-Net provides a potent tool for parallel genetic risk prediction and biological discovery for complex diseases.

Early identification of safe and efficacious disease targets is crucial to alleviating the tremendous cost of drug discovery projects. However, existing experimental methods for identifying new targets are generally labor-intensive and failure-prone. On the other hand, computational approaches, especially machine learning-based frameworks, have shown remarkable application potential in drug discovery. In this work, we propose Progeni, a novel machine learning-based framework for target identification. In addition to fully exploiting the known heterogeneous biological networks from various sources, Progeni integrates literature evidence about the relations between biological entities to construct a probabilistic knowledge graph. Graph neural networks are then employed in Progeni to learn the feature embeddings of biological entities to facilitate the identification of biologically relevant target candidates. A comprehensive evaluation of Progeni demonstrated its superior predictive power over the baseline methods on the target identification task. In addition, our extensive tests showed that Progeni exhibited high robustness to the negative effect of exposure bias, a common phenomenon in recommendation systems, and effectively identified new targets that can be strongly supported by the literature. Moreover, our wet lab experiments successfully validated the biological significance of the top target candidates predicted by Progeni for melanoma and colorectal cancer. All these results suggested that Progeni can identify biologically effective targets and thus provide a powerful and useful tool for advancing the drug discovery process.

MOTIVATION:Drug repositioning (DR), identifying novel indications for approved drugs, is a cost-effective strategy in drug discovery. Despite numerous proposed DR models, integrating network-based features, differential gene expression, and chemical structures for high-performance DR remains challenging. RESULTS:We propose a comprehensive deep pretraining and fine-tuning framework for DR, termed DrugRepPT. Initially, we design a graph pretraining module employing model-augmented contrastive learning on a vast drug-disease heterogeneous graph to capture nuanced interactions and expression perturbations after intervention. Subsequently, we introduce a fine-tuning module leveraging a graph residual-like convolution network to elucidate intricate interactions between diseases and drugs. Moreover, a Bayesian multiloss approach is introduced to balance the existence and effectiveness of drug treatment effectively. Extensive experiments showcase the efficacy of our framework, with DrugRepPT exhibiting remarkable performance improvements compared to SOTA (state of the arts) baseline methods (improvement 106.13% on Hit@1 and 54.45% on mean reciprocal rank). The reliability of predicted results is further validated through two case studies, i.e. gastritis and fatty liver, via literature validation, network medicine analysis, and docking screening. AVAILABILITY AND IMPLEMENTATION:The code and results are available at https://github.com/2020MEAI/DrugRepPT.

This Voices piece will highlight the impact of artificial intelligence on algorithm development among computational biologists. How has worldwide focus on AI changed the path of research in computational biology? What is the impact on the algorithmic biology research community?

Peptides offer great biomedical potential and serve as promising drug candidates. Currently, the majority of approved peptide drugs are directly derived from well-explored natural human peptides. It is quite necessary to utilize advanced deep learning techniques to identify novel peptide drugs in the vast, unexplored biochemical space. Despite various in silico methods having been developed to accelerate peptide early drug discovery, existing models face challenges of overfitting and lacking generalizability due to the limited size, imbalanced distribution and inconsistent quality of experimental data. In this study, we propose PepGB, a deep learning framework to facilitate peptide early drug discovery by predicting peptide-protein interactions (PepPIs). Employing graph neural networks, PepGB incorporates a fine-grained perturbation module and a dual-view objective with contrastive learning-based peptide pre-trained representation to predict PepPIs. Through rigorous evaluations, we demonstrated that PepGB greatly outperforms baselines and can accurately identify PepPIs for novel targets and peptide hits, thereby contributing to the target identification and hit discovery processes. Next, we derive an extended version, diPepGB, to tackle the bottleneck of modeling highly imbalanced data prevalent in lead generation and optimization processes. Utilizing directed edges to represent relative binding strength between two peptide nodes, diPepGB achieves superior performance in real-world assays. In summary, our proposed frameworks can serve as potent tools to facilitate peptide early drug discovery.

Accurate prediction of protein-ligand binding affinities is an essential challenge in structure-based drug design. Despite recent advances in data-driven methods for affinity prediction, their accuracy is still limited, partially because they only take advantage of static crystal structures while the actual binding affinities are generally determined by the thermodynamic ensembles between proteins and ligands. One effective way to approximate such a thermodynamic ensemble is to use molecular dynamics (MD) simulation. Here, an MD dataset containing 3,218 different protein-ligand complexes is curated, and Dynaformer, a graph-based deep learning model is further developed to predict the binding affinities by learning the geometric characteristics of the protein-ligand interactions from the MD trajectories. In silico experiments demonstrated that the model exhibits state-of-the-art scoring and ranking power on the CASF-2016 benchmark dataset, outperforming the methods hitherto reported. Moreover, in a virtual screening on heat shock protein 90 (HSP90) using Dynaformer, 20 candidates are identified and their binding affinities are further experimentally validated. Dynaformer displayed promising results in virtual drug screening, revealing 12 hit compounds (two are in the submicromolar range), including several novel scaffolds. Overall, these results demonstrated that the approach offer a promising avenue for accelerating the early drug discovery process.

Identifying active compounds for target proteins is fundamental in early drug discovery. Recently, data-driven computational methods have demonstrated promising potential in predicting compound activities. However, there lacks a well-designed benchmark to comprehensively evaluate these methods from a practical perspective. To fill this gap, we propose a Compound Activity benchmark for Real-world Applications (CARA). Through carefully distinguishing assay types, designing train-test splitting schemes and selecting evaluation metrics, CARA can consider the biased distribution of current real-world compound activity data and avoid overestimation of model performances. We observed that although current models can make successful predictions for certain proportions of assays, their performances varied across different assays. In addition, evaluation of several few-shot training strategies demonstrated different performances related to task types. Overall, we provide a high-quality dataset for developing and evaluating compound activity prediction models, and the analyses in this work may inspire better applications of data-driven models in drug discovery.