MOTIVATION:AlphaFold has transformed structural biology with an unprecedented accuracy in modelling protein structures and their interactions with biomolecules, with AlphaFold3 (AF3) achieving state-of-the-art performance. However, AF3 and other methods often struggle to accurately predict the structure of protein complexes that lack strong co-evolutionary information, such as antibody-antigen (Ab-Ag) complexes. One of the fundamental issues is that AF3 often generates accurate predictions, but fails to reliably distinguish them from the much larger set of incorrect ones. RESULTS:To address this, we propose ABAG-Rank, a deep neural network that provides an efficient and robust solution for model selection of Ab-Ag interactions from a pool of structural ensembles predicted with AlphaFold. Built on the permutation-invariant DeepSets architecture, ABAG-Rank can process variable-sized ensembles of structural decoys and is directly applicable to prediction settings in which the number of candidates may vary. We train a model on a redundancy-reduced set of all known antibody-antigen complexes and find that simple geometric descriptors, along with confidence scores from AlphaFold, provide rich information about interface quality without requiring intensive physics-based calculations. Our experiments demonstrate that ABAG-Rank significantly outperforms AF3 internal scoring and the ranking performance of existing deep learning baselines. AVAILABILITY AND IMPLEMENTATION:Source code can be found at: https://github.com/tadteo/ABAG-Rank or on Zenodo at https://doi.org/10.5281/zenodo.21132090. SUPPLEMENTARY MATERIAL:Supplementary data are available at Bioinformatics online.
SUMMARY:A growing number of tools enable the analysis of large variant libraries produced by multiplexed assays of variant effects (MAVEs). Experiments using fluorescent reporters and fluorescence-activated cell sorting sequencing (FACS-seq or Sort-seq) can coarsely quantify a given variant's impact on phenotypes such as transcription activity. Existing bioinformatics tools for Sort-seq data broadly fall into two categories: methods that model a genotype-phenotype landscape to infer latent variant phenotypes, and methods that directly estimate individual variant scores from experimental binned counts. Within this second category, sortscore provides an activity score computed directly from observed counts without fitting a model, retaining the original experimental scale when bin median values are known.We present sortscore, a python package that incorporates a standard Sort-seq scoring method and normalization across sorted samples, technical replicates from separate sort times, and across oligos in tiled experiments. It also provides convenient heatmap visualizations. This software was used to analyze DMS experiments for the transcription factor GLI2. This work seeks to lower the barrier to entry and provide a clear starting point for scoring Sort-seq cell-based functional assays. This extends the availability of well-documented and reproducible data analysis protocols to a wider community employing MAVE techniques. AVAILABILITY AND IMPLEMENTATION:The package can be downloaded from PyPI using the pip installer. The code is also freely accessible and available for reuse through a Public GitHub repository (MIT License). Installation instructions, documentation, and tutorials are accessible at the sortscore GitHub repository: https://github.com/dbaldridge-lab/sortscore. A snapshot of the code and data is available on Zenodo: https://doi.org/10.5281/zenodo.22119031. SUPPLEMENTARY INFORMATION:Supplementary figures are available at Bioinformatics online.
MOTIVATION:Modelling long genomic sequences remains challenging due to extreme sequence length, high redundancy, and the need for biological interpretability. Although Transformer-based architectures have achieved strong performance across genomic tasks, their high computational cost and reliance on fixed tokenization strategies limit their scalability and ability to focus on biologically informative regions. RESULTS:We propose ATSFormer, a token-pruning Transformer framework for efficient and biologically informed genomic sequence modelling. ATSFormer incorporates an attention-guided and parameter-free Adaptive Token Sampling (ATS) module into Transformer layers. Guided by attention-derived importance scores, ATS dynamically retains informative tokens while probabilistically discarding redundant ones, thereby reducing sequence length, FLOPs, and memory usage without introducing additional learnable parameters or extra training procedures. Importantly, the retained tokens correspond to key contributors to model predictions, enabling ATSFormer to highlight biologically meaningful sites and sequence motifs. We evaluated ATSFormer on four benchmark RNA modification datasets derived from RMVar 2.0, covering A-to-I, m1A, m5C, and m7G. Experimental results show that ATSFormer consistently outperforms existing state-of-the-art methods while achieving substantial computational savings. Furthermore, structural analysis using AlphaFold3 supports the biological relevance of the motifs identified by ATSFormer. AVAILABILITY AND IMPLEMENTATION:The source data and code are freely available at GitHub (https://github.com/1gao2/ATSFormer) and Zenodo (https://doi.org/10.5281/zenodo.21813541).
SUMMARY:The human genome is dominated by repetitive DNA, whose genetic and epigenetic variation plays a key role in gene regulation, genome stability, and disease. Recent advances in long-read sequencing now enable large-scale, haplotype-resolved, and DNA methylation-informative analysis of the human genome, including on previously inaccessible complex and repetitive regions. However, the comprehensive, simultaneous characterisation of the "human repeatome" remains challenging, largely due to the lack of comprehensive tools integrated in a single pipeline that can capture the full spectrum of variation across diverse types of DNA repeats. Here, we present ECHO, a user-friendly, Snakemake-based pipeline for the "(Epi)genomic Characterisation of Human Repetitive Elements using Oxford Nanopore Sequencing". ECHO provides a reproducible and scalable framework for end-to-end analysis of whole-genome nanopore sequencing data, enabling integrative but also tailored (epi)genetic analyses of the human repeatome. AVAILABILITY AND IMPLEMENTATION:ECHO is freely available at Github: https://github.com/leenput/ECHO-pipeline, with the archived version at Zenodo: https://zenodo.org/records/19068468. SUPPLEMENTARY INFORMATION:Supplementary data are available at Bioinformatics online.
MOTIVATION:As unique drugs positioned between small and macro molecules, anticancer peptides (ACPs) hold great potential in oncotherapy owing to their high selectivity and low toxicity. Nowadays, computational ACP prediction has emerged as a cost-effective alternative to bioassay screening, but most methods are limited to identifying bioactivity and fail to resolve tumor cell-specific targeting, primarily because of the sparse annotated data. RESULTS:To fill this gap, we integrate a hybrid dataset compiled from five well-established peptide databases and propose TargetPC, a deep learning method tailored for cell line-targeted ACP prediction. TargetPC encodes multimodal representations of ACPs and cell lines via pretrained protein and omics models, and combines them via hierarchical intra- and inter-modal fusion for targeting prediction. This combination is further augmented by a mutual learning paradigm that distills domain knowledge from both ACP and cell line, enabling improved generalization under sparse supervision. Experimental results on the hybrid dataset demonstrate the effectiveness of TargetPC, which outperforms the state-of-the-art baselines in terms of prediction accuracy, and maintains strong generalization to unseen ACPs and cell lines. When extended to out-of-distribution samples, TargetPC has successfully screened dozens of novel ACPs targeted to breast cancer cells and uncovered biological motifs underlying its predictions. As a result, our TargetPC is expected to serve as a versatile tool for lead ACP discovery at a lower burden. AVAILABILITY:The source code and data are available at GitHub (https://github.com/liuxuan666/TargetPC).
MOTIVATION:Single-cell multiomics reveals regulatory relationships across biological layers but captures only static snapshots, obscuring the dynamics coordinated across modalities. RNA velocity predicts transcriptome dynamics, yet cannot be extended to other layers such as the regulome, leaving chromatin accessibility dynamics unresolved. RESULTS:We developed mmVelo, a deep generative model that infers cell state dynamics from spliced and unspliced mRNA and projects them onto other modalities, yielding chromatin velocity at single-peak resolution. In developing mouse brain, mmVelo accurately recovered accessibility dynamics; in mouse skin, it identified transcription factors regulating accessibility. Decomposing posterior velocity variability into manifold-aligned and off-manifold components revealed modality-specific uncertainty structure, with chromatin fluctuation elevated near lineage branching. Using multiomics data as a bridge, mmVelo inferred the dynamics of missing modalities from single-modal human brain data. AVAILABILITY AND IMPLEMENTATION:Source code is freely available under the MIT license at https://github.com/nomuhyooon/mmVelo; the version and test data used here are archived at https://doi.org/10.5281/zenodo.20103609. SUPPLEMENTARY INFORMATION:Supplementary data are available at Bioinformatics online.
MOTIVATION:Whole-genome bisulfite sequencing (WGBS/BS-Seq) is the gold standard for single-base resolution DNA methylome profiling. However, the diverse statistical models of existing computational methods lead to limited overlap between their results, highlighting the need for novel methods to detect Differentially Methylated Positions (DMPs) and Differentially Methylated Regions (DMRs). RESULTS:We developed MultiDMPcaller, an automated downstream methylome analysis software. It processes upstream outputs to profile DMPs, Non-DMPs, DMRs, and context-specific (CpG/CHG/CHH) methylation status, alongside visualizing their chromosomal distribution and enrichment. The software features two key innovations: i) An adaptive two-step p-value adjustment strategy based on organism-specific methylation patterns, with raw p-value ≤ 0.05 pre-filtering followed by false discovery rate (FDR) correction, to recover potential DMPs usually missed by standard FDR correction in plant CHG/CHH and animal CpG contexts; ii) A multiple pairwise comparison approach, which performs m×n pairwise comparisons for m control and n experimental replicates, followed by a voting system supporting both user-defined majority thresholds and model-based adaptive thresholds, to identify robust and reliable DMPs (with a stricter voting threshold exclusively for loci with low methylation differences) and DMRs. On real datasets from Arabidopsis, apple, and mouse, as well as simulated human datasets, MultiDMPcaller's results showed good agreement with those of other software, exhibiting high conservativeness and superior precision, which suggested a low false discovery proportion. AVAILABILITY:MultiDMPcaller is available at GitHub (https://github.com/jiantaoyuNWAFU/MultiDMPcaller) and via a web server (https://ciebioinfo.nwafu.edu.cn).
MOTIVATION:Cold-start drug-target affinity prediction remains challenging because static interaction mechanisms cannot adapt to individual drug-target pairs. RESULTS:We propose PCIM-DTA, which constructs pair-level interaction representations and derives a pair-specific condition vector from global drug and target features. The condition vector modulates attention, pair-token features, distribution-aware recalibration, and regression parameters, while graph message passing captures higher-order dependencies. Experiments on Davis and BindingDB-Kd show that PCIM-DTA achieves competitive or superior performance under Warm, Cold-drug, Cold-target, Cold-both, and Scaffold-drug settings. Ablation studies support the contribution of each component. AVAILABILITY AND IMPLEMENTATION:The datasets used in this study are publicly available, including the Davis and BindingDB-Kd datasets. The implementation code of PCIM-DTA is publicly available at https://github.com/1322469934/PCIM-DTA. SUPPLEMENTARY INFORMATION:Supplementary data are available at Bioinformatics online.
SUMMARY:Plotgardener is an R package used for generating high-quality genomic visualizations. Despite its broad range of functions and versatility, its reliance on code presents a barrier for many potential users. To address this, we developed a macOS desktop application version of Plotgardener that enables users to create publication-ready genomic plots with no programming experience. The application employs a modular architecture comprising an Electron.js backend, a React frontend, and a Python parser that dynamically analyzes the Plotgardener package to ensure interface fields remain synchronized with package updates. By lowering the technical barrier to advanced genomic visualization, the Plotgardener desktop application broadens access to powerful visualization workflows for researchers and clinicians. AVAILABILITY:The current release of the Plotgardener App is an open source macOS desktop application built with Electron (Node.js), featuring a React frontend and a Python-based parser. The download link is available at https://phanstiellab.github.io/plotgardener/articles/guides/plotgardenerApp.html and on Zenodo (doi: https://doi.org/10.5281/zenodo.21684228). The source code is hosted on GitHub at https://github.com/rishabhsvemuri/ThePlotgardenerApp. SUPPLEMENTARY INFORMATION:The original R package documentation can be found at https://phanstiellab.github.io/plotgardener/.
MOTIVATION:Genomic sequence-to-activity models can decipher gene regulatory mechanisms and predict the functional impact of regulatory variants. However, current models struggle to integrate information from sequences outside promoters, especially information from cell type specific regulatory elements. RESULTS:Here, we propose incorporating base-pair resolution evolutionary conservation data into genomic sequence-to-expression predictors. We explore two training strategies-training from scratch or fine-tuning an existing sequence-only model with additional conservation input. We find that in both cases, base-pair resolution conservation data improves cell type specific sequence-to-expression prediction, with training from scratch yielding the greatest benefit. The improvement in cell type specific expression prediction can be attributed in part to the fact that models trained on sequence and conservation data learn to better recognize cell type specific regulatory elements than models trained on sequence alone. AVAILABILITY:Code is available at https://github.com/ni-lab/basenji-phyloP. SUPPLEMENTARY INFORMATION:Supplementary data are available at Bioinformatics online.
MOTIVATION:High-throughput genomic technologies have transformed our understanding of biological systems, yet direct comparison and visualisation of these complex datasets remains challenging. Existing normalisation methods often fail to align genomic signal across samples due to sensitivity to sequencing depth differences and localised high-signal artefacts, leading to inconsistent replicate behaviour and increased downstream variability. RESULTS:We introduce Zone Equalisation Normalisation (ZEN), a novel approach designed to improve cross-sample signal alignment of genomic data. ZEN rescales genomic signal based on variance estimated within biologically enriched regions, reducing the influence of extreme outliers while preserving underlying biological structure. Using a diverse collection of data and our new genome-wide benchmarking approach, we reveal that ZEN improves biological and technical replicate alignment across the majority of tested conditions and experimental platforms. We further show that this improved signal comparability is associated with fewer differential accessibility calls between technical replicates and a more conservative set of biological differences. Together, these results demonstrate that ZEN provides a complementary framework to improve the accuracy and reliability of genomic data analysis and that normalisation choice can affect downstream analyses and biological interpretation. AVAILABILITY AND IMPLEMENTATION:ZEN is available as an open-source Python package via conda and PyPI. Source code, documentation, tutorials, and code to reproduce the analyses are available at https://github.com/Genome-Function-Initiative-Oxford/Zone-Equalisation-Normalisation and Zenodo (https://doi.org/10.5281/zenodo.21067751).
SUMMARY:SeqUIaSCOPE is an open-source platform designed for routine clinical oncology diagnostics through case-centric integration and visualization of genomic variants, fusion events, and expression profiles. The platform combines molecular-level validation via embedded genome browsing with systems-level interpretation through dynamic pathway visualization, enabling geneticists to assess how alterations converge across biological networks. Flexible reporting with customizable templates accommodates diverse institutional requirements, while secure cluster-based or local deployment ensures compliance with data protection policies, making advanced multi-omics diagnostics accessible to academic and clinical institutions. AVAILABILITY AND IMPLEMENTATION:SeqUIaSCOPE is freely available on GitHub at https://github.com/BioIT-CEITEC/sequiascope under the MIT license and archived at Zenodo (https://zenodo.org/records/21338445). Due to the sensitive nature of patient data, the repository provides simulated datasets that mimic the structure of real clinical data for testing and exploration. Documentation and a live demo accompany these datasets, allowing users to explore the application without any prior setup. The repository also includes a Helm chart for Kubernetes deployment and Docker containers for local deployment, ensuring compatibility across Linux, macOS, and Windows. No user registration is required, and all data remains on local or institutional infrastructure.
MOTIVATION:Knowledge graphs (KGs) organize complex biomedical knowledge into structured representations of entities and relations. Knowledge graph embedding (KGE) models learn compact representations of KGs, and are widely applied for biomedical link prediction. Despite extensive work on KGE models, current evaluations often overlook the issue of data leakage, which can artificially inflate performance and undermine benchmark validity. Data leakage can arise when (i) there is redundancy between training and test sets, (ii) the model leverages illegitimate features, or (iii) the test set does not accurately reflect real-world inference. RESULTS:We assess the impact of data leakage on KGE-based link prediction across three biomedical KGs, using decoder-only and GNN-based models. We first demonstrate the impact of train-test redundancies and implement a systematic procedure to detect and remove them. Using permutation experiments, we also investigate whether node degree acts as an illegitimate predictive feature, and find no evidence that predictions are driven by degree alone. Finally, we evaluate how well common test set sampling strategies reflect real-world inference in drug repurposing. We compare random and cold-start data splits with an independent test set from Orphanet, and observe a substantial performance drop on the latter, indicating that current benchmarking practices may overestimate how well KGE models generalize to practical applications. Overall, our findings highlight the importance of rigorous benchmark design and careful evaluation of the generalization ability of KGE models for biomedical link prediction. AVAILABILITY AND IMPLEMENTATION:Code and results are openly available on GitHub at https://github.com/galadrielbriere/data_leakage_kge_benchmark.git and archived on Zenodo at https://doi.org/10.5281/zenodo.21885112.
SUMMARY:Recent technological advances have propelled comparative functional genomics into the single-cell era, spurring a rapid development of methods to analyse these complex datasets. However, comparing single-cell gene expression across species to quantify expression similarity and ultimately identify homologous cell types remains an open problem. The ICC algorithm (Iterative Correlation of Coexpression) has been recently proposed as an attractive approach to tackle this challenge, but, to date, no software implementation is available. Here, we introduce Pesci (Pretty Easy Single-cell Comparisons using ICC), an efficient and user-friendly implementation of the ICC algorithm applied to pairwise comparisons of single-cell gene expression atlases across species. AVAILABILITY:Pesci is implemented in Python 3 (≥3.7). It is available for download on Linux, macOS and Windows via pip, conda and GitHub at https://github.com/eparey/pesci. The source code is permanently archived on Zenodo (https://doi.org/10.5281/zenodo.21477543).
MOTIVATION:Optimizing cell-culture media for cultivated meat and cellular agriculture is both experimentally demanding and expensive. Traditional approaches use Design of Experiments (DOE) to iterate through candidate formulations until a satisfactory medium emerges. Data-driven methods, and in particular Bayesian optimization (BO), can reach higher objective function values (here, titer) with fewer evaluations than classical DOE. Prior comparisons of optimizers on cell-media data have, however, relied largely on computational simulations rather than measured formulations, leaving researchers without evidence-based guidance on which strategy to adopt. RESULTS:We introduce two complementary benchmark protocols and apply them to six published cell-culture datasets. In the "Hide-the-Label" protocol each optimizer begins with only a partial view of the data and sequentially reveals hidden formulations, singly or in batches, reproducing the stepwise acquisition of costly laboratory measurements; the "Open Race" instead measures search efficiency of a method under a fixed evaluation budget. Both protocols were run across four generative surrogate classes [Gaussian process (GP), random forest, neural network, and Bayesian neural network], two difficulty regimes, and batch sizes of one to 20. GP-based BO was consistently the most sample-efficient strategy, locating hidden targets in substantially fewer evaluations than classical DOE designs or random search; its advantage narrowed but remained under injected noise and multi-modality, and optimizers built on more complex surrogates (deeper networks or adaptive ensembles) performed worse rather than better on these harder landscapes. We detail the mathematical basis of the benchmark, criteria for selecting suitable datasets, and how to obtain statistically robust conclusions in the presence of measurement bias and heterogeneity, and we provide a roadmap for selecting an optimization strategy in practice. AVAILABILITY AND IMPLEMENTATION:Source code and the benchmark datasets are freely available at https://github.com/Amii-Applied-AI/amii-cell-ag-tools/tree/main/active-learning-for-cell-media and archived at https://doi.org/10.5281/zenodo.21501659.
MOTIVATION:Hi-C is a powerful technology for mapping chromatin interactions genome-wide. However, interaction loops identified from Hi-C contact maps often vary across replicate experiments due to experimental noise, making reproducibility assessment essential. A major challenge lies in the genomic distance dependence of interaction strength, which systematically affects reproducibility but is overlooked by existing methods for reproducibility assessment. RESULTS:We introduce Stratum-Adjusted Irreproducible Discovery Rate (SIDR), a novel statistical model that integrates distance stratification into the widely-used Irreproducible Discovery Rate (IDR) framework. SIDR explicitly models the confounding effect of genomic distance, enabling global control of irreproducibility across interaction ranges. Through simulations and real Hi-C datasets, we demonstrate that SIDR improves discriminative power and recovers more biologically meaningful interactions than existing approaches, making it a valuable tool for robust and reproducible Hi-C analysis. AVAILABILITY:The R package SIDR is freely available on GitHub https://github.com/qunhualilab/SIDR.
SUMMARY:Pangenome variation graphs integrate multiple samples into a unified representation, mitigating the reference bias inherent to linear genomes. However, these graphs can be large and structurally complex. Existing visualization tools are each confined to a fixed scale of resolution, requiring researchers to switch between multiple tools to examine variation at different levels of detail. PangyPlot is an interactive pangenome browser designed for multi-scale exploration of reference variation graphs from full chromosome to nucleotide-level sequence segments. PangyPlot anchors navigation to linear reference coordinates, organizes variation into hierarchical bubble structures, and uses a force-directed layout engine for automatic node arrangement. AVAILABILITY AND IMPLEMENTATION:An instance preloaded with data is available at https://pangyplot.research.sickkids.ca. Source code and documentation are openly available at https://github.com/strug-hub/pangyplot under the MIT License.
MOTIVATION:The accelerating crisis of antimicrobial resistance among the critical so-called ESKAPE pathogens demands the urgent identification of novel molecular targets. However, a substantial fraction of ESKAPE proteomes remains functionally uncharacterized, with many genes annotated as encoding hypothetical proteins. These protein sequences often lack significant similarity to known protein families when conventional homology-based annotation methods are used and thus remain "dark". This limits our ability to explore their roles in pathogenicity, and it is thus crucial to bridge this substantial gap in pathogen biology by developing new strategies to illuminate these "dark" regions of the ESKAPE pan-proteome. RESULTS:We introduce ECLIPSE (ESKAPE Connectome Linkage and Inference for Proteome Sequence Exploration), a network-based computational framework that systematically identifies and prioritizes functionally dark protein families in ESKAPE pan-proteomes. ECLIPSE embeds target ESKAPE pathogen proteomes within the global sequence similarity network of the Protein Universe Atlas. It detects connected components composed entirely of unannotated proteins, called the "dark proteome." As a case study, we applied ECLIPSE to a pan-proteome of 3 460 657 protein sequences from 635 strains of Pseudomonas aeruginosa (PA). ECLIPSE identified 120 985 proteins (4%) residing in completely dark connected components. Furthermore, we have performed a taxonomic diversity analysis using normalized Shannon indices to characterize each dark component by its enrichment in ESKAPE pathogens. The analysis utilized the evenness (E) value (see Methods 2.1), which distinguishes Pseudomonas-specific (target-specific) from ESKAPE-enriched dark components. We then developed the Dark Proteome Prioritization Score (DPPS), a composite multidimensional scoring framework (see Methods 2.5). It ranks these dark components by biological relevance across four orthogonal axes: (i) functional darkness, (ii) P. aeruginosa proportion in the Atlas, (iii) AMR-clade taxonomic restriction, and (iv) conservation across the 635 P. aeruginosa strains. This framework outputs a robust four-tier scoring system; the prioritized Tier I components were validated by weight sensitivity analysis and remained stable across 500 Monte Carlo weight perturbations. Structural characterization of one of the top-ranked ESKAPE-enriched dark components revealed that it belongs to the beta-barrel fold DUF1302 (PF06980) family, for which no experimentally solved three-dimensional structure exists in the PDB. The genomic context analysis indicates that it is co-localized with a LuxR-type transcriptional regulator. Collectively, ECLIPSE identifies evolutionarily conserved, structurally defined, and functionally dark proteins enriched across ESKAPE pathogens; these dark proteins can further be utilized as alternative antimicrobial targets for experimental characterization. AVAILABILITY AND IMPLEMENTATION:The source code and dataset are available for free at: Github: https://github.com/surabhilata/ECLIPSE.git, Zenodo: DOI: 10.5281/zenodo.21064323.
MOTIVATION:Cross-modal translation enables reconstruction of missing modalities in single-cell multi-omics data, supporting integrative analyses of cellular heterogeneity and regulatory relationships. However, existing methods often struggle to disentangle shared biological signals from modality-specific variation and to generalize across datasets. RESULTS:We present scDAU, a deep learning framework that combines conditional diffusion-based feature regularization with multi-scale cross-modal translation networks. scDAU employs a feature decoupling strategy to separate shared semantic representations from modality-specific components, followed by U-Net-based architectures for accurate bidirectional translation between modalities. Across multiple benchmark datasets, scDAU outperforms existing methods in both within-dataset and cross-dataset settings, as well as in predicting modalities for previously unseen cell types. The framework further generalizes to transcriptome-proteome translation, demonstrating flexibility across diverse multi-omics contexts. Application to a human glioblastoma dataset showed that scDAU preserves cell-type-specific gene expression and chromatin accessibility patterns, supporting downstream analyses such as marker identification and functional enrichment. Overall, scDAU provides a robust and extensible approach for cross-modal translation. AVAILABILITY:The source code of scDAU is available at https://github.com/zhyu-lab/scdau and https://doi.org/10.5281/zenodo.19303337.
MOTIVATION:Predicting protein-RNA binding affinity is crucial for understanding cellular regulation and advancing RNA-targeted drug discovery. However, this task remains challenging due to structural complexity, limited labeled data, and insufficient modeling of fine-grained interactions. RESULTS:We propose M2-PRNet, a multi-scale and multi-modal framework that integrates atom-level graphs, residue-level graphs, and tri-view molecular representations to capture complementary structural information. A cross-scale contrastive learning objective is introduced to align representations across different structural resolutions of the same complex. Under a clustering-based five-fold cross-validation setting on benchmark datasets, M2-PRNet achieves state-of-the-art performance. To further assess generalization under reduced sequence homology, we construct homology-aware RNA-cold, protein-cold, and dual-cold evaluations under a stricter 40% sequence identity threshold, where M2-PRNet maintains competitive performance. To account for conformational flexibility, we evaluate the model on MD150-1ns and an extended MD75-10ns subset, demonstrating stable performance under MD-derived structural perturbations. In addition, representative case studies suggest that M2-PRNet can highlight relevant RNA-binding regions and support preliminary discrimination between strong and weak binders when plausible complex structures are available. These results demonstrate the effectiveness of integrating multi-scale and multi-modal representations with cross-scale alignment for protein-RNA affinity prediction. AVAILABILITY AND IMPLEMENTATION:The source code and datasets for M2-PRNet are freely available at https://github.com/CSUBioGroup/M2-PRNet.
