Latest bioRxiv papers

Category: bioinformatics — Showing 50 items

• Generating Structurally Diverse Therapeutic Peptides with GFlowNet

  Wijaya, E.

  AI Summary

  • The study addresses mode collapse in reinforcement learning for therapeutic peptide generation by introducing GFlowNet, which samples sequences proportionally to reward.
  • GFlowNet was compared with GRPO, showing GFlowNet achieves greater sequence diversity without explicit diversity penalties.
  • When diversity penalties were removed, GRPO failed while GFlowNet maintained diversity, highlighting GFlowNet's robustness in drug discovery applications.

  Abstract

  Reinforcement learning approaches for therapeutic peptide generation suffer from mode collapse, converging to narrow regions of sequence space even when explicit diversity penalties are applied. Fine-grained analysis reveals persistent mode-seeking behavior invisible to standard diversity metrics. We propose GFlowNet for peptide generation, which samples sequences proportionally to reward rather than maximizing expected reward. This provides intrinsic diversity without diversity penalties. Comparing against GRPO with explicit diversity enforcement, GFlowNet achieves substantially more uniform sequence sampling and fewer repetitive motifs. Critically, when diversity mechanisms are removed from the reward, GRPO collapses completely while GFlowNet maintains natural diversity. These results demonstrate that proportional sampling is inherently robust to reward function design, offering a key advantage for drug discovery pipelines requiring diverse candidates.

  bioinformatics2026-02-25v4

• Protein Compositional Ratio Representation (PCRR)Systematically Improves Human Disease Prediction

  Madduri, A. V.; Ellis, R. J.; Patel, C. J.

  AI Summary

  • The study proposes using pairwise protein ratios (log(A)-log(B)) in machine learning to better capture the compositional nature of proteomic data for disease prediction.
  • Applied to the ROSMAP cohort, this approach improved Alzheimer's subtype classification with an average AUROC increase of +0.1274.
  • In the UK Biobank dataset, the ratio-based model outperformed raw protein level models in 95.1% of diseases, with significant improvements in 56.7%.

  Abstract

  Plasma proteomics captures a functional snapshot of human physiology; yet, most machine learning models treat protein abundances as independent variables, ignoring the fact that biological systems and proteomic measurements are inherently compositional. Many molecular processes depend not on absolute concentrations but on relative balances: receptor-ligand stoichiometry, enzyme-substrate ratios, and homeostatic feedbacks that govern signaling and metabolism. We propose that these relationships are best captured through pairwise protein ratios, which more faithfully reflect underlying biochemical constraints than raw expression values. We evaluate a machine learning framework that models pairwise log-ratios of proteins (log(A)-log(B)) as features, thereby encoding compositional structure directly into the learning space. Applied to the ROSMAP plasma proteomics cohort (n = 871), this approach substantially improved the classification of Alzheimer's subtypes (NCI, MCI, AD, AD+) with an average AUROC gain of +0.1274 over a strong baseline that incorporated raw proteomics and demographics. The top-ranked ratios(e.g., SEMA3C:TMEM70, IDUA:NPTXR) captured converging pathogenic pillars of Alzheimer's disease, including microglial activation, proteostasis dysregulation, and lipid-clearance imbalance, highlighting that ratio-based features recover biologically coherent axes of disease. To assess generality, we scaled the method to the UK Biobank proteomic dataset (n > 53,000; 587 phenotypes). The ratio-based model outperformed raw-level models in 95.1% of diseases, with statistically significant (FDR < 0.05) gains in 56.7%. Together, these results suggest that proteomic data should be viewed and modeled as compositional systems, where relative protein abundances carry the accurate functional signal. This insight supports the broader utility of ratio-based representations for disease prediction and biomarker discovery.

  bioinformatics2026-02-25v4

• Distilling Protein Language Models with Complementary Regularizers

  Wijaya, E.

  AI Summary

  • The study distills a large 738M-parameter protein language model into smaller models using uncertainty-aware position weighting and calibration-aware label smoothing, which together improve performance despite individual degradation.
  • The distilled models offer up to 5x faster inference, use less memory, and maintain natural amino acid distributions, making them suitable for consumer-grade hardware.
  • When fine-tuned on small protein family datasets, these models outperform the original in generating family-matching sequences, showing higher efficiency and Pfam hit rates.

  Abstract

  Large autoregressive protein language models generate novel sequences de novo, but their size limits throughput and precludes rapid domain adaptation on scarce proprietary data. We distill a 738M-parameter protein language model into compact students using two protein-specific enhancements, uncertainty-aware position weighting and calibration-aware label smoothing, that individually degrade quality yet combine for substantial improvement. We trace this complementary-regularizer effect to information theory: smoothing denoises teacher distributions while weighting amplifies the cleaned signal at biologically variable positions. Students achieve up to 5x inference speedup, preserve natural amino acid distributions, and require as little as 170 MB of GPU memory, enabling deployment on consumer-grade hardware. When fine-tuned on protein families with as few as 50 sequences, students generate more family-matching sequences than the teacher, achieving higher sample efficiency and Pfam hit rates despite their smaller capacity. These results establish distilled protein language models as superior starting points for domain adaptation on scarce data.

  bioinformatics2026-02-25v3

• OriGene: A Self-Evolving Virtual Disease Biologist Automating Therapeutic Target Discovery

  Zhang, Z.; Qiu, Z.; Wu, Y.; Li, S.; Wang, D.; Liu, Y.; Zhou, Z.; Hu, Y.; Chen, Y.; An, D.; Wang, Y.; Li, Y.; Zhong, Z.; Ou, C.; Wang, Z.; Tang, F.; Chen, J. X.; Ma, R.; Li, J.; Wang, X.; Lu, W.; Xue, H.; Zhang, W.; Wei, Z.; Ma, R.; Shi, Z.; Wang, K.; Liu, Q.; Dong, B.; He, Y.; Liu, T.; Gu, J.; Song, S.; Feng, Q.; Zhang, J.; Zhang, B.; Tian, L.; Bai, L.; Gao, Q.; Sun, S.; Zheng, S.

  AI Summary

  • OriGene is a self-evolving multi-agent system designed to automate therapeutic target discovery by integrating diverse biomedical data.
  • It outperforms human experts and other AI models in accuracy, recall, and robustness, particularly with sparse or noisy data.
  • OriGene identified novel therapeutic targets for liver (GPR160) and colorectal cancer (ARG2), showing significant anti-tumor activity in preclinical models.

  Abstract

  Therapeutic target discovery remains a critical yet intuition-driven bottleneck in drug development, typically relying on disease biologists to laboriously integrate diverse biomedical data into testable hypotheses for experimental validation. Here, we present OriGene, a self-evolving multi-agent system that functions as a virtual disease biologist, systematically identifying original and mechanistically grounded therapeutic targets at scale. OriGene coordinates specialized agents that reason over diverse modalities, including genetic data, protein networks, pharmacological profiles, clinical records, and literature evidence, to generate and prioritize target discovery hypotheses. Through a self-evolving framework, OriGene continuously integrates human and experimental feedback to iteratively refine its core thinking templates, tool composition, and analytical protocols, thereby enhancing both accuracy and adaptability over time. To comprehensively evaluate its performance, we established TRQA, a benchmark comprising over 1,900 expert-level question-answer pairs spanning a wide range of diseases and target classes. OriGene consistently outperforms human experts, leading research agents, and state-of-the-art large language models in accuracy, recall, and robustness, particularly under conditions of data sparsity or noise. Critically, OriGene nominated previously underexplored therapeutic targets for liver (GPR160) and colorectal cancer (ARG2), which demonstrated significant anti-tumor activity in patient-derived organoid and tumor fragment models mirroring human clinical exposures. These findings demonstrate OriGene's potential as a scalable and adaptive platform for AI-driven discovery of mechanistically grounded therapeutic targets, offering a new paradigm to accelerate drug development.

  bioinformatics2026-02-25v2

• KuPID: Kmer-based Upstream Preprocessing of Long Reads forIsoform Discovery

  Borowiak, M.; Yu, Y. W.

  AI Summary

  • KuPID is introduced as a method for preprocessing long RNAseq reads to enhance novel isoform discovery by using kmer sketching to pseudo-align reads to known isoforms.
  • This approach reduces the need for full alignment to only relevant reads, speeding up the process by 2-3x and improving the f1 accuracy of isoform discovery by up to 16.7 points.
  • An optional mode allows KuPID to be used for both isoform discovery and transcript quantification.

  Abstract

  Eukaryotic genes can encode multiple protein isoforms based on alternative splicing of their transcribed regions. Most modern novel isoform discovery methods function by identifying and assembling exon splice junctions from an RNAseq sample. However, splice junctions can only be accurately annotated with time-intensive dynamic programming alignment. This manuscript introduces KuPID, a method for preprocessing long RNAseq reads with the goal of better identifying novel isoform transcripts. KuPID utilizes kmer sketching as a pre-filter to quickly pseudo-align reads to known reference isoforms. Full alignment need only then be applied to reads that are most relevant to isoform discovery. Not only does KuPID speed up the discovery pipeline, it also increases downstream accuracy by filtering out extraneous reads. KuPID preprocessing simultaneously increases the f1 accuracy of isoform discovery pipelines by up to 16.7 points while decreasing the runtime by a factor of 2-3x. An optional mode permits a KuPID sample to be paired with both isoform discovery and transcript quantification. Code availability: https://github.com/mboro2000/KuPID.git

  bioinformatics2026-02-25v2

• Integrative Multi-Scale Sequence-Structure Modeling for Antimicrobial Peptide Prediction and Design

  Li, J.; Shao, Y.; Li, Y.; Yu, Q.

  AI Summary

  • The study introduces MultiAMP, a framework that integrates multi-scale sequence and structure information to predict antimicrobial peptides (AMPs), addressing the limitations of current methods that treat these aspects in isolation.
  • MultiAMP significantly outperforms existing methods by over 10% in MCC, particularly in identifying AMPs with low sequence identity to known peptides.
  • Applied to marine organisms, MultiAMP identified 484 novel high-confidence AMPs and provided insights into AMP mechanisms, aiding in the design of peptides with specific motifs.

  Abstract

  Antimicrobial resistance (AMR) is accelerating worldwide, undermining frontline antibiotics and making the need for novel agents more urgent than ever. Antimicrobial peptides (AMPs) are promising therapeutics against multidrug-resistant pathogens, as they are less prone to inducing resistance. However, current AMP prediction approaches often treat sequence and structure in isolation and at a single scale, leading to mediocre performance. Here, we propose MultiAMP, a framework that integrates multi-level information for predicting AMPs. The model captures evolutionary and contextual information from sequences alongside global and fine-grained information from structures, synergistically combining these features to enhance predictive power. MultiAMP achieves state-of-the-art performance, outperforming existing methods by over 10\% in MCC when identifying distant AMPs sharing less than 40\% sequence identity with known AMPs. To discover novel AMPs, we applied MultiAMP to marine organism data, discovering 484 high-confidence peptides with sequences that are highly divergent from known AMPs. Notably, MultiAMP accurately recognizes various structural types of peptides. In addition, our approach reveals functional patterns of AMPs, providing interpretable insights into their mechanisms. Building on these findings, we employed a gradient-based strategy and achieved the design of AMPs with specific motifs. We believe MultiAMP empowers both the rational discovery and mechanistic understanding of AMPs, facilitating future experimental validation and therapeutic design. The codebase is available at \url{https://github.com/jiayili11/multi-amp}.

  bioinformatics2026-02-25v1

• Quantified duplications of proteins within complexes across eukaryotes

  Francis, O.

  AI Summary

  • The study integrates orthology and protein interaction data to map proteins of verified complexes across 31 diverse eukaryotes, identifying 184 universal orthogroups with components from all species.
  • It developed the PCOC suite to analyze duplications and reductions in these orthogroups, revealing both multi-copy and single-copy proteins in various complexes.
  • Case studies on Naegleria gruberi and Guillardia theta demonstrated taxon-specific expansions, enhancing understanding of eukaryotic protein-complex evolution.

  Abstract

  Protein complexes are central to cell biology and typically verified via a combination of interaction data, complete genome sequencing and comprehensive protein-coding gene predictions for reference eukaryotes. However this data is lacking for non-reference eukaryotes. Protein complexes can be predicted in species for which no interaction data is available by mapping orthology of verified protein complex components from reference eukaryotes to predicted proteomes. Studies that map conservation of protein complex components by orthology are often limited to a small number of protein queries, an under-representation of non-reference, microbial eukaryotes and are scattered across the literature. Here, I integrate orthology and protein interaction data by mapping proteins of experimentally verified complexes to orthogroups of proteins spanning 31 diverse eukaryotes. Proteins within complex-harbouring orthogroups are retained and distributed more evenly across taxa than non-complex orthogroups. I identified 184 universal orthogroups that included orthologs of known protein complex components from all 31 eukaryotes, consistent with a conserved core repertoire, likely present in the last eukaryotic common ancestor (LECA). I generated the protein complex orthology cartographer (PCOC) suite to find significant duplications and reductions of proteins in universal orthogroups across and between eukaryotes. This revealed both multi-copy and notably single-copy proteins, in all queried species, from the exosome, spliceosome, proteasome, small-ribosomal processome, tRNA synthetases, MCM complexes and RNA polymerase III. Case analyses of Naegleria gruberi and Guillardia theta highlight taxon-specific expansions and show how broader protist inclusion improves domain-wide inference of eukaryotic protein-complex evolution.

  bioinformatics2026-02-25v1

• Bioactivity-driven discovery of repurposable antivirals as OSCAR inhibitors that promote cartilage protection via transcriptomic reprogramming

  Ryu, G.; Kim, J.; Kim, S.; Lee, S. Y.; Kim, W.

  AI Summary

  • Researchers used sBEAR to identify adefovir (ADV) and brivudine (BRV) as inhibitors of the OSCAR-collagen interaction in chondrocytes, targeting OA treatment.
  • Molecular docking showed both compounds bind to the OSCAR D2 domain's collagen-recognition pocket.
  • In a mouse model, ADV and BRV reduced OA progression, promoted chondrocyte regeneration, and BRV reversed inflammatory and matrix-degrading gene expression.

  Abstract

  Osteoarthritis (OA) is a progressive degenerative joint disorder characterized by cartilage degradation, chronic pain, and impaired joint function. The avascular nature of cartilage isolates chondrocytes from systemic circulation, presenting significant challenges for therapeutic intervention. Despite extensive efforts, no clinically effective disease modifying osteoarthritis drugs (DMOADs) are currently available. Targeting chondrocyte-specific receptors has therefore emerged as a promising strategy. The osteoclast-associated receptor (OSCAR), expressed on chondrocytes, has been implicated in the regulation of cartilage homeostasis and OA pathogenesis. Here, we applied sBEAR (Structurally similar Bioactive compound Enrichment by Assay Repositioning), a bioactivity-driven virtual screening framework independent of target structural information, to identify small molecule inhibitors of the OSCAR collagen interaction. By mining large scale bioactivity profiles, we identified adefovir (ADV) and brivudine (BRV), as candidate OSCAR inhibitors. Molecular docking analyses indicated that both compounds occupy the collagen-recognition pocket within the OSCAR D2 domain. Intra-articular administration of these compounds in a post-traumatic OA mouse model significantly attenuated OA progression and enhanced chondrocyte regeneration. Both compounds increased Sox9 expression, and transcriptomic analyses revealed that BRV reverses inflammatory and extracellular matrix degrading transcriptional programs. Together, these findings establish OSCAR as a therapeutically actionable target in OA and highlight ADV and BRV as potential DMOAD candidates.

  bioinformatics2026-02-25v1

• scDesignPop generates realistic population-scale single-cell RNA-seq for power analysis, benchmarking, and privacy protection

  Dong, C. Y.; Cen, Y.; Song, D.; Li, J. J.

  AI Summary

  • scDesignPop is introduced as a statistical simulator for generating realistic population-scale scRNA-seq data with genetic effects, addressing cost, method selection, and privacy issues in large cohort studies.
  • It models cell- and individual-level covariates, cell-type-specific eQTLs, and uses real or synthetic genotypes, validated against OneK1K and CLUES cohorts.
  • Compared to splatPop, scDesignPop better preserves eQTL effects and gene-gene dependencies, enabling power analysis, method benchmarking, and privacy protection through synthetic data.

  Abstract

  Single-cell RNA sequencing (scRNA-seq) combined with genotyping in large cohorts has enabled the discovery of genetic associations with molecular traits (e.g., eQTLs) at cell-type resolution. However, generating population-scale data remains cost-prohibitive, selecting appropriate analysis methods lacks consensus, and sharing eQTL results alongside scRNA-seq data raises privacy risks. To address these challenges, we introduce scDesignPop, a flexible statistical simulator for generating realistic population-scale scRNA-seq data with genetic effects. scDesignPop models cell- and individual-level covariates, putative cell-type-specific eQTLs (cts-eQTLs), and either real or synthetic genotypes. We validated scDesignPop using the OneK1K and CLUES cohorts across 4 qualitative and 16 quantitative metrics. Unlike splatPop, the only existing population-scale simulator, scDesignPop better preserves eQTL effects and gene-gene dependencies within cell types, closely recapitulating characteristics of the reference data. Leveraging its generative framework, scDesignPop enables power analysis in cell types under multiple eQTL model specifications to guide experimental design; facilitates benchmarking of single-cell eQTL mapping methods through user-defined ground truths; and mitigates re-identification risk using synthetic data while retaining cts-eQTL effects.

  bioinformatics2026-02-25v1

• STRATA: Spatial Regulon Field Theory Reveals Coupling Architecture of Human Skin and Its Homogenization in Melanoma

  Tjiu, J.-W.

  AI Summary

  • STRATA, a new differential-geometric framework, analyzes spatial transcriptomics by constructing continuous regulon activity fields and quantifying local co-regulation in tissues.
  • Applied to human skin melanoma data, STRATA identifies coupling phase boundaries that align with histological tissue architecture.
  • The analysis shows that melanoma homogenizes regulon coupling, reducing variance by 28% and phase boundary intensity by 18% compared to normal epidermal zones.

  Abstract

  Spatial transcriptomics captures gene expression in tissue context, yet current analyses reduce continuous regulatory landscapes to discrete cell clusters, discarding the geometry of intercellular regulation. Here we introduce STRATA (Spatial Transcription-factor Regulatory Architecture of Tissue Analysis), a differential-geometric framework that constructs continuous regulon activity fields from transcript coordinates, computes their coupling tensor to quantify local co-regulation between transcription factor programs, and derives a Regulon Stability Index from the Jacobian singular value decomposition. Applied to Xenium in situ data from human skin melanoma (382 genes, 13.7 million transcripts), STRATA identifies coupling phase boundaries -- positions where the regulatory logic of tissue changes -- that track histological tissue architecture (Pearson r = 0.32 with the dermal-epidermal junction marker KRT-diff, r = 0.51 with maximum principal stretch; P < 10^-10). Within-tissue comparison reveals that the melanoma microenvironment does not abolish regulon coupling but homogenizes it: coupling variance decreases 28% and phase boundary intensity drops 18% relative to the epidermal zone. STRATA transforms spatial transcriptomics from cell cataloguing to continuous field analysis of regulatory tissue architecture.

  bioinformatics2026-02-25v1

• A Comprehensive Analysis of the Electrolytic Hydrogen Water Mechanism via a Feedforward Loop and its Functional Role in Intestinal Cells In Vitro

  LI, J.

  AI Summary

  • This study investigated the molecular mechanisms of electrolytic hydrogen water (EHW) in Caco-2 cells using next-generation sequencing to analyze mRNA and miRNA expression.
  • EHW was found to modulate oxidative stress response and tight junction formation, with bioinformatics revealing its impact on the HIF1 signaling pathway and the expression of genes like CUL5 and GOLGA7.
  • EHW treatment reduced miR-429 and miR-200c-3p levels, enhancing CUL5 and GOLGA7 expression, and promoted cell differentiation, highlighting EHW's regulatory role via feed-forward loops.

  Abstract

  Electrolytic hydrogen water (EHW) plays a critical role in modulating cellular metabolism; yet, the underlying molecular mechanisms remain unclear. This study utilized next-generation sequencing (NGS) to assess mRNA and miRNA expression in EHW treated Caco 2 cells. Bioinformatics analysis identified differentially expressed genes (DEGs) and pathways influenced by EHW and highlighted its involvement in the oxidative stress response and tight junction formation. Protein-protein interaction (PPI) network analysis of the DEGs identified first neighbor genes, supporting the role of EHW in suppressing oxidative stress related genes while also enhancing the expression of the TCEB2 CUL5 COMMD8 (ECS complex) genes, both of which converged on the HIF1 signaling pathway. We also constructed an mRNA and miRNA competing endogenous RNA (ceRNA) network, which revealed four hub genes, two non-coding RNAs (miR-429 and miR-200c-3p) and two protein-coding RNAs (CUL5 and GOLGA7). These genes co-target the transcription factor KLF4 in Caco 2 cells, forming a TF miRNA gene network (TMGN). EHW treatment significantly decreased the levels of miR 429 and miR 200c 3p and stabilized CUL5 and GOLGA7 transcripts post-transcriptionally as compared to ACW. Concurrently, reduced miRNA expression weakened their pretranscriptional competition with mRNAs for KLF4 binding, further enhancing CUL5 and GOLGA7 expression. Phenotypic assays confirmed that continuous EHW treatment promotes Caco 2 cell differentiation. This study underscores the regulatory role of EHW in intestinal cells via feed-forward loops (FFLs), offering novel insights into the molecular mechanisms and functions of EHW.

  bioinformatics2026-02-25v1

• ARCH3D: A foundation model for global genome architecture

  Galioto, N.; Stansbury, C.; Gorodetsky, A. A.; Rajapakse, I.

  AI Summary

  • ARCH3D is introduced as a foundation model for global genome architecture, utilizing a novel masked locus modeling task to incorporate genome-wide contact profiles.
  • The model's embeddings preserve genomic spatial structure, reconstruct interchromosomal interactions under sparse conditions, and identify multi-way interactions.
  • ARCH3D aims to serve as a structural foundation for developing a virtual genome model to simulate genome behavior and dynamics.

  Abstract

  Biological foundation models are transforming scientific discovery by creating information-rich representations that enable inference in low-data settings. Progress on these models has mainly been achieved by increasing input contextual information, e.g., base pairs or genes. Most work, however, focuses on DNA, RNA, and protein, leaving genome architecture, a fundamental component regulating processes like the cell cycle and cell-fate determination, underexplored. Here, we introduce ARCH3D: a foundation model for global genome architecture. ARCH3D uses a novel masked locus modeling task that increases input contextual information to include genome-wide contact profiles of loci spread across the entirety of the genome. We demonstrate this strategy captures global genome structure by showing ARCH3D embeddings preserve genomic spatial structure, reconstruct interchromosomal interactions under extreme sparsity, and enable identification of multi-way interactions. Ultimately, ARCH3D provides a potential structural foundation for building the virtual genome, an artificial intelligence-based model capable of simulating genome behavior and dynamics.

  bioinformatics2026-02-25v1

• RNA foundation models enable generalizable endometriosis disease classification and stable gene-level interpretation

  McConnell, N.; Kelly, J.; Tadikonda, R.; Bettencourt-Silva, J.; Mulligan, N.; Madgwick, M.; Krishna, R.; Strudwick, J.; Evans, A.; Checkley, S.; Carrieri, A. P.; Smyrnakis, M.; Knowles, C. H.; Gardiner, L.-J.

  AI Summary

  • Researchers investigated whether foundation models (FMs) pretrained on large-scale transcriptomic data could improve endometriosis classification across different patient cohorts.
  • Using a 12-cohort RNA-seq benchmark, FM embeddings significantly outperformed traditional TPM baselines, achieving a weighted F1-score of 0.83 versus 0.68.
  • A new interpretability method, classified-aligned integrated gradients (CA-IG), identified a stable set of predictive genes across cohorts, highlighting novel candidates involved in endometriosis pathways.

  Abstract

  Endometriosis is a chronic inflammatory condition with significant diagnostic delays impacting one in ten reproductive age women worldwide. While machine learning (ML) models trained on transcriptomic data show promise for disease prediction, limited generalizability across independent patient cohorts has hindered clinical translation. Foundations models (FMs) pretrained on large-scale transcriptomic data offer promise to learn transferrable, biologically meaningful representations that could support cross-cohort predictions. We assembled a 12-cohort bulk RNA-seq benchmark (334 samples) and developed a computationally efficient pipeline to test whether FMs improve endometriosis classification, an approach not previously applied to this disease. Using AutoXAI4Omics with cohort-aware validation, we compared embeddings derived from five state-of-the-art RNA FMs against TPM baselines. In cross-cohort prediction, FM embeddings significantly improved performance, achieving a weighted F1-score of 0.83 vs. 0.68 for the baseline. To allow gene-level interpretation of FM embedding models, we introduce classified-aligned integrated gradients (CA-IG), an interpretability approach aligning gene-level attributions to the downstream classifier without end-to-end fine-tuning. CA-IG revealed a conserved set of predictive genes from FM embeddings across cohort-validation regimes, contrasting with unstable baseline explainability, suggesting that FM embeddings prioritized transferable disease-related signal over cohort-specific effects. These genes include novel candidates that converge on biologically plausible pathways for endometriosis.

  bioinformatics2026-02-25v1

• Improved multimodal protein language model-driven universal biomolecules-binding protein design with EiRA

  Zeng, W.; Zou, H.; Li, X.; Dou, Y.; Wang, X.; Peng, S.

  AI Summary

  • The study introduces EiRA, a generative model for designing proteins that bind to various biomolecules, using a two-stage post-training process on a multimodal protein language model.
  • EiRA showed state-of-the-art performance in structural confidence, diversity, novelty, and designability across 8 test sets for 6 biomolecule types, and improved downstream task predictions.
  • Experimental validation confirmed a 100% success rate in expressing variants, and EiRA designed a Glucagon peptide binder with micromolar affinity.

  Abstract

  The interactions between proteins and biomolecules form a complex system that supports life activities. Designing proteins capable of targeted biomolecular binding is therefore critical for protein engineering and gene therapy. Here, we propose a new generative model, EiRA, specifically designed for universal biomolecular-binding protein design, which undergo two-stage post-training, i.e., domain-adaptive masking training and binding site-informed preference optimization, based on a general multimodal protein language model. A systemic evaluation reveals the SOTA performance of EiRA, including structural confidence, diversity, novelty, and designability on 8 test sets across 6 biomolecule types. Meanwhile, EiRA provides a better characterization for biomolecular-binding proteins than generic model, thereby improving the predictive performance of various downstream tasks. We also mitigate severe repetition generation in the original language model by optimizing training strategies and loss. Additionally, we introduced DNA information into EiRA to support DNA-conditioned binder design, further expanding the boundaries of the design paradigm. Experimental validation yielded a 100% success rate (20/20) in expressing highly divergent variants. Remarkably, EiRA achieved the one-shot design of a Glucagon peptide binder with SPR-confirmed micromolar affinity.

  bioinformatics2026-02-24v3

• The phylodynamic threshold of measurably evolving populations

  Weber, A.; Kende, J.; Duitama Gonzalez, C.; Oeversti, S.; Duchene, S.

  AI Summary

  • This study investigates the concepts of measurably evolving populations and the phylodynamic threshold, crucial for molecular clock calibration using sampling times.
  • Through simulations and empirical data analysis, it was found that determining these thresholds depends on model assumptions, sampling strategies, and the sensitivity of priors in Bayesian analyses.
  • The study emphasizes the importance of assessing prior sensitivity over tests of temporal signal to enhance molecular clock inferences and highlights sampling limitations.

  Abstract

  The molecular clock is a fundamental tool for understanding the time and pace of evolution, requiring calibration information alongside molecular data. Sampling times are often used for calibration since some organisms accumulate enough mutations over the course of their sampling period. This practice ties together two key concepts: measurably evolving populations and the phylodynamic threshold. Our current understanding suggests that populations meeting these criteria are suitable for molecular clock calibration via sampling times. However, the definitions and implications of these concepts remain unclear. Using Hepatitis B virus-like simulations and analyses of empirical data, this study shows that determining whether a population is measurably evolving or has reached the phylodynamic threshold does not only depend on the data, but also on model assumptions and sampling strategies. In Bayesian applications, a lack of temporal signal due to a narrow sampling window results in a prior that is overly informative relative to the data, such that a prior that is potentially misleading typically requires a wider sampling window than one that is reasonable. In our analyses we demonstrate that assessing prior sensitivity is more important than the outcome of tests of temporal signal. Our results offer guidelines to improve molecular clock inferences and highlight limitations in molecular sequence sampling procedures.

  bioinformatics2026-02-24v2

• Transcriptomic analysis reveals immune signatures associated with specific cutaneous manifestations of lupus in systemic lupus erythematosus

  Lee, E. Y.; Patterson, S.; Cutts, Z.; Lanata, C. M.; Dall'Era, M.; Yazdany, J.; Criswell, L. A.; Haemel, A.; Katz, P.; Ye, C. J.; Langelier, C.; Sirota, M.

  AI Summary

  • This study used transcriptomics from a large cohort of SLE patients to identify molecular pathways associated with ten distinct cutaneous manifestations of SLE.
  • Specific immune signatures were found, such as upregulation of type I interferon, TNF-, and IL6-JAK-STAT3 pathways in subacute cutaneous lupus, suggesting potential therapeutic targets.
  • Unexpected findings included the absence of interferon signaling in patients with skin and mucosal ulcers, and roles for CD14+ monocytes in photosensitivity and NK cells in alopecia, mucosal ulceration, and livedo reticularis.

  Abstract

  Systemic lupus erythematosus (SLE) presents with diverse and heterogenous cutaneous manifestations. However, the molecular and immunologic pathways driving specific cutaneous manifestations of SLE are poorly understood. Here, we leverage transcriptomics from a large well-phenotyped longitudinal cohort of SLE patients to map molecular pathways linked to ten distinct SLE-related rashes. Through whole blood and immune cell-sorted bulk RNA sequencing, we identified immune signatures specific to cutaneous subtypes of SLE. Subacute cutaneous lupus (SCLE) exhibited broad upregulation of type I interferon, TNF-, and IL6-JAK-STAT3, pathways suggesting potential unique therapeutic responses to JAK and type I interferon inhibition. While interferon signaling is prominent in SCLE, discoid lupus, and acute lupus, it is unexpectedly absent in patients with skin and mucosal ulcers. Pathway and cell-type enrichment analysis revealed unexpected roles for CD14+ monocytes in photosensitivity of SLE and NK cells in alopecia, mucosal ulceration, and livedo reticularis. These findings illuminate the immune heterogeneity of rashes in SLE, highlighting subtype-specific mechanistic targets, and presenting opportunities to identify precision therapies for SLE-associated skin phenotypes.

  bioinformatics2026-02-24v2

• Condensate-Driven Transcriptional Reprogramming Defines Core Vulnerabilities in Esophageal and Gastric Cancers

  Alvarez-Carrion, L.; R. Tejedor, A.; Ardura, J. A.; Alonso, V.; Alonso-Moreno, C.; Collepardo-Guevara, R.; Gutierrez-Rojas, I.; Privat, C.; Moreno, V.; Calvo, E.; Gyorffy, B.; Espinosa, J. R.; Ocana, A.

  AI Summary

  • The study investigates how biomolecular condensates contribute to esophageal and gastric cancers using multi-omics profiling, functional genomics, and simulations.
  • Findings show these cancers share a condensate-driven transcriptional program with upregulation of genes like TOPBP1 and CHERP, essential for tumor cell survival.
  • Simulations confirmed that TOPBP1 and CHERP form condensates through phase separation, suggesting these proteins as potential therapeutic targets.

  Abstract

  Biomolecular condensates organize key nuclear functions by compartmentalizing biomolecules, yet their contribution to gastrointestinal tumorigenesis remains poorly defined. Integrating multi-omics profiling, functional genomics, and molecular dynamics simulations, we reveal that esophageal and gastric cancers share a condensate-enriched transcriptional program driven by intrinsically disordered proteins involved in transcription, RNA processing, and replication stress. Transcriptomic analyses identify a hyperactive transcriptional state with upregulation of condensate-associated genes, including TOPBP1 and CHERP. Dependency mapping demonstrates that these proteins are essential for tumor cell viability, defining a conserved condensate core across different tumor types. Machine-learned predictions and residue-resolution coarse-grained simulations confirm that TOPBP1 and CHERP undergo phase separation through homotypic interactions mediated by intrinsically disordered regions, with saturation concentrations below 2 M, consistent with spontaneous condensate formation observed in vitro. Together, these findings establish condensate organization as a fundamental mesoscale principle in upper gastrointestinal cancers and nominate condensate scaffolds as tractable therapeutic vulnerabilities.

  bioinformatics2026-02-24v1

• Beyond alignment: synergistic integration is required for multimodal cell foundation models

  Richter, T.; Zimmermann, E.; Hall, J.; Theis, F. J.; Raghavan, S.; Winter, P. S.; Amini, A. P.; Crawford, L.

  AI Summary

  • The study introduces the Synergistic Information Score (SIS) to measure the information gain from cross-modal interactions in multimodal cell foundation models, addressing the limitation of alignment-based fusion methods which only detect linear redundancies.
  • Benchmarking on spatial transcriptomics datasets showed that tasks with linear redundancies are well-handled by unimodal models, while complex tasks benefit from synergy-aware integration.
  • The analysis suggests that for standard tasks, fine-tuning a dominant unimodal model is sample-efficient, but multimodal frameworks are advantageous when tasks require information from multiple modalities.

  Abstract

  The vision of a "virtual cell" - a computational model that simulates biological function across modalities and scales - has become a defining goal in computational biology. While powerful unimodal foundation models exist, the lack of large-scale paired data prohibits the joint training of multimodal approaches. This scarcity favors compositional foundation models (CFMs): architectures that fuse frozen unimodal experts via a learned interface. However, it remains unclear when this multimodal fusion adds task-relevant information beyond the strongest unimodal representation and when it merely aggregates redundant signal. Here, we introduce the Synergistic Information Score (SIS), a metric grounded in partial information decomposition (PID), that quantifies the information gain achievable only through cross-modal interactions. Extending theoretical results from self-supervised learning, we show that standard alignment-based fusion objectives on frozen encoders inherently collapse to detecting linear redundancies, limiting their ability to capture nonlinear synergistic states. This distinction is directly relevant for tasks aiming to link tissue morphology and gene expression. Benchmarking ten fusion methods on spatial transcriptomics datasets, we use SIS to demonstrate that tasks dominated by linear redundancies are sufficiently served by unimodal baselines, whereas complex niche definitions benefit from synergy-aware integration objectives that enable cross-modal interactions beyond linear alignment. Finally, we perform a scaling analysis which highlights that fine-tuning a dominant unimodal expert is the most sample-efficient path for standard tasks, suggesting that the benefits of multimodal frameworks only emerge when tasks depend on information distributed across modalities. Together, these results establish that building towards a virtual cell will require a fundamental shift from alignment objectives that emphasize shared structure to synergy-maximizing integration that preserves and exploits complementary cross-modal signal.

  bioinformatics2026-02-24v1

• RSTG: Robust Generation of High Quality Spatial Transcriptomics Data using Beta Divergence Based AutoEncoder

  Halder, A.; Ghosh, A.; Bandyopadhyay, S.

  AI Summary

  • The study addresses the challenge of insufficient data in spatial transcriptomics by proposing RSTG, an autoencoder with a β-ELBO loss, to generate high-quality synthetic data.
  • RSTG uses variational inference to uncover the intrinsic structure of the data, enhancing interpretability and robustness.
  • Validation on datasets from the dorsolateral cortex and brain showed RSTG's superior performance in recovering cellular positions and its robustness to data contamination like noise and outliers.

  Abstract

  One of the key challenges in spatial transcriptomics data analysis is the lack of sufficient data to train the models. To address this shortcoming, multiple generative models have been developed to generate synthetic spatial transcriptomics samples in a controlled environment. However, these often fail at out-of-the-box generation in the presence of noise (such as outliers). To tackle this challenge, we propose RSTG (Robust Spatial Transcriptomic Generator), an autoencoder incorporating a {beta}-ELBO loss, to generate high-quality realistic spatial transcriptomic sequences. Our model uncovers the data' intrinsic structure by approximating its underlying distribution through variational inference, resulting in more interpretable and robust density estimation. We validate the effectiveness of RSTG across multiple tasks, including the recovery of cellular positions in both the 2D spatial and location domains. Our method shows improved performance both qualitatively and quantitatively on multiple datasets from the dorsolateral cortex and the brain using MERFISH and Visium technologies. We further illustrate the robustness of our model to outliers by contaminating a portion of the data with possible anomalies (such as white noises, batch effects, and dropouts). Promising results show that our proposal maintains high quality and stability even when the training data are contaminated, across a variety of experimental settings and in comparison with existing approaches.

  bioinformatics2026-02-24v1

• A partition-based spatial entropy for co-occurrence analysis with broad application.

  Otto, T.; Nemri, A.; Claessens, A.; Radulescu, O.

  AI Summary

  • The study introduces Regional Co-occurrence Entropy (RCE), a new spatial entropy measure to analyze how categorical co-occurrences relate to specific environments.
  • RCE was applied to various fields: it identified interactions between immune cells in Alzheimer's Disease, analyzed building diversity in town neighborhoods, and examined bird species distribution in a natural reserve.
  • Key findings include novel interactions in Alzheimer's, potential drivers of social mixing, and vegetation-driven changes in bird community composition.

  Abstract

  Despite the advent of spatial data science, including spatial biology, there exist few methods that study the distribution of points e.g. cells or individuals, accounting for both their own characteristics and environmental factors. We propose a new spatial entropy measure, termed the Regional Co-occurrence Entropy (RCE), that detects when categorical co-occurrences happen preferentially in specific environments. We demonstrate its use over a broad range of application fields. As examples, we study brain cell dynamics in Alzheimer's Disease, identifying both known and likely novel interactions between immune cells around beta-amyloid plaques. We also investigate the diversity of buildings across a town neighborhoods, to detect potential drivers of social mixing at local scale. Finally, we dissect bird species distribution across a natural reserve, identifying potential vegetation-driven changes in community composition. Altogether, the proposed RCE enables rapid insights into interactions with an environmental component, making it a useful addition to the spatial data science toolbox.

  bioinformatics2026-02-24v1

• A functional annotation based integration of different similarity measures for gene expressions

  Misra, S.; Roy, S.; Ray, S. S.

  AI Summary

  • The study developed an integrated similarity score (ISS) by combining various gene expression similarity measures, weighted by biological information, to enhance gene similarity prediction.
  • A fitness function (FFFAG) was used to optimize the weights in ISS by minimizing the difference between functional similarity and ISS.
  • ISS outperformed individual measures in identifying similar gene pairs and predicted functional categories for 40 unclassified yeast genes with high significance (p-value < 10^(-10)).

  Abstract

  Genes with similar expression profiles often exhibit similar functional properties. An integrated similarity score (ISS) is developed by combining different expression similarity measures through weights, obtained using biological information, for improving gene similarity. The expression similarity measures are converted to the common framework of positive predictive value using functional annotation. A fitness function, called fitness function using functional annotation of genes (FFFAG), is also developed by minimizing the difference between functional similarity value and the ISS. The FFFAG is used to determine the weight combination of different similarity measures in ISS. In addition, an existing similarity measure, called TMJ (integrated similarity measure by multiplying Triangle and Jaccard similarity), is also modified to incorporate biological knowledge involving functional annotation. The results demonstrate that ISS is superior to individual similarity measure to find similar gene pairs. Further, the ISS predicts the functional categories of 40 unclassified yeast genes at p-value cutoff of 10^(-10) from 12 clusters. The associated code is accessible at http://www.isical.ac.in/~shubhra/ISS.html.

  bioinformatics2026-02-24v1

• Graph-based RNA structural representation reveals determinants of subcellular localization

  Hao, Y.; Sun, H.; Ran, Z.; Guo, X.; Liu, M.; Bi, Y.; Polo, J.; Liu, N.; Li, F.

  AI Summary

  • The study introduces GRASP, a graph neural network framework for predicting RNA subcellular localization, using a graph representation that includes nucleotide and substructure nodes.
  • GRASP models both base-level interactions and structural context, incorporating multi-label dependency learning for co-localization patterns.
  • It outperforms existing methods in accuracy, F1 score, and AUC across various RNA types, offering insights into structural determinants of RNA localization.

  Abstract

  RNA subcellular localization is a key determinant of RNA function and regulation, yet existing computational approaches rely primarily on sequence or simplified structural descriptors, limiting their scalability to long transcripts, their ability to model inter-label dependencies, and their applicability across RNA types. Here, we present GRASP, a unified graph neural network framework for predicting RNA subcellular localization using a heterogeneous graph representation that is RNA substructure-aware. GRASP presents each RNA as a multi-scale graph comprising nucleotide nodes and secondary-structure-derived substructure nodes, connected by relational edges, enabling joint modeling of base-level interactions and regional structural context. The model further incorporates multi-label dependency learning to capture co-localization patterns across cellular compartments within a unified framework. Across multiple benchmark datasets and RNA types, GRASP consistently outperforms state-of-the-art sequence-based and structure-informed methods, achieving substantial improvements in accuracy, F1 score, and AUC while maintaining strong scalability to long transcripts. In addition, the graph-based representation provides biologically interpretable insights into structural determinants of RNA localization. The source code and data are available at https://github.com/ABILiLab/GRASP, and the web server is accessible at http://grasp.biotools.bio.

  bioinformatics2026-02-24v1

• CAPHEINE, or everything and the kitchen sink: a workflow for automating selection analyses using HyPhy

  Verdonk, H. E.; Callan, D.; Kosakovsky Pond, S. L.

  AI Summary

  • CAPHEINE is a workflow designed to automate evolutionary analysis from unaligned pathogen sequences, using a reference genome.
  • It facilitates studies on site-level selection dynamics, gene-level positive selection, and lineage-specific selective pressure changes.
  • The workflow is compatible with Mac OS, Windows, and Linux, enhancing accessibility for researchers.

  Abstract

  Here we present CAPHEINE, a computational workflow that starts with a set of unaligned pathogen sequences and a reference genome and performs a comprehensive exploratory evolutionary analysis of the input data. CAPHEINE pairs nicely with studies of site-level selection dynamics, gene-level positive selection, and lineage-specific shifts in selective pressure. Our workflow is portable across Mac OS, Windows, and Linux, allowing researchers to focus on results.

  bioinformatics2026-02-24v1

• Systematic identification of DNA methylation biomarkers for tumor-type-specific detection

  Arbona, J. S.; Garcia Samartino, C.; Angeloni, A. R.; Vaquer, C. C.; Wetten, P. A.; Bocanegra, V.; Militello, R. D.; Sanguinetti, G.; Correa, A.; Pellegrini, P.; Carlen, M.; Minatti, W. R.; Vaschalde, G. A.; Perez, R.; Manzino, R. N.; Rodriguez, J. D.; Valdemoros, P.; Sarrio, L.; Ledesma, A.; Campoy, E. M.

  AI Summary

  • The study developed a browser-based platform integrating genome-wide methylomes with transcriptomes to identify DNA methylation biomarkers, addressing issues like shared epigenetic programs and mixed cellular composition in cancer diagnostics.
  • Validation using MSRE-qPCR in colorectal cancer cohorts confirmed effective biomarkers with AUCs of 0.81-1.00.
  • The approach also successfully distinguished hepatocellular carcinoma from cirrhotic liver and identified subtype-specific markers in lung cancers.

  Abstract

  DNA methylation biomarkers for cancer diagnostics often underperform when tumor and background tissues share epigenetic programs, or when complex specimens with mixed cellular composition dilute tumor-derived signals and increase variability. To address these limitations, we developed a gene-centric, browser-based discovery platform that integrates genome-wide methylomes with matched transcriptomes and reference layers spanning pan-cancer tissues and leukocytes, enabling background-aware filtering beyond binary tumor-normal contrasts. Candidate loci are prioritized using combined thresholds on methylation effect size and intra-group variability to penalize stochastic and heterogeneous variation. In colorectal cancer, methylation-sensitive restriction enzyme quantitative PCR (MSRE-qPCR) validation in independent tissue cohorts confirmed multiple candidate loci with AUCs of 0.81-1.00. Using the same framework, MSRE-qPCR validation distinguished hepatocellular carcinoma from cirrhotic liver, and analysis of public tumor methylomes identified subtype-specific markers in lung adenocarcinoma and squamous-cell carcinoma. This resource bridges genome-scale epigenomic discovery with clinically accessible PCR-based methylation assays.

  bioinformatics2026-02-24v1

• OligoGraph: A novel geometric graph-based approach for siRNA efficacy prediction

  Saligram, S. S.; Kasturi, V. V.; Surkanti, S. R.; Basangari, B. C.; Kondaparthi, V.

  AI Summary

  • The study introduces OligoGraph, a graph-based deep learning model for predicting siRNA efficacy against mRNA, addressing the limitations of traditional models by handling variable siRNA lengths.
  • OligoGraph uses RiNALMo embeddings, GATconv, Transformerconv layers, and self-supervised pretraining, showing superior performance on both seen and unseen datasets.
  • Specialized versions for 19- and 21-nucleotide siRNAs outperformed existing models, with significant improvements in AUC-ROC and PCC on various datasets.

  Abstract

  RNA interference (RNAi) is a biological process in which a small interfering RNA (siRNA) prevents the translation of a messenger RNA (mRNA) into a protein by cleaving the mRNA before translation. We exploit this process to prevent the formation of harmful proteins by using an effective siRNA on the target mRNA. The current rapidly emerging RNAi-based drugs show immense potential for therapeutic applications. Traditionally, designing a potent siRNA for an mRNA requires extensive lab experimentation and trials; therefore, there is a need to develop a model that reliably predicts a siRNA's efficacy against mRNA. This saves both cost and time. But designing such models is challenging, as the data available is either scarce or biased. The current models available exhibit limited generalization and are restricted to a fixed siRNA lengths of either 19 or 21 nucleotides, limiting flexible use. To address these challenges, we introduce OligoGraph, a graph-based deep learning architecture that operates on the siRNA-mRNA duplex. It leverages RiNALMo embeddings, multiple GATconv and Transformerconv layers, and self-supervised pretraining, and outperforms all other existing models in our testing on seen and unseen data. We implemented specialized OligoGraph variants for 19- and 21-nucleotide siRNAs, both of which outperformed the current state-of-the-art models on unseen data. The 19-nucleotide model yielded AUC-ROC and PCC increases of 1.1% and 4.6% on the Mixset; 19.07% and 127.3% on the Takayuki dataset, respectively. Furthermore, the 21-nucleotide model improved predictive performance on the Simone dataset by 2.62% (AUC-ROC) and 6.65% (PCC).

  bioinformatics2026-02-24v1

• RevelioPlots: An Interactive Web Application for Fast AI-Based Protein Models Quality Assessment

  Fernandes, L. L. d. S.; Azevedo, A. H. D. d.; Franca, J. V. S. d.; Lima, J. P. M. S.

  AI Summary

  • RevelioPlots is an interactive web application designed to assess the quality of AI-predicted protein structures by integrating statistical pLDDT score analysis with confidence-colored Ramachandran plots.
  • It supports both individual and batch model uploads, using B-factors as a fallback for pLDDT values.
  • Testing with example proteins showed that RevelioPlots effectively highlights correlations between low pLDDT scores and sterically disallowed regions, aiding non-expert researchers in model quality assessment.

  Abstract

  High-accuracy protein structure prediction by deep learning requires rigorous model quality assessment, a process currently hampered by fragmented, non-interactive tools designed for older experimental data formats. We present RevelioPlots, an open-source, interactive web application (Python/Streamlit) that simplifies and streamlines the assessment of AI-predicted protein structure quality. Its key feature is the combination of statistical pLDDT score analysis (mean, median, box plots) with an interactive, confidence-colored Ramachandran plot. This integration establishes a direct visual link between a model's predicted local reliability (pLDDT) and its stereochemical feasibility (backbone geometry). RevelioPlots handles both individual and batch-uploaded models, intelligently falling back to B-factors as a proxy for pLDDT values. Using example model proteins, we demonstrated the tool's effectiveness, revealing differences in reliability and a clear visual correlation between regions of low pLDDT scores and residues in sterically disallowed regions. By unifying these critical metrics, RevelioPlots empowers non-experienced researchers to quickly and intuitively assess, compare, and interpret structural model quality, enabling a more confident and integrated use of predicted data.

  bioinformatics2026-02-24v1

• SPrOUT: A computational and targeted sequencing approach for mixed plant DNA identification with Angiosperms353

  Hu, N.; Bullock, M. R.; Jackson, C.; Miller, C.; Hunter, E.; Huff, C.; Chen, Y.; Handy, S.; Johnson, M.

  AI Summary

  • This study introduces SPrOUT, a method using the Angiosperms353 target sequencing kit to identify plant species in mixed samples through phylogenetic inference.
  • The approach achieves high accuracy (98.1-99.6%) and precision (92.9-100%) for in-silico mixes, and 90.7% accuracy with 98.0% precision for mock supplement mixtures.
  • The method effectively identifies taxa in mixed plant DNA samples, providing a practical framework for various applications.

  Abstract

  Premise: The identification of plant species from mixed samples is crucial in various fields, including ecological surveys, conservation efforts, and food and dietary supplement safety. Traditional methods face potential challenges due to the high costs of DNA sequencing, inefficiencies in computational workflows, and incomplete sequence databases. Methods and Results: This study introduces a novel approach using the Angiosperms353 target sequencing kit for efficient taxonomic identification of angiosperm DNA in mixed samples. Our method assembles short pair-end reads for each mixed sample. Using gene sets of Angiosperms353 from 871 species, we apply phylogenetic inference to categorize the variance in phylogenetic distance across genes to identify the presence of taxa in mixed plant samples. The pipeline reaches 98.1 to 99.6% accuracy, 92.9 to 100% precision for identifying unknown taxa in in-silico mixes, and 90.7% accuracy and 98.0% precision for mock supplement mixtures. We explored the parameter cutoffs of the pipeline to offer an empirical range for different applications. Conclusions: The Angiosperms353 and HybPiper assembly proved effective in sorting mixed plant DNA samples. Our method offers a framework for scientific and practical applications in plant species identification in both single and mixed samples.

  bioinformatics2026-02-23v1

• Comprehensive top-down mass spectral repository enables pan-dataset analysis and top-down spectral prediction

  Li, K.; Liu, K.; Fulcher, J. M.; Tang, H.; Liu, X.

  AI Summary

  • The study introduces TopRepo, a comprehensive repository of over 18 million top-down mass spectrometry (TD-MS) spectra from 12 species, creating a large-scale spectral library with over 5 million annotated spectra.
  • TopRepo facilitates pan-dataset analyses of proteoform characteristics like N-terminal processing and mass shifts.
  • The repository enhances proteoform identification via spectral library searching and supports training deep learning models for accurate TD-MS spectral prediction.

  Abstract

  Mass spectral libraries have become essential resources for training deep learning (DL) models for spectral prediction and de novo sequencing in bottom-up mass spectrometry (BU-MS). Compared with BU-MS, top-down MS (TD-MS) offers unique advantages for characterizing intact proteoforms by analyzing proteoforms without enzymatic digestion. Despite these advantages, large-scale spectral libraries for TD-MS are currently lacking. Here we present TopRepo, the first comprehensive repository of TD-MS spectra, comprising more than 18 million spectra acquired from 12 species across eight types of mass spectrometers. Using TopRepo, we constructed a large-scale top-down spectral library containing over 5 million spectra with curated proteoform and fragment-ion annotations. We demonstrate that TopRepo enables pan-dataset analyses of N-terminal processing, mass shifts, and other proteoform characteristics identified by TD-MS. Furthermore, we show that the TopRepo spectral library substantially improves proteoform identification through spectral library searching and supports the training of DL models for high-accuracy top-down spectral prediction.

  bioinformatics2026-02-23v1

• Inference of cancer driver mutations from tumor microenvironmentcomposition: a pan-cancer study with cross-platform external validation

  Baker, E. A.; Mehaffy, N. S.

  AI Summary

  • This study investigated whether tumor microenvironment (TME) composition can predict cancer driver mutations across glioblastoma, breast, lung adenocarcinoma, and colorectal cancers using machine learning models trained on RNA-seq data from TCGA.
  • The models were externally validated on independent cohorts, achieving AUC >0.65 for 14 out of 15 driver-cancer pairs, with top performance for ERBB2 in breast cancer (AUC=0.980).
  • TME-predicted ERBB2 status was associated with overall survival in breast cancer, and the study highlighted the complexity of predicting KRAS mutations in lung adenocarcinoma due to co-mutant profiles.

  Abstract

  Cancer driver mutations shape the tumor microenvironment (TME), yet whether TME composition alone can predict genotype has not been systematically evaluated across cancers with external validation. We trained machine learning models to predict driver mutation status from TME cell-type composition signatures derived from bulk transcriptomes. Tissue-specific TME signatures (22-28 programs per cancer) were scored from RNA-seq data in TCGA for glioblastoma (GBM, n=157 total; n=90 EGFR-amplification evaluable), breast cancer (BRCA, n=1,082 total; n=994 evaluable), lung adenocarcinoma (LUAD, n=510 total; n=502 evaluable), and colorectal cancer (CRC, n=592 total; n=524 evaluable), then externally validated on independent cohorts spanning different platforms: CPTAC (GBM, n=65), METABRIC (BRCA, n=1,859), GSE72094 (LUAD, n=442), and GSE39582 (CRC, n=585). Of 15 driver-cancer pairs tested, 14 achieved external AUC >0.65, with top performance for ERBB2 amplification in BRCA (AUC=0.980), BRAF mutation in CRC (0.899), and TP53 mutation in BRCA (0.871). TME-predicted ERBB2 status stratified overall survival in METABRIC (Cox HR=1.73, p=7.95x10^-8). Marginal KRAS performance in LUAD (AUC=0.615) reflected opposing TME profiles in KRAS+STK11 versus KRAS+TP53 co-mutant tumors. These results demonstrate that TME composition encodes sufficient information to infer driver mutations across cancers.

  bioinformatics2026-02-23v1

• GlycoForge generates realistic glycomics data under known ground truth for rigorous method benchmarking

  Hu, S.; Bojar, D.

  AI Summary

  • GlycoForge is introduced as a tool for simulating realistic glycomics data with known ground truths, addressing the challenge of simulating data with controlled effects and biases.
  • It supports the creation of synthetic data with specified motif-level effects, batch effects, and realistic missing data scenarios.
  • The utility of GlycoForge was demonstrated by evaluating batch effect correction algorithms, providing guidelines for their application in real-world glycomics data analysis.

  Abstract

  Quantifying all complex carbohydrates in a sample produces glycomics data, which constitutes compositional data and is stymied by biosynthetic dependencies between glycans, requiring dedicated analytic workflows. Properly assessing such methods frequently requires simulated data with known ground truths and injectable effects. However, simulating glycomics data, especially with control over effects and biases, is still unsolved. Here, we present GlycoForge, a feature-complete solution for simulating comparative glycomics data. GlycoForge supports simulating fully synthetic glycomics data, with specified motif-level effects, drawn from Dirichlet distributions, and templated simulations based on real-world data. We further support the injection of batch effects, both mean and variance shifts, via center-log ratio transformations to maintain compositional closure, and realistic missing data simulation. We showcase the utility of GlycoForge by evaluating batch effect correction algorithms for glycomics data, with automated guidelines for when to use such methods on real-world data. GlycoForge is available as an open-access Python package at https://github.com/BojarLab/GlycoForge.

  bioinformatics2026-02-23v1

• Skip-Zeros Variational Inference in the Million-Cell Era of Single-Cell Transcriptomics

  Shimamura, T.; Yuki, S.; Abe, K.

  AI Summary

  • The study introduces UNISON, a scalable framework for matrix factorization in single-cell RNA sequencing, using skip-zeros variational inference to handle the sparsity of large datasets.
  • UNISON performs inference using only nonzero elements, improving efficiency and scalability, and was tested on over one million cells from the Mouse Organogenesis Cell Atlas.
  • Application to cross-species analysis showed UNISON's ability to distinguish conserved and species-specific transcriptional programs, enhancing understanding of biological processes like glaucoma.

  Abstract

  Combinatorial indexing-based single-cell RNA sequencing methods such as sci-RNA-seq and sci-RNA-seq3 now enable the profiling of millions of cells, producing expression matrices that are both extremely sparse and high-dimensional. Conventional nonnegative matrix factorization (NMF) provides an interpretable framework for uncovering latent biological structures but is computationally prohibitive at this scale, as it requires explicit access to the vast number of zero entries. We introduce UNISON (Unified Sparse-Optimized Nonnegative factorization), a scalable framework for matrix and tensor factorization based on skip-zeros variational inference. By reformulating stochastic variational Bayes updates in terms of sufficient statistics, UNISON performs inference using only nonzero elements, while implicitly accounting for zeros through geometric sampling. This strategy enables efficient parameter estimation without matrix expansion and naturally accommodates multiple experimental contexts. Simulation studies show that UNISON is robust to diverse learning-rate schedules and mini-batch sizes, providing practical guidelines for optimization. Application to the Mouse Organogenesis Cell Atlas demonstrates scalability to over one million cells, yielding latent factors that capture developmental trajectories and lineage-specific signatures with improved interpretability compared to existing methods. Cross-species analysis of aqueous humor outflow pathways across five vertebrate species further highlights UNISON's ability to disentangle conserved from species-specific transcriptional programs and to recover biologically meaningful gene-gene and gene-phenotype relationships relevant to glaucoma. By efficiently exploiting sparsity while preserving interpretability, UNISON establishes a principled and practical solution for integrative, large-scale single-cell transcriptomics.

  bioinformatics2026-02-23v1

• What makes a banana false? How the genome of Ethiopian orphan staple Ensete ventricosum differs from the banana A and B sub-genomes

  Muzemil, S.; Paul, P.; Baxter, L.; Dominguez-Ferreras, A.; Sahu, S. K.; Van Deynze, A.; Mai, G.; Yemataw, Z.; Tesfaye, K.; Ntoukakis, V.; Studholme, D. J.; Grant, M.

  AI Summary

  • The study sequenced the genome of the Ensete ventricosum landrace Mazia, identifying 38,940 protein-coding genes, with the assembly being more complete than the previously published Bedadeti genome.
  • Comparative analysis showed about 25% of the Mazia genome is unique to enset, with distinct functional signatures related to DNA integration, carbohydrate metabolism, disease resistance, and transcriptional regulation.
  • The research highlights the potential for marker-assisted breeding in enset, providing a foundation for improving agronomically important traits through comparative genomics within the Musaceae family.

  Abstract

  Background: Ensete ventricosum, also known as the "tree against hunger" plays a key role in Ethiopian food security and farming systems, feeding more than 20 million people. Since domestication via clonal selection in the south-west Ethiopian highlands, today's diverse enset landraces contribute multiple benefits including food, fibre by-product, animal bedding and cattle fodder to farmers and local communities. Improved genomic resources for this highly drought-tolerant plant are essential to supplement the conventional clonal selection-based breeding programme and pave the way towards targeted breeding. Results: We sequenced the genome of enset landrace Mazia, which is partially resistant/tolerant to Xanthomonas wilt and predicted 38,940 protein-coding genes. The Mazia assembly (540.14 Mb) is more complete than the previously published genome assembly of landrace Bedadeti (451.28 Mb) and displayed 1.41% heterozygosity and 64.64% repetitive DNA content. Comparative analyses with the Bedadeti assembly and chromosome-level genome sequences of the two main banana progenitors (Musa acuminata, AA genome; Musa balbisiana, BB genome) unexpectedly revealed ~25% of the Mazia genome is unique to enset. Gene Ontology (GO) and sequence similarity search analysis of enset-specific protein-coding genes identified distinct functional signatures that underpin the lifestyle, adaptation, and corm productive quality of enset, including functions related to DNA integration, carbohydrate metabolism, disease resistance and transcriptional regulation. In contrast, Musa-specific genes showed enrichment for defence response, protein phosphorylation and fruit development pathways. Focusing on the classical nucleotide binding site leucine rich repeat (NLR) disease resistance genes, we identified and characterised NLRs in enset and Musa species genomes, revealing a considerable expansion in the Musa acuminata genome. We also identified unique genes in enset and banana genomes whose functional and evolutionary roles are yet to be determined. Conclusions: Here, we report a de novo genome assembly for the enset (Ensete ventricosum) landrace Mazia and provide a high-quality annotation of both Mazia and the previously published assembly of the landrace Bedadeti. Collectively, these genomic resources provide a valuable foundation for comparative genomics within the Musaceae family and open new opportunities for the development of marker-assisted breeding strategies to accelerate the improvement of agronomically important traits in enset. Keywords: Ensete ventricosum, Musa, gene families, nucleotide binding site leucine rich repeat (NLR), orphan crop.

  bioinformatics2026-02-23v1

• An Integrated and Configurable End-to-End Pipeline for Longitudinal Cell Painting Analysis

  Zhao, G.

  AI Summary

  • The study introduces SCALE, an end-to-end pipeline for analyzing longitudinal cell painting data, addressing challenges like imaging variability and time consistency.
  • SCALE integrates nucleus-centered segmentation, quality control, feature extraction, and signal aggregation in a modular framework.
  • The pipeline's effectiveness was demonstrated with a chronic radiation exposure dataset, showing its capability for consistent longitudinal analysis.

  Abstract

  Cell painting assays generate high-dimensional, multi-channel imaging data that enable systematic characterization of cellular phenotypes. Increasingly, such assays are performed in longitudinal settings and under chronic perturbations, introducing additional challenges related to imaging variability, focus-field heterogeneity, and consistency across time points. Existing analysis workflows often require substantial manual adaptation to handle these complexities, limiting scalability and reproducibility. In this paper, we propose SCALE (Stable Cell painting Analysis for Longitudinal Experiments), an integrated, end-to-end analysis pipeline designed for robust longitudinal analysis of cell painting data. The pipeline combines nucleus-centered segmentation, automated quality control, feature extraction, and signal aggregation within a modular and configurable framework. Once assay-specific configurations are specified, the pipeline executes in a fully automated manner from raw images to downstream summary statistics and analysis-ready outputs. We demonstrate the utility of the pipeline using a chronic radiation exposure cell painting dataset, illustrating its ability to support consistent longitudinal comparisons across conditions and time points.

  bioinformatics2026-02-23v1

• LinkDTI: Drug-Target Interactionsprediction through a Link Predictionframework on Biomedical KnowledgeGraph

  Mondal, M.; Arunachalam, S.; Wu, S.; Datta, A.

  AI Summary

  • LinkDTI is a computational framework that predicts drug-target interactions (DTIs) by analyzing connections in a heterogeneous biomedical knowledge graph using a modified GraphSAGE model.
  • It employs negative sampling to balance data and outperforms baseline methods by at least 2.5% in AUROC and AUPRC.
  • The framework identified 945 new potential DTIs, a 49.14% increase over known interactions.

  Abstract

  Computational drug-target interactions (DTI) prediction serves as a valuable tool for drug discovery and repurposing by cost-effectively narrowing down the potential drug-target space. This paper presents LinkDTI, a computational framework that predicts DTIs by identifying connections within a heterogeneous knowledge graph of drugs, proteins, diseases, and side effects. Unlike methods that rely on mathematical techniques like matrix completion or similarity-based scoring, LinkDTI uses an advanced graph-based approach to capture relationships between biomedical entities. Specifically, LinkDTI applies a modified version of the multilayer GraphSAGE model that learns from the heterogeneous knowledge graph and predicts potential drug-target interactions. Our model incorporates negative sampling that balances the data to address the issue of having more negative than positive interactions. Our results show that LinkDTI consistently performs better in AUROC and AUPRC than baseline methods by at least 2.5% across different sampling ratios and conditions. Subsequently, it identifies approximately 945 new potential DTIs, marking a 49.14% increase over known DTIs. Overall, LinkDTI offers a simple yet effective method for integrating diverse biomedical data to identify potential drug-target interactions.

  bioinformatics2026-02-23v1

• Structure-Based TCR-pMHC Binding Prediction and Generalization to Unseen Peptides

  Abeer, A. N. M. N.; Roy, R. S.; Qian, X.; Yoon, B.-J.

  AI Summary

  • This study investigates the generalization performance of graph neural network (GNN)-based classifiers for predicting TCR-pMHC binding, focusing on their accuracy with unseen peptides.
  • The research assesses factors like interaction features and structural uncertainty that affect classifier performance.
  • By designing classifier architecture with auxiliary training objectives, the study shows improved generalization to novel peptides.

  Abstract

  The interaction between T-cell receptors (TCRs) with the peptide-bound major histocompatibility complex (MHC) intricately impacts the functional specificity of T-cell-mediated adaptive immune response. Consequently, implication in immunotherapy has contributed to the ever-growing computational methods for TCR recognition, which have recently attracted structure-based approaches due to advancements in protein structure modeling. Despite access to structural information of the predicted binding interface, graph neural network (GNN)-based TCR-pMHC binding specificity classifiers tend to show poor accuracy for samples with unseen peptides. In this work, we comprehensively assess the potential factors that critically impact the generalization performance of classifiers trained with computationally predicted structures. Specifically, our experiments focus on analyzing the sensitivity of such predictors to the interaction features in the TCR-pMHC interface and the structural uncertainty. Building on the analysis, we demonstrate how the design of classifier architecture with auxiliary training objectives can improve the generalization performance to novel peptides not yet seen during model training. Overall, our work highlights the challenges of unseen peptide generalization from different perspectives of the GNN-based classifier paradigm, showcasing the strengths and weaknesses of the current state-of-the-art approaches in the generalization landscape.

  bioinformatics2026-02-23v1

• A Spatio-Temporal Analysis Framework for Characterizing Radiation-Induced Genomic Instability

  Chopra, K.; Cucinell, C.; Weinberg, R.; Forrester, S.; Brettin, T.; Kilic, O.; Yoon, B.-J.

  AI Summary

  • This study developed an analytical framework to investigate the coupling between structural variants and point mutations in human endothelial cells exposed to chronic low-dose gamma radiation.
  • The framework revealed a 7.13-fold enrichment of doublet base substitutions (DBS) near inversion breakpoints, with this enrichment diminishing with distance.
  • Temporal analysis showed inversions were transient, while DBS persisted, affecting genes critical for genomic stability like DNA damage response and chromatin regulation.

  Abstract

  Chronic low-dose ionizing radiation induces complex genomic instability encompassing both structural variants and point mutations, yet these alterations are typically analyzed as independent events limiting detection of mechanistic coupling between rearrangement formation and localized mutagenesis at breakpoint junctions. This gap is particularly consequential given the widespread occupational and environmental exposure contexts; nuclear energy, medical imaging, and environmental contamination, where coupled genomic alterations may contribute to cancer risk through mechanisms invisible to type-agnostic analyses. We developed an integrated analytical framework combining temporal pattern tracking, breakpoint-proximal mutation enrichment analysis, and systematic testing across all structural variant types to resolve these coupled dynamics across dose and time. Applying this framework to whole-genome sequencing data from primary human endothelial cells (HUVEC) exposed to chronic low-dose gamma radiation (0.001 - 2 mGy/hr) over three weeks, we discovered 7.13-fold enrichment of doublet base substitutions (DBS) within 10bp of inversion breakpoints, a signal absent from other structural variant types. This enrichment decayed sharply with distance (to [~]1.9 fold at 100bp), indicating localized mutagenesis at these junctions. Temporal analysis revealed divergent fates: inversions appeared transiently (100% single-timepoint) while DBS showed greater persistence (9.0% multi-timepoint). Among the INV-DBS events identified, affected genes include 16 high-constraint loci (pLI [&ge;] 0.9) involved in DNA damage response, signal transduction, and chromatin regulation; pathways critical for maintaining genomic stability. Our framework provides a generalizable approach for investigating structural variant-mutation relationships, with applications to radiation biology, cancer genomics, and mechanistic studies of DNA repair fidelity.

  bioinformatics2026-02-23v1

• art_modern: An Accelerated ART Simulator of Diverse Next-Generation Sequencing Reads

  YU, Z.

  AI Summary

  • The study introduces art_modern, an accelerated version of the ART simulator for next-generation sequencing (NGS) data, enhanced with updated algorithms, SIMD instructions, and parallel processing.
  • art_modern supports simulation of transcriptome profiling with contig-specific coverage and strand information.
  • Benchmarking showed art_modern reduces CPU time by 75-77% and accelerates wall-clock time by 15-24 times compared to the original ART on multi-core systems._

  Abstract

  Fast simulation of next-generation sequencing (NGS) data is vital for software development and benchmarking. Here we describe art_modern, an accelerated ART simulator that can simulate various NGS data. We accelerated ART using updated sampling algorithms, single-instruction multiple-data (SIMD) instruction-set extensions (ISEs), thread- and node-level parallelism, and an asynchronous output writer, while enabling simulation of transcriptome profiling data by supporting contig-specific coverage with strand information. The new implementation was benchmarked against popular performance-oriented NGS simulators, revealing a 75--77% reduction in CPU time and a 15--24 times acceleration in wall-clock time on a multi-core machine compared to the original implementation. With this simulator, the process of developing and benchmarking NGS sequence analysis algorithms can be largely accelerated. Availability and Implementation: The software is implemented in C++17 with CMake as the building system. It can be built and executed on a modern GNU/Linux operating system with Boost, Zlib, and a C++17 compiler, with further acceleration available using Intel OneAPI C++/DPC++ compilers and Intel oneAPI MKL random generators. The software is available at https://github.com/YU-Zhejian/art_modern under the GNU General Public License v3.

  bioinformatics2026-02-23v1

• Universal physical principles govern the deterministic genesis of protein structure

  Chuanyang, L.; Liu, J.; Qiu, X.; Wu, X.; Li, W.; Min, L.; Zhang, G.; Zhang, S.; Zhu, L.

  AI Summary

  • The study introduces ProtGenesis, a framework that models protein genesis as a deterministic process within a discrete structural space, governed by three universal principles: Assembly, Emergence, and Phase-Transition.
  • These principles describe how amino acids form fractal-like structures, how peptides follow spatial trajectories, and how mutations lead to topological phase shifts in protein structure.
  • ProtGenesis provides a mathematical foundation to interpret deep learning models and offers a basis for engineering protein structures.

  Abstract

  The origin of functional proteins remains a fundamental biological enigma. Although Anfinsen's dogma established sequence as the determinant of structure, and deep learning models can predict structures with high fidelity, the physical principles governing protein genesis itself, from prebiotic condensation to functional protein emergence, remain unresolved. This gap leaves a critical disconnect between mechanistic biological insights and artificial intelligence. Herein, we introduce a unified methodological framework ProtGenesis that recasts genesis of protein as a structured, deterministic navigation within a discrete structural space. We identify three universal principles governing this hierarchical organization: the Assembly Principle directs amino acids condensation into multilayer fractal-like architectures; the Emergence Principle ensures nascent peptides' emergence follow deterministic spatial trajectories; and the Phase-Transition Principle describes wherein incremental residue accrual or mutations drives precise topological phase shifts from short-range to long-range order. By quantifying these trajectories with novel tripartite spatial metrics, we reveal that protein genesis is not an abstract continuum but a principle-governed physical process with measurable coordinates. ProtGenesis thus provides an universal interpretable mathematical foundation for decoding "black-box" of deep learning models and establishes a rigorous basis for exploring, understanding, and engineering the molecular blueprint of life.

  bioinformatics2026-02-23v1

• MetaTracer: A nucleotide alignment-based framework for high-resolution taxonomic and transcript assignment in metatranscriptomic data

  Furstenau, T.; Shaffer, I.; Hsu, K.-L. C.; Pearson, T.; Ernst, R. K.; Fofanov, V.

  AI Summary

  • MetaTracer is a tool that performs nucleotide alignment for metatranscriptomic analysis, assigning reads to both taxonomic groups and genes in one pass.
  • It offers improved accuracy and species-level resolution compared to k-mer or protein-based methods by mapping reads to annotated genomic features.
  • Testing on simulated and real dental plaque data showed high accuracy in taxonomic and gene assignment, revealing species-specific transcriptional differences in children with early childhood caries versus healthy controls.

  Abstract

  Summary: MetaTracer is a nucleotide alignment-based tool for metatranscriptomic analysis of complex bacterial communities that assigns sequence reads to both taxonomic groups and expressed genes in a single pass. Full nucleotide-level alignment improves accuracy relative to k-mer-based classifiers and preserves species-level resolution that is often lost in protein-based approaches. By retaining alignment coordinates and mapping reads directly to annotated genomic features, MetaTracer enables direct attribution of gene expression to specific microbial species. On simulated datasets, MetaTracer achieves high accuracy for both taxonomic and gene assignment. Applied to real dental plaque metatranscriptomic datasets, MetaTracer resolves species-specific transcriptional activity and detects reproducible differences in microbial gene expression between children with early childhood caries and healthy controls. Availability and implementation: MetaTracer is implemented as a Python-based workflow wrapper (metatracer v0.1.0) that depends on the mtsv-tools core engine (v2.1.0), which is written in Rust. The required functionality is supported by the v2.1.0 release of mtsv-tools. Both packages are open source under the MIT license and are available at github.com/FofanovLab/metatracer and github.com/FofanovLab/mtsv-tools. Versioned releases are archived at Zenodo (DOI: 10.5281/zenodo.18665766 and DOI: 10.5281/zenodo.18718002). Installation is supported via Bioconda.

  bioinformatics2026-02-23v1

• CellAwareGNN: Single-Cell Enhanced Knowledge Graph Foundation Model for Drug Indication Prediction

  Zhang, X.; Jeong, E.; Yan, C.; Feng, Y.; Lyu, L.; Guo, X.; Chen, Y.

  AI Summary

  • The study introduces CellAwareGNN, a model integrating single-cell genomics into a biomedical knowledge graph (scPrimeKG) to enhance drug indication prediction.
  • CellAwareGNN was evaluated on all diseases in the graph, achieving an AUPRC of 0.826, surpassing TxGNN-U (0.816) and TxGNN (0.799).
  • For autoimmune diseases, CellAwareGNN showed significant improvement, with an AUPRC of 0.864, and suggested specific drug repurposing candidates like Ocrelizumab for Pemphigus.

  Abstract

  Graph foundation models have emerged as powerful tools for drug repurposing by enabling the prediction of novel drug-disease indications from large biomedical knowledge graphs. A representative example is TxGNN, which was previously developed and trained on PrimeKG, a comprehensive biomedical knowledge graph covering over 17,000 diseases. While TxGNN demonstrates strong performance, existing biomedical knowledge graphs largely lack fine-grained, cell-type-specific genomic context. It limits their ability to capture disease mechanisms driven by dysregulated cellular programs, such as immune cell-specific pathways in autoimmune diseases. Moreover, prior evaluations typically test only randomly selected subsets of diseases, leaving many diseases unexamined and limiting conclusions about model performance across the full disease spectrum. To address these limitations, we first update PrimeKG to PrimeKG-U by incorporating expanded and curated biomedical knowledge and then develop TxGNN-U as a stronger graph-based baseline. Building on this foundation, we introduce CellAwareGNN, a graph foundation model that integrates single-cell genomics into PrimeKG-U. We construct a single-cell-enhanced knowledge graph, scPrimeKG, by incorporating cell-type-specific genetic associations from the OneK1K dataset, expanding PrimeKG from approximately 8.1 million edges and 129k nodes to over 14 million edges and 140k nodes. CellAwareGNN is pre-trained on all relation types in scPrimeKG and evaluated on drug indication prediction with explicit coverage of all diseases in the knowledge graph. CellAwareGNN consistently outperforms TxGNN and TxGNN-U. For drug indication prediction, CellAwareGNN achieves an AUPRC of 0.826, representing a 1.2% improvement over TxGNN-U (0.816) and a 3.4% improvement over TxGNN (0.799). Notably, for autoimmune diseases, CellAwareGNN attains an AUPRC of 0.864, improving by 2.0% over TxGNN-U (0.847) and 6.0% over TxGNN (0.815). Importantly, CellAwareGNN prioritizes promising repurposing candidates, including Ocrelizumab for Pemphigus via CD20-expressing B cells, Methotrexate for Pemphigus through DHFR and ATIC activity in T and B cells, and Rosiglitazone for Rheumatoid Arthritis through PPAR-{gamma} activation. These results demonstrate the value of incorporating cell-type-specific genomic context to improve both predictive performance and biological interpretability in graph- based drug repurposing.

  bioinformatics2026-02-23v1

• Hierarchical Multi-Omics Trajectory Prediction forFecal Microbiota Transplantation: A Novel MachineLearning Framework for Small-Sample LongitudinalMulti-Omics Integration

  Zhou, Y.-H.; Sun, G.

  AI Summary

  • The study introduces Hierarchical Multi-Omics Trajectory Prediction (HMOTP), a machine learning framework for predicting patient trajectories post-fecal microbiota transplantation (FMT) using small-sample, longitudinal multi-omics data.
  • HMOTP integrates lipidomics and metagenomics data through hierarchical feature construction and multi-level attention, achieving 96.67% accuracy in predicting treatment outcomes in a cohort of 15 patients.
  • The framework identified key biomarkers and cross-omics associations, demonstrating its utility in personalized medicine and biological discovery in FMT applications.

  Abstract

  Fecal microbiota transplantation (FMT) has emerged as a highly effective treatment for recurrent Clostridioides difficile infection and is being actively investigated for numerous other conditions. While multi-omics studies have revealed dynamic changes in microbial communities and host metabolism following FMT, existing approaches are primarily descriptive and lack the ability to predict individual patient trajectories or identify early biomarkers of treatment response. Small-sample, multi-omics, longitudinal prediction problems present unique computational challenges: high dimensionality, multi-omics integration, temporal dynamics, and interpretability. Here, we present Hierarchical Multi-Omics Trajectory Prediction (HMOTP), a novel machine learning framework specifically designed for small-sample, multi-omics, longitudinal prediction that addresses these challenges through hierarchical feature construction using domain knowledge, multi-level attention mechanisms, and patient-specific trajectory prediction. HMOTP integrates multi-omics data at multiple biological levels (raw features, aggregated classes/categories, and cross-level interactions) while preserving biological interpretability. The framework employs multi-head attention to learn feature importance at different hierarchy levels and integrates information across omics layers. Patient-specific trajectory prediction enables personalized predictions despite limited sample sizes through transfer learning. We evaluated HMOTP on a cohort of 15 patients with recurrent Clostridioides difficile infection who underwent fecal microbiota transplantation, with comprehensive lipidomics (397 features) and metagenomics (10,634 pathways) profiling at four timepoints spanning six months. Using leave-one-patient-out cross-validation, HMOTP achieved 96.67% {+/-} 10.54% accuracy, outperforming baseline methods including Random Forest (91.33% {+/-} 21.33%) and Logistic Regression (86.33% {+/-} 24.67%). The framework demonstrated robust generalization across timepoints. Through hierarchical interpretability, HMOTP identified key biomarkers and revealed mechanistically informative cross-omics associations, including 324 strong correlations (|r| > 0.7) involving top-predictive biomarkers, demonstrating its utility for both prediction and biological discovery in FMT applications. HMOTP provides a generalizable framework applicable to other small-sample multi-omics problems, offering a powerful tool for personalized medicine applications.

  bioinformatics2026-02-23v1

• A PLUM Job: Peptide modeLs for Understanding and engineering antiMicrobial therapeutics

  Banerjee, P.; Friedberg, I.; Rued, B. E.; Eulenstein, O.

  AI Summary

  • The study addresses the need for new antimicrobial strategies by developing PLUM, a structured conditional Variational Autoencoder for designing antimicrobial peptides (AMPs).
  • PLUM allows for controlled generation of AMPs by disentangling sequence, function, and length, producing peptides 5-35 amino acids long.
  • Compared to HydrAMP, PLUM achieved a 7% higher AMP yield, 14% increased diversity, and 37% more AMPs in prototype-conditioned generation, with low predicted toxicity.

  Abstract

  Motivation: Antibiotic-resistant infections in humans and animals are rising, creating an urgent need for new antimicrobial strategies. This challenge extends beyond clinical settings to food production systems; the Centers for Disease Control and Prevention estimates that foodborne pathogens cause over 48 million illnesses annually in the U.S. alone. Antimicrobial peptides (AMPs) are a promising alternative due to their broad activity and lower risk of resistance. However, rational design remains challenging, particularly when simultaneously controlling sequence, function, and peptide length. Results: We introduce Peptide modeLs for Understanding and engineering antiMicrobial therapeutics (PLUM), a structured conditional Variational Autoencoder for controlled AMP generation. PLUM disentangles sequence, function, and length within its latent space, enabling both *de novo* and prototype-conditioned generation of peptides 5-35 amino acids long, thereby capturing larger functional domains. Across 45,000 generated peptides, PLUM: * Achieved the highest AMP yield (0.885), **7% higher** than HydrAMP * Increased AMP diversity by 14%compared with HydrAMP * Maintained the highest non-AMP sequence yield (0.895), 19% higher than HydrAMP. For prototype-conditioned generation, PLUM produced 37% more AMPs than HydrAMP, generating sequences that closely matched real peptide compositions while exhibiting low predicted toxicity. Integrated AMP classifiers enabled robust evaluation of identity and potency across diverse bacterial targets. These results establish PLUM as a scalable and versatile platform for designing antimicrobial peptides and next-generation therapeutics. Availability: https://github.com/priyamayur/PLUM

  bioinformatics2026-02-23v1

• EES-Transformer: A Dual-Path Transformer for Tissue Classification and Gene Representation Learning from Extreme Expression Sets

  Park, J.-S.; Lee, Y.; Kang, Y. J.

  AI Summary

  • The EES-Transformer V2 uses a dual-path architecture to classify tissues and learn gene representations from extreme gene expression sets in Arabidopsis thaliana RNA-seq data.
  • It achieves 91-92% accuracy in classifying 47 tissue types, significantly outperforming the 2.1% random baseline.
  • Analysis revealed tissue-specific gene relationships, with attention highlighting 3,524 gene-tissue associations and showing that tissue identity is encoded in gene pairs rather than single genes.

  Abstract

  Accurate tissue classification from gene expression data is fundamental to transcriptomic analysis. Here we introduce EES-Transformer V2, a dual-path transformer architecture that learns from Extreme Expression Sets (EES); sequences of genes at expression extremes (above 95th or below 5th percentile). The architecture separates tissue classification from masked language modeling through independent branches: the classification branch operates without tissue label information, while the generative branch receives tissue conditioning. This design enables fair evaluation of classification performance while learning tissue-specific gene relationships. Applied to 12,212 Arabidopsis thaliana RNA-seq samples spanning 47 tissue types, EES-Transformer achieves 91-92% classification accuracy (varying across evaluation runs due to stochastic input masking); substantially above the 2.1% random baseline. Attention-based analysis identifies 3,524 gene-tissue high-attention associations whose importance patterns reflect known biology. Critically, while individual high-attention genes appear broadly across tissues, gene pairs from attention-derived regulatory networks show higher tissue specificity: pollen gene pairs show a 2.7-fold enrichment over single-gene rates, and root and leaf pairs each show 1.5-fold enrichment. This finding reveals that tissue identity is encoded in combinatorial gene expression patterns rather than individual genes. Attention-derived gene regulatory networks exhibit scale-free topology and biologically coherent hub gene programs, with pollen networks consisting entirely of DOWN-DOWN interactions among silenced vegetative genes. EES-Transformer provides accurate tissue classification, interpretable gene importance scores, and attention-derived regulatory networks for biological discovery.

  bioinformatics2026-02-23v1

• Pro-GAT: Reconnecting Fragmented PROTACs Using Graph Attention Transformer

  Vemuri, S.; Bijigiri, L. P.; Gogte, S.; Kondaparthi, V.

  AI Summary

  • The study addresses the issue of chemically invalid or disconnected linker structures in PROTACs generated by diffusion-based models, which fail to maintain local bonding requirements.
  • Pro-GAT, a graph attention-based framework, was developed to repair these disconnected PROTAC candidates by predicting coordinate corrections and atom-type modifications.
  • When integrated with DiffPROTACs and DiffLinker, Pro-GAT increased the percentage of chemically valid PROTAC candidates from 76.70% to 83.92% and 63.16% to 68.73%, respectively, while maintaining high uniqueness levels.

  Abstract

  PROTACs work by bringing together a protein-of-interest ligand and an E3 ligase recruiter to trigger targeted degradation. However, Diffusion-based generative models frequently produce chemically invalid or disconnected linker structures that satisfy global geometric constraints but violate local bonding requirements. These models operate in continuous coordinate space and therefore lack explicit mechanisms for enforcing discrete chemical connectivity under fixed-anchor constraints. Invalid, disconnected outputs recur rather than being a rare exception, such that naive resampling is not an effective method to obtain valid chimeras. Pro-GAT is a graph attention-based framework for geometry-preserving molecular graph repair, capable of functioning on chemically disconnected diffusion-generated PROTAC candidates by predicting bounded coordinate corrections and constrained atom-type modifications using geometry-aware graph attention network (GAT) layers. The proposed model is trained on PROTAC datasets with added disconnections to overcome systematic connectivity failures in diffusion-based PROTAC generation with fixed anchors. When combined with DiffPROTACs and DiffLinker, Pro-GAT improves the percentage of chemically valid candidates in the aggregated output from 76.70% to 83.92% and 63.16% to 68.73% while maintaining 80.18% and 63.80% uniqueness levels of valid candidates respectively, thus facilitating the generation of usable PROTAC candidates from invalid diffusion samples. Pro-GAT was used in a case study of the 7Z76 ternary complex to repair DiffPROTACs and DiffLinker generated samples, which gave rise to connected chimeras whose docking scores were comparable to the original 7Z76 structure.

  bioinformatics2026-02-23v1

• BacTaxID: A universal framework for standardized bacterial classification

  Fernandez-de-Bobadilla, M. D.; Lanza, V. F.

  AI Summary

  • BacTaxID is a universal framework for bacterial classification using whole-genome k-mer-based sketches, which organizes strains into hierarchical clusters based on user-defined similarity thresholds.
  • It provides a direct quantitative link to Average Nucleotide Identity (ANI), showing universal concordance with species and sub-species classifications across 2.3 million genomes from 67 genera.
  • BacTaxID outperforms traditional methods in capturing strain-level diversity and replicates SNP and cgMLST-based definitions in surveillance and outbreak scenarios.

  Abstract

  Bacterial strain typing is key to surveillance, outbreak investigation and microbial ecology, yet current systems remain species-specific, reference-dependent and lack a universal, interpretable metric of genomic relatedness. Here, we introduce BacTaxID, a fully configurable, whole-genome k-mer-based framework that encodes each genome as a numeric sketch and organizes strains into hierarchical clusters with user-defined similarity thresholds. BacTaxID distances are strictly proportional to Average Nucleotide Identity (ANI), providing a direct quantitative link between vectorial typing and genome-wide divergence. Applied to 2.3 million genomes from All the Bacteria across 67 genera, BacTaxID demonstrates universal concordance species and sub-species classification systems, while capturing finer strain-level diversity than traditional reference-based approaches. In simulated surveillance and real outbreak datasets, BacTaxID reproduces SNP and cgMLST-based definitions while enabling rapid, scalable screening. Precomputed genus-level schemes and an open implementation provide a practical, genus-agnostic alternative to classical typing systems for standardized bacterial classification.

  bioinformatics2026-02-22v3

• Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction

  Zhang, Y.; Gong, C.; Zhang, H.; Ma, W.; Liu, Z.; Chen, X.; Guan, J.; Wang, L.; Yang, Y.; Xia, Y.; Xiao, W.

  AI Summary

  • Protenix-v1 (PX-v1) is an open-source biomolecular structure prediction model that outperforms AlphaFold3 with the same constraints, showing improved accuracy with increased sampling budget.
  • It includes features like protein template integration and RNA MSA support, with a variant, Protenix-v1-20250630, trained on a larger dataset for enhanced accuracy.
  • The study also addresses benchmarking limitations by providing updated evaluation tools and year-stratified benchmarks for more reliable assessments.

  Abstract

  We introduce Protenix-v1 (PX-v1), the first open-source structure prediction model to attain superior performance to AlphaFold3 while strictly adhering to the same training data cutoff, model size, and inference budget. Beyond standard evaluations, we highlight the effectiveness of inference-time scaling behavior, demonstrating that increasing the sampling budget yields consistent improvements in prediction quality--a behavior previously seen in AlphaFold3 but not in other open-source models. In addition to improved accuracy, Protenix-v1 incorporates key capabilities including protein template integration and RNA MSA support. Furthermore, to better support real-world applications such as drug discovery, we additionally release Protenix-v1-20250630, a variant trained on a larger dataset (cutoff: June 30, 2025), delivering further improved prediction accuracy. Finally, we identify the limitations of current benchmarking tools and we provide updated evaluation tools and year-stratified benchmarks to facilitate more reliable and transparent assessment within the community. Collectively, these contributions provide a robust foundation for the Protenix series and the broader field.

  bioinformatics2026-02-22v3

• Deciphering Features of Metalloprotease Cleavage Targets Using Protein Structure Prediction

  Chung, D. S.; Park, J.; Choi, W.; Hong, D.

  AI Summary

  • The study developed a computational model using protein structure prediction to identify and classify substrates of ADAM10, a metalloprotease, and predict its cleavage sites.
  • The model analyzed predicted protein complexes, focusing on protein-protein interactions, structural information of cleavage sites, and spatial relationships with metal ions.
  • Results showed the model's effectiveness in substrate classification and cleavage site prediction, with potential applications to other ADAM family members and metalloproteases.

  Abstract

  Metalloproteases are a class of enzymes that utilize metal ions within their active sites to catalyze the hydrolysis of peptide bonds in proteins. Among these, ADAM10 (A Disintegrin and Metalloproteinase 10), a member of the ADAM family, plays a crucial role in mediating intracellular signaling by cleaving specific substrates, thereby influencing a variety of physiological and pathological processes. The mechanisms underlying the activity of ADAM10 present significant opportunities for the development of novel therapeutic strategies aimed at disease intervention. However, the information available to identify the substrate and cleavage sites of ADAM10 is still insufficient. Therefore, it is essential to identify and classify the features of substrates and to elucidate cleavage sites through experimental approaches. However, these studies across numerous proteins present significant challenges. To address the promise of these investigations, we developed a model that leverages protein structure prediction to decipher substrate features, classify substrates, and predict cleavage sites. Through the analysis of predicted protein complexes between ADAM10 and its substrates using PDB files, we evaluated protein-protein interaction (PPI) data, the structural information of cleavage sites, and the spatial relationships between the cleavage sites and metal ions. Finally, we present a computational model that effectively classifies substrates and accurately predicts cleavage sites in this study. Our study demonstrates the potential for application not only to ADAM10 but also to other members of the ADAM family and, more broadly, to additional metalloproteases. By leveraging computed protein structural information, our approach offers a novel perspective for substrate classification.

  bioinformatics2026-02-22v2

• Bias in genome-wide association test statistics due to omitted interactions

  Yelmen, B.; Güler, M. N.; Estonian Biobank Research Team, ; Kollo, T.; Möls, M.; Charpiat, G.; Jay, F.

  AI Summary

  • This study investigates how omitting interaction terms in linear models used for GWAS can bias test statistics, specifically by altering the mean and variance of the null statistic.
  • Through mathematical derivation and simulation using Estonian Biobank data, the study shows that ignoring epistasis can lead to an anti-conservative regime, inflating significance.
  • The findings suggest that GWAS results based on linear models might include spurious associations, urging caution in their interpretation.

  Abstract

  Over the past two decades, genome-wide association studies (GWAS) enabled the discovery of thousands of variants associated with many complex human traits. However, conventional GWAS are still widely performed with linear models with the assumption that the genetic effects are predominantly additive. In this work, we investigate the test statistic behavior when linear models are used to obtain significant genotype-phenotype associations without accounting for epistasis. We first algebraically derive mean and variance shift in the null statistic due to the omitted interaction term, and define the boundary between conservative (i.e., deflated statistic tail) and anti-conservative (i.e., inflated statistic tail) regimes for the common GWAS significance threshold. We then perform phenotype simulation analyses using the Estonian Biobank genotypes and validate the mathematical model. We demonstrate that the anti-conservative regime is plausible under realistic parameter settings and models omitting interaction terms can produce spurious significance. Our findings suggest caution when interpreting statistically significant signals reported in the literature based on linear models, especially for large-scale GWAS.

  bioinformatics2026-02-22v2

• Cellects, a software to quantify cell expansion and motion

  Boussard, A.; Petit, M.; Arrufat, P.; Dussutour, A.; Perez-Escudero, A.

  AI Summary

  • Cellects is an open-source software designed for automated quantification of cell expansion, motion, and morphology from 2D and time-lapse images across various biological systems.
  • It features a graphical interface for interactive parameter tuning, visualization, and batch processing, with customization options via a Python API.
  • The software outputs quantitative data like area and trajectory in .csv format, facilitating statistical analysis and integration into existing research workflows.

  Abstract

  Cellects is a user-friendly and open-source software for automated quantification of biological growth, motion, and morphology from 2D image data and time-lapse sequences (2D + t), acquired under a wide range of experimental conditions and biological systems (from fungal colonies to unicellular branching networks). The software is available as a stand-alone version, featuring a graphical interface that supports interactive parameter tuning, visualization, validation, and batch processing. The analysis pipeline can be extended and customized using a dedicated Python API. The typical inputs and outputs are as follows. Cellects is designed to process grayscale or color images originating from standard microscopy, macroscopic imaging setups, or camera-based platforms. The software supports single or multiple organisms growing or moving in one or several arenas and can analyze multiple folders sequentially. All quantitative results (area, circularity, orientation axes, centroid trajectories, oscillations, network topology) are exported as standardized .csv files suitable for downstream statistical analysis, ensuring reproducibility and integration into existing workflows.

  bioinformatics2026-02-22v2

• Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction

  Zhang, Y.; Gong, C.; Zhang, H.; Ma, W.; Liu, Z.; Chen, X.; Guan, J.; Wang, L.; Yang, Y.; Xia, Y.; Xiao, W.

  AI Summary

  • Protenix-v1 (PX-v1) is introduced as an open-source biomolecular structure prediction model that outperforms AlphaFold3 with the same constraints.
  • It shows improved prediction quality with increased sampling budget and includes features like protein template integration and RNA MSA support.
  • A variant, Protenix-v1-20250630, trained on a larger dataset, offers enhanced accuracy, and new evaluation tools are provided for better benchmarking.

  Abstract

  We introduce Protenix-v1 (PX-v1), the first open-source structure prediction model to attain superior performance to AlphaFold3 while strictly adhering to the same training data cutoff, model size, and inference budget. Beyond standard evaluations, we highlight the effectiveness of inference-time scaling behavior, demonstrating that increasing the sampling budget yields consistent improvements in prediction quality--a behavior previously seen in AlphaFold3 but not in other open-source models. In addition to improved accuracy, Protenix-v1 incorporates key capabilities including protein template integration and RNA MSA support. Furthermore, to better support real-world applications such as drug discovery, we additionally release Protenix-v1-20250630, a variant trained on a larger dataset (cutoff: June 30, 2025), delivering further improved prediction accuracy. Finally, we identify the limitations of current benchmarking tools and we provide updated evaluation tools and year-stratified benchmarks to facilitate more reliable and transparent assessment within the community. Collectively, these contributions provide a robust foundation for the Protenix series and the broader field.

  bioinformatics2026-02-22v2