Bioinformatics Advances

MotivationAs biomedical datasets and knowledge graphs continue to grow in size, complexity, and heterogeneity, navigating and extracting actionable insights from them presents a major bottleneck for researchers. There is a clear need for autonomous analytical solutions that can utilize recent advancements in agentic AI such as agent harnessing and loop engineering without introducing hallucination or workflow fragmentation. Researchers, regardless of technical expertise, need tools that streamline complex data analysis and deliver meaningful, actionable insights grounded in both data and established biomedical knowledge. EcoXAI addresses this by introducing a modular, customizable, containerized multi-agent system that structures analysis into explicit pipeline execution stages, lowering the computational barrier for clinical and translational researchers. ResultEcoXAI replaces monolithic AI text interfaces with an autonomous execution-driven framework with specialized bioinformatics agents for delivering proactive, data-driven insights grounded in established biological knowledge. Unlike purely LLM-driven or less integrated AI solutions prone to hallucinations or biologically implausible outcomes, EcoXAIs multi-agent framework, which leverages modern agentic management and explicit knowledge graph integration, provides greater transparency and verifiability in its reasoning. In our use case in drug repurposing for Alzheimers Disease, EcoXAI evaluated 103 drug candidates and identified 79 novel candidates whose predictive models exceeded a randomized baseline, including the CCR5 antagonist Maraviroc, whose generated hypothesis was subsequently supported by the literature. These results demonstrate the potential of knowledge graph-grounded AI agents to accelerate hypothesis-driven biomedical research. Availability and implementationEcoXAI is available on GitHub at: https://github.com/EpistasisLab/EcoXAI. Contactjason.moore@csmc.edu

Drug discovery remains one of the costliest and most time-intensive endeavors in the pharmaceutical pipeline, with average development costs exceeding $2.3 billion per drug, timelines spanning more than a decade, and attrition rates above 90% in clinical trials. While computational methods have expanded the searchable chemical space, current pipelines remain fragmented and largely inaccessible to researchers without deep interdisciplinary expertise. Here we present F.A.D.E. (Fully Agentic Drug Engine), a multi-agent, open-source platform that converts natural language queries into potential drug candidates, substantially lowering the expertise barrier to advanced computational drug discovery. F.A.D.E. employs a three-branch hierarchical architecture that adapts to the level of available structural data for any protein target, integrating structure prediction, binding pocket detection, equivariant diffusion-based de novo ligand generation, and binding affinity estimation into a single automated pipeline. We validate F.A.D.E. on two structurally distinct targets: the epidermal growth factor receptor kinase domain (EGFR), a well-established oncology target, and cellular retinol-binding protein 1 (CRBP1), a lipid-binding protein involved in retinoid metabolism. For EGFR, our generated candidates achieved QED scores of 0.85 compared to 0.46 for the co-crystallised reference ligand, demonstrating marked improvement in predicted drug-likeness. Results across both targets confirm that F.A.D.E. can reliably generate chemically tractable, drug-like hit compounds across diverse protein classes from simple natural language input.

pylimma is a faithful Python port of limma, intended to bring one of the most widely used tools for differential expression analysis to the developing Python ecosystem for transcriptomics and proteomics. We validated pylimma against the existing R implementation through 227 function-level comparisons and across six real world datasets spanning microarray, RNAseq, proteomics and single-cell transcriptomics. pylimma reproduces limmas numerical output to a median agreement of 13 significant figures and calls identical sets of differentially expressed features and gene sets. This supports its use as a drop-in replacement for the R implementation.

MotivationIdentifying bacterial antimicrobial resistance (AMR) is critical for diagnostics and treatment, but resistance is a complex trait arising from myriad mechanisms spanning multiple molecular scales. Existing computational approaches often function as black boxes and rarely explore cross-species or multi-drug patterns. We developed amR, an integrated R package suite that provides a complete framework from bacterial genome data curation to interpretable AMR predictions, enabling identification of resistance mechanisms across species and drugs. ResultsThe amR R package suite contains three modular packages. amRdata downloads genomes and paired antimicrobial susceptibility testing data from BV-BRC and processes them, constructs pangenomes, and extracts features at gene/protein cluster, protein domain, annotated Clusters of Orthologous Groups and ResFinder AMR-associated features, and structural variant scales; data are stored in memory-efficient formats (Parquet, DuckDB). amRml trains interpretable machine learning models per species-drug combination, calculates feature importance and performance metrics, and provides rich ground for hypothesis generation and mechanism discovery. amRviz provides an interactive Shiny dashboard to explore metadata distributions and model performance across species and drugs, visualize top predictive AMR features, and analyze cross-model patterns across geographic/temporal strata. We apply the suite to Shigella sonnei, achieving a median Matthews Correlation Coefficient of 0.89 across 23 drugs and drug classes. With thousands of genomes, multi-scale features, and interpretable models, amR provides an accessible, comprehensive framework for AMR research. The amR package suite is installable via GitHub (https://github.com/JRaviLab/amR; BSD-3-Clause license).

Motivation: Deep learning has rapidly become essential for predicting biomolecular interactions; however, most web-tools expose only a single, pre-built model with a fixed, non-configurable architecture that users cannot redesign, retrain on their own data, or compare; they are typically dedicated to one interaction type and often one species, and report prediction scores with little interpretability. These constraints force researchers across several disconnected, single-purpose tools and limit the flexibility, reproducibility, and long-term usability of existing platforms. Results: We present Deep-Interact Studio, a unified, web-based deep-learning platform that addresses these limitations by shifting interaction prediction from a model-centric to a user-driven, comparative, and interpretable paradigm. Within a single interface spanning all four interaction classes, namely protein-protein, drug-target, RNA-protein, and protein-DNA, users design their own model architectures layer by layer, configure training hyperparameters, and train them on their own data, including custom, species-specific datasets. Multiple user-built models can then be trained under identical conditions and compared side by side at both the training and inference levels, while integrated interpretability, including SHAP-based feature attribution, embedding-space visualization, and interaction hub analysis, turns predictions into auditable, mechanistically grounded results. Deep-Interact Studio is, to our knowledge, the only such platform to combine fine-grained per-layer model customization with multi-model comparison and interpretability, offering a flexible and transparent alternative to fixed, single-purpose tools.

Prediction of B-cell epitopes can assist in reducing costly wet-lab screening in vaccine design, diagnostics, and antibody discovery. However, current predictors often suffer from noisy labels, weak generalization, and structure-dependent workflows. Here we present EO_SCPLOWPIC_SCPLOWESM-GA, an efficient sequenceonly pipeline for linear B-cell epitope prediction. Positive and negative peptide examples are collected from IEDB, which provides experimentally tested epitopes and distinguishes positive and negative epitope records based on assay evidence(Vita et al., 2019). Each peptide is encoded with a frozen ESM-2 protein language model: a bidirectional transformer producing amino acid embeddings for downstream structure and function tasks (Lin et al., 2023). Mean-pooled embeddings are further compressed into a compact 420-feature representation with a genetic algorithm and classified with lightweight Random Forest, XGBoost, or MLP heads. This avoids foundation-model fine-tuning, reduces the number of trainable parameters, improves interpretability, and enables low-resource deployment. On an IEDB-derived benchmark, EO_SCPLOWPIC_SCPLOWESM-GA attains 0.880{+/-} 0.004 AUC-ROC, 0.852{+/-} 0.005 PR-AUC, 82.0 {+/-} 0.6% accuracy, 0.79 {+/-} 0.01 F1, and 0.74{+/-} 0.01 MCC, outperforming dense ESM-2 features and baselines LBCE-XGB, EpitopeVec, and BepiPred-2.0 (mean{+/-} std over five independent random seeds). The framework shows how frozen protein foundation models can enable pandemic preparedness, peptide vaccine prioritization, diagnostic antigen screening, and equitable computational immunology.

Drug repurposing offers a practical strategy to identify new therapeutic uses for approved drugs, potentially reducing the time and cost associated with conventional drug development. We present a novel three-stage drug repurposing pipeline that integrates knowledge graph-based gene prediction, network-based drug-disease association analysis, and systematic classification of candidate drugs by therapeutic class. The pipeline integrates DGLinker to predict novel disease-associated genes, SAveRUNNER to identify drug repurposing candidates, and ATC Category Enrichment Analysis (ATCEA) to prioritise candidates by pharmacological class. We benchmarked the pipeline across twelve diseases using DrugBank and MEDI2-HPS as validation resources. Utilising DGLinker-expanded disease-gene sets as input increased the number of predicted repurposed drugs, while overall discriminative performance remained stable across diseases (AUROC 0.71-0.77). Application of ATCEA consistently improved precision, F1-score, and specificity, while reducing recall, reflecting a conservative prioritisation strategy that contracts the candidate space while retaining pharmacologically coherent drug-disease candidates. We further applied the pipeline to amyotrophic lateral sclerosis (ALS), a neurodegenerative disease with limited therapeutic options, and performed a deeper literature-based validation of the results. Incorporation of DGLinker-predicted genes substantially increased the number of significant candidate drugs and uncovered enriched ATC categories not identified using known ALS genes alone, including antidepressants and antipsychotics. Moreover, several drugs with supporting evidence available in the literature were identified only when DGLinker-predicted genes were used. Overall, 77 candidate drugs were prioritised within significantly enriched ATC categories, several of which are supported by previously published studies. To provide exploratory real-world support for these findings, we further evaluated candidate drugs in a longitudinal electronic health record (EHR) dataset of 2361 patients with ALS from King's College Hospital. Although the number of evaluable drugs was limited due to sample size, the EHR analysis provided additional clinically relevant context for selected prioritised drugs and pharmacological classes. Our pipeline demonstrates potential to accelerate drug repurposing by integrating complementary computational approaches to each step of the process, providing an end-to-end framework that showed robust performance across benchmarking experiments and use cases.

Determining the structural basis of antigen recognition by antibodies and T cell receptors (TCRs) provides critical insights into effective immune targeting and can inform design of biotherapeutics and vaccines. Accurate computational modeling of antibodies and TCRs in complex with their targets poses a major challenge for predictive methods, including AlphaFold, which is generally accurate for modeling protein complexes but has shown limited success for immune recognition. In this study we assessed the performance of AlphaFold2, AlphaFold3, increased sampling protocols, and related deep learning methods for modeling antibody-protein, antibody-peptide, and TCR-peptide-major histocompatibility complex (pMHC) recognition. We show that increased sampling and AlphaFold3 generally improve performance relative to default sampling and AlphaFold2, however predictive accuracy and improvement levels varied considerably among interface classes, with antibody-peptide complexes representing a challenge despite their small antigen size. Comparing per-case success across methods showed some complementarity, indicating opportunities for increased success through model pooling approaches, for instance increasing antibody-peptide near-native success from 41% to 59%. Analysis of AlphaFold confidence scores and modeling of a noncanonical complex provided further insights into predictive performance. These results highlight considerations for predictive antibody and TCR complex modeling efforts, while revealing key distinctions among protocols, scoring, and immune complex classes.

Accurate gene-protein-reaction (GPR) associations are essential for the predictive performance of genome-scale metabolic models (GEMs),as they define the mapping between genes, enzymes, and metabolic reactions. However, GPR rules are often incomplete or inconsistent due to limitations in annotation transfer and the ambiguous representation of multi-subunit protein complexes, leading to errors in downstream analyses such as gene essentiality prediction. Here, I introduce Comp2GPR, an automated pipeline for reconstructing GPR rules that integrates curated protein complex information with sequence-level evidence. Protein complexes were sourced from the Complex Portal and subjected to an AI-assisted curation workflow to retain only metabolically relevant assemblies. Comp2GPR combines deterministic sequence similarity mapping with explicit rule construction to generate Boolean GPR expressions that accurately represent obligate subunit relationships and isoenzyme redundancy. I evaluated the impact of the reconstructed GPR rules by integrating them into the Yeast9 metabolic model and comparing gene essentiality predictions with the original model. While global performance metrics remained largely unchanged, the updated model achieved a net improvement in prediction accuracy through gene-level corrections. Overall, Comp2GPR demonstrates that combining curated protein complex data with sequence-based validation improves the accuracy, interpretability, and reproducibility of GPR rules. The method provides a robust framework for enhancing metabolic model annotations and supports more reliable simulation-based analyses.

Variant effect prediction (VEP) models can be used to select promising novel enzymes from a pool of candidates. Most supervised VEP models are framed as regression tasks, placing more emphasis on getting the predicted quantities correct than on the relative comparison of individual candidates. Preferential or contrastive models may better align with the goal of selection, or acquisition, especially when informed by predictive uncertainty. Here, we introduce a probabilistic preferential learning model based on the Kermut Gaussian process (PKermut) that we designed with the ambition to increase the hit rate among selected variants. We benchmark PKermut against established models, including the original Kermut, the RITA regressor, and an augmented Potts model, on 69 curated ProteinGym datasets across various assay categories. To evaluate acquisition performance, we propose a novel quantile cross-validation scheme that ensures the evaluation of a models ability to extrapolate by reserving high-performing variants exclusively for the test set. We assess models using Spearman correlation and evaluate their acquisition performance using five different acquisition functions, encompassing both uncertainty-aware and unaware strategies. Our experimental results indicate that uncertainty estimates improve the acquisition ability of our models, and that strategies that reward uncertainty generally result in better outcomes than those that do not on single-mutation variant datasets. We observe that PKermuts Spearman scores and ability to acquire improved variants are greatly affected by the number of variant comparisons sampled in the training set. Kermut achieves the highest Spearman correlation in 54/69 datasets (78%), compared to 12/69 (17%) for PKermut. For acquisition performance, Kermut leads in 44/69 datasets (64%), while PKermut leads in 15/69 (22%). While at this stage PKermut is not a recommended alternative to Kermut, its contrastive nature offers several conceptual opportunities. We share our findings to inspire further development aimed at improving the alignment between training objectives of VEP models and their downstream application in protein engineering.

Aptamers are emerging as important molecular recognition ligands in oncology, playing significant roles in cancer diagnostics, targeted therapies, drug delivery systems, and molecular imaging. Numerous aptamers have advanced to clinical trials, indicating their potential for real-world applications; however, existing databases fail to capture that. To bridge this critical gap, we developed AptCancerDB (https://webs.iiitd.edu.in/raghava/aptcancerdb/), a comprehensive, manually curated database of experimentally verified anticancer aptamers. The current release contains 1,941 entries collected from studies published between 2000 and 2025, covering 29 cancer types, approximately 200 cancer cell lines, and direct links to 22 clinical trials. Each entry is annotated with sequence information, target details, cancer type, cell line, SELEX methodology, affinity determination data, chemical modifications, and biological activities. The dataset is dominated by 82.7% ssDNA, reflecting its superior stability and ease of synthesis, while only 16.6% is ssRNA and appears primarily in studies targeting complex intracellular or protein-protein interactions. To facilitate structural analysis, predicted secondary structures, dot-bracket notations, specific structural elements, and minimum free energy values were also included. AptCancerDB integrates a MySQL backend with an ArcadeDB/OpenCypher-based Knowledge Graph, enabling exploration of relationships among aptamers, targets, cancer types, cell lines, and functional applications. The platform provides advanced search and browsing facilities, BLASTn-based similarity searching, and GC Calculator. Built on a modern, responsive frontend (React/TypeScript/Tailwind CSS), the platform includes a REST API for data retrieval. By integrating fragmented experimental data into a unified cancer-focused resource, AptCancerDB serves as a valuable resource for comparative analysis, aptamer discovery, and the development of next-generation aptamer-based diagnostics and therapeutics. HighlightsO_LICurated knowledge base of experimentally validated anticancer aptamers. C_LIO_LIAptCancerDB contain therapeutic, tumor-homing and cell-penetrating aptamers. C_LIO_LISummarizes clinical progress and translational trends in anticancer aptamer research. C_LIO_LISupports rational aptamer design using molecular, functional, and clinical annotations C_LIO_LIDisease-focused resource for cancer diagnosis, therapy, and drug delivery C_LI TeaserAptCancerDB maintains experimentally validated anticancer aptamers relevant to diagnosis, drug delivery, and therapy.

Protein language models (pLMs) offer great potential for protein sequence analysis, yet the scarcity of labeled data often limits their effectiveness in fine-tuning. Data augmentation is a promising remedy, but systematic evaluation of augmentation strategies for protein sequences remains limited, and the conditions under which augmentation confers downstream benefits are not well understood. In this paper, we systematically investigate pLM-guided substitution-based augmentation across seven protein prediction tasks. We propose ProtAug, a framework that leverages encoder-based (ESM-2) and autoregressive (ProtGPT2) pLMs to generate augmented sequences with user-controlled variation levels. Our investigation focuses on four questions: (Q1) whether pLM-synthesized sequences preserve more original signals than simpler methods, (Q2) to what extent augmentation improves prediction performance, (Q3) how variation levels affect downstream accuracy across tasks and models, and (Q4) whether biological plausibility is a necessary condition for achieving improvement. Our experimental results show that: (1) ProtAug Esm generally preserves motifs and structural similarity better than simple substitution, often comparable to homology retrieval; (2) augmentation yields consistent but task-dependent improvements, with ProtAug Esm achieving the best or second-best performance in 5 out of 7 tasks at 10% variation; (3) low-to-moderate variation levels (2-30%) perform best overall, although high-variation augmentation can benefit certain structure-related tasks; (4) the necessity of biological plausibility is task- and variation-dependent--while semantic preservation correlates with performance at low-to-moderate variation levels, improved generalization at high variation levels suggests that regularization effects, rather than label preservation, can also drive performance gains.

Heterogeneous biomedical knowledge networks (hetnets) integrate disparate data types, drugs, genes, diseases, and pathways, across independent sources; Hetionet (https://het.io) is a widely used example. A standard approach for assessing connectivity significance is XSwap, which permutes the hetnet P times and fits a gamma-hurdle null model to the degree-weighted path count (DWPC), pooling permuted values across pairs with matching source and target degrees to increase the effective sample size. This permutation approach has been highly successful in practice, but it faces four practical constraints in large graphs: (1) a finite resolution for the smallest reportable p-values, (2) computational cost that grows prohibitive at path lengths L [≥] 4 or 5, (3) a variance model (Var {propto} {micro}2) that departs from the configuration-model form (1 +{kappa} ){micro}, and (4) O(P 10m L) runtime. To complement this approach, we present HetNetEX (Heterogeneous Network EXact inference), which computes the null DWPC distribution analytically from degree sequences using the configuration model in O(Ln) time. In simulations at P = 200 across L = 1-4, HetNetEX achieves Spearman{rho} > 0.96 concordance with XSwap rankings while being >10,000x faster and providing analytical p-values without a resolution ceiling. High-degree pairs show larger XSwap sampling error than low-degree pairs, reflecting the finite-sample nature of permutation that analytical computation avoids.

BackgroundLarge language model (LLM) agents increasingly automate bioinformatics analyses, but most existing bioinformatics tools were built for standalone use by human experts. An agent driving such a tool must reason about its installation, configuration, and execution from documentation for human, spending many turns, tokens, and tool calls per result. How a method is exposed to an agent can therefore matter as much as the method itself. By designing agentic interfaces for these tools, agent can reduce such overhead and improve the reliability of agent-driven analyses. FindingsTo test this design, we re-architected Clair3, a widely used deep-learning-based long-read variant caller, into a client-server system, Clair3-Connect. The client performs all genomics related processing and holds the identifiable data. The server runs only neural-network inference, and the client sends only feature tensors to the server, while sample identifiers and genomic context remain on the client. The client exposes schema-defined agent-facing tools that an agent invokes through single structured calls. On an APOE diplotyping task, all 60 agent runs were correct. The agentic tools used 12K tokens in 3 turns, 6.8 to 14 times fewer tokens than the shell-driven baselines (81K-163K tokens), at about a quarter the wall-clock time and far more stably (4% versus 35% token usage variation). Dropping the pileup and phasing stages to keep the client light left SNP F1 within 0.1-0.3 points of standard Clair3 by 50x coverage, while mutual TLS and AES-256-GCM encryption added 7.2% to end-to-end runtime. ConclusionsRecasting an established algorithm as developer-built, agentic tools behind a secure client-server boundary makes it more efficient, reliable, and easier to deploy for an LLM agent than a third-party wrapper, which cannot recover the defaults and conventions only its developers know. Agentic interfaces should be a first-class deliverable of bioinformatics tool development.

MotivationRNA inverse folding, the design of RNA sequences that fold into specified target structures, is a central problem in RNA design, with applications in functional RNA engineering, synthetic biology, and nucleic-acid therapeutics. This task becomes especially challenging for pseudoknotted target structures because pseudoknots disrupt the nested structure assumed by standard thermodynamic folding models. Existing pseudoknot inverse-folding methods often rely on structure-predictor-based objectives. Direct optimization of the thermodynamic folding probability of a specified pseudoknotted target remains limited. This requires an evaluator that can assign target-specific folding probabilities within a pseudoknot-aware ensemble and can be used as an optimization signal. ResultsWe present PKProbDesign, a sampling-based inverse-folding framework that directly optimizes a thermodynamic folding-probability objective for pseudoknotted targets. For each target, candidate sequences are scored by combining the folding probability of a pseudoknot-free scaffold with the conditional folding probability of the remaining extension component. On 354 PseudoBase++ targets, PKProbDesign achieved the highest folding probability on 221 targets, compared with 117 for DesiRNA and 16 for MODENA. ConclusionsPKProbDesign demonstrates that pseudoknot inverse folding can be formulated around target folding probabilities rather than structure-prediction agreement alone. By combining scaffold decomposition with HFold/CParty-consistent conditional-ensemble evaluation, the method provides a practical probability-based framework for designing sequences for density-2 pseudoknotted targets. AvailabilityThe source code of PKProbDesign is available at https://github.com/TakumiOtagaki/PKProbDesign.

Germinal is a recently described computational pipeline for de novo antibody design that combines AlphaFold-Multimer hallucination with antibody language model guidance to generate epitope-targeted antibodies. Germinal identified binders with nanomolar-to-low-micromolar affinities by testing only 43-101 designs per target across four diverse antigens, establishing it as a practical tool for epitope-directed antibody design accessible to standard academic laboratories. As this architecture is itself very recent, systematic replacement and benchmarking of its individual components remains largely unexplored, yet offers a valuable opportunity to probe the robustness of the underlying design. We present OpenGerminal, which replaces PyRosetta with a fully open-source stack comprising OpenMM 8.5.1, FreeSASA, FASPR, Biopython, and sc-rs v1.0.0, and adopts AbLang1 (ablang2 v0.2.1) as the sole antibody language model in place of IgLM. Benchmarking on two VHH targets (PD-L1 and IL-3) reveals that OpenGerminal achieves a markedly higher cofolding pass rate (PD-L1: 33.7% vs. 18.6%; IL-3: 24.6% vs. 8.0%) with equivalent or improved Chai-1 structural confidence metrics in accepted designs, at the cost of a modest increase in per-trajectory computation time (>=1.5x). Multi-chain target support is also extended and verified to run without error on the official insulin example. OpenGerminal provides the first systematic benchmarking of IgLM versus AbLang1 within the Germinal architecture, and its fully open-source component stack broadens the range of deployment contexts in which the pipeline can be used.

MotivationComputational enzyme classification has previously utilised sequence homology features and protein language model embeddings. The Kolmogorov-Arnold Network (KAN) paradigm, which uses learnable edge functions rather than fixed ones, has shown promising results in biological sequence tasks. ResultsA fully reproducible investigation of KAN variants for seven-class EC classification on up to 9,516 labelled sequences from the CLEAN benchmark [1] (9,386 for language model experiments). In the sequence only settings, fixed basis KAN variants outperformed an MLP baseline moderately (macro F1 = 0.17-0.29). Utilisation of ESM-2 650M embeddings [2] greatly improved results via 5-fold cross-validation: MLP macro F1 = 0.750 {+/-} 0.009, accuracy = 0.823 {+/-} 0.009; learnable SineKAN macro F1 = 0.716 {+/-} 0.023, accuracy = 0.788 {+/-} 0.019. MLP performed comparably but did not exceed conventional baselines. As an aside, we introduce but do not investigate an approach to EC oxidoreductase sub-classification through the use of a Marcus theory-based electron transfer feature framework. AvailabilityCode and result files are available at https://github.com/sanjuz-cas/ENZYKAN.

BackgroundSingle-cell foundation models are increasingly used for perturbation prediction and gene network inference, but their learned gene representations are rarely audited directly. In natural language processing, geometric analyses of token embeddings have revealed anomalous "glitch tokens" associated with erratic model behaviour. Whether analogous representational anomalies exist in biological foundation models remains unknown. ResultsThis study introduces a weight-only geometric audit framework that scores genes by embedding norm, centroid distance, cosine similarity, and isolation to identify representational outliers. Applied to Geneformer, scGPT, and scFoundation, the analysis identifies hundreds of outliers in discrete-tokenisation models. Shared Geneformer-scGPT outliers are enriched for loss-of-function intolerance (OR=12.0) and disease association (OR=3.7), whereas scFoundations continuous value embeddings form a near-isotropic space with no detectable enrichment under the annotation panels tested. In Geneformer, geometric anomaly predicts perturbation sensitivity ({rho} = 0.725); the signal is supported by mask-in-place experiments, shows rank agreement in real PBMC cells, and correlates with Replogle perturb-seq effect sizes ({rho} = 0.645). Metric decomposition separates magnitude-driven outliers, enriched for highly expressed housekeeping genes, from isolation-driven outliers enriched for tissue-restricted genes. ConclusionsTokenisation strategy helps determine which genes are represented reliably. Embedding geometry provides a rapid, model-agnostic diagnostic that requires only an embedding matrix and can flag genes whose representations warrant caution before downstream use.

Large language models can annotate cell types from marker gene lists, but they typically operate after preprocessing and clustering are complete, treating annotation as a terminal labeling step rather than controlling the analytical decisions that produce the evidence for cell identity. We present CellPilot, an agentic framework that guides a locally deployable small language model through the full single-cell analysis workflow, from raw count matrices to cluster-level annotation. CellPilot combines standard single-cell analysis tools with structured workflow control and observation-guided reasoning, allowing the model to plan analyses, execute tools, inspect intermediate results and revise decisions within a traceable session. On GTEx, structured workflow orchestration raised the same 8B model from 0.39 in a prompt-only setting to 0.89, closing most of the gap to GPT-4o (0.92) within the same framework; the framework gain was substantially larger for the smaller backbone across datasets (+0.35 versus +0.19). Across GTEx, Tabula Sapiens, and Mouse Cell Atlas, CellPilot achieves cluster-level annotation accuracies of 0.891, 0.750, and 0.773, outperforming representative reference-based, marker-based, and LLM-based methods. CellPilot confidence scores were associated with annotation correctness and supported post hoc filtering, while complete execution traces were retained for each analysis. These results suggest that structured workflow orchestration can be a critical determinant of performance in multi-step single-cell analysis, enabling locally deployable small language models to approach larger proprietary models while preserving transparency and practical usability.

Single-cell foundation models are increasingly positioned as virtual cells, yet their capabilities are assessed by fragmented, largely single-task benchmarks that obscure where these models improve on simple baselines. VCBench addresses this by synthesizing four independent virtual-cell frameworks into seven capability dimensions: perturbation response prediction, cross-species universality, gene regulatory network (GRN) inference, modality integration, temporal dynamics, multi-scale integration, and in silico experimentation. Each dimension is assessed for operational testability under current architectures and datasets: five admit direct or proxy evaluation, while multi-scale integration and in silico experimentation are structurally untestable as end-to-end tasks. We evaluate five foundation models (Geneformer, scGPT, UCE, TranscriptFormer, Arc State) against pre-registered linear and nearest-neighbor baselines across the five testable dimensions, and report three findings. First, the baselines match or exceed every foundation model on four of the five scored dimensions, replicating the reported competitiveness of linear baselines on perturbation prediction and extending it to cross-species transfer, GRN inference, and temporal ordering. Second, TranscriptFormer alone exceeds the strongest baseline on cross-modal RNA-to-protein prediction (53% Pearson improvement, with a documented contamination caveat) and is the only model to reach Level 2 in the pre-registered Virtual Cell (VC) Level rubric; the architectural choice behind this advantage simultaneously causes a spectral collapse that destroys its temporal-ordering performance, a tradeoff invisible to single-task benchmarks. Third, no foundation model publishes a complete cell-level training manifest, leaving data contamination undetectable to users. Alongside the benchmark, VCBench releases a Contamination Reporting Schema and contributes two further methodological tools: a common-label-set protocol that controls for class-count confounds in cross-species transfer, and a spread-error correlation probe for epistemic calibration.