Nature Computational Science

Mapping T-cell receptor (TCR) sequences to their cognate peptide-major histocompatibility complex (pMHC) ligands underlies both basic immunology and T-cell target discovery, yet current models aimed at predicting TCR specificity are limited by sparse labels, viral-biased training data, and an inability to recognize receptors outside their training distribution. We present PRISM, an uncertainty-aware metric-learning framework for TCR{beta} sequence representation. PRISM embeds receptors into a peptide-organized latent space, returns top-k peptides by nearest-neighbor retrieval, and abstains on out-of-distribution receptors by modeling an intrinsic uncertainty that tracks annotation correctness. To offset the viral bias of public databases, PRISM augments training data with structure-guided synthetic receptors that diversify TCR sequences while preserving the energetics of the TCR-pMHC interface. Across a held-out set of 923 peptides and the independent IMMREP23 benchmark, PRISM matches or exceeds sequence-based models, with largest gains on rare epitopes. Finally, PRISM learns attention weights on TCR residues that concentrate on the CDR3{beta} salt-bridge and hydrophobic contacts central to peptide recognition, linking PRISM's positional focus to the biochemical properties of TCR-pMHC structures.

Summary: The X chromosome comprises approximately 5% of the human genome and encodes over 800 protein-coding genes, many of which exhibit sex-differentiated expression patterns due to escape from X chromosome inactivation (XCI) mechanisms. Despite its relevance to sex differences in complex traits, the X chromosome is routinely excluded from genome-wide association studies due to analytical challenges, and when analyzed, the impact of escape from XCI or sex is limitedly explored. No dedicated, publicly accessible browser for X chromosome-wide association study (XWAS) summary statistics currently exists, creating a barrier to systematic investigation of X-linked contributions to human traits. Here, we present geneXplore, an interactive web browser based on the PheWeb2 implementation, tailored for XWAS summary statistics across 1,944 phenotypes while distinguishing random XCI (rXCI), escape from XCI (eXCI), and sex-stratified analyses. Users can explore results via interactive plots (Manhattan and Miami, PheWAS and LocusZoom), searchable tables and access to cross-database lookup, with full summary statistics available for download. Availability and Implementation: geneXplore is freely available at https://genexplore.wustl.edu/ with no registration required and will be maintained for a minimum of two years following publication. Source code is available at https://github.com/Belloy-Lab/geneXplore_XWAS_Browser under an MIT license.

Biobank-scale genomic analyses remain computationally expensive, CPU-bound workflows, particularly when adjusting for confounding. Here, we present CuGen, a GPU-accelerated framework for large-scale genomics. CuGen uses UltraLasso, a novel hierarchical application of univariate-guided sparse regression (uniLasso), to select a compact, phenotype-informed active set of fewer than 30,000 variants. This achieves robust leave-one-chromosome-out (LOCO) confounding control, enabling both downstream GWAS and in-sample fine-mapping. Additionally, we introduce the .cugen file format, a genotype representation designed for memory-optimized, high-throughput streaming and random access on GPU hardware. Building on this substrate, we provide a general GPU-accelerated genomics toolkit handling polygenic prediction, data manipulation, quality control, analysis, and visualization. We demonstrate CuGen's efficacy in the UK Biobank with up to 408,624 individuals, where the full GWAS pipeline and fine-mapping against 6.8 million imputed variants completes in approximately 10 minutes on a single high-throughput GPU with 80 GB of memory. The pipeline scales efficiently to massive phenome-wide analyses with sublinear resource consumption.

Personalized whole-brain modeling aims to transform treatment planning for neurological disorders by enabling patient-specific simulations of brain network dynamics. Neural mass models (NMMs) offer a tractable compromise between biophysical detail and computational cost and can be directly linked to macroscopic observables such as EEG. However, scaling NMMs to whole-brain networks with realistic connectivity, conduction delays, and cortical surface resolution--and fitting them to individual patient data--imposes computational demands that existing frameworks cannot meet at clinically relevant timescales. Here we introduce BraiNN, a JAX-based Python framework for large-scale neural mass modeling that achieves speedups of up to two to three orders of magnitude over existing tools by leveraging GPU/TPU-accelerated, XLA-compiled array computation. BraiNN combines a region-level Jansen-Rit network with a subject-specific cortical surface mesh of coupled neural mass models and biophysically grounded EEG forward modeling via reciprocity-based lead fields. Its fully differentiable computational graph enables a hybrid personalization pipeline that pairs Bayesian optimization for global parameter exploration with gradient-based refinement, completing EEG-driven spectral fitting of an eight-dimensional parameter space in approximately 2-3 hours on a single consumer GPU--compared to multiple days with conventional neural mass modeling software. Numerical verification against established benchmarks confirms that BraiNN faithfully reproduces canonical synchronization and bifurcation dynamics of Jansen-Rit networks. By reducing the time requirements for personalizing a high-detail whole-brain surface model from days to a few hours on consumer-grade hardware, BraiNN brings personalized brain network modeling closer to practical use in clinical contexts. We anticipate that BraiNN will serve as a foundation for patient-specific digital twins and EEG-guided neuromodulation planning.

Volumetric electron microscopy (EM) has become a critical approach to generating high-resolution reconstructions of brain tissue. As the size of EM volumes increase, use of automated image segmentation within the reconstruction pipeline has become essential, although it generates errors that need correction. The proofreading and correcting of these errors has since become the dominant cost driver in the pipeline, but precisely estimating the sufficient number of proofreading edits to enable meaningful scientific analyses of the reconstructed neuronal networks remains a challenge. We present a fast, computationally inexpensive way to estimate the progress of a connectomic proofreading effort without requiring a priori knowledge of ground truth. We show that simple global graph invariants converge predictably to asymptotic limits with increasing numbers of proofreading edits, informing a quantitative "pencils down" criterion for proofreading completeness. We illustrate our method on two datasets in different stages of proofreading progress, a zebrafish spinal cord and the hemibrain Drosophila melanogaster dataset. Our method reduces the uncertainty associated with the planning and prioritization of proofreading activities and enables data owners to accurately predict and budget the amount of proofreading necessary for their scientific questions.

Mapping cell-type-specific expression quantitative trait loci (ct-eQTLs) is essential for interpreting disease-associated variants, yet studies in underrepresented populations are hindered by limited statistical power. Here, we present traceCB, a statistical framework that enhances ct-eQTL mapping in target ancestries by integrating summary statistics from single-cell and bulk-tissue eQTL studies across diverse populations. By explicitly modeling trans-ancestry genetic architecture and accounting for cellular heterogeneity in bulk tissues, traceCB optimizes information borrowing from well-powered European cohorts while robustly controlling for type I error. Simulation studies demonstrate that traceCB achieves superior statistical power compared to original ct-eQTL, particularly when leveraging tissue-level data. In an application to immune cells in East Asian and African cohorts, traceCB increased the effective sample size by up to 2.9-fold and identified approximately 40% more eGenes than single-ancestry analyses, with a replication rate exceeding 90% in independent datasets. Furthermore, traceCB improved the colocalization of regulatory variants with GWAS signals for blood and immune-related traits, revealing cell-type-specific mechanisms underlying complex diseases. These findings establish traceCB as a powerful and scalable tool for leveraging global genomic resources to improve regulatory variant discovery at the cellular level across diverse populations.

Neural systems exhibit multiple firing states that reflect an organism's internal state and modulate the relationship between external environmental stimuli and behavior. Several studies have inferred these latent states by supplementing the traditional hidden Markov Model (HMM) with generalized linear models (GLMs) with non-Poisson behavioral observations. However, understanding the relationship between internal brain states and behavior also requires modeling the neural activity. Nonetheless, fitting multi-neuron GLM-HMMs is non-trivial due to high sparsity, collinearity, and low trial counts in neuronal datasets. Therefore, we built a robust multi-neuron GLM-HMM framework that uncovers latent states from population activity while incorporating the influence of time-stamped task variables and spike histories. To obtain reliable model parameters, we employ a modified expectation-maximization procedure. Specifically, we show that incorporating neuron-adaptive penalization in the maximization step overcomes the covariate co-linearity issues typical of time-stamped events and sparse spiking, yielding stable estimates of Poisson GLM coefficients. Furthermore, we incorporate a trust-region algorithm to ensure stable M-step convergence in the presence of ill-conditioned Hessians that can lead to unstable Newton-Raphson updates. We further demonstrate the utility of leave-one-out cross-validation analysis for evaluating model performance on datasets with low trial counts and without breaking their temporal structure. We evaluate our framework on three electrophysiological datasets from primates and rodents as they perform a decision-making task, demonstrate stable model convergence, and discuss the behavioral relevance of the inferred states.

Spatial transcriptomics (ST) technologies enable the study of gene expression within the spatial context of tissues, providing insights into tissue structure, cellular interactions, and disease progression. However, existing dimension reduction methods often overlook spatial information or struggle to distinguish spatial gene patterns from those driven by cell-type differences, limiting biological interpretability by convolving differences in gene expression patterns with differences in cell-type proportions. To address these challenges, we introduce the scalable multi-group nonnegative spatial factorization (smNSF), a computationally-tractable probabilistic framework that integrates spatial coordinates and cell-type labels into a unified matrix factorization model. By using multi-group Gaussian processes (MGGPs) as priors, our model captures complex spatial variation in a cell-type specific way while enforcing nonnegativity to enhance interpretability. We develop a variational inference framework for MGGPs that supports scalable optimization and improves the numerical stability of smNSF. Across seven spatial transcriptomics datasets spanning diverse technologies and tissues, smNSF recovers sparse, interpretable spatial factors and, through its cell-type conditional posteriors, organizes them into cell-type enriched, cell-type specific, and universal spatial programs that are not apparent from marginal factors alone. Given cell-type labels in ST data, smNSF enables cell-type aware spatial decompositions and supports cell-type conditional posteriors for in silico exploration of relationships between spatial patterns and cellular identity.

Abstract-Large language models (LLMs) [1], [2] have demon strated remarkable capabilities across general domains, yet their application in specialized medical contexts demands rigorous domain adaptation [3], [4]. We present Infoxmed2.0-27B, a medical foundation model built upon Qwen3.5-27B [5] through a comprehensive multi-stage post-training pipeline: (1) proprietary medical data synthesis from a MySQL database with MedicalCategoryTree organization, medical PhD team validation, Chinese RoBERTa [6] semantic deduplication, and API-assisted language refinement; (2) instruction supervised fine-tuning of Qwen3.5- 27B via LoRA [7] (r = 8, = 32) using MS-Swift [8], producing iterations Infoxmed2.0.0[->]2.0.2[->]2.0.4; (3) Direct Preference Optimization (DPO) [9] on 6,283 curated medical preference pairs [10] using DPO-RPO loss ({beta} = 0.3, RPO = 0.1) across eight progressive training iterations (v0-v7); and (4) parallel Group Relative Policy Optimization (GRPO) [11]-based medical reward model training on Qwen3.5 combining internal rule-based reward functions with external DeepSeek signals. Comprehensive evaluations under a uniform LLM-as-Judge [12] framework with GPT-5.4 demonstrate 77.0% accuracy (mean quality score +7.18) on MedMCQA [10] and +2.59 on HLE, with pipeline progression from +6.69 (base) to +7.06 (SFT) to +7.18 (final).

Rare genetic variants provide critical insight into the mechanisms underlying complex diseases, yet their study is limited by inherent statistical challenges, particularly in the noncoding genome where functional prioritization remains difficult. Here, we introduce parmigiano, an empirical Bayesian framework that systematically integrates functional annotations into existing rare variant association tests (RVATs), jointly learning annotation weights and a variant filter threshold to enable trait-informed variant prioritization. We apply parmigiano to Alzheimer's disease (AD) whole-genome sequencing data (12,900 cases and 23,846 controls) and perform both coding and noncoding RVATs, leveraging AD-relevant cell-type-specific predictions of variant regulatory effect. Integrating parmigiano significantly increases association yield across five existing RVATs, uncovering 23 candidate AD genes -- 19 uniquely detected by our framework -- including SIGLEC10 and HUNK. Associations detected by parmigiano replicate more reliably in held-out data than those from the original RVATs and show higher overlap with known AD associations. parmigiano offers a unified approach to variant prioritization, enabling scalable, interpretable rare variant analyses across coding and noncoding regions.

Long runtime, high memory demands, and reliance on high-performance computing increasingly limit the evolutionary analysis of long phylogenomic datasets. We review a scalable framework based on phylogenomic subsampling and upsampling (PSU), in which many small subsamples of sites from a long concatenated sequence alignment are extended by upsampling prior to inference, and the resulting analyses are then aggregated to obtain stable evolutionary estimates. PSU exploits a useful distinction between the computational burden and the inferential power of statistical methods in molecular phylogenetics: computational cost is strongly influenced by the number of distinct site patterns in the concatenated alignment, whereas statistical power depends primarily on the amount of evolutionary information represented by sites and substitutions. By reducing the former while restoring the latter through upsampling, PSU can approximate many full-data analyses at substantially lower computational cost. Evidence from simulated and empirical datasets shows that PSU can accurately estimate bootstrap support values, select optimal substitution models, test evolutionary hypotheses, and infer branch lengths, divergence times, and associated uncertainty measures, while often reducing runtime and memory requirements by orders of magnitude. The same subsampling-upsampling-aggregation principle underlies all of these applications. PSU also provides distributions of inferred clade support across independent subsamples, enabling detection of concordant and conflicting phylogenetic signals that may remain hidden in conventional concatenated phylogenomic analyses. Adaptive procedures for selecting the subsample size, the number of subsamples, and the number of upsampling replicates make the framework practical across diverse datasets. We suggest that PSU is a general strategy for scalable phylogenomic inference across a broad range of statistical methods. By enabling rigorous analyses of genome-scale alignments on standard computing hardware, PSU expands access to computationally intensive evolutionary methods while reducing the environmental and infrastructural costs of big-data phylogenomics.

Genomic epidemiology workflows typically require expert curation of multiple specialized tools, extensive manual parameter tuning, and access to heterogeneous compute infrastructure. While standard generative AI models often hallucinate in complex biological domains, we introduce biomeStat: an autonomous AI agent that functions as a strict deterministic orchestrator. By automatically writing code to execute established bioinformatics tools in sandboxed environments, biomeStat dynamically provisions compute resources (CPU and GPU) and guarantees reproducibility, making it immediately useful for scientists without requiring command-line expertise. To demonstrate the platform, we performed a fully autonomous genomic epidemiology and structural analysis of 1,000 Dengue virus (DENV) genomes sampled from 16 Asian countries between 2000 and 2025. The agent seamlessly orchestrated phylogenetic reconstruction (IQ-TREE, TreeTime), Bayesian phylodynamics (BEAST2 via NVIDIA H200 GPU), selection pressure analysis (HyPhy), and structural mapping (PyMOL). The analysis was completed in under 24 hours of wall-clock time, revealing endemic stability (R_e [~]1.0) and identifying 1,869 candidate immune escape sites structurally colocalized with B-cell and T-cell epitopes. Furthermore, the agent validated 176 highly conserved drug target residues across the viral replication complex, confirming that resistance-associated positions for emerging antivirals JNJ-1802 and NITD-688 remain absolutely conserved across all four serotypes. By bridging the gap between natural language intent and deterministic computational execution, biomeStat reduces weeks of expert effort into a single-session analysis with full methodological transparency.

Ghost introgression from unsampled hominin lineages has emerged as an increasingly important component of human evolutionary history. Recent studies suggest that deeply divergent hominin lineages may have contributed ancestry either directly to modern humans or indirectly through Denisovan introgression, while inference remains difficult due to few reference genomes, weak signal, and uncertainty in reconstructing deep genealogies. Here we show analytically and through simulations that Denisovan-mediated superarchaic introgression produces predictable shifts in local coalescent depth that can be approximated by scalable summary statistics, particularly pairwise sequence divergence, suggesting that substantial information regarding deeply divergent ancestry is preserved in sequence variations without explicit reconstruction of genealogies. Leveraging this insight, we develop DEEP (\textbf{D}eep ancestry \textbf{E}stimation through \textbf{E}fficient \textbf{P}roxies), an ARG-free neural-network framework for identifying candidate regions of superarchaic ancestry. DEEP retains detectable power at low false positive rates across a broad range of demographic parameter space, remains scalable and recovers signals from small sample sizes. Applying DEEP to Oceanians, Tibetans, and Han Chinese, we identify approximately 0.4-0.6\% of genomic windows with evidence of superarchaic ancestry. Candidate regions show both substantial overlap and notable heterogeneity across populations, with repeated enrichment near the HLA locus across all populations, suggesting immune-related regions recurrently retain deeply divergent ancestry.

We propose a study protocol for routine clinical electroencephalograms (EEGs) from public hospitals, which represents a vast resource for neuroscience research. These non-invasive measures of brain function, paired with rich clinical annotations from large and diverse patient populations, are critical for developing robust artificial intelligence (AI) models and conducting population-level studies. This protocol presents a scalable methodology for curating and harmonizing extensive clinical EEG datasets, encompassing over 40,000 individual studies, to facilitate research applications. Key steps include: (i) integration of raw EEG recordings with corresponding clinical records, including neurological reports, diagnostic codes, and potentially medication data; and (ii) spatial standardization of EEG signals by mapping them to a common brain space defined by functional and anatomical landmarks. The resulting harmonized datasets enable the development of large-scale EEG foundation models, the discovery of novel EEG waveform representations, and the creation of normative "brain charts" for electrophysiological assessment across the lifespan. By enabling standardised, large-scale analyses of real-world clinical EEG data, this protocol supports data-intensive solutions for EEG applications and addresses the challenge of generalising AI models. Our approach promotes the translation of AI tools from research to diverse patient populations, advancing population neuroscience.

Interpretation of genomic variants, positions, and regions depends on reliable annotation--adding evidence such as predicted effect, conservation, population frequency, and gene-level context--yet the underlying resources are numerous, versioned, and assembly-specific. We present the Genomic Annotation Infrastructure (GAIn), a platform that generates transparent, reproducible annotations via declarative pipelines that define annotation tasks as ordered lists of components, called annotators, that produce annotation attributes using genomic resources from Genomic Resource Repositories (GRRs). We provide two public GRRs: a main repository containing more than 250 heterogeneous genomic resources, and a separate GRR-ENCODE repository containing resources derived from thousands of ENCODE (Encyclopedia of DNA Elements) project experiments. Users can use the annotation pipelines we made available, author custom annotation pipelines, and execute annotation tasks with these pipelines via GAIns web and command-line interfaces. The web interface can be used without any setup, but it relies on shared computational infrastructure and imposes limits on the size of annotation tasks. The command-line interface requires setup but supports arbitrarily large annotation tasks through simple-to-use parallelization and offers a broader set of features. For example, command-line GAIn can be extended by using custom GRRs or creating custom annotators via its plugin architecture. In addition, GAIns re-annotation feature, which updates annotations as they evolve, substantially simplifies maintaining annotations in a large genomics analysis project. GAIns resource management, explicit versioning, and pipeline abstraction provide an auditable, maintainable, and efficient foundation for modern genomic annotation across reference assemblies and use cases.

Protein function prediction traditionally relies on structured gene ontology (GO) labels or multi-label classifiers. However, these labels or classifiers cannot flexibly describe molecular function, biological process, cellular component, and free-text functional narratives in a single output. In comparison, generation-based approaches offer an intuitive paradigm for flexible free-text protein annotation, with large language models (LLMs) as a representative method for protein-text modeling. Recent efforts on utilizing LLMs for protein semantic understanding and annotation generation have adopted sequence-only encoding or sequence-text contrastive alignment paradigms, yet without explicit consideration of three-dimensional structural information. To address these limitations in current protein function prediction methods, we present ProtBLIP2-SST, a two-stage framework built on the BLIP2 model architecture that bridges protein sequence, structure, and text for open-ended protein functional caption generation. Specifically, we first integrate sequence and structure information through SaProt, a protein language model (PLM) with a structure-aware vocabulary that fuses residue tokens with Foldseek-derived 3Di structural tokens. To empower the LLM to understand protein semantics, we employ a Q-Former (a querying transformer in BLIP2) with learnable query tokens as the cross-modal projector to align protein features from the frozen SaProt encoder and text features from a frozen BiomedBERT via protein-text contrasting, protein-text matching, and protein captioning objectives. After alignment, the protein features are linearly projected and prepended to the prompt embeddings of the LLM for protein captioning fine-tuning with LoRA. Trained on 441k protein-text pairs from Swiss-Prot with corresponding structures from the AlphaFold Database, our ProtBLIP2-SST outperforms sequence-only and sequence-text alignment baselines on protein captioning metrics, with ablation studies demonstrating the effectiveness of integrating structure with sequence information for improved protein understanding. Through a unified two-stage alignment-and-generation pipeline, ProtBLIP2-SST integrates protein sequence and structural information, overcomes the rigidity of traditional GO-centric classification, generating open-ended captions that jointly describe molecular function, subcellular location, and homology context in one single output.

Precise prognostication in acute brain injury is limited by a lack of reliable biomarkers of consciousness available to clinicians at the bedside. The ABCD framework is a method of classifying resting-state clinical EEG into categories that reflect levels of thalamocortical network function. ABCD classifications in the intensive care unit (ICU) have been shown to provide diagnostic and prognostic utility for patients with severe brain injuries, but the current gold standard for ABCD classification is visual inspection of power spectra, which is labor-intensive and requires expertise in spectral analysis. Using 4,611 manually classified EEG power spectra, we developed an automated, highly accurate, and well-calibrated convolutional neural net-based classifier of EEG into ABCD categories. The classifier has performance comparable to that of the current gold standard and that outperforms an alternative method of automated spectral analysis. As proof-of-principle for clinical implementation, we apply the classifier to a continuous EEG record from a patient with acute severe traumatic brain injury in the ICU, demonstrating its ability to yield continuous ABCD classifications that capture state fluctuations with high temporal and spatial resolution. The automated ABCD classifier allows for efficient analysis of continuous EEG records, facilitating the translation of the ABCD framework to the bedside for patients with acute severe brain injuries. The ABCD classifier also creates new opportunities to efficiently analyze large EEG datasets and generate new insights into the electrophysiological properties of human consciousness.

Intrinsically disordered proteins and regions (IDPs) are ubiquitous cellular regulators. Uncovering how their transient, multivalent interactions organize and fine-tune cellular processes requires transferable methods capable of deriving dynamic conformational ensembles across diverse environments at scale. Here, we present HyRes, a physics-based, hybrid-resolution protein model with atomistic backbones and intermediate-resolution sidechains that bridges the gap between atomistic accuracy and computational efficiency. Evaluated across [~]100 IDPs, HyRes generates atomistic ensembles that match or outperform state-of-the-art all-atom force fields in reproducing experimental chain dimensions, transient tertiary contacts, and local secondary structures. Demonstrating exceptional transferability, HyRes accurately captures dynamic IDP interactions in dilute phases, condensed phases, and amyloid fibril fuzzy coats. Finally, we leverage HyRes scalability to generate disordered ensembles for [~]30,000 IDPs from the human proteome and DisProt, revealing strong correlation between residual structures and cellular function and localization. HyRes and this open-access database provide unprecedented resources for IDP biology and deep learning.

Dense electron-microscopy connectomes provide synaptic-resolution maps of neuronal structure and wiring, but learning scalable representations that integrate structure and connectivity for connectome discovery with minimal human intervention remains difficult. Here we present a self-supervised framework for structure-connectivity representation learning in dense connectomes. A hierarchical graph neural network with skeleton decomposition enables contrastive learning from finely sampled FlyWire neuronal skeletons, showing that fine skeletons preserve substantially richer identity information than coarse representations. Coordinate-free topology reduces developmental and geometric confounds, improving clustering and label-efficient inference. We then use learned structural embeddings as continuous descriptors of synaptic partners to construct structure-driven connectivity representations, improving subtype discrimination without predefined partner-type labels. Iterative multi-hop learning further reveals higher-order organization, including hemispheric connectivity lateralization and connectivity-defined subgroups. Attention analysis links these differences to specific synaptic partners. Together, these results establish a self-supervised and scalable framework for discovering neuronal identity and connectome organization in a large-scale dense connectome.

Motivation: Chimeric metagenome-assembled genomes (MAGs) that pool DNA from multiple organisms contaminate downstream analyses. Marker-gene tools such as CheckM2 miss low-level chimerism, and DNA foundation models have been proposed as a sequence-composition alternative, but whether large autoregressive models (Evo2, 7B parameters) outperform smaller contrastive models (DNABERT-S, 117M) has not been rigorously tested.