Pigeon
Pigeon Your Daily Science Feed

Latest Articles

45 articles
Active filters: Machine learning ×
N
Nature Chemical Biology · Dec 05, 2025

Accurate single-domain scaffolding of three nonoverlapping protein epitopes using deep learning

De novo protein design has seen major success in scaffolding single functional motifs; however, in nature, most proteins present multiple functional sites. Here, we describe an approach to simultaneously scaffold multiple functional sites in a single-domain protein using deep learning. We designed small single-domain immunogens, under 130 residues, that present three distinct and irregular motifs from respiratory syncytial virus. These motifs together comprise nearly half of the designed proteins; hence, the overall folds are quite unusual with little global similarity to proteins in the Protein Data Bank. Despite this, X-ray crystal structures confirmed the accuracy of presentation of each of the motifs and the multiepitope design yields improved cross-reactive titers and neutralizing response compared to a single-epitope immunogen. The successful presentation of three distinct binding surfaces in a small single-domain protein highlights the power of generative deep learning methods to solve complex protein design problems.

Machine learning Protein design Proteins Vaccines biology
N
Nature Methods · Dec 04, 2025

Deep Imputation for Skeleton data (DISK) for behavioral science

Pose estimation methods and motion capture systems have opened doors to quantitative measurements of animal kinematics. While animal behavior experiments are expensive and complex, tracking errors sometimes make large portions of the experimental data unusable. Here our deep learning method, Deep Imputation for Skeleton data (DISK), uncovers dependencies between keypoints and their dynamics to impute missing tracking data without the help of any manual annotations. We demonstrate the utility and performance of DISK on seven animal skeletons including multi-animal setups. The imputed recordings allow us to detect more episodes of motion, such as steps, and obtain more statistically robust results when comparing these episodes between experimental conditions. In addition, by learning to impute the missing content, DISK learns meaningful representations of the data capturing, for example, underlying actions. This stand-alone imputation package, available athttps://github.com/bozeklab/DISK.git/, is applicable to outputs of tracking methods (marker-based or markerless) and allows for varied types of downstream analysis.

Computational neuroscience Machine learning
N
Nature Methods · Nov 27, 2025

A comprehensive foundation model for cryo-EM image processing

Cryogenic electron microscopy (cryo-EM) has become a premier technique for determining high-resolution structures of biological macromolecules. However, its broad application is constrained by the demand for specialized expertise. Here, to address this limitation, we introduce the Cryo-EM Image Evaluation Foundation (Cryo-IEF) model, a versatile tool pre-trained on ~65 million cryo-EM particle images through unsupervised learning. Cryo-IEF performs diverse cryo-EM processing tasks, including particle classification by structure, pose-based clustering and image quality assessment. Building on this foundation, we developed CryoWizard, a fully automated single-particle cryo-EM processing pipeline enabled by fine-tuned Cryo-IEF for efficient particle quality ranking. CryoWizard resolves high-resolution structures across samples of varied properties and effectively mitigates the prevalent challenge of preferred orientation in cryo-EM.

Cryoelectron microscopy Machine learning Proteins biology
N
Nature Methods · Nov 24, 2025

Helixer: ab initio prediction of primary eukaryotic gene models combining deep learning and a hidden Markov model

The accurate identification of genes is vital for understanding biological function, yet this remains challenging across many newly sequenced or less-studied species. Here we present Helixer, an artificial intelligence-based tool for ab initio gene prediction that delivers highly accurate gene models across fungal, plant, vertebrate and invertebrate genomes. Unlike traditional methods, Helixer operates without requiring additional experimental data such as RNA sequencing, making it broadly applicable to diverse species. We show that Helixer’s pretrained models achieve accuracy on par with or exceeding current tools, producing gene annotations that closely match expert-curated references across multiple evaluation metrics. Its design enables immediate use on genomes without retraining, providing an efficient, accessible solution for genome annotation in both research and applied settings. The tool is available as an open-source software for local installation via GitHub. An online web interface is also available as well as through the Galaxy ToolShed.

Computational biology and bioinformatics Genome informatics Machine learning biology
N
Nature · Nov 19, 2025

Semantic design of functional de novo genes from a genomic language model

Generative genomic models can design increasingly complex biological systems1. However, controlling these models to generate novel sequences with desired functions remains challenging. Here, we show that Evo, a genomic language model, can leverage genomic context to perform function-guided design that accesses novel regions of sequence space. By learning semantic relationships across prokaryotic genes2, Evo enables a genomic ‘autocomplete’ in which a DNA prompt encoding genomic context for a function of interest guides the generation of novel sequences enriched for related functions, which we refer to as ‘semantic design’. We validate this approach by experimentally testing the activity of generated anti-CRISPR proteins and type II and III toxin–antitoxin systems, including de novo genes with no significant sequence similarity to natural proteins. In-context design of proteins and non-coding RNAs with Evo achieves robust activity and high experimental success rates even in the absence of structural priors, known evolutionary conservation or task-specific fine-tuning. We then use Evo to complete millions of prompts to produce SynGenome, a database containing over 120 billion base pairs of artificial intelligence-generated genomic sequences that enables semantic design across many functions. More broadly, these results demonstrate that generative genomics with biological language models can extend beyond natural sequences.

Computational models Genetic databases Machine learning Protein design
N
Nature Methods · Nov 18, 2025

ImmunoMatch learns and predicts cognate pairing of heavy and light immunoglobulin chains

The development of stable antibodies formed by compatible heavy (H) and light (L) chain pairs is crucial in both in vivo maturation of antibody-producing cells and ex vivo designs of therapeutic antibodies. We present ImmunoMatch, a machine-learning framework trained on paired H and L sequences from human B cells to identify molecular features underlying chain compatibility. ImmunoMatch distinguishes cognate from random H–L pairs and captures differences associated withκandλlight chains, reflecting B cell selection mechanisms in the bone marrow. We apply ImmunoMatch to reconstruct paired antibodies from spatial VDJ sequencing data and study the refinement of H–L pairing across B cell maturation stages in health and disease. We find further that ImmunoMatch is sensitive to sequence differences at the H–L interface. These insights provide a computational lens into the broader biological principles governing antibody assembly and stability.

Adaptive immunity Lymphocytes Machine learning Software
N
Nature Biotechnology · Nov 11, 2025

Multimodal learning enables chat-based exploration of single-cell data

Single-cell sequencing characterizes biological samples at unprecedented scale and detail, but data interpretation remains challenging. Here, we present CellWhisperer, an artificial intelligence (AI) model and software tool for chat-based interrogation of gene expression. We establish a multimodal embedding of transcriptomes and their textual annotations, using contrastive learning on 1 million RNA sequencing profiles with AI-curated descriptions. This embedding informs a large language model that answers user-provided questions about cells and genes in natural-language chats. We benchmark CellWhisperer’s performance for zero-shot prediction of cell types and other biological annotations and demonstrate its use for biological discovery in a meta-analysis of human embryonic development. We integrate a CellWhisperer chat box with the CELLxGENE browser, allowing users to interactively explore gene expression through a combined graphical and chat interface. In summary, CellWhisperer leverages large community-scale data repositories to connect transcriptomes and text, thereby enabling interactive exploration of single-cell RNA-sequencing data with natural-language chats.

Gene regulation in immune cells Machine learning Preclinical research Software Transcriptomics
N
Nature Methods · Nov 11, 2025

Universal consensus 3D segmentation of cells from 2D segmented stacks

Cell segmentation is the foundation of a wide range of microscopy-based biological studies. Deep learning has revolutionized two-dimensional (2D) cell segmentation, enabling generalized solutions across cell types and imaging modalities. This has been driven by the ease of scaling up image acquisition, annotation and computation. However, three-dimensional (3D) cell segmentation, requiring dense annotation of 2D slices, still poses substantial challenges. Manual labeling of 3D cells to train broadly applicable segmentation models is prohibitive. Even in high-contrast images annotation is ambiguous and time-consuming. Here we develop a theory and toolbox, u-Segment3D, for 2D-to-3D segmentation, compatible with any 2D method generating pixel-based instance cell masks. u-Segment3D translates and enhances 2D instance segmentations to a 3D consensus instance segmentation without training data, as demonstrated on 11 real-life datasets, comprising >70,000 cells, spanning single cells, cell aggregates and tissue. Moreover, u-Segment3D is competitive with native 3D segmentation, even exceeding when cells are crowded and have complex morphologies.

Cellular imaging Image processing Machine learning Software
N
Nature Biomedical Engineering · Nov 06, 2025

A vision–language pretrained transformer for versatile clinical respiratory disease applications

General artificial intelligence models have unique challenges in clinical practice when applied to diverse modalities and complex clinical tasks. Here we present MedMPT, a versatile, clinically oriented pretrained model tailored for respiratory healthcare, trained on 154,274 pairs of chest computed-tomography scans and radiograph reports. MedMPT adopts self-supervised learning to acquire medical insights and is capable of handling multimodal clinical data and supporting various clinical tasks aligned with clinical workflows. We evaluate the performance of MedMPT on a broad spectrum of chest-related pathological conditions, involving common medical modalities such as computed tomography images, radiology reports, laboratory tests and drug relationship graphs. MedMPT consistently outperforms the state-of-the-art multimodal pretrained models in the medical domain, achieving significant improvements in diverse clinical tasks. Extensive analysis indicates that MedMPT effectively harnesses the potential of medical data, showing both data and parameter efficiency and offering explainable insights for decision-making. MedMPT highlights the potential of multimodal pretrained models in the realm of general-purpose artificial intelligence for clinical practice.

Biomedical engineering Machine learning Respiratory tract diseases
N
Nature Medicine · Nov 05, 2025

A multimodal whole-slide foundation model for pathology

The field of computational pathology has been transformed with recent advances in foundation models that encode histopathology region-of-interests (ROIs) into versatile and transferable feature representations via self-supervised learning. However, translating these advancements to address complex clinical challenges at the patient and slide level remains constrained by limited clinical data in disease-specific cohorts, especially for rare clinical conditions. We propose Transformer-based pathology Image and Text Alignment Network (TITAN), a multimodal whole-slide foundation model pretrained using 335,645 whole-slide images via visual self-supervised learning and vision-language alignment with corresponding pathology reports and 423,122 synthetic captions generated from a multimodal generative AI copilot for pathology. Without any fine-tuning or requiring clinical labels, TITAN can extract general-purpose slide representations and generate pathology reports that generalize to resource-limited clinical scenarios such as rare disease retrieval and cancer prognosis. We evaluate TITAN on diverse clinical tasks and find that it outperforms both ROI and slide foundation models across machine learning settings, including linear probing, few-shot and zero-shot classification, rare cancer retrieval, cross-modal retrieval and pathology report generation.

Machine learning Pathology
N
Nature Biomedical Engineering · Nov 05, 2025

A pre-trained large generative model for translating single-cell transcriptomes to proteomes

Measuring protein abundance at the single-cell level can facilitate a high-resolution understanding of biological mechanisms in cellular processes and disease progression. However, current single-cell proteomic technologies face challenges such as limited coverage, constrained throughput and sensitivity, batch effects, high costs and stringent experimental operations. Inspired by the translation procedure in both natural language processing and the genetic central dogma, we propose a pre-trained, large generative model named single-cell translator (scTranslator). scTranslator can generate multi-omics data by inferring the missing single-cell proteome based on the transcriptome. Through systematic benchmarking and validation on independent datasets, we have confirmed the accuracy, stability and flexibility of scTranslator across various profiling techniques (for example, CITE-seq, spatial CITE-seq, REAP-seq, NEAT-seq), cell types (for example, monocytes, macrophages, T cells, B cells), tissues (for example, blood, lung, brain) and a wide range of disease contexts, including infectious, metabolic and oncologic conditions. Furthermore, scTranslator shows its superiority in assisting various downstream analyses and applications, including gene/protein interaction inference, perturbation prediction, cell clustering, batch correction and cell origin recognition in pan-cancer data.

Machine learning Proteome informatics
N
Nature Methods · Nov 03, 2025

Squidiff: predicting cellular development and responses to perturbations using a diffusion model

Single-cell sequencing has revolutionized our understanding of cellular heterogeneity and responses to environmental stimuli. However, mapping transcriptomic changes across diverse cell types in response to various stimuli and elucidating underlying disease mechanisms remains challenging. Here we present Squidiff, a diffusion model-based generative framework that predicts transcriptomic changes across diverse cell types in response to environmental changes. We demonstrate the robustness of Squidiff across cell differentiation, gene perturbation and drug response prediction. Through continuous denoising and semantic feature integration, Squidiff learns transient cell states and predicts high-resolution transcriptomic landscapes over time and conditions. Furthermore, we applied Squidiff to model blood vessel organoid development and cellular responses to neutron irradiation and growth factors. Our results demonstrate that Squidiff enables in silico screening of molecular landscapes and cellular state transitions, facilitating rapid hypothesis generation and providing valuable insights into the regulatory principles of cell fate decisions.

Biotechnology Computational models Machine learning Stem-cell differentiation
N
Nature Methods · Oct 30, 2025

Nicheformer: a foundation model for single-cell and spatial omics

Tissue makeup depends on the local cellular microenvironment. Spatial single-cell genomics enables scalable and unbiased interrogation of these interactions. Here we introduce Nicheformer, a transformer-based foundation model trained on both human and mouse dissociated single-cell and targeted spatial transcriptomics data. Pretrained on SpatialCorpus-110M, a curated collection of over 57 million dissociated and 53 million spatially resolved cells across 73 tissues on cellular reconstruction, Nicheformer learns cell representations that capture spatial context. It excels in linear-probing and fine-tuning scenarios for a newly designed set of downstream tasks, in particular spatial composition prediction and spatial label prediction. Critically, we show that models trained only on dissociated data fail to recover the complexity of spatial microenvironments, underscoring the need for multiscale integration. Nicheformer enables the prediction of the spatial context of dissociated cells, allowing the transfer of rich spatial information to scRNA-seq datasets. Overall, Nicheformer sets the stage for the next generation of machine-learning models in spatial single-cell analysis.

Computational models Machine learning Software Transcriptomics
N
Nature Methods · Oct 29, 2025

Annotating the genome at single-nucleotide resolution with DNA foundation models

Genome annotation models that directly analyze DNA sequences are indispensable for modern biological research, enabling rapid and accurate identification of genes and other functional elements. Current annotation tools are typically developed for specific element classes and trained from scratch using supervised learning on datasets that are often limited in size. Here we frame the genome annotation problem as multilabel semantic segmentation and introduce a methodology for fine-tuning pretrained DNA foundation models to segment 14 different genic and regulatory elements at single-nucleotide resolution. We leverage the self-supervised pretrained model Nucleotide Transformer to develop a general segmentation model, SegmentNT, capable of processing DNA sequences up to 50-kb long and that achieves state-of-the-art performance on gene annotation, splice site and regulatory elements detection. We also integrated in our framework the foundation models Enformer and Borzoi, extending the sequence context up to 500 kb and enhancing performance on regulatory elements. Finally, we show that a SegmentNT model trained on human genomic elements generalizes to different species, and a multispecies SegmentNT model achieves strong generalization across unseen species. Our approach is readily extensible to additional models, genomic elements and species.

Genomics Machine learning Software
N
Nature Medicine · Oct 27, 2025

A full life cycle biological clock based on routine clinical data and its impact in health and diseases

Aging research has primarily focused on adult aging clocks, leaving a critical gap in understanding a biological clock across the full life cycle, particularly during infancy and childhood. Here we introduce LifeClock, a biological clock model that predicts biological age across all life stages using routine electronic health records and laboratory test data. To enhance individualized predictions, we integrated virtual patient representations from 24,633,025 heterogeneous longitudinal clinical visits across 9,680,764 individuals and projected them into a latent space. Our approach leverages EHRFormer, a time-series transformer-based model, to analyze developmental and aging dynamics with high precision and develop accurate biological age clocks spanning infancy to old age. Our findings reveal distinct biological clock patterns across different life stages. The pediatric clock is strongly associated with children’s development and accurately predicts current and future risks of major pediatric diseases, including malnutrition, growth and developmental abnormalities. The adult clock is strongly associated with aging and accurately predicts current and future risks of major age-related diseases, such as diabetes, renal failure, stroke and cardiovascular diseases. This work therefore distinguishes pediatric development from adult aging, establishing a novel framework to advance precision health by leveraging routine clinical data across the entire lifespan.

Ageing Data mining Machine learning
N
Nature Methods · Oct 27, 2025

Improved reconstruction of single-cell developmental potential with CytoTRACE 2

While single-cell RNA sequencing has advanced our understanding of cell fate, identifying molecular hallmarks of potency—a cell’s ability to differentiate into other cell types—remains a challenge. Here we introduce CytoTRACE 2, an interpretable deep learning framework for predicting absolute developmental potential from single-cell RNA sequencing data. Across diverse platforms and tissues, CytoTRACE 2 outperformed previous methods in predicting developmental hierarchies, enabling detailed mapping of single-cell differentiation landscapes and expanding insights into cell potency.

Cancer genomics Machine learning Software Stem cells Transcriptomics
N
Nature Biotechnology · Oct 24, 2025

Deep-learning-based virtual screening of antibacterial compounds

The increase in multidrug-resistant bacteria underscores an urgent need for additional antibiotics. Here, we integrate small-molecule high-throughput screening with a deep-learning-based virtual screening approach to uncover new antibacterial compounds. We screen ~2 million small molecules against a sensitizedEscherichia colistrain, yielding thousands of hits. We use these data to train a deep learning model, GNEprop, to predict antibacterial activity, retrospectively validating robustness with respect to out-of-distribution generalization and activity cliff prediction. Virtual screening of over 1.4 billion synthetically accessible compounds identifies potential candidates, of which 82 exhibit antibacterial activity on the same strain, illustrating a 90-fold improved hit rate over the high-throughput screening experiment used for training. Many newly identified compounds exhibit high dissimilarity to known antibiotics, potency beyond the training bacterial strain and selectivity. Biological characterization identifies specific, validated targets, indicating promising avenues for further exploration in antibiotic discovery.

High-throughput screening Machine learning Virtual drug screening Virtual screening
N
Nature Biomedical Engineering · Oct 24, 2025

Implanted microelectrode arrays in reinnervated muscles allow separation of neural drives from transferred polyfunctional nerves

Targeted muscle reinnervation surgery reroutes residual nerve signals into spare muscles, enabling the recovery of neural information through electromyography (EMG). However, EMG signals are often overlapping, making the interpretation of limb functions complicated. Regenerative peripheral nerve interfaces surgically partition the nerve into individual fascicles that reinnervate specific muscle grafts, isolating distinct neural sources for precise control and interpretation of EMG signals. Here we combine targeted muscle reinnervation surgery of polyvalent nerves with a high-density microelectrode array implanted at a single site within a reinnervated muscle, and via mathematical source separation methods, we separate all neural signals that are redirected into a single muscle. In participants with upper-limb amputation, the deconvolution of EMG signals from four reinnervated muscles into motor unit spike trains revealed distinct clusters of motor neurons associated with diverse functional tasks. Our method enabled the extraction of multiple neural commands within a single reinnervated muscle, eliminating the need for surgical nerve division. This approach holds promises for enhancing control over prosthetic limbs and for understanding how the central nervous system encodes movement after reinnervation.

Biomedical engineering Computational neuroscience Machine learning Microarrays Motor neuron
N
Nature Methods · Oct 22, 2025

scooby: modeling multimodal genomic profiles from DNA sequence at single-cell resolution

Understanding how regulatory sequences shape gene expression across individual cells is a fundamental challenge in genomics. Joint RNA sequencing and epigenomic profiling provides opportunities to build models capturing sequence determinants across steps of gene expression. However, current models, developed primarily for bulk omics data, fail to capture the cellular heterogeneity and dynamic processes revealed by single-cell multimodal technologies. Here, we introduce scooby, a framework to model genomic profiles of single-cell RNA-sequencing coverage and single-cell assay for transposase-accessible chromatin using sequencing insertions from sequence at single-cell resolution. For this, we leverage the pretrained multiomics profile predictor Borzoi and equip it with a cell-specific decoder. Scooby recapitulates cell-specific expression levels of held-out genes and identifies regulators and their putative target genes. Moreover, scooby allows resolving single-cell effects of bulk expression quantitative trait loci and delineating their impact on chromatin accessibility and gene expression. We anticipate scooby to aid unraveling the complexities of gene regulation at the resolution of individual cells.

Computational models Machine learning Software Transcriptomics
N
Nature Biotechnology · Oct 16, 2025

Accurate somatic small variant discovery for multiple sequencing technologies with DeepSomatic

Somatic variant detection is an integral part of cancer genomics analysis. While most methods have focused on short-read sequencing, long-read technologies offer potential advantages in repeat mapping and variant phasing. We present DeepSomatic, a deep-learning method for detecting somatic small nucleotide variations and insertions and deletions from both short-read and long-read data. The method has modes for whole-genome and whole-exome sequencing and can run on tumor–normal, tumor-only and formalin-fixed paraffin-embedded samples. To train DeepSomatic and help address the dearth of publicly available training and benchmarking data for somatic variant detection, we generated and make openly available the Cancer Standards Long-read Evaluation (CASTLE) dataset of six matched tumor–normal cell line pairs whole-genome sequenced with Illumina, PacBio HiFi and Oxford Nanopore Technologies, along with benchmark variant sets. Across samples, both cell line and patient-derived, and across short-read and long-read sequencing technologies, DeepSomatic consistently outperforms existing callers.

Genome informatics Machine learning
N
Nature Methods · Oct 15, 2025

gReLU: a comprehensive framework for DNA sequence modeling and design

Deep learning models trained on DNA sequences can predict cell-type-specific regulatory activity, reveal cis-regulatory grammar, prioritize genetic variants and design synthetic DNA. However, building and interpreting these models correctly remains difficult, and models and software built by different groups are often not interoperable. Here we present gReLU, a comprehensive software framework that enables advanced sequence modeling pipelines, including data preprocessing, modeling, evaluation, interpretation, variant effect prediction and regulatory element design. gReLU advances deep-learning-based modeling and analysis of DNA sequences with comprehensive toolsets and versatile applications.

Genomics Machine learning Software
N
Nature Methods · Oct 13, 2025

Deep generative modeling of sample-level heterogeneity in single-cell genomics

Single-cell genomic studies were recently conducted on hundred of samples exhibiting complex designs. These data have tremendous potential for discovering how sample- or tissue-level phenotypes relate to cellular and molecular composition. However, current analyses are often based on simplified representations of these data by averaging information across cells. We present multi-resolution variational inference (MrVI), a deep generative model designed to realize the potential of cohort studies at the single-cell level. MrVI tackles two fundamental, intertwined problems: stratifying samples into groups and evaluating the cellular and molecular differences between groups, without requiring predefined cell states. Leveraging its single-cell perspective, MrVI detects clinically relevant stratifications of cohorts of people with COVID-19 or inflammatory bowel disease that are manifested in only certain cellular subsets, enabling new discoveries that would otherwise be overlooked. MrVI can de novo identify groups of small molecules with similar biochemical properties and evaluate their effects on cellular composition and gene expression in large-scale perturbation studies. MrVI is an open-source tool atscvi-tools.org.

Machine learning Software Statistical methods Transcriptomics
N
Nature Methods · Oct 08, 2025

Automated classification of cellular expression in multiplexed imaging data with Nimbus

Multiplexed imaging offers a powerful approach to characterize the spatial topography of tissues in both health and disease. To analyze such data, the specific combination of markers that are present in each cell must be enumerated to enable accurate phenotyping, a process that often relies on unsupervised clustering. We constructed the Pan-Multiplex (Pan-M) dataset containing 197 million distinct annotations of marker expression across 15 different cell types. We used Pan-M to create Nimbus, a deep learning model to predict marker positivity from multiplexed image data. Nimbus is a pretrained model that uses the underlying images to classify marker expression of individual cells as positive or negative across distinct cell types, from different tissues, acquired using different microscope platforms, without requiring any retraining. We demonstrate that Nimbus predictions capture the underlying staining patterns of the full diversity of markers present in Pan-M, and that Nimbus matches or exceeds the accuracy of previous approaches that must be retrained on each dataset. We then show how Nimbus predictions can be integrated with downstream clustering algorithms to robustly identify cell subtypes in image data. We have open-sourced Nimbus and Pan-M to enable community use athttps://github.com/angelolab/Nimbus-Inference.

Image processing Machine learning Software
N
Nature · Oct 08, 2025

Enzyme specificity prediction using cross attention graph neural networks

Enzymes are the molecular machines of life, and a key property that governs their function is substrate specificity—the ability of an enzyme to recognize and selectively act on particular substrates. This specificity originates from the three-dimensional (3D) structure of the enzyme active site and complicated transition state of the reaction1,2. Many enzymes can promiscuously catalyze reactions or act on substrates beyond those for which they were originally evolved1,3-5. However, millions of known enzymes still lack reliable substrate specificity information, impeding their practical applications and comprehensive understanding of the biocatalytic diversity in nature. Herein, we developed a cross-attention-empowered SE(3)-equivariant graph neural network architecture named EZSpecificity for predicting enzyme substrate specificity, which was trained on a comprehensive tailor-made database of enzyme-substrate interactions at sequence and structural levels. EZSpecificity outperformed the existing machine learning models for enzyme substrate specificity prediction, as demonstrated by both an unknown substrate and enzyme database and seven proof-of-concept protein families. Experimental validation with eight halogenases and 78 substrates revealed that EZSpecificity achieved a 91.7% accuracy in identifying the single potential reactive substrate, significantly higher than that of the state-of-the-art model ESP (58.3%). EZSpecificity represents a general machine learning model for accurate prediction of substrate specificity for enzymes related to fundamental and applied research in biology and medicine.

Biocatalysis Machine learning Protein function predictions
N
Nature Methods · Oct 03, 2025

All-at-once RNA folding with 3D motif prediction framed by evolutionary information

Structural RNAs exhibit a vast array of recurrent short three-dimensional (3D) elements found in loop regions involving non-Watson–Crick interactions that help arrange canonical double helices into tertiary structures. Here we present CaCoFold-R3D, a probabilistic grammar that predicts these RNA 3D motifs (also termed modules) jointly with RNA secondary structure over a sequence or alignment. CaCoFold-R3D uses evolutionary information present in an RNA alignment to reliably identify canonical helices (including pseudoknots) by covariation. Here we further introduce the R3D grammars, which also exploit helix covariation that constrains the positioning of the mostly noncovarying RNA 3D motifs. Our method runs predictions over an almost-exhaustive list of over 50 known RNA motifs (‘everything’). Motifs can appear in any nonhelical loop region (including three-way, four-way and higher junctions) (‘everywhere’). All structural motifs as well as the canonical helices are arranged into one single structure predicted by one single joint probabilistic grammar (‘all-at-once’). Our results demonstrate that CaCoFold-R3D is a valid alternative for predicting the all-residue interactions present in a RNA 3D structure. CaCoFold-R3D is fast and easily customizable for novel motif discovery and shows promising value both as a strong input for deep learning approaches to all-atom structure prediction as well as toward guiding RNA design as drug targets for therapeutic small molecules.

Computational models Machine learning Non-coding RNAs Riboswitches
N
Nature Methods · Oct 02, 2025

Foundation model for efficient biological discovery in single-molecule time traces

Single-molecule fluorescence microscopy (SMFM) can reveal important biological insights. However, uncovering rare but critical intermediates often demands manual inspection of time traces and iterative ad hoc approaches. To facilitate systematic and efficient discovery from SMFM time traces, we introduce META-SiM, a transformer-based foundation model pretrained on diverse SMFM analysis tasks. META-SiM rivals best-in-class algorithms on a broad range of tasks including trace classification, segmentation, idealization and stepwise photobleaching analysis. Additionally, the model produces embeddings that encapsulate detailed information about each trace, which the web-based META-SiM Projector (https://www.simol-projector.org) casts into lower-dimensional space for efficient whole-dataset visualization, labeling, comparison and sharing. Combining this Projector with the objective metric of local Shannon entropy enables rapid identification of condition-specific behaviors, even if rare or subtle. Applying META-SiM to an existing single-molecule Förster resonance energy transfer dataset, we discover a previously undetected intermediate state in pre-mRNA splicing. META-SiM removes bottlenecks, improves objectivity and both systematizes and accelerates biological discovery in single-molecule data.

Machine learning Single-molecule biophysics
N
Nature Methods · Oct 01, 2025

Fourier-based three-dimensional multistage transformer for aberration correction in multicellular specimens

High-resolution tissue imaging is often compromised by sample-induced optical aberrations that degrade resolution and contrast. Although wavefront sensor-based adaptive optics (AO) can measure these aberrations, such hardware solutions are typically complex, expensive to implement and slow when serially mapping spatially varying aberrations across large fields of view. Here we introduceAOViFT(adaptive optical vision Fourier transformer)—a machine learning-based aberration sensing framework built around a three-dimensional multistage vision transformer that operates on Fourier domain embeddings.AOViFTinfers aberrations and restores diffraction-limited performance in puncta-labeled specimens with substantially reduced computational cost, training time and memory footprint compared to conventional architectures or real-space networks. We validatedAOViFTon live gene-edited zebrafish embryos, demonstrating its ability to correct spatially varying aberrations using either a deformable mirror or postacquisition deconvolution. By eliminating the need for the guide star and wavefront sensing hardware and simplifying the experimental workflow,AOViFTlowers technical barriers for high-resolution volumetric microscopy across diverse biological samples.

Computational models Machine learning
N
Nature Biomedical Engineering · Oct 01, 2025

Peptide design through binding interface mimicry with PepMimic

Peptides offer advantages for targeted therapy, including oral bioavailability, cellular permeability and high specificity, setting them apart from conventional small molecules and biologics. Here we develop an artificial intelligence algorithm, PepMimic, to transform a known receptor or an existing antibody of a target into a short peptide binder by mimicking the binding interfaces between targets and known binders. We apply PepMimic to drug targets PD-L1, CD38, BCMA, HER2 and CD4. Surface plasmon resonance imaging results show that 8% of the peptides exhibit dissociation constant (KD) values at the 10−8M level, and 26 peptides achievingKDvalues as low as 10−9M, substantially higher than random library screening conducted under identical conditions. We apply PepMimic to target proteins lacking available binders by first using existing algorithms to design protein binders, followed by designing peptide through simulating these artificial interfaces. We extensively validate the top-ranked peptides using tail vein injections in breast, myeloma and lung tumour mouse models. Experimental results demonstrate effective membrane binding and highlight their strong potential for clinical diagnostic imaging and targeted therapeutic applications.

Machine learning Peptides
N
Nature Biomedical Engineering · Sep 30, 2025

Brain–heart–eye axis revealed by multi-organ imaging genetics and proteomics

Multi-organ research investigates interconnections among multiple human organ systems, enhancing our understanding of human aging and disease mechanisms. Here we use multi-organ imaging, individual- and summary-level genetics, and proteomics data consolidated via the MULTI Consortium to delineate a brain–heart–eye axis using brain patterns of structural covariance (PSCs), heart imaging-derived phenotypes (IDPs) and eye IDPs. We find that proteome-wide associations of the PSCs and IDPs show within-organ specificity and cross-organ interconnections. Pleiotropic effects of common single-nucleotide polymorphisms are observed across multiple organs, and key genetic parameters are estimated for single-nucleotide polymorphism-based heritability, polygenicity and selection signatures across the three organs. A gene–drug–disease network shows the potential of drug repurposing for cross-organ diseases. Co-localization and causal analyses reveal cross-organ causal relationships between PSC/IDP and chronic diseases, such as Alzheimer’s disease, heart failure and glaucoma. Finally, integrating multi-organ/omics features improves prediction for systemic disease categories and cognition compared with single-organ/omics features, providing future avenues for modelling human aging and disease.

Genetics research Heritable quantitative trait Machine learning
N
Nature Methods · Sep 25, 2025

Merging conformational landscapes in a single consensus space with FlexConsensus algorithm

Structural heterogeneity analysis in cryogenic electron microscopy is experiencing a breakthrough in estimating more accurate, richer and interpretable conformational landscapes derived from experimental data. The emergence of new methods designed to tackle the heterogeneity challenge reflects this new paradigm, enabling users to gain a better understanding of protein dynamics. However, the question of how intrinsically different heterogeneity algorithms compare remains unsolved, which is crucial for determining the reliability, stability and correctness of the estimated conformational landscapes. Here, to overcome the previous challenge, we introduce FlexConsenus: a multi-autoencoder neural network able to learn the commonalities and differences among several conformational landscapes, enabling them to be placed in a shared consensus space with enhanced reliability. The consensus space enables the measurement of reproducibility in heterogeneity estimations, allowing users to either focus their analysis on particles with a stable estimation of their structural variability or concentrate on specific particle subsets detected by only certain methods.

Image processing Machine learning Software
N
Nature Methods · Sep 25, 2025

EpiAgent: foundation model for single-cell epigenomics

Although single-cell assay for transposase-accessible chromatin using sequencing (scATAC-seq) enables the exploration of the epigenomic landscape that governs transcription at the cellular level, the complicated characteristics of the sequencing data and the broad scope of downstream tasks mean that a sophisticated and versatile computational method is urgently needed. Here we introduce EpiAgent, a foundation model pretrained on our manually curated large-scale Human-scATAC-Corpus. EpiAgent encodes chromatin accessibility patterns of cells as concise ‘cell sentences’ and captures cellular heterogeneity behind regulatory networks via bidirectional attention. Comprehensive benchmarks show that EpiAgent excels in typical downstream tasks, including unsupervised feature extraction, supervised cell type annotation and data imputation. By incorporating external embeddings, EpiAgent enables effective cellular response prediction for both out-of-sample stimulated and unseen genetic perturbations, reference data integration and query data mapping. Through in silico knockout ofcis-regulatory elements, EpiAgent demonstrates the potential to model cell state changes. EpiAgent is further extended to directly annotate cell types in a zero-shot manner.

Computational models Data integration Machine learning Software
N
Nature Aging · Sep 15, 2025

Systems Age: a single blood methylation test to quantify aging heterogeneity across 11 physiological systems

Aging occurs at different rates across individuals and physiological systems, but most epigenetic clocks provide a single age estimate, overlooking within-person variation. Here we developed systems-based DNA methylation clocks that measure aging in 11 distinct physiological systems—Heart, Lung, Kidney, Liver, Brain, Immune, Inflammatory, Blood, Musculoskeletal, Hormone and Metabolic—using data from a single blood draw. By integrating supervised and unsupervised machine learning with clinical biomarkers, functional assessments and mortality risk, we derived system-specific scores that outperformed existing global clocks in predicting relevant diseases and aging phenotypes. We also created a composite Systems Age score to capture overall multisystem aging. Clustering individuals based on these scores revealed distinct biological aging subtypes, each associated with unique patterns of health decline and disease risk. This framework enables a more granular and clinically relevant assessment of biological aging and may support personalized approaches to monitor and target system-specific aging processes.

Ageing Machine learning Prognostic markers
N
Nature Methods · Sep 15, 2025

Spatial gene expression at single-cell resolution from histology using deep learning with GHIST

The increased use of spatially resolved transcriptomics provides new biological insights into disease mechanisms. However, the high cost and complexity of these methods are barriers to broader application. Consequently, methods have been created to predict spot-based gene expression from routinely collected histology images. Recent benchmarking showed that current methodologies have limited accuracy and spatial resolution, constraining translational capacity. Here, we introduce GHIST, a deep learning-based framework that predicts spatial gene expression at single-cell resolution by leveraging subcellular spatial transcriptomics and synergistic relationships between multiple layers of biological information. We validated GHIST using public datasets and The Cancer Genome Atlas data, demonstrating its flexibility across different spatial resolutions and superior performance. Our results underscore the utility of in silico generation of single-cell spatial gene expression measurements and the capacity to enrich existing datasets with a spatially resolved omics modality, paving the way for scalable multi-omics analysis and biomarker identification.

Computational models Functional genomics Gene expression Image processing Machine learning
N
Nature Methods · Sep 15, 2025

Integrating diverse experimental information to assist protein complex structure prediction by GRASP

Protein complex structure prediction is crucial for understanding of biological activities and advancing drug development. While various experimental methods can provide structural insights into protein complexes, the knowledge obtained is often sparse or approximate. A general tool is needed to integrate limited experimental information for high-throughput and accurate prediction. Here we introduce GRASP to efficiently and flexibly incorporate diverse forms of experimental information. GRASP outperforms existing tools in handling both simulated and real-world experimental restraints including those from crosslinking, covalent labeling, chemical shift perturbation and deep mutational scanning. For example, GRASP excels at predicting antigen–antibody complex structures, even surpassing AlphaFold3 when using experimental deep mutational scanning or covalent-labeling restraints. Beyond its accuracy and flexibility in restrained structure prediction, GRASP’s ability to integrate multiple forms of restraints enables integrative modeling. We also showcase its potential in modeling protein structural interactome under near-cellular conditions using previously reported large-scale in situ crosslinking data for mitochondria.

Cryoelectron microscopy Machine learning Protein structure predictions Solution-state NMR
N
Nature Methods · Sep 15, 2025

Scaling up spatial transcriptomics for large-sized tissues: uncovering cellular-level tissue architecture beyond conventional platforms with iSCALE

Recent advances in spatial transcriptomics (ST) technologies have transformed our ability to profile gene expression while preserving crucial spatial context within tissues. However, existing ST platforms are constrained by high costs, long turnaround times, low resolution, limited gene coverage and inherently small tissue capture areas, which hinder their broad applications. Here we present iSCALE, a method that reconstructs large-scale, super-resolution gene expression landscapes and automatically annotates cellular-level tissue architecture in samples exceeding capture areas of current ST platforms. The performance of iSCALE was assessed by comprehensive evaluations involving benchmarking experiments, immunohistochemistry staining and manual annotations by pathologists. When applied to multiple sclerosis human brain samples, iSCALE uncovered lesion-associated cellular characteristics undetectable by conventional ST experiments. Our results demonstrate the utility of iSCALE in analyzing large tissues by enabling unbiased annotation, resolving cell type composition, mapping cellular microenvironments and revealing spatial features beyond the reach of standard ST analysis or routine histopathological assessment.

Gene expression analysis Machine learning RNA sequencing Transcriptomics
N
Nature Methods · Sep 11, 2025

Biophysics-based protein language models for protein engineering

Protein language models trained on evolutionary data have emerged as powerful tools for predictive problems involving protein sequence, structure and function. However, these models overlook decades of research into biophysical factors governing protein function. We propose mutational effect transfer learning (METL), a protein language model framework that unites advanced machine learning and biophysical modeling. Using the METL framework, we pretrain transformer-based neural networks on biophysical simulation data to capture fundamental relationships between protein sequence, structure and energetics. We fine-tune METL on experimental sequence–function data to harness these biophysical signals and apply them when predicting protein properties like thermostability, catalytic activity and fluorescence. METL excels in challenging protein engineering tasks like generalizing from small training sets and position extrapolation, although existing methods that train on evolutionary signals remain powerful for many types of experimental assays. We demonstrate METL’s ability to design functional green fluorescent protein variants when trained on only 64 examples, showcasing the potential of biophysics-based protein language models for protein engineering.

Machine learning Protein design
N
Nature Microbiology · Sep 09, 2025

Learning ecosystem-scale dynamics from microbiome data with MDSINE2

Although dynamical systems models are a powerful tool for analysing microbial ecosystems, challenges in learning these models from complex microbiome datasets and interpreting their outputs limit use. We introduce the Microbial Dynamical Systems Inference Engine 2 (MDSINE2), a Bayesian method that learns compact and interpretable ecosystems-scale dynamical systems models from microbiome timeseries data. Microbial dynamics are modelled as stochastic processes driven by interaction modules, or groups of microbes with similar interaction structure and responses to perturbations, and additionally, noise characteristics of data are modelled. Our open-source software package provides multiple tools for interpreting learned models, including phylogeny/taxonomy of modules, and stability, interaction topology and keystoneness. To benchmark MDSINE2, we generated microbiome timeseries data from two murine cohorts that received faecal transplants from human donors and were then subjected to dietary and antibiotic perturbations. MDSINE2 outperforms state-of-the-art methods and identifies interaction modules that provide insights into ecosystems-scale interactions in the gut microbiome.

Bayesian inference Dynamical systems Ecological networks Machine learning
N
Nature Biomedical Engineering · Sep 09, 2025

Combinatorial prediction of therapeutic perturbations using causally inspired neural networks

Phenotype-driven approaches identify disease-counteracting compounds by analysing the phenotypic signatures that distinguish diseased from healthy states. Here we introduce PDGrapher, a causally inspired graph neural network model that predicts combinatorial perturbagens (sets of therapeutic targets) capable of reversing disease phenotypes. Unlike methods that learn how perturbations alter phenotypes, PDGrapher solves the inverse problem and predicts the perturbagens needed to achieve a desired response by embedding disease cell states into networks, learning a latent representation of these states, and identifying optimal combinatorial perturbations. In experiments in nine cell lines with chemical perturbations, PDGrapher identifies effective perturbagens in more testing samples than competing methods. It also shows competitive performance on ten genetic perturbation datasets. An advantage of PDGrapher is its direct prediction, in contrast to the indirect and computationally intensive approach common in phenotype-driven models. It trains up to 25× faster than existing methods, providing a fast approach for identifying therapeutic perturbations and advancing phenotype-driven drug discovery.

Drug development Machine learning Molecular medicine Target identification
N
Nature Methods · Sep 08, 2025

Scvi-hub: an actionable repository for model-driven single-cell analysis

The growing availability of single-cell omics datasets presents new opportunities for reuse, while challenges in data transfer, normalization and integration remain a barrier. Here we present scvi-hub: a platform for efficiently sharing and accessing single-cell omics datasets using pretrained probabilistic models. It enables immediate execution of fundamental tasks like visualization, imputation, annotation and deconvolution on new query datasets using state-of-the-art methods, with massively reduced storage and compute requirements. We show that pretrained models support efficient analysis of large references, including the CZI CELLxGENE Discover Census. Scvi-hub is built within the scvi-tools open-source environment and integrated into scverse. Scvi-hub offers a scalable and user-friendly framework for accessing and contributing to a growing ecosystem of ready-to-use models and datasets, thus putting the power of atlas-level analysis at the fingertips of a broad community of users.

Machine learning Software Statistical methods Transcriptomics
N
Nature Biomedical Engineering · Sep 05, 2025

A generalist foundation model and database for open-world medical image segmentation

Vision foundation models have demonstrated vast potential in achieving generalist medical segmentation capability, providing a versatile, task-agnostic solution through a single model. However, current generalist models involve simple pre-training on various medical data containing irrelevant information, often resulting in the negative transfer phenomenon and degenerated performance. Furthermore, the practical applicability of foundation models across diverse open-world scenarios, especially in out-of-distribution (OOD) settings, has not been extensively evaluated. Here we construct a publicly accessible database, MedSegDB, based on a tree-structured hierarchy and annotated from 129 public medical segmentation repositories and 5 in-house datasets. We further propose a Generalist Medical Segmentation model (MedSegX), a vision foundation model trained with a model-agnostic Contextual Mixture of Adapter Experts (ConMoAE) for open-world segmentation. We conduct a comprehensive evaluation of MedSegX across a range of medical segmentation tasks. Experimental results indicate that MedSegX achieves state-of-the-art performance across various modalities and organ systems in in-distribution (ID) settings. In OOD and real-world clinical settings, MedSegX consistently maintains its performance in both zero-shot and data-efficient generalization, outperforming other foundation models.

Imaging Machine learning
N
Nature Methods · Aug 30, 2025

Machine learning approaches for protein structure prediction

Accurate protein structure prediction remains one of the most challenging problems in computational biology. We present a new machine learning framework that combines deep neural networks with physics-based constraints to achieve unprecedented accuracy in protein folding predictions. Our model, trained on a comprehensive dataset of protein structures, demonstrates superior performance compared to existing methods and provides new insights into protein folding mechanisms.

Structural Biology Machine Learning Protein Science
N
Nature Medicine · Aug 28, 2025

Influenza vaccine strain selection with an AI-based evolutionary and antigenicity model

Current vaccines provide limited protection against rapidly evolving viruses. For example, Centers for Disease Control and Prevention estimates show that the overall influenza vaccine effectiveness against outpatient illness in the United States averaged below 40% between 2012 and 2021. Moreover, the clinical outcomes of a vaccine can be assessed only retrospectively. Here we propose an in silico method named VaxSeer that predicts the antigenic match of vaccine candidates with circulating viruses, in the context of the viruses’ relative dominance in the future influenza season. Based on 10 years of retrospective evaluation using sequencing and antigenicity data, our approach consistently selects strains with better empirical antigenic matches to circulating viruses than annual recommendations. Finally, our predicted estimate of antigenic match exhibits a strong correlation with influenza vaccine effectiveness and reduction in disease burden, highlighting the promise of this framework to drive the vaccine selection process.

Epidemiology Machine learning Vaccines
N
Nature Methods · Aug 26, 2025

A realistic phantom dataset for benchmarking cryo-ET data annotation

Cryo-electron tomography (cryo-ET) is a powerful technique for imaging molecular complexes in their native cellular environments. However, identifying the vast majority of molecular species in cellular tomograms remains prohibitively difficult. Machine learning (ML) methods provide an opportunity to automate the annotation process, but algorithm development has been hindered by the lack of large, standardized datasets. Here we present an experimental phantom dataset with comprehensive ground-truth annotations for six molecular species to spur new algorithm development and benchmark existing tools. This annotated dataset is available on the CryoET Data Portal with infrastructure to streamline access for methods developers across fields.

Cryoelectron tomography Data acquisition Machine learning Protein databases Proteins
N
Nature Methods · Aug 26, 2025

DeepMVP: deep learning models trained on high-quality data accurately predict PTM sites and variant-induced alterations

Post-translational modifications (PTMs) are critical regulators of protein function, and their disruption is a key mechanism by which missense variants contribute to disease. Accurate PTM site prediction using deep learning can help identify PTM-altering variants, but progress has been limited by the lack of large, high-quality training datasets. Here, we introduce PTMAtlas, a curated compendium of 397,524 PTM sites generated through systematic reprocessing of 241 public mass-spectrometry datasets, and DeepMVP, a deep learning framework trained on PTMAtlas to predict PTM sites for phosphorylation, acetylation, methylation, sumoylation, ubiquitination and N-glycosylation. DeepMVP substantially outperforms existing tools across all six PTM types. Its application to predicting PTM-altering missense variants shows strong concordance with experimental results, validated using literature-curated variants and cancer proteogenomic datasets. Together, PTMAtlas and DeepMVP provide a robust platform for PTM research and a scalable framework for assessing the functional consequences of coding variants through the lens of PTMs.

Genomics Machine learning Post-translational modifications Proteome informatics Proteomics
N
Nature Medicine · Aug 20, 2025

AI-based diagnosis of acute aortic syndrome from noncontrast CT

The accurate and timely diagnosis of acute aortic syndrome (AAS) in patients presenting with acute chest pain remains a clinical challenge. Aortic computed tomography (CT) angiography is the imaging protocol of choice in patients with suspected AAS. However, due to economic and workflow constraints in China, the majority of suspected patients initially undergo noncontrast CT as the initial imaging testing, and CT angiography is reserved for those at higher risk. Although noncontrast CT can reveal specific signs indicative of AAS, its diagnostic efficacy when used alone has not been well characterized. Here we present an artificial intelligence-based warning system, iAorta, using noncontrast CT for AAS identification in China, which demonstrates remarkably high accuracy and provides clinicians with interpretable warnings. iAorta was evaluated through a comprehensive step-wise study. In the multicenter retrospective study (n= 20,750), iAorta achieved a mean area under the receiver operating curve of 0.958 (95% confidence interval 0.950–0.967). In the large-scale real-world study (n= 137,525), iAorta demonstrated consistently high performance across various noncontrast CT protocols, achieving a sensitivity of 0.913–0.942 and a specificity of 0.991–0.993. In the prospective comparative study (n= 13,846), iAorta demonstrated the capability to significantly shorten the time to correct diagnostic pathway for patients with initial false suspicion from an average of 219.7 (115–325) min to 61.6 (43–89) min. Furthermore, for the prospective pilot deployment that we conducted, iAorta correctly identified 21 out of 22 patients with AAS among 15,584 consecutive patients presenting with acute chest pain and under noncontrast CT protocol in the emergency department. For these 21 AAS-positive patients, the average time to diagnosis was 102.1 (75–133) min. Finally, iAorta may help prevent delayed or missed diagnoses of AAS in settings where noncontrast CT remains the only feasible initial imaging modality—such as in resource-limited regions or in patients who cannot receive, or did not receive, intravenous contrast.

Aortic diseases Computed tomography Machine learning