AI

A comprehensive review from Cell envisioning 'AI scientists' as systems capable of skeptical learning and reasoning that empower biomedical research through collaborative agents integrating AI models and biomedical tools with expert human oversight. Covers the landscape of AI agents for biomedical discovery, including virtual laboratories, autonomous experimentation, and human-AI collaboration frameworks.

We envision 'AI scientists' as systems capable of skeptical learning and reasoning that empower biomedical research through collaborative agents that integrate AI models and biomedical tools with expert human oversight. This review covers the landscape of AI agents for biomedical discovery, including virtual laboratories, autonomous experimentation, and human-AI collaboration frameworks.

A comprehensive review from Nature Reviews Cancer examining how AI agents (beyond traditional ML classifiers) are transforming cancer research and oncology. Covers LLM-powered agents for clinical decision support, drug discovery, treatment planning, and patient care, including agentic systems capable of logical reasoning, multi-step planning, and tool use in oncology contexts.

Since 2022, artificial intelligence (AI) methods have progressed far beyond their established capabilities of data classification and prediction. Large language models (LLMs) can perform logical reasoning, multi-step planning, and tool use, enabling a new paradigm of AI agents for cancer research and oncology. This review examines how AI agents are transforming clinical decision support, drug discovery, treatment planning, and patient care in oncology.

The AI Scientist is the first fully autonomous AI system to generate a paper that passed peer review (ICLR 2025 ICBINB workshop). It creates research ideas, writes code, runs experiments, analyses data, writes manuscripts, and performs peer review — all end-to-end. Template-free mode uses agentic tree search for open-ended scientific exploration. An automated reviewer achieves balanced accuracy comparable to human reviewers (69%). Paper quality scales with foundation model capability and test-time compute.

The automation of science is a long-standing ambition in artificial intelligence research. Although the community has made substantial progress in automating individual components of the scientific process, a system that autonomously navigates the entire research life cycle from conception to publication has remained out of reach. Here we present a pipeline for automating the entire scientific process end to end. We present The AI Scientist, which creates research ideas, writes code, runs experiments, plots and analyses data, writes the entire scientific manuscript, and performs its own peer review. Its ideas, execution and presentation are of sufficient quality that the manuscript generated by this AI system passed the first round of peer review for a workshop of a top-tier machine learning conference.

AI-MARRVEL (AIM) is a knowledge-driven AI system for diagnosing Mendelian disorders that uses a random-forest machine-learning classifier trained on over 3.5 million variants from thousands of diagnosed cases. It incorporates expert-engineered features to recapitulate the complex decision-making processes in molecular diagnosis. AIM doubled the rate of accurate genetic diagnosis across three independent real-world cohorts compared to benchmarked methods. Its confidence metric achieved 98% precision and identified 57% of diagnosable cases from 871 unsolved cases. AIM also demonstrated potential for novel disease gene discovery.

Diagnosing genetic disorders requires extensive manual curation and interpretation of candidate variants, a labor-intensive task even for trained geneticists. Although artificial intelligence (AI) shows promise in aiding these diagnoses, existing AI tools have only achieved moderate success for primary diagnosis. AI-MARRVEL (AIM) uses a random-forest machine-learning classifier trained on over 3.5 million variants from thousands of diagnosed cases. AIM additionally incorporates expert-engineered features into training to recapitulate the intricate decision-making processes in molecular diagnosis. AIM improved the rate of accurate genetic diagnosis, doubling the number of solved cases as compared with benchmarked methods, across three distinct real-world cohorts. AIM achieved a precision rate of 98% and identified 57% of diagnosable cases out of a collection of 871 cases. AIM demonstrated potential for novel disease gene discovery by correctly predicting two newly reported disease genes from the Undiagnosed Diseases Network.

AlphaFold2 by DeepMind achieved atomic-level accuracy in protein structure prediction, solving a 50-year grand challenge in biology. Its deep learning architecture (Evoformer + structure module) predicts protein 3D structures from amino acid sequences with accuracy rivaling experimental methods. Transformed drug discovery by enabling structure-based design for previously intractable targets. 44,000+ citations, awarded the 2024 Nobel Prize in Chemistry.

Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort, the structures of around 100,000 unique proteins have been determined, but this represents a small fraction of the billions of known protein sequences. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even where no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods.

Unified deep learning model that predicts molecular phenotypes from DNA sequence—including gene expression, chromatin accessibility, histone marks, TF binding, splicing, and contact maps—at single-nucleotide resolution for variant effect interpretation.

variant effect prediction, regulatory genomics, deep learning, single-nucleotide resolution, DNA, chromatin, gene expression, TF binding

Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code. Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance1-5. We present AlphaGenome, a unified DNA sequence model, which takes as input 1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities. The modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction. The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene6. To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence.

AutoResearchClaw is an open-source 23-stage autonomous research pipeline from UNC Chapel Hill that turns a research idea into a conference-ready LaTeX paper. Features: multi-agent debate for hypothesis generation, self-healing executor with Pivot/Refine decision loop, verifiable result reporting preventing hallucinations, human-in-the-loop with 7 intervention modes, and cross-run evolution. Outperforms AI Scientist v2 by 54.7% on ARC-Bench. MIT licensed.

Automating scientific discovery requires more than generating papers from ideas. Real research is iterative: hypotheses are challenged from multiple perspectives, experiments fail and inform the next attempt, and lessons accumulate across cycles. We present AutoResearchClaw, a multi-agent autonomous research pipeline built on five mechanisms: structured multi-agent debate for hypothesis generation and result analysis, a self-healing executor with a Pivot/Refine decision loop that transforms failures into information, verifiable result reporting that prevents fabricated numbers and hallucinated citations, human-in-the-loop collaboration with seven intervention modes, and cross-run evolution that converts past mistakes into future safeguards. On ARC-Bench, AutoResearchClaw outperforms AI Scientist v2 by 54.7%.

BioMedAgent is a self-evolving LLM multi-agent framework that learns to use diverse bioinformatics tools and chain them into executable workflows through interactive exploration and memory retrieval algorithms. It allows biomedical users to initiate tasks using natural language, without requiring computational expertise. Evaluated on BioMed-AQA benchmark (327 biomedical data tasks), BioMedAgent achieved a 77% success rate, surpassing other LLM agents, and generalized robustly to the external BixBench dataset. Beyond benchmarks, it autonomously performs cross-omics analysis, machine-learning modelling and pathology image segmentation.

Artificial intelligence agents are emerging as powerful applications of large language models (LLMs), automating complex tasks and enabling scientific data exploration. However, their use in biomedical data analysis remains limited by the difficulty of handling specialized tools and multistep reasoning. Here we introduce BioMedAgent, a self-evolving LLM multi-agent framework, which learns to use diverse bioinformatics tools and chain them into executable workflows through interactive exploration and memory retrieval algorithms. It allows biomedical users to initiate tasks using natural language, without requiring computational expertise. Evaluated on our newly released BioMed-AQA benchmark comprising 327 biomedical data tasks, BioMedAgent achieved a 77% success rate, surpassing other LLM agents, and generalized robustly to the external BixBench dataset. Beyond benchmarks, it autonomously performs cross-omics analysis, machine-learning modelling and pathology image segmentation, highlighting its potential to advance biomedical research and extend to other scientific domains requiring complex tool integration and multistep reasoning.

Biomni is a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields. It employs an action discovery agent to mine tools, databases, and protocols from tens of thousands of publications across 25 biomedical domains, creating the first unified agentic environment (Biomni-E1). Its generalist agentic architecture (Biomni-A1) integrates LLM reasoning with retrieval-augmented planning and code-based execution, dynamically composing complex workflows without predefined templates. Systematic benchmarking demonstrates strong zero-shot generalization across heterogeneous tasks including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning.

Biomedical research underpins progress in our understanding of human health and disease, drug discovery, and clinical care. However, with the growth of complex lab experiments, large datasets, many analytical tools, and expansive literature, biomedical research is increasingly constrained by repetitive and fragmented workflows that slow discovery and limit innovation. Here, we introduce Biomni, a general-purpose biomedical AI agent designed to autonomously execute a wide spectrum of research tasks across diverse biomedical subfields. To systematically map the biomedical action space, Biomni first employs an action discovery agent to create the first unified agentic environment, mining essential tools, databases, and protocols from tens of thousands of publications across 25 biomedical domains. Built on this foundation, Biomni features a generalist agentic architecture that integrates LLM reasoning with retrieval-augmented planning and code-based execution, enabling it to dynamically compose and carry out complex biomedical workflows entirely without relying on predefined templates or rigid task flows. Systematic benchmarking demonstrates that Biomni achieves strong generalization across heterogeneous biomedical tasks including causal gene prioritization, drug repurposing, rare disease diagnosis, microbiome analysis, and molecular cloning without any task-specific prompt tuning. Real-world case studies further showcase Biomni's ability to interpret complex, multi-modal biomedical datasets and autonomously generate experimentally testable protocols.

BioReason is a DNA-LLM model that incentivizes multimodal biological reasoning by integrating DNA sequence representations with biological knowledge. Published at NeurIPS 2025, it achieves state-of-the-art on several genomic reasoning benchmarks.

DNA-LLM, multimodal reasoning, biological reasoning, foundation model, NeurIPS 2025, long-context, interpretability

Unlocking deep, interpretable biological reasoning from complex genomic data is a major AI challenge hindering scientific discovery. Current DNA foundation models excel at encoding sequence information but lack the ability to perform explicit reasoning over biological concepts. We present BioReason, a DNA-LLM model that integrates a genome foundation model with a large language model to enable multimodal biological reasoning. BioReason is trained to reason over DNA sequences and biological text jointly, enabling interpretable predictions and natural language explanations of genomic functions. The model achieves state-of-the-art performance across multiple genomic reasoning tasks.

BixBench by FutureHouse and ScienceMachine is a benchmark designed to evaluate AI agents on real-world bioinformatics tasks. Features 61 real-world analytical scenarios with 205 associated questions, supporting both open-answer and multiple-choice evaluation. Tests agents on data analysis, insight generation, and result interpretation in bioinformatics. Current frontier models achieve only ~21% accuracy, highlighting significant room for improvement.

Artificial intelligence (AI) is changing scientific research at a rapid pace and is beginning to enable the automation of complex analytical tasks. One of the most promising fields for AI-driven automation is bioinformatics, where data-focused research lends itself to purely computational analysis. We introduce BixBench, a benchmark designed to evaluate AI agents on real-world bioinformatics tasks. BixBench challenges AI models with open-ended analytical research scenarios, requiring them to analyze data, generate insights, and interpret results autonomously. The benchmark comprises over 50 real-world scenarios with nearly 300 associated open-answer questions.

Borzoi is a deep learning model from Calico that predicts cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence. It scores variant effects across transcription, splicing, and polyadenylation, and extracts cis-regulatory motifs driving RNA expression. Published in Nature Genetics.

RNA-seq, gene regulation, splicing, polyadenylation, variant effect, CNN, deep learning, eQTL

Sequence-based machine-learning models trained on genomics data improve genetic variant interpretation by providing functional predictions describing their impact on the cis-regulatory code. However, current tools do not predict RNA-seq expression profiles because of modeling challenges. Here, we introduce Borzoi, a model that learns to predict cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence. Using statistics derived from Borzoi's predicted coverage, we isolate and accurately score DNA variant effects across multiple layers of regulation, including transcription, splicing and polyadenylation. Evaluated on quantitative trait loci, Borzoi is competitive with and often outperforms state-of-the-art models trained on individual regulatory functions. By applying attribution methods to the derived statistics, we extract cis-regulatory motifs driving RNA expression and post-transcriptional regulation in normal tissues. The wide availability of RNA-seq data across species, conditions and assays profiling specific aspects of regulation emphasizes the potential of this approach to decipher the mapping from DNA sequence to regulatory function.

BRFSS (Behavioral Risk Factor Surveillance System) is the nation's premier system of health-related telephone surveys conducted by the CDC and state health departments. Established in 1984, it collects state-level data on US adult residents regarding health-related risk behaviors, chronic health conditions, and use of preventive services. Conducts over 400,000 interviews annually across all 50 states, DC, and US territories. Topics include: high blood pressure, cholesterol, diabetes, asthma, BMI/obesity, smoking/tobacco use, alcohol consumption, physical activity, diet, cancer screening (breast, cervical, colorectal), immunizations, HIV/AIDS, mental health, and healthcare access. Data includes demographic variables (age, sex, race/ethnicity, income, education) and survey weights for population-representative estimates. Completely free, publicly available on the CDC website — no registration or data use agreement required. Available in SAS, ASCII, and CSV formats. Often used alongside NHANES for complementary population health analyses (BRFSS = larger sample, broader coverage; NHANES = deeper phenotyping with physical exams and lab measurements).

BRFSS, CDC, behavioral risk, telephone survey, public health, chronic disease, risk factors, US surveillance

Caduceus is the first family of reverse-complement (RC) equivariant bi-directional long-range DNA language models, built on the Mamba state space model backbone with BiMamba and MambaDNA blocks. Published at ICML 2024, it excels at long-range variant effect prediction.

Mamba, RC equivariance, bi-directional, long-range DNA, variant effect prediction, state space model, ICML 2024

Large-scale sequence modeling has sparked rapid advances that now extend into biology and genomics. However, modeling genomic sequences introduces challenges such as the need to model long-range token interactions, the effects of upstream and downstream regions of the genome, and the reverse complementarity (RC) of DNA. Here, we propose an architecture motivated by these challenges that builds off the long-range Mamba block, and extends it to a BiMamba component that supports bi-directionality, and to a MambaDNA block that additionally supports RC equivariance. We use MambaDNA as the basis of Caduceus, the first family of RC equivariant bi-directional long-range DNA language models, and we introduce pre-training and fine-tuning strategies that yield Caduceus DNA foundation models. Caduceus outperforms previous long-range models on downstream benchmarks; on a challenging long-range variant effect prediction task, Caduceus exceeds the performance of prior models.

A Perspective from Nature Genetics delineating the application of causal machine learning to single-cell genomics. Discusses causal models, challenges in inferring causative roles of genes from single-cell omics data combined with perturbation screens, and the potential for integrating causal ML with GWAS to understand disease mechanisms at single-cell resolution.

Advances in single-cell '-omics' allow unprecedented insights into the transcriptional profiles of individual cells and, when combined with large-scale perturbation screens, enable measuring of the effect of targeted perturbations on the whole transcriptome. In this Perspective, we delineate the application of causal machine learning to single-cell genomics and its associated challenges, presenting the causal model most commonly applied to single-cell biology.

OpenAI’s consumer and API chat lineup (GPT family), including multimodal and agent-style capabilities.

ChemCrow is an LLM-based agent system that augments large language models with 13 expert-designed chemistry tools, enabling autonomous chemical reasoning and experiment design. It integrates tools for organic synthesis, drug discovery, and materials design, allowing the agent to plan syntheses, analyze chemical properties, and execute complex chemical tasks through natural language interaction.

Large language models (LLMs) have shown strong performance in tasks across domains but struggle with chemistry-related problems. These models also lack access to external knowledge sources, limiting their usefulness in scientific applications. Here we introduce ChemCrow, an LLM-based agent that integrates 13 expert-designed chemistry tools, enabling autonomous chemical reasoning and experiment design. ChemCrow successfully plans syntheses, analyzes chemical properties, and executes complex chemical tasks through natural language interaction.

CHIEF (Clinical Histopathology Imaging Evaluation Foundation) is a general-purpose weakly supervised machine learning framework from Harvard Medical School. Trained on 60,530 WSIs spanning 19 anatomical sites (44TB data), CHIEF leverages two complementary pretraining methods: unsupervised pretraining for tile-level feature identification and weakly supervised pretraining for whole-slide pattern recognition. Validated on 19,491 WSIs from 32 independent slide sets across 24 hospitals internationally. Outperforms SOTA deep learning methods by up to 36.1%, demonstrating strong generalization across diverse populations and slide preparation methods.

Histopathology image evaluation is indispensable for cancer diagnoses and subtype classification. Standard AI methods for histopathology image analyses have focused on optimizing specialized models for each diagnostic task, often with limited generalizability. To address this challenge, we devised CHIEF, a general-purpose weakly supervised machine learning framework to extract pathology imaging features for systematic cancer evaluation. CHIEF leverages two complementary pretraining methods to extract diverse pathology representations: unsupervised pretraining for tile-level feature identification and weakly supervised pretraining for whole-slide pattern recognition. Developed using 60,530 whole-slide images spanning 19 anatomical sites, CHIEF outperformed SOTA deep learning methods by up to 36.1%, showing its ability to address domain shifts observed in samples from diverse populations.

Anthropic’s Claude family of assistants and API models, emphasizing long context, safety, and agentic workflows.

Anthropic’s terminal-based coding agent for large-scale refactors, test runs, and repo navigation using Claude models.

Autonomous coding agent for VS Code (and compatible editors) that plans, edits files, runs commands, and uses browser tools with user approval.

Co-Scientist is a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation. It aims to help scientists discover new original knowledge by formulating demonstrably novel research hypotheses for experimental validation, conditioned on research objectives and prior scientific evidence.

Scientific discovery is driven by scientists generating novel hypotheses for complex problems that undergo rigorous experimental validation. To augment this process, we introduce Co-Scientist, a multi-agent AI system built on Gemini for structured scientific thinking and hypothesis generation. Co-Scientist aims to help scientists discover new original knowledge. Conditioned on their research objectives and prior scientific evidence, it formulates demonstrably novel research hypotheses for experimental validation.

CodeScientist is an autonomous scientific discovery system from Ai2 (Allen Institute for AI) that frames ideation and experiment construction as genetic search over research articles and code blocks. It conducted hundreds of automated experiments on agents and virtual environments, returning 19 discoveries, 6 of which were judged minimally sound and incrementally novel after multi-faceted evaluation (conference review, code review, replication). Discoveries span new tasks, agents, metrics, and data. Published at ACL 2025 Findings.

Despite the surge of interest in autonomous scientific discovery (ASD) of software artifacts, current ASD systems face two key limitations: they largely explore variants of existing codebases, and they produce large volumes of research artifacts typically evaluated using conference-style paper review with limited evaluation of code. In this work we introduce CodeScientist, a novel ASD system that frames ideation and experiment construction as a form of genetic search jointly over combinations of research articles and codeblocks defining common actions in a domain. We use this paradigm to conduct hundreds of automated experiments, with the system returning 19 discoveries, 6 of which were judged as being both at least minimally sound and incrementally novel after multi-faceted evaluation.

OpenAI’s cloud and local coding agent (CLI and integrations) for building, reviewing, and shipping code with GPT-based models.

CONCH (CONtrastive learning from Captions for Histopathology) is a vision-language foundation model from Mahmood Lab (Harvard/BWH). Pretrained on 1.17M histopathology image-text pairs from diverse sources (PubMed, educational resources, textbooks). Evaluated across 14 clinically relevant tasks including zero-shot cancer classification, text-to-image retrieval, image-to-text retrieval, caption generation, and tissue segmentation. Outperforms standard models including CLIP and PLIP. CONCH also works on non-H&E stains (IHC, special stains), demonstrating broad applicability. Available as an open-source model for academic use.

We introduce CONCH, a visual-language foundation model developed using diverse sources of histopathology images and text. Trained on 1.17 million pathology image-text pairs, CONCH achieves state-of-the-art performance across 14 clinically relevant tasks, including zero-shot cancer classification, text-to-image and image-to-text retrieval, caption generation, and tissue segmentation. CONCH outperforms standard models like CLIP and PLIP, and generalizes to non-H&E stains including immunohistochemistry and special stains, demonstrating its versatility as a foundation model for computational pathology.

CONCH (CONtrastive learning from Captions for Histopathology) is a vision-language foundation model pretrained on 1.17 million histopathology image-caption pairs. It achieves state-of-the-art performance across 14 diverse benchmarks including histology image classification, segmentation, captioning, text-to-image and image-to-text retrieval. As a multimodal model bridging visual pathology data with biomedical text, CONCH enables zero-shot transfer and minimal fine-tuning for diverse computational pathology tasks.

The accelerated adoption of digital pathology and advances in deep learning have enabled the development of robust models for various pathology tasks. However, model training is often difficult due to label scarcity. Additionally, most models in histopathology leverage only image data. We introduce CONCH, a visual-language foundation model developed using diverse sources of histopathology images, biomedical text, and over 1.17 million image-caption pairs via task-agnostic pretraining. Evaluated on 14 diverse benchmarks, CONCH achieves state-of-the-art performance on histology image classification, segmentation, captioning, and cross-modal retrieval.

CRISPR-GPT is an LLM-based agent system for automating gene-editing experiments. It leverages large language models to guide researchers through the entire CRISPR experiment workflow, including guide RNA design, off-target prediction, experimental protocol generation, and result interpretation, making gene-editing more accessible to non-expert researchers.

Performing effective gene-editing experiments requires a deep understanding of both the CRISPR technology and the biological system involved. Meanwhile, despite their versatility and promise, large language models have not been fully leveraged for automated experimental design in molecular biology. Here we present CRISPR-GPT, an LLM-based agent system for automating gene-editing experiments across the entire CRISPR workflow, including guide RNA design, off-target prediction, experimental protocol generation, and result interpretation.

AI-native code editor (VS Code fork) with inline chat, agent mode, and codebase-wide context for editing and refactoring.

Pioneering deep learning framework for EHR data by Google Research, using Fast Healthcare Interoperability Resources (FHIR) format to represent patients' raw EHR records. Trained on 216,221 patients across 2 US medical centers with 46.8 billion data points including clinical notes. Achieved AUROC 0.93-0.94 for in-hospital mortality, 0.75-0.76 for 30-day readmission, and 0.90 for final discharge diagnoses, outperforming traditional clinical predictive models. 2,800+ citations, widely considered the landmark paper for deep learning on EHR data.

Predictive modeling with electronic health record (EHR) data is anticipated to drive personalized medicine and improve healthcare quality. We propose a representation of patients' entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format. We demonstrate that deep learning methods using this representation are capable of accurately predicting multiple medical events from multiple centers without site-specific data harmonization. Deep learning models achieved high accuracy for predicting in-hospital mortality (AUROC 0.93-0.94), 30-day unplanned readmission (AUROC 0.75-0.76), prolonged length of stay (AUROC 0.85-0.86), and all of a patient's final discharge diagnoses (AUROC 0.90).

DeepNull is a method that identifies and adjusts for non-linear and interactive covariate effects in GWAS using a deep neural network. It maintains tight control of type I error while increasing statistical power by up to 20% in the presence of non-linear covariate effects. Published in Nature Communications.

Genome-wide association studies (GWASs) examine the association between genotype and phenotype while adjusting for a set of covariates. Although the covariates may have non-linear or interactive effects, due to the challenge of specifying the model, GWAS often neglect such terms. Here we introduce DeepNull, a method that identifies and adjusts for non-linear and interactive covariate effects using a deep neural network, maintaining tight control of the type I error while increasing statistical power by up to 20%.

DeepRare is a multi-agent system for rare disease differential diagnosis that integrates large language models with structured medical knowledge to provide traceable reasoning. It aims to reduce the diagnostic odyssey for rare disease patients by leveraging AI agents to analyze clinical phenotypes, genomic data, and medical literature in a transparent, interpretable manner.

Rare diseases affect more than 300 million people worldwide, yet timely and accurate diagnosis remains an urgent challenge. Patients often endure a prolonged 'diagnostic odyssey' exceeding 5 years, marked by repeated referrals, misdiagnoses and unnecessary interventions, leading to delayed treatment and substantial emotional and economic burden. Here we present DeepRare, a multi-agent system for rare disease differential diagnosis decision support with traceable reasoning.

DeepSEA is a foundational deep learning model that predicts the chromatin effects of noncoding variants directly from DNA sequence, including DNase I sensitivity, histone mark profiles, and transcription factor binding. Published in Nature Methods, it was one of the first deep learning approaches for noncoding variant interpretation.

noncoding variant, chromatin, epigenetics, deep learning, DNase I, histone marks, TF binding, regulatory effect

Noncoding variants are of tremendous biological importance, and their functional interpretation is a critical challenge in genomics. Here we introduce DeepSEA, a deep learning-based sequence model that directly predicts the chromatin effects of sequence alterations at single-nucleotide sensitivity. DeepSEA captures regulatory sequence context and learns a wide range of regulatory features including DNase I sensitivity, histone mark profiles, and transcription factor binding. The model achieves state-of-the-art accuracy for predicting the functional consequences of noncoding variants and can be applied to prioritize disease-associated variants from large-scale sequencing studies.

DeepSeek’s R1 / V3 and related open-weights and API models focused on reasoning, coding, and efficiency.

Denario is an AI multi-agent system from the Flatiron Institute (Simons Foundation) designed to serve as a scientific research assistant across disciplines. It can generate ideas, check literature for novelty, develop research plans, write and execute code, make plots, and draft and review scientific papers. Demonstrated across 11 AI-generated paper drafts in astrophysics, biology, biophysics, chemistry, material science, medicine, neuroscience and more. Excels at combining ideas across disciplines (e.g., quantum physics + ML applied to astrophysics).

We present Denario, an AI multi-agent system designed to serve as a scientific research assistant. Denario can perform many different tasks, such as generating ideas, checking the literature, developing research plans, writing and executing code, making plots, and drafting and reviewing a scientific paper. In this work, we describe in detail Denario and its modules, and illustrate its capabilities by presenting multiple AI-generated papers generated by it in many different scientific disciplines such as astrophysics, biology, biophysics, biomedical informatics, chemistry, material science, mathematical physics, medicine, neuroscience and planetary science. Denario also excels at combining ideas from different disciplines.

A comprehensive survey of deep learning approaches for polygenic risk scores (PRS). Reviews how neural networks can model non-linear relationships between genetic variants and disease risk, going beyond traditional linear PRS methods, and assesses their performance across different traits and architectures. Published in Briefings in Bioinformatics.

Polygenic risk scores (PRS) combine the effects of multiple genetic variants to predict an individual's genetic predisposition to a disease. PRS typically rely on linear models, which assume that all genetic variants act independently. There is growing interest in applying deep learning neural networks to model PRS given their ability to model non-linear relationships. We conducted a survey of the literature to investigate how neural networks are being applied to PRS.

Applied a pretrained CNN (transferred from natural image recognition, e.g. ResNet/Inception-like architecture) to 4.6 million cardiac MRI images from UK Biobank, trained on only 116 manually annotated samples to regress ascending and descending thoracic aorta dimensions. GWAS identified 82 ascending and 47 descending aorta loci. Demonstrates transfer learning from natural images to medical imaging for rapid biobank-scale phenotyping.

deep learning, CNN, ImageNet transfer learning, cardiac MRI, thoracic aorta, image regression, UK Biobank

Enlargement or aneurysm of the aorta predisposes to dissection, an important cause of sudden death. We trained a deep learning model to evaluate the dimensions of the ascending and descending thoracic aorta in 4.6 million cardiac magnetic resonance images from the UK Biobank. We then conducted genome-wide association studies in 39,688 individuals, identifying 82 loci associated with ascending and 47 with descending thoracic aortic diameter. Transcriptome-wide analyses, rare-variant burden tests and human aortic single nucleus RNA sequencing prioritized genes including FBN1 and MFAP5.

First comprehensive, unbiased benchmark of five DNA foundation models (DNABERT-2, Nucleotide Transformer V2, HyenaDNA, Caduceus-Ph, GROVER) across 57 datasets spanning sequence classification, gene expression prediction, variant effect quantification, and TAD recognition using zero-shot embeddings. Key finding: mean token embedding pooling consistently outperforms other strategies. Model choice should align with task — Caduceus-Ph excels at TFBS, NT-v2 at pathogenic variants, HyenaDNA scales to long sequences. Specialized models (Enformer, Sei) still outperform general DNA models on QTL prediction.

The rapid evolution of DNA foundation models promises to revolutionize genomics, yet comprehensive evaluations are lacking. Here, we present a comprehensive, unbiased benchmark of five models (DNABERT-2, Nucleotide Transformer V2, HyenaDNA, Caduceus-Ph, and GROVER) across diverse genomic and genetic tasks including sequence classification, gene expression prediction, variant effect quantification, and TAD region recognition, using zero-shot embeddings. Our analysis reveals that mean token embedding consistently and significantly improves sequence classification performance. Model performance varies among tasks and datasets; while general purpose DNA foundation models showed competitive performance in pathogenic variant identification, they were less effective in predicting gene expression and identifying putative causal QTLs compared to specialized models.

DNABERT is a pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in the genome. Uses k-mer tokenization and masked language modeling to learn global and transferrable understanding of genomic DNA sequences. After fine-tuning, achieves SOTA on promoters, splice sites and TF binding sites prediction.

BERT, Transformer, pre-trained, promoter prediction, splice site, transcription factor binding site, k-mer

Deciphering the language of non-coding DNA is one of the fundamental problems in genome research. Gene regulatory code is highly complex due to the existence of polysemy and distant semantic relationship, which previous informatics methods often fail to capture especially in data-scarce scenarios. To address this challenge, we developed a novel pre-trained bidirectional encoder representation, named DNABERT, to capture global and transferrable understanding of genomic DNA sequences based on up and downstream nucleotide contexts. We compared DNABERT to the most widely used programs for genome-wide regulatory elements prediction and demonstrate its ease of use, accuracy and efficiency. We show that the single pre-trained transformers model can simultaneously achieve state-of-the-art performance on prediction of promoters, splice sites and transcription factor binding sites, after easy fine-tuning using small task-specific labeled data. Further, DNABERT enables direct visualization of nucleotide importance for model prediction, contributing to better interpretability. DNABERT is publicly available at https://github.com/jerryji1993/DNABERT.

DNABERT-2 is an efficient foundation model for multi-species genome understanding, improving upon DNABERT with Byte-Pair Encoding (BPE) tokenization, attention head pruning, and improved training techniques. Published at ICLR 2024, it achieves SOTA across 28 genome prediction tasks while being significantly more efficient than previous models.

BERT, Transformer, BPE tokenization, multi-species genome, foundation model, ICLR 2024

Deciphering the language of non-coding DNA is a critical challenge in genome research. Existing approaches often rely on k-mer based tokenization and single-species datasets, limiting their effectiveness for multi-species genome understanding. Here, we introduce DNABERT-2, a foundation model that leverages Byte-Pair Encoding (BPE) tokenization to capture meaningful DNA units across species, combined with attention head pruning and other optimizations for efficient training and inference. We also present a comprehensive multi-species genome benchmark (Genome Understanding Evaluation, GUE) covering 28 tasks across 7 species. DNABERT-2 achieves state-of-the-art performance across diverse genome prediction tasks, demonstrating superior efficiency and generalization. The model and benchmark are publicly available.

DNABERT-S builds upon DNABERT-2 to develop species-aware DNA embeddings via Manifold Instance Mixup (MI-Mix) contrastive learning and Curriculum Contrastive Learning (C2LR), enabling unsupervised species differentiation from DNA sequences. Published in Bioinformatics (ISMB 2025 proceedings).

species awareness, contrastive learning, manifold instance mixup, curriculum learning, long-read sequencing, ISMB 2025

We introduce DNABERT-S, a tailored genome model that develops species-aware embeddings to naturally cluster and segregate DNA sequences of different species in the embedding space. Differentiating species from genomic sequences (i.e. DNA and RNA) is vital yet challenging, since many real-world species remain uncharacterized, lacking known genomes for reference. Embedding-based methods are therefore used to differentiate species in an unsupervised manner. DNABERT-S builds upon a pre-trained genome foundation model named DNABERT-2. To encourage effective embeddings to error-prone long-read DNA sequences, we introduce Manifold Instance Mixup (MI-Mix), a contrastive objective that mixes the hidden representations of DNA sequences at randomly selected layers and trains the model to recognize and differentiate these mixed proportions at the output layer. We further enhance it with the proposed Curriculum Contrastive Learning (C2LR) strategy. Empirical results on 28 diverse datasets show DNABERT-S achieves state-of-the-art performance across multiple species classification and clustering tasks.

The eICU Collaborative Research Database (eICU-CRD) is a large multi-center intensive care unit database from Philips Healthcare's eICU telehealth program, in partnership with MIT Laboratory for Computational Physiology. Contains de-identified data for over 200,000 admissions from 139,000 unique patients across 335 ICU units at 208 US hospitals (2014-2015). Includes: demographics, vital signs, care plan documentation, severity of illness measures (APACHE IV), diagnoses (3,933 unique active problems), laboratory measurements (158 lab types), medications, continuous infusions, intake/output, microbiology, nurse charting, and structured notes. Data access follows the same process as MIMIC: free of charge, requires PhysioNet credentialed access (CITI Data or Specimens Only Research course + Data Use Agreement). Complements MIMIC-IV (single-center, Boston) with multi-center US coverage for external validation and generalizability of ICU ML models.

eICU, critical care, ICU, multicenter, Philips, PhysioNet, vital signs, severity of illness

Critical care patients are monitored closely through the course of their illness. Philips Healthcare has developed a telehealth system, the eICU Program, which leverages these data to support management of critically ill patients. Here we describe the eICU Collaborative Research Database, a multi-center intensive care unit (ICU) database with high granularity data for over 200,000 admissions to ICUs monitored by eICU Programs across the United States.

Enformer is a DeepMind model that integrates long-range DNA interactions (up to 200 kb) using a CNN + Transformer architecture to predict gene expression, chromatin profiles, and TF binding from DNA sequence. Achieves SOTA on variant effect prediction and regulatory element annotation.

gene expression prediction, long-range interactions, CNN, Transformer, DeepMind, chromatin, variant effect, regulatory genomics

Quantitative gene expression measurements across cell types and tissues can provide a complete picture of the dynamic functions of the genome. However, gene expression is challenging to predict from sequence alone because of the enormous distances over which regulatory elements act. Enformer combines CNN and Transformer architectures to integrate information from up to 200 kb of DNA sequence, enabling accurate prediction of gene expression, chromatin state, and transcription factor binding. Enformer achieves state-of-the-art predictions across diverse genomic assays and accurately predicts the effects of genetic variants on gene expression.

ENLIGHT-DeepPT (Deep Phenotyping of Tumors) is a deep-learning framework (ResNet50 + MLP) that predicts genome-wide tumor mRNA expression from routine H&E histopathology images across 16 TCGA cancer types. The imputed transcriptomics then drive treatment response prediction, achieving odds ratio of 2.28 across 5 independent treatment cohorts. Directly links medical imaging (histopathology) with genomics/transcriptomics via AI, enabling precision oncology from standard pathology slides.

Predicting cancer treatment response from routinely collected clinical material is a central challenge in precision oncology. Here we present ENLIGHT-DeepPT, a deep-learning framework that predicts genome-wide tumor mRNA expression from routine H&E histopathology images. Using a two-stage approach (image-to-transcriptomics via ResNet50 + MLP, then transcriptomics-to-treatment response), ENLIGHT-DeepPT achieves an odds ratio of 2.28 across 5 independent treatment cohorts spanning multiple cancer types and drug classes.

ESMFold from Meta FAIR uses a protein language model (ESM-2) trained on 65 million protein sequences via masked language modeling to predict protein 3D structures directly from sequence, without requiring multiple sequence alignments (MSAs). 60-100x faster than AlphaFold2 while maintaining high accuracy. Enables structure prediction at evolutionary scale — 617 million predicted structures released. Represents a paradigm shift combining protein language models with structure prediction.

Protein language models can learn evolutionary patterns from sequences without explicit alignment or structural data. Here we demonstrate that direct inference of structure at scale is possible by training a language model, ESM-2, on 65 million protein sequences and using it to predict structure. ESMFold predicts structure 60x faster than AlphaFold2 while maintaining high accuracy, enabling evolutionary-scale structural biology. We release 617 million predicted protein structures, covering the majority of sequences in the UniRef50 database.

A genomic foundation model using the StripedHyena 2 architecture, trained autoregressively on OpenGenome2 (trillions of nucleotides across prokaryotic, eukaryotic, archaeal, and phage genomes) at single-nucleotide resolution with long context (up to about one megabase). Supports generalist prediction and design tasks spanning DNA, RNA, and proteins; code and weights are open source with Hugging Face checkpoints.

DNA foundation model, autoregressive, StripedHyena 2, prokaryotic, eukaryotic, genome design, variant effect, long context, open source

All of life encodes information with DNA. Although tools for genome sequencing, synthesis and editing have transformed biological research, we still lack sufficient understanding of the immense complexity encoded by genomes to predict the effects of many classes of genomic changes or to intelligently compose new biological systems. Artificial intelligence models that learn information from genomic sequences across diverse organisms have increasingly advanced prediction and design capabilities1,2. Here we introduce Evo 2, a biological foundation model trained on 9 trillion DNA base pairs from a highly curated genomic atlas spanning all domains of life to have a 1 million token context window with single-nucleotide resolution. Evo 2 learns to accurately predict the functional impacts of genetic variation-from noncoding pathogenic mutations to clinically significant BRCA1 variants-without task-specific fine-tuning. Mechanistic interpretability analyses reveal that Evo 2 learns representations associated with biological features, including exon-intron boundaries, transcription factor binding sites, protein structural elements and prophage genomic regions. The generative abilities of Evo 2 produce mitochondrial, prokaryotic and eukaryotic sequences at genome scale with greater naturalness and coherence than previous methods. Evo 2 also generates experimentally validated chromatin accessibility patterns when guided by predictive models3,4 and inference-time search. We have made Evo 2 fully open, including model parameters, training code5, inference code and the OpenGenome2 dataset, to accelerate the exploration and design of biological complexity.

ExPecto (Basenji2) is a deep learning framework that predicts the causal effects of both coding and noncoding genetic variants on gene expression levels and disease risk directly from DNA sequence, without requiring any prior knowledge of regulatory elements or annotations.

variant effect, gene expression, disease risk, ab initio, deep learning, noncoding, tissue-specific

The complexity and scale of human genetics studies present a significant challenge for interpreting the functional consequences of genetic variants. Here we introduce ExPecto, a deep learning framework that can predict the causal effects of genetic variants on gene expression levels and disease risk directly from DNA sequence. ExPecto uses a modular architecture with a core deep convolutional neural network to learn regulatory features from sequence, followed by spatial transformation and tissue-specific prediction layers. We demonstrate that ExPecto can accurately predict variant effects on expression across a wide range of tissues and can identify pathogenic variants from population-scale sequencing data. The framework enables ab initio prediction of expression and disease risk without relying on prior annotations of regulatory elements.

Flashzoi is an enhanced Borzoi model that replaces relative positional encodings with rotary positional encodings (RoPE) and uses FlashAttention-2, achieving over 3x faster training/inference and up to 2.4x reduced memory usage while maintaining or improving accuracy on RNA-seq coverage prediction, variant effects, and enhancer-promoter linking.

Borzoi, FlashAttention, RoPE, accelerated inference, regulatory genomics, variant effect, enhancer-promoter

Accurately predicting how DNA sequence drives gene regulation and how genetic variants alter gene expression is a central challenge in genomics. Borzoi, which models over ten thousand genomic assays including RNA-seq coverage from over half a megabase of sequence context alone promises to become an important foundation model in regulatory genomics, both for massively annotating variants and for further model development. However, the currently used relative positional encodings limit Borzoi's computational efficiency. We present Flashzoi, an enhanced Borzoi model that leverages rotary positional encodings and FlashAttention-2. This achieves over 3-fold faster training and inference and up to 2.4-fold reduced memory usage, while maintaining or improving accuracy in modeling various genomic assays including RNA-seq coverage, predicting variant effects, and enhancer-promoter linking. Flashzoi's improved efficiency facilitates large-scale genomic analyses and opens avenues for exploring more complex regulatory mechanisms and modeling.

Google DeepMind’s Gemini model family for chat, search, code, and multimodal tasks across consumer and Vertex / AI Studio APIs.

Genos is a human-centric genomic foundation model using mixture-of-experts architecture (Genos-1.2B/Genos-10B) for million-basepair sequence modeling, trained on high-quality human de novo assemblies including the Human Pangenome Reference Consortium data. Published in GigaScience.

mixture of experts, million-basepair context, human pangenome, human-centric, variant effect, structural variation

The rapid expansion of human genomic data demands foundation models that manage ultra-long sequences and capture population diversity, limitations common in existing models that lack human-specific representation, and clinical inference efficiency. Here, we introduce Genos (Genos-1.2B/Genos-10B), a human-centric genomic foundation model engineered for million-basepair sequence modeling. Genos utilizes a large-scale mixture of experts structure, optimized for a 1-Mb context, trained on high-quality human de novo assemblies from datasets such as the Human Pangenome Reference Consortium and the Human Genome Structural Variation Consortium, representing diverse global populations. A suite of optimization strategies was implemented to ensure training stability and enhance computational efficiency, which collectively reduces costs and facilitates million-basepair context modeling. Functionally, Genos performs single-nucleotide resolution analysis and dynamically simulates the cascade effects of genetic variation on molecular phenotypes, including the influence of both common and rare single-nucleotide variants as well as structural variants.

Zhipu AI’s ChatGLM / GLM family of bilingual LLMs and coding assistants, with open and commercial variants.

GPN-MSA is a DNA language model leveraging whole-genome multiple-sequence alignments across species to predict the effects of genome-wide variants, achieving outstanding performance on deleteriousness prediction for both coding and noncoding variants. Published in Nature Biotechnology.

multiple sequence alignment, variant effect prediction, noncoding variants, ClinVar, COSMIC, gnomAD, DNA language model

Protein language models have demonstrated remarkable performance in predicting the effects of missense variants but DNA language models have not yet shown a competitive edge for complex genomes such as that of humans. This limitation is particularly evident when dealing with the vast complexity of noncoding regions that comprise approximately 98% of the human genome. To tackle this challenge, we introduce GPN-MSA (genomic pretrained network with multiple-sequence alignment), a framework that leverages whole-genome alignments across multiple species while taking only a few hours to train. Across several benchmarks on clinical databases (ClinVar, COSMIC and OMIM), experimental functional assays (deep mutational scanning and DepMap) and population genomic data (gnomAD), our model for the human genome achieves outstanding performance on deleteriousness prediction for both coding and noncoding variants. We provide precomputed scores for all ~9 billion possible single-nucleotide variants in the human genome.

xAI’s Grok models integrated with X (Twitter) and standalone apps, aimed at real-time, conversational assistance.

GWANN (Genome-Wide Association Neural Networks) is a novel approach that uses neural networks to perform gene-level association studies. Applied to Alzheimer's disease in UK Biobank, GWANN identifies genes linked to family history of AD by aggregating SNP-level information at the gene level through neural network architectures. Published in Briefings in Bioinformatics.

Augmenting traditional genome-wide association studies (GWAS) with advanced machine learning algorithms can allow the detection of novel signals in available cohorts. We introduce "genome-wide association neural networks (GWANN)", a novel approach that uses neural networks (NNs) to perform a gene-level association study with family history of Alzheimer's disease (AD) in UK Biobank.

Developed an abdominal MRI-based machine-learning regression model (gradient-boosted regression on raw MRI signal intensities) to accurately estimate liver fat from UK Biobank abdominal MRI scans (correlation 0.97-0.99 with ground truth). Trained on 4,511 participants with gold-standard MRI biomarker measurements and applied to 32,192 additional individuals. GWAS identified 8 associated variants (5 novel: MTARC1, ADH1B, TRIB1, GPAM, MAST3) and a polygenic score strongly associated with future chronic liver disease risk (HR>1.32 per SD, p<9e-17).

MRI signal regression, liver fat quantification, abdominal MRI, hepatic steatosis, gradient boosting, UK Biobank

Excess liver fat, called hepatic steatosis, is a leading risk factor for end-stage liver disease and cardiometabolic diseases but often remains undiagnosed in clinical practice because of the need for direct imaging assessments. We developed an abdominal MRI-based machine-learning algorithm to accurately estimate liver fat from a truth dataset of 4,511 middle-aged UK Biobank participants, enabling quantification in 32,192 additional individuals. A genome-wide association study of common genetic variants and liver fat replicated three known associations and identified five newly associated variants.

Hermes Agent is an open-source autonomous AI agent built by Nous Research with a built-in closed learning loop — it creates skills from experience, improves them during use, persists knowledge, searches past conversations, and builds a user model across sessions. Supports 14 messaging platforms (Telegram, Discord, Slack, WhatsApp, Signal, etc.) and 6 execution backends (local, Docker, SSH, Modal, Singularity). Features include persistent memory (FTS5 + Honcho), automated skill creation, cron scheduling, parallel sub-agent delegation, full browser automation, MCP integration, and model-agnostic provider switching via CLI. MIT licensed, first released February 2026.

HyenaDNA is a genomic foundation model using implicit convolution operators (Hyena) to achieve up to 1 million token context length at single nucleotide resolution, overcoming the quadratic scaling limitations of Transformer-based models. Published at NeurIPS 2023.

Hyena, implicit convolution, long-range context, single nucleotide resolution, SNP, regulatory elements, NeurIPS 2023

Genomic (DNA) sequences encode an enormous amount of information for gene regulation and protein synthesis. Similar to natural language models, researchers have proposed foundation models in genomics to learn generalizable features from unlabeled genome data that can then be fine-tuned for downstream tasks such as identifying regulatory elements. Due to the quadratic scaling of attention, previous Transformer-based genomic models have used 512 to 4k tokens as context (<0.001% of the human genome), significantly limiting the modeling of long-range interactions in DNA. In addition, these methods rely on tokenizers or fixed k-mers to aggregate meaningful DNA units, losing single nucleotide resolution where subtle genetic variations can completely alter protein function via single nucleotide polymorphisms (SNPs). Recently, Hyena, a large language model based on implicit convolutions was shown to match attention in quality while allowing longer context lengths and lower time complexity. Leveraging Hyena, we present HyenaDNA, a genomic foundation model that scales context length to 1 million tokens at single nucleotide resolution, an up to 500x increase over previous dense attention-based models. HyenaDNA achieves state-of-the-art on 12 of 18 benchmarks and excels at long-range regulatory element prediction and SNP effect prediction.

iGWAS uses self-supervised contrastive learning (SimCLR-style framework with a CNN encoder backbone) to extract a 128-dimensional phenotype vector directly from retinal fundus images without any manual labels. Trained on 40,000 EyePACS images via instance discrimination, then applied to 130,329 UK Biobank fundus images. GWAS on these 128 learned phenotypes identified 14 genome-wide significant loci. First demonstration of unsupervised deep phenotyping for image-based GWAS — discovering genetic associations without predefined human annotations.

self-supervised contrastive learning, SimCLR, CNN, retinal fundus, deep phenotyping, image-based GWAS, UK Biobank

Existing imaging genetics studies have been mostly limited in scope by using imaging-derived phenotypes defined by human experts. Here, leveraging new breakthroughs in self-supervised deep representation learning, we propose a new approach, image-based genome-wide association study (iGWAS), for identifying genetic factors associated with phenotypes discovered from medical images using contrastive learning. Using retinal fundus photos, our model extracts a 128-dimensional vector representing features of the retina as phenotypes. We identified 14 loci with genome-wide significance.

A comprehensive review of imaging-derived phenotypes (IDPs) for genetic analysis. Covers MRI, CT, X-ray, OCT, and other imaging modalities; traditional (FSL, FreeSurfer) and deep learning (U-Net, nnU-Net, self-supervised contrastive learning) pipelines for IDP extraction; biobank-scale cohorts (UK Biobank, All of Us); GWAS of brain, cardiac, retinal, abdominal, and skeletal IDPs; Mendelian randomization for causal inference linking organ structure to disease; and emerging challenges in multi-organ phenomics, longitudinal imaging, and clinical translation of IDP-based PRS.

Imaging-derived phenotypes (IDPs) developed from medical imaging data, such as magnetic resonance imaging, computed tomography and X-ray scans, are traits that provide quantitative information on anatomical and functional properties of organs and tissues. IDPs are powerful tools for identifying biomarkers and studying disease mechanisms. When coupled with genetic data, IDPs can be analysed as heritable phenotypes using modern gene mapping methods to uncover genotype-phenotype relationships. The field of imaging genomics is rapidly maturing, with the emergence of high-quality imaging datasets collected in biobank-scale cohorts and sophisticated computational methods for extracting IDPs from imaging data, including tools that leverage machine learning. Here we review common imaging modalities and analytical approaches for developing IDPs, discuss biological insights gleaned from the large-scale genetic analysis of imaging traits and highlight emerging areas and remaining challenges that must be overcome to realize the full potential of IDPs for genetic analysis.

InsightGWAS is a Transformer-based model that enhances genetic discovery for migraine GWAS by integrating functional annotations and leveraging transfer learning from GWAS datasets of major depressive disorder. It identified 293 previously unreported loci from 53,109 cases and 230,876 controls. Published in Nature Communications.

Migraine is a complex neurological disorder with substantial heritability, yet genome-wide association studies (GWAS) have explained only a fraction of its genetic component. We developed InsightGWAS, a Transformer-based model, to enhance genetic discovery for migraine by integrating functional annotations and leveraging transfer learning from GWAS datasets of major depressive disorder (MDD), identifying 293 previously unreported loci.

KEEP (KnowledgE-Enhanced Pathology) is a vision-language foundation model from Shanghai AI Lab / SJTU that systematically integrates disease knowledge into pretraining for cancer diagnosis. Uses a comprehensive disease knowledge graph with 11,454 diseases and 139,143 attributes from DO and UMLS to reorganize millions of pathology image-text pairs into 143,000 semantically structured groups aligned with disease ontology hierarchies. Across 18 public benchmarks (14,000+ WSIs) and 4 institutional rare cancer datasets (926 cases), KEEP consistently outperforms existing foundation models (CHIEF, CONCH, UNI), with substantial gains for rare subtypes (+8.5 pts balanced accuracy vs CONCH on 30 rare brain cancers). Published in Cancer Cell, Feb 2026.

Vision-language foundation models have shown great promise in computational pathology but remain primarily data-driven, lacking explicit integration of medical knowledge. We introduce KEEP, a foundation model that systematically incorporates disease knowledge into pretraining for cancer diagnosis. KEEP leverages a comprehensive disease knowledge graph encompassing 11,454 diseases and 139,143 attributes to reorganize millions of pathology image-text pairs into 143,000 semantically structured groups aligned with disease ontology hierarchies. Across 18 public benchmarks (over 14,000 WSIs) and 4 institutional rare cancer datasets (926 cases), KEEP consistently outperformed existing foundation models, showing substantial gains for rare subtypes.

Applied a CNN-based segmentation model (U-Net style architecture) to automatically segment left ventricular myocardium from cardiac MRI in 43,230 UK Biobank participants. The segmented contours were used to compute left ventricular mass indexed to body surface area (LVMI), enabling GWAS that identified 12 associations (11 novel) implicating genes associated with cardiac contractility and cardiomyopathy. The LVMI polygenic risk score validated in independent Mass General Brigham cohort.

deep learning, cardiac MRI segmentation, U-Net, left ventricular mass, CNN, cardiomyopathy, UK Biobank

Left ventricular mass is a risk marker for cardiovascular events, and may indicate an underlying cardiomyopathy. Cardiac magnetic resonance is the gold-standard for left ventricular mass estimation, but is challenging to obtain at scale. Here, we use deep learning to enable genome-wide association study of cardiac magnetic resonance-derived left ventricular mass indexed to body surface area within 43,230 UK Biobank participants. We identify 12 genome-wide associations (1 known at TTN and 11 novel for left ventricular mass).

Moonshot AI’s Kimi chat and model lineup, known for long-context reasoning and Chinese–English bilingual use.

Applied a U-Net-based CNN segmentation pipeline to over 38,000 abdominal MRI scans from UK Biobank. The deep learning model automatically segmented 7 organs/tissues (liver, pancreas, kidneys, spleen, lungs, visceral adipose tissue, subcutaneous adipose tissue) and quantified their volume, fat content (via signal intensity), and iron content (via T2* mapping). GWAS on these 11 DL-derived traits identified 93 independent genome-wide significant associations (heritability 8-44%), including 4 novel liver trait associations.

deep learning, abdominal MRI segmentation, U-Net, organ volume quantification, liver fat, pancreas iron, UK Biobank

Cardiometabolic diseases are an increasing global health burden. While socioeconomic, environmental, behavioural, and genetic risk factors have been identified, a better understanding of the underlying mechanisms is required to develop more effective interventions. Magnetic resonance imaging (MRI) has been used to assess organ health, but biobank-scale studies are still in their infancy. Using over 38,000 abdominal MRI scans in the UK Biobank, we used deep learning to quantify volume, fat, and iron in seven organs and tissues, and demonstrate that imaging-derived phenotypes reflect health status. We identify 93 independent genome-wide significant associations.

Meta’s open-weights Llama family for research and product fine-tuning, from dense LLMs to multimodal stacks.

A natural-language interface that enables large language models (LLMs) to perform end-to-end variant interpretation for Mendelian diseases via structured tool access (MCP). MARRVEL-MCP equips LLMs with 44 tools spanning gene and variant utilities, pathogenicity databases, phenotype resources, expression atlases, ortholog data, and literature APIs. Without hard-coded workflows, LLMs infer which tools to invoke and in what sequence, performing named-entity recognition, identifier normalization, and multi-database synthesis from clinical queries. A 20B-parameter model achieved 94% pass rate on 100 expert-curated questions (vs 41% without MARRVEL-MCP), approaching state-of-the-art proprietary performance. Establishes context engineering as a core principle for biomedical AI.

Variant interpretation in rare diseases requires navigating multiple genomic databases, each with strict input formats, while synthesizing heterogeneous evidence. To address these usability barriers, we developed MARRVEL-MCP, a natural-language interface that enables large language models (LLMs) to perform end-to-end variant interpretation via structured tool access. MARRVEL-MCP equips LLMs with 44 tools spanning gene and variant utilities, pathogenicity databases, phenotype resources, expression atlases, ortholog data, and literature APIs. Without hard-coded workflows, LLMs infer which tools to invoke and in what sequence, performing named-entity recognition, identifier normalization, and multi-database synthesis from clinical queries. Using 100 expert-curated questions, lightweight models (3B-20B parameters) with MARRVEL-MCP matched or outperformed larger models without tool access. A 20B-parameter model achieved a 94% pass rate, versus 41% without MARRVEL-MCP, approaching state-of-the-art proprietary performance. These findings establish context engineering as a core principle for biomedical AI and support scalable integration of LLMs with curated genomic resources.

MILTON is an ensemble machine learning framework that utilizes biomarkers and multi-omics data to predict 3,213 diseases in the UK Biobank. It predicts incident disease cases undiagnosed at time of recruitment and demonstrates utility in augmenting genetic association discovery by empowering case-control GWAS with predicted phenotypes. Published in Nature Genetics.

The emergence of biobank-level datasets offers new opportunities to discover novel biomarkers and develop predictive algorithms for human disease. Here, we present an ensemble machine-learning framework (machine learning with phenotype associations, MILTON) utilizing a range of biomarkers to predict 3,213 diseases in the UK Biobank. MILTON predicts incident disease cases undiagnosed at time of recruitment, largely outperforming available polygenic risk scores, and augments genetic association discovery.

MIMIC-IV is a large, freely-available de-identified clinical database comprising over 300,000 patients admitted to the Beth Israel Deaconess Medical Center (2008-2019). It includes comprehensive ICU and Emergency Department data: demographics, vital signs, laboratory measurements, medications, procedures, diagnoses (ICD codes), imaging reports, nursing notes, and mortality outcomes. The relational database (BigQuery or local PostgreSQL) links hospital admissions (ADMISSIONS), patient stays (ICUSTAYS), charted observations (CHARTEVENTS), lab events (LABEVENTS), microbiology data (MICROBIOLOGYEVENTS), prescriptions (PRESCRIPTIONS), and discharge summaries. MIMIC-IV replaces MIMIC-III (2001-2012) with a modernized schema, cleaner data model, and expanded coverage. Widely used for developing and benchmarking clinical AI models (mortality prediction, sepsis detection, phenotyping, NLP), it requires credentialed access via PhysioNet (CITI Data or Specimens Only course). Supporting datasets include MIMIC-CXR (chest X-ray images) and MIMIC-NOTE (de-identified clinical notes).

EHR, ICU, clinical database, de-identified, Beth Israel, critical care, MIMIC, medical informatics

MIMIC-IV is a publicly available database of de-identified electronic health records for patients admitted to the Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts. The database is updated annually and is freely available to credentialed researchers. MIMIC-IV contains information on patient demographic characteristics, vital signs, laboratory measurements, medications, and diagnoses. We describe the process of creating the database, the structure of the data, and the tools available to users. MIMIC-IV is a valuable resource for researchers in critical care, clinical informatics, and machine learning.

MiniMax’s text, voice, and video model ecosystem for chat, APIs, and creative / agent applications.

MIRA is an autonomous AI agent powered by GPT-4o (T=0.01) with o1-preview for structured reasoning, operating within a sandboxed HL7 FHIR-based EHR environment. It navigates 85,000+ clinical decision options across 8 emergency department diagnoses, using 11 FHIR-compliant tools (PatientHistory, PhysicalExam, Lab/Urine/Microbiology/Radiology requests, Medication/Procedure ordering, Plan, Admission). Evaluated on 574 real MIMIC-IV patient cases, MIRA outperformed two independent physician cohorts in diagnostic accuracy, guideline-concordant treatment, medication safety, and appropriate admission decisions. All tool parameter validity is enforced through token masking, making hallucination of non-existent options programmatically impossible.

Large language models (LLMs) show great potential for clinical decision-making, yet most applications remain narrow, task-specific chat tools rather than systems integrated into clinical workflows. However, building physician copilots will require models that operate within the electronic health record (EHR), with governed access to patient data and the ability to initiate permitted EHR actions within defined safety constraints. Here we show that MIRA (Medical Intelligence for Reasoning and Action), an autonomous artificial intelligence agent operating in a sandboxed EHR environment, can navigate a large clinical action space to obtain patient histories; order and interpret laboratory, imaging and microbiology tests; generate differential diagnoses; and formulate treatment plans such as prescribing medications, scheduling surgical procedures and planning admissions. In simulations on real patient cases spanning multiple diagnoses, MIRA outperformed physicians in diagnostic accuracy and made guideline-concordant, medication-safe and appropriate admission decisions.

Mistral AI’s open and commercial Mistral / Mixtral models for chat, code, and EU-focused deployments.

MixEHR-SAGE is a PheCode-guided multi-modal topic model that integrates diagnoses, procedures, and medications from EHR to enhance phenotyping for GWAS. By combining expert-informed priors with probabilistic inference, it identifies over 1000 interpretable phenotype topics from UK Biobank data and improves disease incidence prediction and GWAS discovery. Published in Briefings in Bioinformatics.

Phenome-wide association studies rely on disease definitions derived from diagnostic codes, often failing to leverage the full richness of electronic health records (EHR). We present MixEHR-SAGE, a PheCode-guided multi-modal topic model that integrates diagnoses, procedures, and medications to enhance phenotyping from large-scale EHRs. Applied to 350,000 individuals with high-quality genetic data, MixEHR-SAGE-derived risk scores accurately predicted disease incidence and improved GWAS discovery.

mSTAR (Multimodal Self-TAught PRetraining) is a pathology foundation model from HKUST/SJTU that integrates three modalities: pathology slides (WSIs), expert pathology reports, and gene expression (RNA-Seq) data. Curates the largest multimodal dataset of 26,169 slide-level modality pairs across 32 cancer types from 10,275 TCGA patients (>116M patch images). Uses a two-stage paradigm: (1) slide-level contrastive learning across WSI-report-gene modalities, (2) self-taught training that propagates multimodal knowledge from slide aggregator (teacher) to patch extractor (student). Evaluated on 97 tasks across 15 application types, outperforming UNI, CONCH, CHIEF, and GigaPath. Key finding: multimodal integration yields greater improvements than simply expanding vision-only datasets (53x data efficiency vs Virchow). Published in Nat Commun, Dec 2025.

Computational pathology has advanced through foundation models, yet faces challenges in multimodal integration and capturing whole-slide context. We present mSTAR, the pathology foundation model that incorporates three modalities: pathology slides, expert-created reports, and gene expression data, within a unified framework. Our dataset includes 26,169 slide-level modality pairs across 32 cancer types, comprising over 116 million patch images. This approach injects multimodal whole-slide context into patch representations, expanding modeling from single to multiple modalities and from patch-level to slide-level analysis. Across 97 tasks, mSTAR outperforms previous SOTA models, particularly in molecular prediction, revealing that multimodal integration yields greater improvements than simply expanding vision-only datasets.

NHANES is a program of studies by the US CDC designed to assess the health and nutritional status of adults and children in the United States. It combines interviews, physical examinations, and laboratory tests across 2-year continuous cycles (1999-present). The survey includes: demographics (age, sex, race/ethnicity, income); ~40+ questionnaire components (diabetes DIQ, cardiovascular CDQ, depression PHQ-9/DPQ, sleep SLQ, smoking SMQ, alcohol ALQ, kidney KIQ, physical function PFQ, osteoporosis OSQ, weight history WHQ); physical examinations (anthropometry BMX, blood pressure BPX, DXA bone density, spirometry SPX, hearing AUXAR, vision VIX, accelerometry PAXRAW); and ~50+ laboratory biomarkers (comprehensive metabolic panel BIOPRO, complete blood count CBC, HbA1c GHB, glucose GLU, insulin INS, lipids TCHOL/HDL, CRP/HSCRP, vitamin D VID, ferritin FERTIN, testosterone TST, hepatitis/HIV/HPV serology, PFAS chemicals, heavy metals, cotinine). Dietary intake data includes 2-day 24-hour recall (DR1IFF/DR2IFF). NHANES is multi-ethnic and US-representative with survey weights. Total: ~1,600 SAS XPT files, ~5 GB. Available at no cost, no registration required. Widely used for epidemiological research, biomarker GWAS replication, and cross-population comparisons with biobank data (UKBB, BBJ).

NHANES, CDC, health survey, biomarkers, nutrition, demographics, epidemiology, US representative, public health

Built a CNN-based deep learning algorithm for automated segmentation of left ventricular myocardium from cardiac MRI, enabling precise calculation of 12 regional wall thickness (LVRWT) measurements in 42,194 UK Biobank participants. GWAS of these 12 CNN-derived LVRWT traits identified 72 significant genetic loci involved in heart development and contraction pathways. Mendelian randomization confirmed causal relationships with hypertrophic cardiomyopathy. The PRS of inferoseptal LVRWT enabled identification of high-risk individuals.

deep learning, cardiac MRI, CNN, left ventricular wall thickness segmentation, hypertrophic cardiomyopathy, UK Biobank

Left ventricular regional wall thickness (LVRWT) is an independent predictor of morbidity and mortality in cardiovascular diseases (CVDs). To identify specific genetic influences on individual LVRWT, we established a novel deep learning algorithm to calculate 12 LVRWTs accurately in 42,194 individuals from the UK Biobank with cardiac magnetic resonance (CMR) imaging. Genome-wide association studies of CMR-derived 12 LVRWTs identified 72 significant genetic loci associated with at least one LVRWT phenotype.

A family of transformer foundation models (from tens of millions to multi-billion parameters) pretrained on thousands of human and other-species genomes to learn DNA sequence representations. Embeddings support fine-tuning for tasks such as splice-site prediction, enhancer activity, histone marks, and transcription-factor binding, with benchmarks and weights released openly.

Transformer, foundation model, human genome, multi-species, DNA embeddings, splice-site, enhancer, histone marks, TF binding

The prediction of molecular phenotypes from DNA sequences remains a longstanding challenge in genomics, often driven by limited annotated data and the inability to transfer learnings between tasks. Here, we present an extensive study of foundation models pre-trained on DNA sequences, named Nucleotide Transformer, ranging from 50 million up to 2.5 billion parameters and integrating information from 3,202 human genomes and 850 genomes from diverse species. These transformer models yield context-specific representations of nucleotide sequences, which allow for accurate predictions even in low-data settings. We show that the developed models can be fine-tuned at low cost to solve a variety of genomics applications. Despite no supervision, the models learned to focus attention on key genomic elements and can be used to improve the prioritization of genetic variants. The training and application of foundational models in genomics provides a widely applicable approach for accurate molecular phenotype prediction from DNA sequence.

Open-source, local-first autonomous AI assistant: file/shell access, browser automation, skills/plugins, and optional chat-app bridges; model-agnostic (e.g. Claude, OpenAI, local/Ollama) with your own API keys.

PathOrchestra is a versatile pathology foundation model from Shanghai AI Lab and multiple Chinese institutions, trained via self-supervised learning on 287,424 H&E-stained WSIs from 21 tissue types across 3 independent clinical centers. Evaluated on the largest known clinical task benchmark (112 tasks: 61 private + 51 public) spanning digital slide preprocessing, pan-cancer classification (17 cancer types), lesion identification, multi-cancer subtype classification (36 tasks), biomarker assessment (36 tasks), gene expression prediction, and structured report generation. Achieves over 0.950 accuracy in 47 tasks. First model to generate structured pathology reports for colorectal cancer and lymphoma. Apache 2.0 open-source license.

The complexity and variability of high-resolution pathological images present significant challenges in computational pathology. We present PathOrchestra, a versatile pathology foundation model trained via self-supervised learning on 287,424 slides from 21 tissue types across three centers. Evaluated on 112 tasks from 61 private and 51 public datasets, covering digital slide preprocessing, pan-cancer classification, lesion identification, multi-cancer subtype classification, biomarker assessment, gene expression prediction, and structured report generation. Across 27,755 WSIs and 9,415,729 ROI images, it achieved over 0.950 accuracy in 47 tasks. It is the first to generate structured reports for colorectal cancer and lymphoma.

PoPS (Polygenic Priority Score) is a method that learns trait-relevant gene features, such as cell-type-specific expression, to prioritize genes at GWAS loci. It leverages polygenic enrichments across multiple gene features to predict causal genes underlying complex traits and diseases. Published in Nature Genetics.

Genome-wide association studies (GWASs) are a valuable tool for understanding the biology of complex human traits and diseases, but associated variants rarely point directly to causal genes. In the present study, we introduce a new method, polygenic priority score (PoPS), that learns trait-relevant gene features, such as cell-type-specific expression, to prioritize genes at GWAS loci. PoPS and the closest gene individually outperform other gene prioritization methods.

A deep learning model from Illumina that scores how variants in gene promoter regions alter predicted gene expression, trained on chromatin and expression-related signals at nucleotide resolution. Distributed as a Python package with precomputed genome-wide scores to support rare-disease and research variant interpretation alongside other splice and protein effect tools.

promoter, variant effect, deep learning, rare disease, noncoding, gene expression, Illumina

Only a minority of patients with rare genetic diseases are presently diagnosed by exome sequencing, suggesting that additional unrecognized pathogenic variants may reside in noncoding sequence. In this work, we describe PromoterAI, a deep neural network that accurately identifies noncoding promoter variants that dysregulate gene expression. We show that promoter variants with predicted expression-altering consequences produce outlier expression at both the RNA and protein levels in thousands of individuals and that these variants experience strong negative selection in human populations. We observed that clinically relevant genes in patients with rare diseases are enriched for such variants and validated their functional impact through reporter assays. Our estimates suggest that promoter variation accounts for 6% of the genetic burden associated with rare diseases.

Prov-GigaPath by Microsoft Research, Providence, and UW is a whole-slide pathology foundation model pretrained on 1.3 billion 256x256 image tiles from 171,189 whole slides across 28 cancer centers (>30,000 patients, 31 tissue types). Uses a novel GigaPath vision transformer with dilated self-attention (LongNet) for gigapixel-level context. Achieves SOTA on 25/26 benchmark tasks including cancer subtyping, mutation prediction, and TMB classification. The first large-scale whole-slide foundation model trained on real-world clinical data.

Digital pathology poses unique computational challenges, as a standard gigapixel slide may comprise tens of thousands of image tiles. Prior models have often resorted to subsampling a small portion of tiles for each slide, thus missing important slide-level context. Here we present Prov-GigaPath, a whole-slide pathology foundation model pretrained on 1.3 billion pathology image tiles in 171,189 whole slides from Providence, a large US health network comprising 28 cancer centres. To pretrain Prov-GigaPath, we propose GigaPath, a novel vision transformer for pretraining gigapixel pathology slides using dilated self-attention. Prov-GigaPath attains state-of-the-art performance on 25 out of 26 benchmark tasks.

PTB-XL is the largest freely accessible clinical 12-lead ECG-waveform dataset, comprising 21,837 records from 18,885 patients of 10 seconds length. Annotated by up to two cardiologists with 71 SCP-ECG diagnostic, form, and rhythm statements organized into 5 superclasses (NORM, CD, MI, HYP, STTC) and 24 subclasses. Includes raw waveforms at 500Hz and downsampled 100Hz, plus rich metadata: demographics (age, sex, height, weight), signal quality annotations (noise, baseline drift, electrodes), and recommended stratified 10-fold cross-validation splits. Fully open access on PhysioNet — no registration or training required. Widely used as the standard benchmark for automated ECG interpretation, arrhythmia detection, and deep learning in cardiology.

ECG, electrocardiography, 12-lead, cardiovascular, waveform, signal processing, PhysioNet

Electrocardiography (ECG) is a key non-invasive diagnostic tool for cardiovascular diseases which is increasingly supported by algorithms based on machine learning. Major obstacles for the development of automatic ECG interpretation algorithms are both the lack of public datasets and well-defined benchmarking procedures to allow comparisons of different algorithms. To address these issues, we put forward PTB-XL, the to-date largest freely accessible clinical 12-lead ECG-waveform dataset comprising 21837 records from 18885 patients of 10 seconds length.

Agentic coding platform with IDE, JetBrains plugin, and CLI—quest mode, repo-aware chat, and parallel expert-style agents.

Quickdraws is a method that increases association power in quantitative and binary traits for GWAS without sacrificing computational efficiency, leveraging a spike-and-slab prior on variant effects, stochastic variational inference, and graphics processing unit acceleration. Published in Nature Genetics.

The rapid growth of modern biobanks is creating new opportunities for large-scale genome-wide association studies (GWASs) and the analysis of complex traits. However, performing GWASs on millions of samples often leads to trade-offs between computational efficiency and statistical power, reducing the benefits of large-scale data collection efforts. We developed Quickdraws, a method that increases association power in quantitative and binary traits without sacrificing computational efficiency, leveraging a spike-and-slab prior on variant effects, stochastic variational inference and graphics processing unit acceleration.

Open-weight and API language / multimodal models (Qwen family) from Alibaba, widely used in research and products.

Robin is the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology. By integrating literature search agents with data analysis agents, Robin can generate testable hypotheses from literature and design experiments to validate them, automating the entire scientific discovery cycle for biological research.

Scientific discovery is driven by the iterative process of observation, hypothesis generation, experimentation, and data analysis. Despite recent advancements in applying artificial intelligence to biology, no system has yet automated all these stages. Here, we introduce Robin, the first multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology. By integrating literature search agents with data analysis agents, Robin can generate testable hypotheses from literature and design experiments to validate them.

scGPT is a generative pretrained transformer foundation model for single-cell biology, pretrained on over 33 million human cells from 51 organs across 441 studies. Uses a GPT architecture adapted for gene expression data with a specialized attention mask. Outperforms traditional methods on cell type annotation, multi-batch integration, multi-omic integration, perturbation response prediction, and gene network inference. Represents a foundational AI model for cellular biology analogous to GPT for natural language.

Generative pretrained models have achieved remarkable success in various domains such as language and computer vision. Using burgeoning single-cell sequencing data, we have constructed a foundation model for single-cell biology, scGPT, based on a generative pretrained transformer across a repository of over 33 million cells. Our findings illustrate that scGPT effectively distills critical biological insights concerning genes and cells. Through further adaptation of transfer learning, scGPT can be optimized to achieve superior performance across diverse downstream applications including cell type annotation, multi-batch integration, multi-omic integration, perturbation response prediction and gene network inference.

SciSciGPT is an open-source, prototype AI collaborator that uses the domain of science of science as a testbed to explore LLM-powered scientific collaboration. It assists researchers in analyzing scientific literature, identifying research trends, generating hypotheses about scientific dynamics, and facilitating human-AI collaborative research in metascience and scientometrics.

We introduce SciSciGPT, an open-source, prototype artificial intelligence (AI) collaborator that uses the domain of science of science as a testbed to explore the potential of large language model-powered scientific collaboration. SciSciGPT assists researchers in analyzing scientific literature, identifying research trends, generating hypotheses about scientific dynamics, and facilitating human-AI collaborative research in metascience and scientometrics.

SciToolAgent is a knowledge-graph-driven scientific agent that integrates multiple computational tools for scientific research. It leverages a knowledge graph to understand tool capabilities and relationships, enabling automated multi-tool workflow composition for complex scientific tasks across domains including bioinformatics, cheminformatics, and materials science.

Scientific research increasingly relies on specialized computational tools, yet effectively utilizing these tools requires substantial domain expertise. While large language models show promise in tool use, they struggle with complex multi-tool workflows. Here we introduce SciToolAgent, a knowledge-graph-driven scientific agent that integrates multiple computational tools for scientific research. By leveraging a knowledge graph to understand tool capabilities and relationships, SciToolAgent enables automated multi-tool workflow composition for complex scientific tasks.

ByteDance Seed LLM lineup (e.g. Seed 2.0 Pro/Lite/Mini, prior Seed 1.x) for chat, agents, multimodal, and production APIs—sibling to consumer products such as Doubao.

Sei is a deep learning model that generates a comprehensive map of regulatory activity from DNA sequence, predicting 21,907 chromatin features across cell types and contexts. Enables interpretation of noncoding variants in terms of specific regulatory functions and tissues, providing a global atlas of human cis-regulation.

regulatory activity map, chromatin, transcriptional regulation, deep learning, sequence-to-activity, human genetics, 21,907 features

Deciphering the impact of noncoding variants on gene regulation is a major challenge in human genetics. While deep learning models can accurately predict regulatory features from DNA sequence, interpreting these predictions to understand variant effects across diverse contexts remains difficult. Here we present Sei, a deep learning model that produces a comprehensive, sequence-based global map of regulatory activity. Sei predicts 21,907 chromatin profiles encompassing a wide range of cell types and regulatory features, and organizes them into 40 tissue-agnostic regulatory activities using a hierarchical model. We demonstrate that Sei accurately predicts regulatory effects and can identify disease-relevant variants across diverse conditions, enabling a more complete understanding of human genetic variation.

SPARK (System of Pathology Agents for Research and Knowledge) is a foundational agentic AI framework that uses language as a universal interface to autonomously generate biologically driven concepts for tumor analysis. It functions as a pathology 'brain' — an interconnected system of AI agents that autonomously reason, generate and implement biologically meaningful hypotheses as analytical tools without additional model training. SPARK uses four linked modules: idea generation (OpenAI o1), idea refinement, parameter coding (Claude Sonnet 3.5), and parameter verification. Evaluated across 18 patient cohorts spanning 5 cancer types and >5,400 patients, SPARK produced clinically and biologically relevant concepts correlated with prognosis, pathological variables, and predictive biomarkers, including patterns of tumor progression inferred from static images.

Artificial intelligence has advanced cancer pathology, but many systems still depend on hand-crafted features, are hard to explain and rely on fragmented workflows. We introduce SPARK (System of Pathology Agents for Research and Knowledge), a foundational agentic artificial intelligence approach that uses language as a universal interface to autonomously generate biologically driven concepts for tumor analysis. SPARK turns biological ideas into analytical tools and works directly with complex pathology data without extra model training. We evaluated SPARK across 18 patient cohorts spanning five cancer types (lung adenocarcinoma, lung squamous cell carcinoma, colorectal cancer, breast cancer and oropharyngeal squamous cell carcinoma) and more than 5,400 patients with available histopathology images and clinical/follow-up information, in both prognostic and predictive settings and on a well characterized spatial biology breast cancer dataset (n=625). We found that SPARK produced clinically and biologically relevant concepts correlated with prognosis, known pathological variables and predictive biomarkers, including patterns of tumor progression and temporal change inferred from static images. A dedicated module allows for human interaction with SPARK. All code, parameters and results are openly released.

SynSurr (Synthetic Surrogate analysis) is a method that makes GWAS on imputed phenotypes robust to imputation errors. Rather than replacing missing values, SynSurr jointly analyzes the observed and imputed data to provide calibrated association statistics, improving power for genome-wide association studies of partially missing phenotypes in population biobanks. Published in Nature Genetics.

Within population biobanks, incomplete measurement of certain traits limits the power for genetic discovery. Machine learning is increasingly used to impute the missing values from the available data. However, performing GWAS on imputed traits can introduce spurious associations. Here we introduce SynSurr analysis, which makes GWAS on imputed phenotypes robust to imputation errors by jointly analyzing observed and imputed data.

Therapeutics Data Commons (TDC) is a coordinated initiative providing AI-ready datasets and curated benchmarks across the full spectrum of therapeutic modalities (small molecules, biologics, gene therapy) and stages (target identification, hit discovery, lead optimization, manufacturing). Features 100+ datasets across 50+ learning tasks, with standardized evaluation protocols, data splits, and public leaderboards. Supports systematic evaluation of AI methods for drug discovery and development.

Artificial intelligence is poised to enable breakthroughs and discoveries in therapeutic science. Therapeutics Data Commons is a coordinated initiative to access and evaluate AI capability across therapeutic modalities and stages of discovery. The Commons is a resource with AI-solvable tasks, AI-ready datasets, and curated benchmarks, providing an ecosystem of tools, libraries, leaderboards, and community resources.

TITAN (Transformer-based pathology Image and Text Alignment Network) is a multimodal whole-slide foundation model from Mahmood Lab (Harvard/BWH). Pretrained on 335,645 WSIs via visual self-supervised learning and vision-language alignment with 423K synthetic captions from PathChat + 183K pathology reports. Without any fine-tuning, TITAN produces general-purpose slide representations for zero-shot classification, rare cancer retrieval, cross-modal retrieval, and pathology report generation. Outperforms both ROI and slide foundation models across diverse clinical tasks.

The field of computational pathology has been transformed with recent advances in foundation models that encode histopathology region-of-interests into versatile feature representations. However, translating these advancements to address complex clinical challenges at the patient and slide level remains constrained by limited clinical data. We propose TITAN, a multimodal whole-slide foundation model pretrained using 335,645 whole-slide images via visual self-supervised learning and vision-language alignment with pathology reports and 423,122 synthetic captions. Without any fine-tuning, 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.

AI-native IDE (ByteDance) with builder mode, multi-model chat, and agent workflows for full-stack development.

transferGWAS performs GWAS directly on full medical images using deep transfer learning: (1) a pretrained CNN (ResNet-based architecture, pretrained on ImageNet) extracts feature embeddings from raw images; (2) these learned representations are used as quantitative phenotypes for genetic association testing. Applied to UK Biobank retinal fundus images, identified 60 genomic regions including 7 novel candidate loci for eye-related traits. First demonstration of direct GWAS on whole images without predefined phenotype engineering.

deep transfer learning, pretrained CNN, ResNet, retinal fundus, whole-image GWAS, representation learning, UK Biobank

MOTIVATION: Medical images can provide rich information about diseases and their biology. However, investigating their association with genetic variation requires non-standard methods. We propose transferGWAS, a novel approach to perform genome-wide association studies directly on full medical images. First, we learn semantically meaningful representations of the images based on a transfer learning task, during which a deep neural network is trained on independent but similar data. Then, we perform genetic association tests with these representations. RESULTS: We validate the type I error rates and power of transferGWAS in simulation studies of synthetic images. Then we apply transferGWAS in a genome-wide association study of retinal fundus images from the UK Biobank. This first-of-a-kind GWAS of full imaging data yielded 60 genomic regions associated with retinal fundus images, of which 7 are novel candidate loci for eye-related traits and diseases.

UNI is a general-purpose self-supervised foundation model for computational pathology from Mahmood Lab (Harvard/BWH), pretrained on >100 million images from >100,000 H&E-stained WSIs (>77 TB) across 20 tissue types. Evaluated on 34 representative CPath tasks — outperforming prior models across cancer classification, organ transplant assessment, and rare disease analysis. Demonstrates resolution-agnostic classification, few-shot slide classification, and generalization to 108 cancer types in the OncoTree system. 1,300+ citations.

Quantitative evaluation of tissue images is crucial for computational pathology tasks. The high resolution of WSIs and the variability of morphological features present significant challenges. We introduce UNI, a general-purpose self-supervised model for pathology, pretrained using more than 100 million images from over 100,000 diagnostic H&E-stained WSIs across 20 major tissue types. The model was evaluated on 34 representative CPath tasks. UNI outperforms previous state-of-the-art models and demonstrates new capabilities including resolution-agnostic tissue classification, few-shot slide classification, and disease subtyping generalization to 108 cancer types.

Virchow is the first million-slide foundation model for computational pathology, developed by Paige in collaboration with Microsoft. A 632M-parameter ViT-H model trained using DINOv2 on 1.5 million H&E-stained WSIs from MSKCC (17 tissue types). Demonstrates clinical-grade pan-cancer detection with 0.95 AUC across nine common and seven rare cancers. With less training data, the pan-cancer detector built on Virchow achieves similar performance to tissue-specific clinical-grade models in production, outperforming them on rare cancer variants. Serves as the foundation for Paige's Virchow2 (3M WSIs, multimodal) and Virchow2G (1.8B parameters) models.

The analysis of histopathology images with artificial intelligence aims to enable clinical decision support systems and precision medicine. We present Virchow, the largest foundation model for computational pathology to date. In addition to the evaluation of biomarker prediction and cell identification, we demonstrate that a large foundation model enables pan-cancer detection, achieving 0.95 specimen-level AUC across nine common and seven rare cancers. With less training data, the pan-cancer detector built on Virchow achieved similar performance to tissue-specific clinical-grade models in production and outperformed them on some rare variants of cancer.

The Virtual Lab is an AI agent framework that uses LLM-powered researchers in a simulated laboratory environment to collaboratively design and test scientific hypotheses. It was demonstrated by successfully designing new SARS-CoV-2 nanobodies, with AI agents specializing in different scientific roles working together to propose, evaluate, and refine experimental designs.

Science frequently benefits from teams of interdisciplinary researchers, but many scientists do not have easy access to experts from multiple fields. Although large language models (LLMs) have shown an impressive ability to aid researchers across diverse domains, their uses have been largely limited to answering specific scientific questions rather than performing open-ended research. Here we expand the capabilities of LLMs for science by introducing the Virtual Lab, a framework where LLM-powered AI agents collaborate in a simulated laboratory to design and test scientific hypotheses. The Virtual Lab successfully designed new SARS-CoV-2 nanobodies, demonstrating the potential of multi-agent AI systems for open-ended scientific discovery.

Agentic IDE from Codeium with Cascade flow for multi-step edits, terminal integration, and deep workspace awareness.