Biology subjects

Leskovec, J.

Publications and source records attributed to Leskovec, J..

At least 19 recordsLinked to original sources

bioRxiv2026

AutoScreen: AI Co-Scientist System for Target Discovery in Functional Genomics

Target discovery in functional genomics remains largely manual and time-consuming, lacking systematic tools for efficient and reproducible gene-level hypothesis generation. We introduce AutoScreen, an AI co-scientist system supporting target discovery through both Pre-screen Design, which constructs perturbation libraries de novo from free-text research descriptions, and Post-screen Analysis, which re-ranks experimental screen hits by integrating statistical scores with biological context. AutoScreen leverages the complementary strengths of multiple specialist agents that perform deep research into multi-modal evidence, information restructuring, parallel searches across >26 biomedical databases, evidence synthesis, and target review to provide transparent, explainable gene prioritization with pipeline provenance. Across 320 genome-scale CRISPR screens as expert-curated benchmarks, AutoScreen achieved a ~19% increase in validated-hit recovery among its top 100 predictions, relative to the strongest agent baseline, and required a 1.2-fold smaller library to recover the same number of hits at the top-500 reference point. AutoScreen reached mean average precision more than two orders-of-magnitude above random baseline. Further, we validated AutoScreen in cancer immune-evasion case studies focusing on natural killer (NK) and T-cell therapeutics. AutoScreen recovered NK-resistance genes that were initially lower-ranked in a leukemia screen, moving MUC1, PDPN, and LRRC15 from raw ranks of 118, 81, 1384 to 5, 44, 659, respectively. In follow-up tumor killing assay with primary human NK cells, individual perturbation validated all three hits successfully. Next, in prospective benchmarking across two cytotoxic T-cell-killing screens, AutoScreen recovered 77.1% of ground-truth hits called by the consensus of two gold-standard analysis pipelines (FDR<0.1), achieving 8% improvement over the strongest general-purpose LLM agent baseline. Finally, we re-analyzed >500 public datasets to construct the AutoScreen Resource Hub, a growing knowledge base of pre-computed, annotated reports for CRISPR screens, RNA-seq differential expression, and gene-level UK Biobank genome-wide association studies. This agentic AI approach enables real-time genomics benchmark construction that are continuously updated, expanded, for evaluating frontier AI co-scientists. Overall, AutoScreen enables AI-powered target discovery to be more auditable, scalable, extensible, and reproducible.

bioengineering↗

bioRxiv2026

Hypothalamic Cell Type Evolution across Vertebrates

The hypothalamus regulates vital functions and is highly conserved among vertebrates occupying diverse niches, but little is known about the evolution of its cell types. Using comparative single-cell transcriptomics across major vertebrate classes, we identified deeply conserved and clade-specific cell types. While homologous cell types express conserved transcription factors and neuropeptides, their regulatory linkages are often rewired across evolution. The paraventricular nucleus largely retains its neuropeptide expression and responds to dehydration with fine cellular resolution from amphibians to mammals, whereas lamina terminalis neurons are specific to terrestrial vertebrates and exhibit gene expression divergence in cell types activated by dehydration and heat. We propose that a conserved set of keystone cell types sustain essential functions across vertebrates, while flexible gene regulatory programs alongside numerous clade-specific cell types offer evolutionary fluidity.

evolutionary biology↗

bioRxiv2026

Perturb-ME: Scalable mechanism discovery from phenotype-enriched genome-wide screens

Perturb-seq enables pooled genetic screens with rich single-cell profiling readouts, but genome-scale profiling remains costly and may not be associated with other established functional characteristics. Moreover, as screens grow in size and complexity, interpreting the resulting data comprehensively is challenging and slow. Here, we introduce Perturb-seq with Marker Enrichment (Perturb-ME), which combines genome-scale CRISPR screening, phenotype-based enrichment and multimodal single-cell profiling. Applied to MHC-I cell surface protein expression in melanoma, Perturb-ME profiled HLA-low and HLA-high cells with matched RNA, surface-protein and guide measurements. A regulatory model with 221 impactful regulators affecting 1,998 responsive genes recovered seven coherent co-functional regulatory modules governing nine gene programs, including the canonical IFN{gamma}-MHC-I axis regulating an antigen-presentation and interferon-response program. Agentic interpretation of the entire model with an AI co-scientist linked additional modules to trafficking, proteostasis and chromatin regulation. Perturb-ME, along with agentic interpretation, provide a scalable framework for comprehensive functional discovery from phenotype-enriched genetic screens.

cell biology↗

bioRxiv2026

Bio-BLIP: A Multimodal Architecture for Transferable Reasoning in Genomic Variant Interpretation

Developing scientific hypotheses in biology requires integrating heterogeneous evidence across DNA sequence, gene context, protein function, and prior literature. Existing multimodal AI systems expose biological evidence to reasoning models through textification or by projecting biological embeddings into fine-tuned language models. However, these models are typically highly optimized the specific set of tasks for which they are fine-tuned. Here we present Bio-BLIP, a multimodal Q-former based architecture which leverages biological embeddings and a LLM to generalize to complex reasoning tasks without task-specific fine-tuning. The key to Bio-BLIP is a new neural network architecture that integrates four data modalities - DNA, genes, proteins, and text - through a master Qformer model, which integrates the modality-specific information into a fixed-length prefix for the LLM backbone. Bio-BLIP is pretrained on the task of human genetic variant annotation and achieves a 29.8% increase in generating accurate variant features over frontier LLMs. We evaluate Bio-BLIP zero-shot on downstream genomic tasks of variant prioritization and target gene prediction. Bio-BLIP outperforms two alignment-free genomic language models on regulatory variant prioritization for Mendelian disease. Across the target gene prediction task, Bio-BLIP improves accuracy over LLMs by leveraging learned genomic variant knowledge in difficult cases. Our model produces rich, transparent reasoning traces. In biological domains characterized by multiple scales of data and varied downstream tasks, Bio-BLIP offers a step toward natively multimodal, generalizable reasoning.

genomics↗

bioRxiv2026

BiomniBench: Process-level Evaluation of LLM Agents for Real-world Biomedical Research

LLM agents now perform real biomedical research, but evaluating them rigorously is hard. Outcome-only benchmarks fail in two ways. First, a correct final answer can come from memorization, reward hacking, or wrong reasoning that produces the right number by chance. Second, valid alternative analyses are marked wrong simply because they differ from the reference. We introduce BiomniBench, a process-level evaluation framework that scores the full agent trajectory against expert-designed, task-specific rubrics. Our first release, BiomniBench-DA, contains 100 data-analysis tasks across 17 task types, 5 disease areas, and a general-biology category, each based on a paper from journals such as Nature, Cell, and Science and co-developed with an original author or a domain expert. Benchmarking frontier and open-weight models across four agent harnesses reveals three findings. Frontier and open-weight bases cluster within a few points of each other, with substantial headroom for all models. The agent harness shifts scores by more than the gap between successive model generations. Agents reliably ground claims in real sources yet consistently fall short on method selection, biological interpretation, and scientific reasoning. BiomniBench is the first process-level benchmark for LLM agents in biomedical research, providing the dimension-level diagnostics that outcome scoring cannot. Datasethuggingface.co/datasets/phylobio/BiomniBench-DA

bioinformatics↗

bioRxiv2026

TxConformal: Controlling False Discoveries in AI-Driven Therapeutic Discovery

Artificial Intelligence (AI) is transforming therapeutic discovery by scoring a large set of promising candidates and prioritizing a shortlist for further investigation. Quantifying the reliability of AI scores and preventing false positives among selected candidates is key to the efficiency of the discovery process. Conformal prediction (CP) has emerged as a popular tool for guiding such prioritization, especially via the conformal selection framework to control false discovery rates (FDR) in selecting top-ranked candidates under distributional shift1, 2. However, deploying these advances in real-world therapeutic discovery remains challenging: distribution shifts are difficult to quantify and correct in high-dimensional biomedical data, and practical workflows often require flexible error metrics. Here, we present TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW, a general framework for trustworthy decision making when building shortlists using AI scores. TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW adjusts for distribution shift by balancing the hidden representations in AI models and then provides confidence measures for true discoveries of target biological properties. These confidence measures, interpretable as p-values, can be used in conjunction with statistical multiple testing procedures to derive selection decisions with limited false positives or to estimate the errors in given selection decisions. TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW controls the false positive rate in six real-world tasks spanning various therapeutic discovery stages, modalities, and AI models with realistic data splits. When selecting promising combinatorial genetic perturbations, TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW nearly halves false-positive selections compared to baseline methods, substantially reducing unnecessary experimental costs by tens of thousands of dollars. When selecting stable protein structures under mutant shifts, TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW identifies about 10 times more proteins than baseline methods at stringent thresholds when running at a target FDR level of 10%, recovering over 90% of valuable candidates that baseline methods miss due to unaccounted distribution shifts. Furthermore, we demonstrate that TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW robustly supports various alternative error metrics suitable for resource-constrained settings. Finally, in a prospective fixed-budget virtual screening campaign for novel antibiotic discovery, TO_SCPLOWXC_SCPLOWCO_SCPLOWONFORMALC_SCPLOW predicted false positives in close agreement with experimental outcomes, with substantial improvements over simple baselines.

bioinformatics↗

bioRxiv2026

Are Current AI Virtual Cell Models Useful for Scientific Discovery?

AI models are increasingly developed to predict the effect of perturbations on gene expression, but current benchmarks fail to reliably measure model performance. Here, we argue that new benchmarks that directly measure the value of model predictions for specific scientific discovery outcomes are needed to address this gap. We present PerturbHD, an evaluation framework for AI-enabled hit discovery, to demonstrate the benefits our proposed approach.

bioinformatics↗

bioRxiv2025

PULSAR: a Foundation Model for Multi-scale and Multicellular Biology

Biology emerges from interactions across physical scales, where molecular interactions drive cellular states, which in turn orchestrate multicellular tissue functions that collectively define health and disease. However, current computational models are often constrained to single scales in isolation, failing to integrate the biology that emerges from lower to higher levels [1]. Here we present PULSAR (Patient Understanding Leveraging Single-cell universAl Representation), a multi-scale and multicellular foundation model architecture that explicitly enables information flow from genes to cells to multicellular systems. Applied to the human peripheral immune system, PULSAR extracts a unified donor representation that supports rapid disease classification, biomarker prediction, and forecasting of future clinical events, such as Rheumatoid arthritis onset. As a generative model, PULSAR enables the simulation of cytokine perturbation response across physical resolutions, while its interpretability reveals the key cell types driving disease. Overall, PULSAR opens new avenues for precision medicine by enabling computational reasoning that connects molecular biology to clinical phenotypes.

bioinformatics↗

bioRxiv2025

A deep reinforcement learning platform for antibiotic discovery

Antimicrobial resistance (AMR) is projected to cause up to 10 million deaths annually by 2050, underscoring the urgent need for new antibiotics. Here we present ApexAmphion, a deep-learning framework for de novo design of antibiotics that couples a 6.4-billion-parameter protein language model with reinforcement learning. The model is first fine-tuned on curated peptide data to capture antimicrobial sequence regularities, then optimised with proximal policy optimization against a composite reward that combines predictions from a learned minimum inhibitory concentration (MIC) classifier with differentiable physicochemical objectives. In vitro evaluation of 100 designed peptides showed low MIC values (nanomolar range in some cases) for all candidates (100% hit rate). Moreover, 99 our of 100 compounds exhibited broad-spectrum antimicrobial activity against at least two clinically relevant bacteria. The lead molecules killed bacteria primarily by potently targeting the cytoplasmic membrane. By unifying generation, scoring and multi-objective optimization with deep reinforcement learning in a single pipeline, our approach rapidly produces diverse, potent candidates, offering a scalable route to peptide antibiotics and a platform for iterative steering toward potency and developability within hours.

molecular biology↗

bioRxiv2025

Predicting cellular responses to perturbation across diverse contexts with STATE

Cellular responses to perturbations are a cornerstone for understanding biological mechanisms and selecting drug targets. While machine learning models offer tremendous potential for predicting perturbation effects, they currently struggle to generalize to unobserved cellular contexts. Here, we introduce SO_SCPLOWTATEC_SCPLOW, a transformer model that predicts perturbation effects while accounting for cellular heterogeneity within and across experiments. SO_SCPLOWTATEC_SCPLOW predicts perturbation effects across sets of cells and is trained using gene expression data from over 100 million perturbed cells. SO_SCPLOWTATEC_SCPLOW improved discrimination of effects on large datasets by more than 30% and identified differentially expressed genes across genetic, signaling and chemical perturbations with significantly improved accuracy. Using its cell embedding trained on observational data from 167 million cells, SO_SCPLOWTATEC_SCPLOW identified strong perturbations in novel cellular contexts where no perturbations were observed during training. We further introduce Cell-Eval, a comprehensive evaluation framework that highlights SO_SCPLOWTATEC_SCPLOWs ability to detect cell type-specific perturbation responses, such as cell survival. Overall, the performance and flexibility of SO_SCPLOWTATEC_SCPLOW sets the stage for scaling the development of virtual cell models.

systems biology↗

bioRxiv2025

Biomni: A General-Purpose Biomedical AI Agent

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, underscoring the need for a fundamentally new way to scale scientific expertise. 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 large language model (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, micro-biome analysis, and molecular cloning - without any task-specific prompt tuning. Real-world case studies further showcase Biomnis ability to interpret complex, multi-modal biomedical datasets and autonomously generate experimentally testable protocols. Biomni envisions a future where virtual AI biologists operate alongside and augment human scientists to dramatically enhance research productivity, clinical insight, and healthcare. Biomni is ready to use at https://biomni.stanford.edu, and we invite scientists to explore its capabilities, stress-test its limits, and co-create the next era of biomedical discoveries.

bioinformatics↗

bioRxiv2025

SpatialAgent: An Autonomous AI Agent for Spatial Biology

Advances in AI are transforming scientific discovery, yet spatial biology, a field that deciphers the molecular organization within tissues, remains constrained by labor-intensive workflows. Here, we present SpatialAgent, an autonomous AI agent for spatial biology research. SpatialAgent couples large language models with a Plan-Act-Conclude architecture, dynamic tool and skill retrieval, multimodal interpretation, and verification modules that audit generated claims. It supports the full discovery loop, from gene-panel design and multimodal annotation to trajectory inference, cell-cell communication analysis, imputation, and hypothesis generation. Across human and mouse brain, heart, tonsil, colon, and prostate datasets, SpatialAgent outperformed established computational baselines and matched or surpassed expert scientists in key tasks. In open-ended case studies, it recovered known tissue organization and generated spatially grounded hypotheses. In a prospective mouse prostate cancer Xenium study, it designed a compact 100-gene add-on panel that profiled 4.2 million cells across 21 samples, improved cell-type and malignant-state prediction, and captured spatially structured tumor and microenvironment programs. SpatialAgent establishes a framework for autonomous and collaborative discovery in spatial biology.

bioinformatics↗

bioRxiv2024

PreciCE: Precision engineering of cell fates via data-driven multi-gene control of transcriptional networks

The directed differentiation of stem cells into specific cell types is critical for regenerative medicine and cell-based applications. However, current methods for cell fate control are inefficient, imprecise, and rely on laborious trial-and-error. To address these limitations, we present a method for data-driven multi-gene modulation of transcriptional networks. We develop bidirectional CRISPR-based tools based on dCas12a, Cas13d, and dCas9 for simultaneously activating and repressing many genes. Due to the vast combinatorial complexity of multi-gene regulation, we introduce a machine learning-based computational algorithm that uses single-cell RNA sequencing data to predict multi-gene perturbation sets for converting a starting cell type into a desired target cell type. By combining these technologies, we establish a unified workflow for data-driven cell fate engineering and demonstrate its efficacy in controlling early stem cell differentiation while suppressing alternative lineages through logic-based cell fate operations. This approach represents a significant advancement in the use of synthetic biology to engineer cell identity.

synthetic biology↗

bioRxiv2024

Metric Mirages in Cell Embeddings

Although biological studies increasingly rely on embeddings of single cell profiles, the quality of these embeddings can be challenging to assess. Such evaluations are especially important for avoiding misleading biological interpretations, assessing the accuracy of integration methods, and establishing the zero-shot capabilities of foundational models. Here, we posit that current evaluation metrics can be highly misleading. We show this by training a three-layer perceptron, Islander , which outperforms all 11 leading embedding methods on a diverse set of cell atlases, but in fact distorts biological structures, limiting its utility for biological discovery. We then present a metric, scGraph, to flag such distortions. Our work should help learn more robust and reliable cell embeddings.

bioinformatics↗

bioRxiv2023

Universal Cell Embeddings: A Foundation Model for Cell Biology

Developing a universal representation space for cells which encompasses the tremendous molecular diversity of cell types within the human body and more generally, across species, would be transformative for cell biology. Recent work using single-cell transcriptomic approaches to create molecular definitions of cell types in the form of cell atlases has provided the necessary data for such an endeavor. Here, we present the Universal Cell Embedding (UCE) foundation model. UCE was trained on a corpus of cell atlas data from human and other species in a completely self-supervised way without any data annotations. UCEs modeling approach is to create a unified biological latent space that can represent cells across diverse tissues and species. This universal cell embedding captures important biological variation despite the presence of experimental noise across diverse datasets. An important aspect of UCEs universality is that new cells can be mapped to this embedding space with no additional data labeling, model training or fine-tuning. We applied UCE to create the Integrated Mega-scale Atlas, embedding 36 million cells, with more than 1,000 uniquely named cell types, from hundreds of experiments, dozens of tissues and eight species. We uncovered new insights about the organization of cell types and tissues within this universal cell embedding space, and leveraged it to infer function of newly discovered cell types. UCEs embedding space exhibits emergent behavior, uncovering biology that it was never explicitly trained for, such as identifying developmental lineages and embedding data from novel species not included in the training set. Overall, by enabling a universal representation for every cell state and type, UCE provides a valuable tool for analysis, annotation and hypothesis generation over single cell data.

cell biology↗

bioRxiv2023

Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN

Analysis of single-cell datasets generated from diverse organisms offers unprecedented opportunities to unravel fundamental evolutionary processes of conservation and diversification of cell types. However, inter-species genomic differences limit the joint analysis of cross-species datasets to homologous genes. Here, we present SATURN, a deep learning method for learning universal cell embeddings that encodes genes biological properties using protein language models. By coupling protein embeddings from language models with RNA expression, SATURN integrates datasets profiled from different species regardless of their genomic similarity. SATURN has a unique ability to detect functionally related genes co-expressed across species, redefining differential expression for cross-species analysis. We apply SATURN to three species whole-organism atlases and frog and zebrafish embryogenesis datasets. We show that cell embeddings learnt in SATURN can be effectively used to transfer annotations across species and identify both homologous and species-specific cell types, even across evolutionarily remote species. Finally, we use SATURN to reannotate the five species Cell Atlas of Human Trabecular Meshwork and Aqueous Outflow Structures and find evidence of potentially divergent functions between glaucoma associated genes in humans and other species.

bioinformatics↗

bioRxiv2022

Aging Fly Cell Atlas Identifies Exhaustive Aging Features at Cellular Resolution

Aging is characterized by a decline in tissue function, but the underlying changes at cellular resolution across the organism remain unclear. Here, we present the Aging Fly Cell Atlas, a single-nucleus transcriptomic map of the whole aging Drosophila. We characterize 163 distinct cell types and perform an in-depth analysis of changes in tissue cell composition, gene expression, and cell identities. We further develop aging clock models to predict the fly age and show that ribosomal gene expression is a conserved predictive factor for age. Combining all aging features, we find unique cell type-specific aging patterns. This atlas provides a valuable resource for studying fundamental principles of aging in complex organisms.

developmental biology↗

bioRxiv2022

GEARS: Predicting transcriptional outcomes of novel multi-gene perturbations

Cellular response to genetic perturbation is central to numerous biomedical applications from identifying genetic interactions involved in cancer to methods for regenerative medicine. However, the combinatorial explosion in the number of possible multi-gene perturbations severely limits experimental interrogation. Here, we present GEARS, a method that can predict transcriptional response to both single and multi-gene perturbations using single-cell RNA-sequencing data from perturbational screens. GEARS is uniquely able to predict outcomes of perturbing combinations consisting of novel genes that were never experimentally perturbed by leveraging geometric deep learning and a knowledge graph of gene-gene relationships. GEARS has higher precision than existing approaches in predicting five distinct genetic interaction subtypes and can identify the strongest interactions more than twice as well as prior approaches. Overall, GEARS can discover novel phenotypic outcomes to multi-gene perturbations and can thus guide the design of perturbational experiments.

bioinformatics↗