Biology subjects

Zitnik, M.

Publications and source records attributed to Zitnik, M..

At least 19 recordsLinked to original sources

bioRxiv2026

Medea: An omics AI agent for therapeutic discovery

Therapeutic hypotheses can transfer across diseases but their relevance depends on biological context. The same target, perturbation, or treatment can produce different effects across cell types, disease states, genetic backgrounds, and patients. Therapeutic reasoning therefore requires methods that preserve context, test when evidence supports transfer, and identify where context-specific effects limit it. Although AI agents can perform therapeutic analyses, existing systems often fail to preserve biological context over long workflows, verify intermediate computational steps, or reconcile conflicting evidence across datasets and literature. Here, we present Medea, an AI agent for therapeutic reasoning across biological contexts. Medea executes multi-step analyses using biological tools, machine learning models, and literature retrieval while enforcing verification during planning, execution, and evidence synthesis. We evaluate Medea across 5,673 open-ended analyses in three domains: cell type specific therapeutic target nomination in five diseases and 29 cell types, synthetic lethality prediction in 7 cancer cell lines, and immunotherapy response prediction from multimodal patient profiles. Using a previously unpublished epistatic miniarray profiling screen performed under two DNA-damaging treatments, we evaluate Medea on predicting synthetic lethality among 238,046 gene-gene pairs in yeast. Medea predicts these experimentally measured synthetic lethal interactions, indicating that its performance reflects biological relevance rather than information leakage from benchmark datasets. Across these evaluations, Medea improves performance over large language models, reasoning models, biomedical agents, and specialized machine learning models while maintaining low failure rates and calibrated abstention. These results show that verifiable AI agents can perform therapeutic analyses across biological contexts.

bioinformatics↗

bioRxiv2025

CONCERT predicts niche-aware perturbation responses in spatial transcriptomics

Spatial perturbation transcriptomics measures how genetic or chemical edits alter gene expression while preserving tissue context. Perturbation outcomes depend on a cells intrinsic state and also on how effects propagate across cellular microenvironments. We present CONCERT, a niche-aware generative model that embeds perturbation context and learns spatial kernels with a Gaussian process variational autoencoder to predict perturbation effects across tissue. We formalize three tasks: patch, border, and niche, predicting responses in nearby unperturbed regions, at tissue interfaces, and as a function of surrounding microenvironments. We evaluate CONCERT on Perturb-map lung datasets. CONCERT outperforms state-of-the-art models (dissociated counterfactuals, spatialized perturbation models, and kNN), reducing E-distance by up to 33.77% (patch), 26.05% (border), and 33.74% (niche) versus the next best, with mean absolute error down by up to 23.28% and Pearson correlation up by up to 9.10%. Two case studies go beyond benchmarking. In dextran sodium sulfate-induced colitis, CONCERT reconstructs spatial gene expression at unmeasured time points, produces longitudinal comparisons across unpaired mice, resolves intermouse heterogeneity, and recovers consistent temporal declines of inflammation-associated genes across regions. In ischemic stroke, CONCERT predicts responses under variable lesion sizes and in a 3D formulation across brain sections, capturing lesion-core and peri-lesion patterns. CONCERT performs niche-aware counterfactual prediction, reconstructs missing spatial data, and models perturbation responses across tissues.

bioinformatics↗

bioRxiv2025

ATOMICA: Learning Universal Representations of Intermolecular Interactions

Molecular interactions underlie nearly all biological processes, yet most representation models describe isolated entities or specialize in a single molecular setting. Here, we introduce ATOMICA, an interaction-centered geometric deep learning model designed to learn transferable representations of intermolecular interfaces across proteins, small molecules, metal ions, and nucleic acids. Self-supervised pretraining on 2,037,972 interaction complexes yields representations spanning atoms, molecular building blocks, and complete interfaces. The latent space captures molecular identity and interaction context, supporting sequence recovery and zero-shot prioritization of residues involved in non-covalent interactions. ATOMICA provides structural information complementary to sequence representations on RNA and protein-pocket ligand classification. Across protein-pocket analyses, ATOMICA distinguishes ATP- and ADP-associated pocket states and retrieves ligand-matched pockets across proteins without detectable structural alignment. The latent space also enables cross-modal comparison, with orthosteric inhibitor embeddings retrieving regions proximal to native peptide and protein interfaces. Applied to the dark proteome, ATOMICA-Ligand predicts candidate ions or cofactors for 2,646 pockets, and five heme candidates show Soret-band shifts consistent with heme association. Together, these results show how interaction-centered molecular representations can transfer structural information across molecular interaction types and generate experimentally testable hypotheses.

bioinformatics↗

bioRxiv2025

Fleming: An AI Agent for Antibiotic Discovery in Mycobacterium Tuberculosis

Antibiotic development is challenged by high costs and failure rates. Artificial intelligence (AI) holds promise to overcome these challenges by predicting inhibitory properties of novel compounds, generating new candidates, and contextualizing property predictions in the biological background. Fleming is an integrative AI agent that explores novel chemical space to identify lead compounds meeting multiple criteria. The discriminative and generative AI models for Mycobacterium tuberculosis (Mtb) inhibition were trained on a set of 114,900 diverse compounds and fragments based on in vitro growth inhibition. We combined both models as well as molecular optimization, ADMET prediction and literature search functions to make Fleming an integrated agent for Mtb preclinical lead identification. Fleming has 17% higher discrimination between known Mtb leads and leads for other diseases than a generic LLM agent along with 13% higher discrimination than molecular property prediction alone on challenging ADMET tasks. Fleming demonstrates an 83% in vitro hit rate of predicted inhibition and a 100% hit rate of de novo generative design. Flemings generative designs also demonstrate an 83% rate of favorable ADMET profiles. Fleming is an integrative AI agent able to explore new regions of the chemical space to select lead compounds that simultaneously meet several desirable criteria.

bioinformatics↗

bioRxiv2025

Phyla: Towards a Foundation Model for Phylogenetic Inference

Protein language models (PLMs) are often assumed to capture evolutionary information by training on large protein sequence datasets. Yet it remains unclear whether PLMs can reason about evolution--that is, infer evolutionary relationships between sequences. We test this capability by evaluating whether standard PLM usage, frozen or fine-tuned embeddings with distance-based comparison, supports evolutionary reasoning. Existing PLMs consistently fail to recover phylogenetic structure, despite strong performance on sequence-level tasks such as masked-token and contact prediction. We present PO_SCPLOWHYLAC_SCPLOW, a hybrid state-space and transformer model that jointly processes multiple sequences and is trained using a tree-based objective across 3,000 phylogenies spanning diverse protein families. PO_SCPLOWHYLAC_SCPLOW outperforms the next-best PLM by 9% on tree reconstruction and 23% on taxonomic clustering while remaining alignment- and guide-tree-free. Although classical alignment pipelines achieve higher absolute accuracy, PO_SCPLOWHYLAC_SCPLOW narrows the gap and achieves markedly lower end-to-end runtime. Applied to real data, PO_SCPLOWHYLAC_SCPLOW reconstructs biologically accurate clades in the tree of life and resolves genome-scale relationships among Mycobacterium tuberculosis isolates. These findings suggest that, under standard usage, evolutionary reasoning does not reliably emerge from large-scale sequence modeling. Instead, PO_SCPLOWHYLAC_SCPLOW shows that models trained with phylogenetic supervision can reason about evolution more effectively, offering a biologically grounded path toward evolutionary foundation models.

bioinformatics↗

bioRxiv2024

ProCyon: A multimodal foundation model for protein phenotypes

Characterizing human proteins remains a major challenge: approximately 29% of human proteins lack experimentally validated functions and even well-annotated proteins often lack context-specific phenotypic insights. To enable universal modeling of protein phenotypes, we present PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW, a multimodal foundation model that utilizes protein sequence, structure, and natural language for generating and predicting protein phenotypes across diverse knowledge domains. PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW is trained on our novel dataset, PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW-IO_SCPLOWNSTRUCTC_SCPLOW, with 33 million protein phenotype instructions. On dozens of benchmarking tasks, PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW performs competitively against single-modal and multimodal models. Further, PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW conditionally retrieves proteins via mechanisms of action of small molecule drugs and disease contexts, and it generates candidate phenotypic descriptions for poorly characterized proteins, including those implicated in Parkinsons disease that were identified after PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOWs knowledge cutoff date. We experimentally confirm PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOWs predictions in multiple sclerosis using post-mortem brain RNA-seq, identifying novel MS genes and elucidating associated pathway mechanisms consistent with cortical pathology. PO_SCPLOWROC_SCPLOWCO_SCPLOWYONC_SCPLOW paves the way toward a general approach to generate functional insights into the human proteome.

bioinformatics↗

bioRxiv2024

FoldMark: Protecting Protein Generative Models with Watermarking

Proteins are the principal architects of life, fueling advances in bioengineering, drug discovery, and synthetic biology. The integration of generative AI with computational protein science has revolutionized protein design while also posing dual-use risks, such as enabling the creation of pandemic-capable proteins that require strong biosecurity safeguards. Here, we introduce FoldMark, a first-of-its-kind watermarking strategy leveraging distributional and evolutionary principles tailored for protein generative models, balancing watermark capacity and structural quality. FoldMark achieves over 95% watermark bit accuracy at 32 bits with minimal impact on structural quality (>0.9 scTM scores) for leading models including AlphaFold3, ESMFold, RFDiffusion, and RFDiffusionAA. For user tracing, FoldMark can successfully trace up to 1 million users. To validate FoldMark in wet lab, we applied it to structure-based design of EGFP and CRISPR-Cas13, showing wildtype-level function (98% fluorescence, 95% editing efficiency) and >90% watermark detection, demonstrating its practical utility for safeguarding AI-driven protein research.

bioinformatics↗

bioRxiv2024

TDC-2: Multimodal Foundation for Therapeutic Science

Drug discovery AI datasets and benchmarks have not traditionally included single-cell analysis biomarkers. While benchmarking efforts in single-cell analysis have recently released collections of single-cell tasks, they have yet to comprehensively release datasets, models, and benchmarks that integrate a broad range of therapeutic discovery tasks with cell-type-specific biomarkers. Therapeutics Commons (TDC-2) presents datasets, tools, models, and benchmarks integrating cell-type-specific contextual features with ML tasks across therapeutics. We present four tasks for contextual learning at single-cell resolution: drug-target nomination, genetic perturbation response prediction, chemical perturbation response prediction, and protein-peptide interaction prediction. We introduce datasets, models, and benchmarks for these four tasks. Finally, we detail the advancements and challenges in machine learning and biology that drove the implementation of TDC-2 and how they are reflected in its architecture, datasets and benchmarks, and foundation model tooling.

bioinformatics↗

bioRxiv2024

PocketGen: Generating Full-Atom Ligand-Binding Protein Pockets

Designing protein-binding proteins is critical for drug discovery. However, the AI-based design of such proteins is challenging due to the complexity of ligand-protein interactions, the flexibility of ligand molecules and amino acid side chains, and sequence-structure dependencies. We introduce PocketGen, a deep generative model that simultaneously produces both the residue sequence and atomic structure of the protein regions where ligand interactions occur. PocketGen ensures consistency between sequence and structure by using a graph transformer for structural encoding and a sequence refinement module based on a protein language model. The bilevel graph transformer captures interactions at multiple scales, including atom, residue, and ligand levels. To enhance sequence refinement, PocketGen integrates a structural adapter into the protein language model, ensuring that structure-based predictions align with sequence-based predictions. PocketGen can generate high-fidelity protein pockets with superior binding affinity and structural validity. It operates ten times faster than physics-based methods and achieves a 95% success rate, defined as the percentage of generated pockets with higher binding affinity than reference pockets. Additionally, it attains an amino acid recovery rate exceeding 64%.

bioinformatics↗

bioRxiv2024

Evaluating generalizability of artificial intelligence models for molecular datasets

Deep learning has made rapid advances in modeling molecular sequencing data. Despite achieving high performance on benchmarks, it remains unclear to what extent deep learning models learn general principles and generalize to previously unseen sequences. Benchmarks traditionally interrogate model generalizability by generating metadata based (MB) or sequence-similarity based (SB) train and test splits of input data before assessing model performance. Here, we show that this approach mischaracterizes model generalizability by failing to consider the full spectrum of cross-split overlap, i.e., similarity between train and test splits. We introduce SPECTRA, a spectral framework for comprehensive model evaluation. For a given model and input data, SPECTRA plots model performance as a function of decreasing cross-split overlap and reports the area under this curve as a measure of generalizability. We apply SPECTRA to 18 sequencing datasets with associated phenotypes ranging from antibiotic resistance in tuberculosis to protein-ligand binding to evaluate the generalizability of 19 state-of-the-art deep learning models, including large language models, graph neural networks, diffusion models, and convolutional neural networks. We show that SB and MB splits provide an incomplete assessment of model generalizability. With SPECTRA, we find as cross-split overlap decreases, deep learning models consistently exhibit a reduction in performance in a task- and model-dependent manner. Although no model consistently achieved the highest performance across all tasks, we show that deep learning models can generalize to previously unseen sequences on specific tasks. SPECTRA paves the way toward a better understanding of how foundation models generalize in biology.

bioinformatics↗

bioRxiv2024

Combinatorial prediction of therapeutic perturbations using causally-inspired neural networks

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

bioinformatics↗

bioRxiv2023

Contextualizing protein representations using deep learning on protein networks and single-cell data

Understanding protein function and developing molecular therapies require deciphering the cell types in which proteins act as well as the interactions between proteins. However, modeling protein interactions across biological contexts remains challenging for existing algorithms. Here, we introduce PO_SCPLOWINNACLEC_SCPLOW, a geometric deep learning approach that generates context-aware protein representations. Leveraging a multi-organ single-cell atlas, PO_SCPLOWINNACLEC_SCPLOW learns on contextualized protein interaction networks to produce 394,760 protein representations from 156 cell type contexts across 24 tissues. PO_SCPLOWINNACLEC_SCPLOWs embedding space reflects cellular and tissue organization, enabling zero-shot retrieval of the tissue hierarchy. Pretrained protein representations can be adapted for downstream tasks: enhancing 3D structure-based representations for resolving immuno-oncological protein interactions, and investigating drugs effects across cell types. PO_SCPLOWINNACLEC_SCPLOW outperforms state-of-the-art models in nominating therapeutic targets for rheumatoid arthritis and inflammatory bowel diseases, and pinpoints cell type contexts with higher predictive capability than context-free models. PO_SCPLOWINNACLEC_SCPLOWs ability to adjust its outputs based on the context in which it operates paves way for large-scale context-specific predictions in biology.

bioinformatics↗

bioRxiv2022

Building a knowledge graph to enable precision medicine

Developing personalized diagnostic strategies and targeted treatments requires a deep understanding of disease biology and the ability to dissect the relationship between molecular and genetic factors and their phenotypic consequences. However, such knowledge is fragmented across publications, non-standardized research repositories, and evolving ontologies describing various scales of biological organization between genotypes and clinical phenotypes. Here, we present PrimeKG, a precision medicine-oriented knowledge graph that provides a holistic view of diseases. PrimeKG integrates 20 high-quality resources to describe 17,080 diseases with 4,050,249 relationships representing ten major biological scales, including disease-associated protein perturbations, biological processes and pathways, anatomical and phenotypic scales, and the entire range of approved and experimental drugs with their therapeutic action, considerably expanding previous efforts in disease-rooted knowledge graphs. In addition, PrimeKG supports artificial intelligence analyses of how drugs might target disease-associated molecular perturbations by containing an abundance of indications, contradictions, and off-label use drug-disease edges lacking in other knowledge graphs. We accompany PrimeKGs graph structure with text descriptions of clinical guide-lines to enable multimodal analyses.

bioinformatics↗

bioRxiv2021

Multidose evaluation of 6,710 drug repurposing library identifies potent SARS-CoV-2 infection inhibitors In Vitro and In Vivo.

Identification of host factors contributing to replication of viruses and resulting disease progression remains a promising approach for development of new therapeutics. Here, we evaluated 6710 clinical and preclinical compounds targeting 2183 host proteins by immunocytofluorescence-based screening to identify SARS-CoV-2 infection inhibitors. Computationally integrating relationships between small molecule structure, dose-response antiviral activity, host target and cell interactome networking produced cellular networks important for infection. This analysis revealed 389 small molecules, >12 scaffold classes and 813 host targets with micromolar to low nanomolar activities. From these classes, representatives were extensively evaluated for mechanism of action in stable and primary human cell models, and additionally against Beta and Delta SARS-CoV-2 variants and MERS-CoV. One promising candidate, obatoclax, significantly reduced SARS-CoV-2 viral lung load in mice. Ultimately, this work establishes a rigorous approach for future pharmacological and computational identification of novel host factor dependencies and treatments for viral diseases.

microbiology↗

bioRxiv2020

Discovery of disease treatment mechanisms through the multiscale interactome

Most diseases disrupt multiple proteins, and drugs treat such diseases by restoring the functions of the disrupted proteins. How drugs restore these functions, however, is often unknown as a drugs therapeutic effects are not limited only to the proteins that the drug directly targets. Here, we develop the multiscale interactome, a powerful approach to explain disease treatment. We integrate disease-perturbed proteins, drug targets, and biological functions into a multiscale interactome network, which contains 478,728 interactions between 1,661 drugs, 840 diseases, 17,660 human proteins, and 9,798 biological functions. We find that a drugs effectiveness can often be attributed to targeting proteins that are distinct from disease-associated proteins but that affect the same biological functions. We develop a random walk-based method that captures how drug effects propagate through a hierarchy of biological functions and are coordinated by the protein-protein interaction network in which drugs act. On three key pharmacological tasks, we find that the multiscale interactome predicts what drugs will treat a given disease more effectively than prior approaches, identifies proteins and biological functions related to treatment, and predicts genes that interfere with treatment to alter drug efficacy and cause serious adverse reactions. Our results indicate that physical interactions between proteins alone are unable to explain the therapeutic effects of drugs as many drugs treat diseases by affecting the same biological functions disrupted by the disease rather than directly targeting disease proteins or their regulators. We provide a general framework for identifying proteins and biological functions relevant in treatment, even when drugs seem unrelated to the diseases they are recommended for.

bioinformatics↗

bioRxiv2020

Discovering novel cell types across heterogeneous single-cell experiments

Although tremendous effort has been put into cell type annotation and classification, identification of previously uncharacterized cell types in heterogeneous single-cell RNA-seq data remains a challenge. Here we present MARS, a meta-learning approach for identifying and annotating known as well as novel cell types. MARS overcomes the heterogeneity of cell types by transferring latent cell representations across multiple datasets. MARS uses deep learning to learn a cell embedding function as well as a set of landmarks in the cell embedding space. The method annotates cells by probabilistically defining a cell type based on nearest landmarks in the embedding space. MARS has a unique ability to discover cell types that have never been seen before and annotate experiments that are yet unannotated. We apply MARS to a large aging cell atlas of 23 tissues covering the life span of a mouse. MARS accurately identifies cell types, even when it has never seen them before. Further, the method automatically generates interpretable names for novel cell types. Remarkably, MARS estimates meaningful cell-type-specific signatures of aging and visualizes them as trajectories reflecting temporal relationships of cells in a tissue.

bioinformatics↗

bioRxiv2018

Evolution of resilience in protein interactomes across the tree of life

Phenotype robustness to environmental fluctuations is a common biological phenomenon. Although most phenotypes involve multiple proteins that interact with each other, the basic principles of how such interactome networks respond to environmental unpredictability and change during evolution are largely unknown. Here we study interactomes of 1,840 species across the tree of life involving a total of 8,762,166 protein-protein interactions. Our study focuses on the resilience of interactomes to network failures and finds that interactomes become more resilient during evolution, indicating that a species position in the tree of life is predictive of how robust its interactome is to network failures. In bacteria, we find that a more resilient interactome is in turn associated with the greater ability of the organism to survive in a more complex, variable and competitive environment. We find that at the protein family level, proteins exhibit a coordinated rewiring of interactions over time and that a resilient interactome arises through gradual change of the network topology. Our findings have implications for understanding molecular network structure both in the context of evolution and environment.\n\nSignificance StatementThe interactome network of protein-protein interactions captures the structure of molecular machinery that underlies organismal complexity. The resilience to network failures is a critical property of the interactome as the breakdown of interactions may lead to cell death or disease. By studying interactomes from 1,840 species across the tree of life, we find that evolution leads to more resilient interactomes, providing evidence for a longstanding hypothesis that interactomes evolve favoring robustness against network failures. We find that a highly resilient interactome has a beneficial impact on the organisms survival in complex, variable, and competitive habitats. Our findings reveal how interactomes change through evolution and how these changes affect their response to environmental unpredictability.

systems biology↗

bioRxiv2018

Network Enhancement: a general method to denoise weighted biological networks

Networks are ubiquitous in biology where they encode connectivity patterns at all scales of organization, from molecular to the biome. However, biological networks are noisy due to the limitations of technology used to generate them as well as inherent variation within samples. The presence of high levels of noise can hamper discovery of patterns and dynamics encapsulated by these networks. Here we propose Network Enhancement (NE), a novel method for improving the signal-to-noise ratio of undirected, weighted networks, and thereby improving the performance of downstream analysis. NE applies a novel operator that induces sparsity and leverages higher-order network structures to remove weak edges and enhance real connections. This iterative approach has a closed-form solution at convergence with desirable performance properties. We demonstrate the effectiveness of NE in denoising biological networks for several challenging yet important problems. Our experiments show that NE improves gene function prediction by denoising interaction networks from 22 human tissues. Further, we use NE to interpret noisy Hi-C contact maps from the human genome and demonstrate its utility across varying degrees of data quality. Finally, when applied to fine-grained species identification, NE outperforms alternative approaches by a significant margin. Taken together, our results indicate that NE is widely applicable for denoising weighted biological networks, especially when they contain high levels of noise.

bioinformatics↗