Computational Biology¶
Computational approaches applied to problems in biology.
Source
400 resources imported from inoue0426/awesome-computational-biology, dedicated to the public domain (CC0 / Unlicense). Thanks to its maintainers and contributors. Part of the KEEBOX Library · Category: Miscellaneous.
Databases¶
scRNA¶
- CZ CELLxGENE — Single-cell dataset repository and interactive explorer from the Chan Zuckerberg Initiative.
- Gene Expression Omnibus — Public functional genomics database.
- Human Cell Atlas — Open global atlas of all cells in the human body.
- Single Cell PORTAL — Public database for single-cell RNA.
- Single Cell Expression Atlas — Public database for single-cell RNA.
Compound¶
- PubChem — One of the largest chemical databases (compounds, genes, and proteins).
- ChEBI — Database focused on small chemical compounds.
- ChEMBL — Bioactive molecules with drug-like properties.
- ChemSpider — Chemical structure database.
- DrugTargetCommons — Community platform for curating and integrating experimental bioactivity data across drugs and targets.
- HMDB (Human Metabolome Database) — Comprehensive database of small molecule metabolites found in the human body.
- KEGG COMPOUND — Collection of small molecules and biopolymers.
- LIPID MAPS — Database of lipids.
- Rhea — Database of chemical reactions.
- DrugCentral — Online drug compendium with drug mode of action and indication information.
- Drug Repurposing Hub — Collections of drug repurposing data (drug, MoA, target, etc).
- Therapeutic Target Database — Drug-target, target-disease, and drug-disease datasets.
- ZINC ligand discovery database — Free database of commercially-available compounds for virtual screening.
Pathway¶
- PathwayCommons — Database of pathways and interactions.
- KEGG PATHWAY — Collection of pathway maps.
- WikiPathways — Database of biological pathways.
- Reactome — Expert-curated, peer-reviewed pathway database with detailed reaction mechanisms.
- BioCyc — Collection of pathway/genome databases across thousands of organisms.
- OmniPath — Comprehensive resource integrating protein interactions, signaling pathways, gene regulatory networks, and miRNA targets from over 100 databases.
- SIGNOR 2.0 — Database of causal signaling interactions and pathways, with signed and directed relationships between proteins.
- MSigDB (Molecular Signatures Database) — Curated gene sets derived from pathways and biological processes.
Mass Spectra¶
- MassBank — Open source databases and tools for mass spectrometry reference spectra.
- MoNA MassBank of North America — Meta-database of metabolite mass spectra, metadata, and associated compounds.
Protein¶
- THE HUMAN PROTEIN ATLAS — Comprehensive human protein database (cells, tissues, organs).
- PROTEIN DATA BANK (PDB) — 3D structures of proteins, nucleic acids, complexes.
- UniProt — Functional information on proteins.
- AlphaFold Protein Structure Database — 3D protein structure predictions.
- RCSB Protein Data Bank — Repository for structural data of biological molecules.
- Critical Assessment of Structure Prediction (CASP) — Assessing methods for protein structure prediction.
- Uniclust — Clustered protein sequence databases.
- UniRef — Non-redundant sequence database clustering UniProtKB entries at multiple sequence identity thresholds.
- CATH database — Hierarchical classification of protein domain structures.
- SAbDab — Structural Antibody Database containing all antibody structures in the PDB.
- OADB (Observed Antibody Space Database) — Database of antibody sequences from immune repertoire sequencing.
- InterPro — Protein families, domains, and functional sites database integrating 14 member databases including Pfam and PROSITE.
- Pfam — Database of protein families described by multiple sequence alignments and hidden Markov models.
- NeXtProt — Expert knowledge base on human proteins with deep functional annotation, complementary to UniProt.
Genome¶
- ENCODE — Encyclopedia of DNA Elements; regulatory and functional genomic elements across the genome.
- Ensembl — Genome browser and annotation database for vertebrate and other eukaryotic genomes.
- Human Genome Resources at NCBI — Database for genomics, proteomics, transcriptomics, and systems biology.
- GenBank — NCBI's database of genetic sequences.
- UCSC Genome Browser — UCSC's genome browser.
- cBioPortal — Cancer genomics database; aggregating many patient datasets.
- OncoKB — Precision oncology knowledge base of cancer genes, variants, and therapeutic implications.
- 10x Genomics Dataset — Collection of single-cell datasets.
- The Genotype-Tissue Expression (GTEx) — Human gene expression and regulation resource.
- Dependency Map (DepMap) — CRISPR-Cas9 screens in cancer cell lines.
- Catalogue Of Somatic Mutations In Cancer (COSMIC) — Resource on somatic mutations in cancers.
- MGnify — Resource for metagenomic and metatranscriptomic data.
- JASPAR — Database of transcription factor binding profiles.
- gnomAD — Genome Aggregation Database; genetic variation from large-scale sequencing projects.
- Rfam — Database of RNA families with sequence alignments and consensus structures.
- ROADMAP Epigenomics — Reference epigenome maps for 111 primary human cell types and tissues, including histone modifications, chromatin accessibility, and DNA methylation.
- FANTOM5 — Functional annotation of mammalian genome; comprehensive atlas of active enhancers, promoters, and transcription start sites across human and mouse cell types.
Disease¶
- KEGG DRUG — Comprehensive, approved drug information.
- DrugBank — Database of drugs and targets (University of Alberta).
- DisGeNET — Database of gene-disease associations integrating expert-curated and GWAS data.
- OMIM (Online Mendelian Inheritance in Man) — Comprehensive database of human genes and genetic disorders.
- Open Targets Platform — Systematic target identification and prioritization platform integrating genetics, genomics, and drug data for drug discovery.
- Human Phenotype Ontology (HPO) — Standardized vocabulary of phenotypic abnormalities in human disease, linking genes, variants, and clinical features.
- DISEASES — Gene–disease association database integrating evidence from text mining, curated databases, and experimental data.
Interaction¶
Drug-Gene Interaction¶
- DGIdb — Drug-gene interactions and the druggable genome.
- Comparative Toxicogenomics Database — Chemical-gene interactions, chemical-disease and gene-disease associations, chemical-phenotype associations.
- SNAP — Dataset of drug-gene interactions.
Drug (Cell Line) Response¶
- NCI60 — Focuses on 60 cancer cell lines and many drugs.
- Genomics of Drug Sensitivity in Cancer (GDSC) — Drug sensitivity for ~1000 human cancer cell lines and hundreds of compounds.
- Cancer Cell Line Encyclopedia — Database of ~1000 cancer cell lines.
- CellMiner Cross Database (CellMinerCDB) — Integrates multiple cancer cell line databases.
Chemical-Protein Interaction¶
- STITCH — Chemical-protein interactions.
- BindingDB — Compounds and target database.
- Davis kinase inhibitors DB — Experimental kinase inhibitor binding affinity dataset for protein–ligand interaction research.
- Kinase Inhibitor Bioactivity Data (KIBA) — Integrated bioactivity scores for kinase inhibitors combining Ki, Kd, and IC50 measurements.
- PDBBind — Binding affinity data for biomolecular complexes.
Protein-Protein Interaction¶
- STRING — PPI networks for multiple organisms.
- BioGRID — Protein, genetic, and chemical interactions.
- HIPPIE — Human protein-protein interaction database.
- IntAct — Open-source molecular interaction database and analysis system from EMBL-EBI.
Knowledge Graph¶
- Drug Mechanism Database (DrugMechDB) — Mechanisms of action from drug to disease.
- DRKG — Large-scale biological knowledge graph for drug discovery.
- Hetionet — Heterogeneous network integrating genes, diseases, drugs, pathways, and more.
- PrimeKG — Multi-modal precision medicine knowledge graph integrating clinical, genetic, and drug data.
Gene Regulatory Network¶
- TRRUST v2 — Manually curated database of human and mouse transcriptional regulatory interactions between transcription factors and their target genes, expanded with literature-derived evidence.
- RegNetwork — Database of gene regulatory networks covering transcription factor–target gene and miRNA–gene interaction data across multiple species.
- miRBase — Reference repository for microRNA gene annotations, sequences, and experimentally validated targets.
Clinical Trial¶
- ClinicalTrials.gov — Privately and publicly funded clinical studies.
- ICD10 — International Classification of Diseases, 10th revision.
- EU Drug Regulating Authorities Clinical Trials DB (EudraCT) — European clinical trial database.
- MIMIC-IV — Freely accessible critical care database.
Benchmarks & Datasets¶
Drug Response & Perturbation¶
- scCT-DB — Cancer treatment single-cell database with 266 paired pre/post-treatment datasets, 6.19M cells, 1,142 samples, 27 cancer types, and 102 drugs.
- scICB — Pan-cancer ICB therapy single-cell database spanning 338 patients, 807 samples, 3.69M cells, and 13 cancer types, including 247 matched pre/post-treatment patients.
- CellResDB — Cancer therapy response single-cell database containing 4.7M cells from 1,391 patient samples, 72 datasets, and 24 cancer types.
- Gondal et al. 2025 ICB scRNA-seq collection — Integrated ICB-treated cancer single-cell collection from eight studies and 223 patients, with 90,270 cancer cells and 265,671 other cells.
- BEAT AML — Functional ex vivo drug sensitivity measurements paired with genomics for acute myeloid leukemia.
- Cancer Therapeutics Response Portal (CTRP) — Drug sensitivity profiles across ~900 cancer cell lines for >400 compounds.
- Chem-PerturBridge — Harmonized compendium and processing pipelines for small-molecule perturbation transcriptomics across heterogeneous assays and gene panels.
- GSE191127 Breast Cancer Pre/Post Chemotherapy — Bulk RNA-seq and genomics from matched pre- and post-neoadjuvant chemotherapy breast tumors with treatment outcomes.
- Genomics of Drug Sensitivity in Cancer (GDSC) — Drug sensitivity for ~1000 human cancer cell lines and hundreds of compounds.
- JUMP Cell Painting Datasets — Consortium-scale cell imaging perturbation datasets (chemical and genetic) for phenotypic profiling and drug discovery research.
- LINCS L1000 — Gene expression profiles (978 landmark genes) for >20,000 chemical and genetic perturbations across cell lines.
- LINCS L1000 Phase 1 — Phase-1 L1000 benchmark view used by Bison with 692,787 profiles, 9,233 molecules, 70 contexts, and 978 landmark genes.
- LINCS L1000 Phase 2 — Phase-2 L1000 benchmark view used by Bison with 333,263 profiles, 1,760 molecules, 30 contexts, and 978 landmark genes.
- MIX-Seq — Multiplexed single-cell transcriptional profiling of chemical and genetic perturbation responses across pools of cancer cell lines.
- NCI60 — Drug sensitivity benchmark across 60 diverse human cancer cell lines.
- Novartis Perturbation Dataset — Large chemical perturbation transcriptomic screen used in Chem-PerturBridge; the Bison benchmark view contains 46,748 profiles and 3,770 molecules.
- OP3 — Chemical perturbation dataset used by Chem-PerturBridge and the Bison unseen-compound benchmark; the Bison benchmark view contains 1,813 profiles, 138 molecules, and 4 cellular contexts.
- PDX-Atlas — Portal integrating clinical, genomic, transcriptomic, and drug-response data from patient-derived xenograft models.
- PharmGKB — Curated pharmacogenomics dataset linking genetic variants to drug response phenotypes across thousands of drugs.
- PRISM — Cancer drug sensitivity profiling of >4,500 drugs across >900 cancer cell lines using pooled-cell-line barcoding.
- sci-Plex — Single-cell chemical perturbation screen of ~650,000 transcriptomes across three cancer cell lines, 188 compounds, and four doses.
- scPerturb — Curated and continuously updated single-cell perturbation data resource spanning CRISPR and drug perturbation studies.
- Tahoe-100M — Giga-scale single-cell perturbation atlas with >100 million profiles from 50 cancer cell lines exposed to ~1,100 small molecules.
- VCPI-0001 — Bulk chemical perturbation dataset used in Chem-PerturBridge; the Bison benchmark view contains 27,517 profiles and 2,272 molecules in one cellular context.
- VCPI-0002 — Bulk chemical perturbation dataset used in Chem-PerturBridge; the Bison benchmark view contains 18,139 profiles and 1,488 molecules in one cellular context.
Genetic Perturbation & Perturb-seq¶
- scPerturbBench (TianGzlab / PerturbArena) — Benchmark of single-cell genetic perturbation prediction covering unseen perturbations, combinations, cell-state transfer, and 24 evaluation metrics.
- scPerturBench (BM2 Lab) — Benchmark comparing 27 single-cell perturbation-response prediction methods across 29 genetic and chemical perturbation datasets.
- Dixit et al. 2016 Perturb-seq — Pooled CRISPR knockout with single-cell RNA-seq across K562 and dendritic-cell screens, including stimulated conditions and combinatorial perturbations.
- Adamson et al. 2016 Perturb-seq — CRISPRi Perturb-seq in K562 cells for systematic dissection of the unfolded protein response.
- Datlinger et al. 2017 CROP-seq — Pooled CRISPR knockout screen with single-cell transcriptomic readout in Jurkat cells under T-cell receptor stimulation.
- Norman et al. 2019 CRISPRa Perturb-seq — Large-scale CRISPR activation Perturb-seq in K562 cells spanning single-gene and combinatorial perturbations.
- Gasperini et al. 2019 CRISPRi enhancer screen — High-MOI single-cell CRISPRi screen in K562 cells targeting thousands of candidate enhancers to map enhancer-gene regulation.
- Replogle et al. 2022 K562 Genome-wide Perturb-seq — Genome-scale CRISPRi Perturb-seq targeting nearly all expressed genes in K562 cells.
- Replogle et al. 2022 K562 Essential Perturb-seq — CRISPRi Perturb-seq focused on essential genes in K562 cells.
- Replogle et al. 2022 RPE1 Essential Perturb-seq — CRISPRi Perturb-seq focused on essential genes in RPE1 cells.
- Frangieh et al. 2021 Perturb-CITE-seq — Multimodal CRISPR knockout screen with RNA and surface-protein readouts in patient-derived melanoma models under immune-related conditions.
Single-Cell & Spatial¶
- HEST Xenium virtual spatial transcriptomics — DeepSpot-M predicted transcriptome-wide ST for 59 HEST-1k 10x Xenium samples (~13.3M cells) (gated). Paper: DeepSpot-M.
- scIB (Single-cell Integration Benchmarks) — Comprehensive benchmarking framework for single-cell data integration methods.
- Tabula Muris — Comprehensive single-cell atlas of 20 mouse organs and tissues, enabling cross-tissue and cross-species comparisons.
- Tabula Sapiens — Comprehensive human single-cell atlas of ~500K cells from 24 organs and tissues across multiple donors.
- TCGA virtual spatial transcriptomics atlas — DeepSpot-M predicted transcriptome-wide ST for TCGA H&E (FF + FFPE; 28,664 slides / 32 cancer types; gated). Paper: DeepSpot-M.
Molecular, Protein & Drug Discovery¶
- BACE — Binary classification and regression dataset for β-secretase 1 (BACE-1) inhibitor binding affinity.
- Bento — Protein-ligand docking benchmark covering rigid, flexible, de novo, blind, induced-fit, and covalent docking tasks.
- BindingDB Curated Sets — Curated binding affinity datasets for protein–ligand interaction benchmarking.
- ClinTox — Clinical toxicity dataset contrasting FDA-approved drugs with those that failed clinical trials due to toxicity.
- CrossDocked2020 — Large-scale dataset for structure-based virtual screening.
- DUD-E (Directory of Useful Decoys, Enhanced) — Structure-based virtual screening benchmark with active ligands and challenging decoy sets across diverse protein targets.
- FLIP (Fitness Landscape Inference for Proteins) — Benchmark collection of protein fitness landscape datasets for evaluating protein ML models.
- GuacaMol — Benchmark suite for generative molecular design models.
- MoleculeNet — Benchmark datasets for molecular machine learning.
- MOSES — Benchmarking platform for molecular generation models.
- PK-DB — Open database of experimental pharmacokinetics (PK) and ADME data from clinical and preclinical studies.
- ProteinGym — Large-scale benchmark of deep mutational scanning assays for evaluating protein fitness landscape models.
- QM9 — Quantum chemistry properties for 134K stable small organic molecules computed at DFT level.
- SIDER (Side Effect Resource) — Database of 1,430 approved drugs with their recorded adverse drug reactions across 27 system-organ classes.
- TAPE (Tasks Assessing Protein Embeddings) — Benchmark suite of five biologically meaningful semi-supervised learning tasks for evaluating protein representations.
- Therapeutics Data Commons (TDC) — Unified benchmark suite covering ADMET, drug-target interaction, drug response, and more.
- Tox21 — 12,707 compounds tested in 12 nuclear receptor and stress-response pathway biochemical assays for toxicity prediction.
Genomics, Cancer & Biomedical Cohorts¶
- 1000 Genomes Project — Reference panel of human genetic variation from 2,504 individuals across 26 populations.
- CPTAC (Clinical Proteomic Tumor Analysis Consortium) — Multi-omic proteogenomic datasets for multiple cancer types linking proteomics with genomics.
- The Cancer Genome Atlas (TCGA) — Comprehensive multi-omics (genomics, transcriptomics, proteomics, methylation) dataset for 33 cancer types across ~11,000 patients.
- UK Biobank — Large-scale biomedical database of ~500K participants with genetic, imaging, and health data for population genetics and disease studies.
General Biological ML & Knowledge Graph Benchmarks¶
- CellBench (CellVoyager) — Benchmark for evaluating LLMs and data-analysis agents on reproducible single-cell research tasks.
- OGB (Open Graph Benchmark) — Large-scale graph ML benchmark suite including biological datasets such as ogbl-ppa (protein-protein associations) and ogbg-molhiv.
- OpenBioLink — Benchmark datasets for biological knowledge graph completion.
API¶
- PubMed E-utilities (esearch/efetch) — APIs for searching and retrieving biomedical literature from PubMed.
- NCBI E-utilities — Unified APIs for accessing NCBI databases (Gene, GEO, SRA, PubChem, etc).
- UniProt REST API — Programmatic access to protein sequence and functional annotation data.
- Ensembl REST API — API for genomic annotations, variants, genes, and comparative genomics.
- KEGG REST API — API for accessing KEGG pathways, compounds, genes, and reactions.
- ChEMBL Web Services — REST API for bioactive molecules, targets, and bioassays.
- Open Targets Platform API — API for target–disease associations integrating genetics, genomics, and drug data.
- ClinicalTrials.gov API — API for querying clinical trial metadata and results.
Preprocessing Tools¶
- SpatialData — Interoperable data structures and tooling for spatial omics, microscopy images, coordinates, segmentation shapes, and annotations.
- Chemistry Development Kit — Cheminformatics software & machine learning tools.
- Biopython — Collection of Python tools for biological computation including sequence analysis, structure parsing, and database access.
- FlashDeconv — High-performance spatial transcriptomics deconvolution (~1M spots in ~3 min).
- RDKit — Cheminformatics software & machine learning toolkit.
- DeepChem — Deep learning library for drug discovery, quantum chemistry, and materials science.
- ChatSpatial — MCP server for spatial transcriptomics analysis via natural language.
- Scanpy — Python library for scRNA-seq analysis.
- Seurat — R library for scRNA-seq analysis.
- scvi-tools — Probabilistic models for single-cell omics data analysis.
- CellTypist — Automated cell type annotation for scRNA-seq.
- Squidpy — Python library for spatial single-cell analysis.
- GROMACS — Molecular dynamics simulation package for biochemical molecules.
- MDAnalysis — Python library for analyzing and altering molecular dynamics simulation trajectories.
- OpenMM — High-performance toolkit for molecular simulation and GPU-accelerated MD.
- scVelo — RNA velocity estimation for single-cell transcriptomics, inferring the direction and speed of cell differentiation.
- STAR — Ultrafast universal RNA-seq aligner with support for spliced alignment and single-cell quantification via STARsolo.
- kallisto — Near-optimal RNA-seq quantification using pseudoalignment for fast transcript abundance estimation.
- Harmony — Fast and scalable integration of single-cell data across datasets, conditions, technologies, and species.
- Monocle3 — Single-cell trajectory analysis tool for learning developmental trajectories and ordering cells in pseudotime.
- CellChat — Inference and analysis of cell-cell communication ligand-receptor networks from single-cell transcriptomics data.
- SCENIC — Single-cell regulatory network inference and clustering linking transcription factors to co-expressed gene modules.
- DoubletFinder — Machine learning approach for detecting multiplet (doublet) artifacts in single-cell RNA-seq data.
- Numbat — Haplotype-aware copy number variation inference from single-cell RNA-seq using hidden Markov models.
- CaSpER — CNV identification and visualization by integrative analysis of single-cell or bulk RNA-seq data.
- CellCharter — Identification and characterization of spatial cell niches from spatial transcriptomics using VAEs and Gaussian mixture models.
- STAGATE — Adaptive graph attention auto-encoder for spatial domain identification in spatial transcriptomics.
- NCEM — GNN-based model for learning intercellular communication from spatial graphs of cells.
- DeepTalk — Graph attention network for deciphering cell-cell communication from spatial transcriptomics.
- COMMOT — Optimal transport-based framework for screening cell-cell communication in spatial transcriptomics.
- TIGON — Neural optimal transport method for reconstructing growth and dynamic trajectories from single-cell transcriptomics.
- LINGER — Neural network for gene regulatory network inference from single-cell multiome (RNA+ATAC-seq) data with bulk data pretraining.
- sciPENN — RNN-based method for simultaneous protein expression prediction, uncertainty estimation, and cell-type label transfer from CITE-seq and scRNA-seq data.
- MOGONET — Multi-omics graph convolutional network framework for patient classification and biomarker identification.
- AutoZyme — Autonomous agentic framework that speeds up bioinformatics software (e.g. Scanpy, Seurat) on CPUs while preserving the original results.
- SeqBench — Web-based molecular biology sequence workbench for primer design, cloning simulation (Gibson, Golden Gate, restriction digest), CRISPR guide RNA design, and sequence analysis, with a public REST API, OpenAPI 3.1 spec, and MCP server.
Machine Learning Tasks and Models¶
Drug Discovery¶
Drug Response Prediction¶
- drGAT — Attention-based model for drug response prediction with gene explainability.
- MOFGCN — GCN + heterogeneous network.
- DeepDSC — Autoencoder + fully connected NN.
- DGDRP — Multi-view embedding neural network.
- DeepAEG — GNN embedding + attention mechanism.
- RECOVER — Machine learning framework for predicting synergistic drug combination responses across cell lines.
- TGSA — Tumor gene set and attention-based model leveraging biological pathway knowledge for drug response prediction.
- HiDRA — Hierarchical network model incorporating gene and pathway-level information for cancer drug response prediction.
- DRUML — Ensemble machine learning framework combining standard ML with deep learning to systematically rank anti-cancer drugs from proteomics and RNA-seq data.
- PASO — Pathway-aware multi-omics drug response model combining pathway-difference features, multi-scale convolutions, Transformer encoding, and drug SMILES.
- DTLCDR — Target-based multimodal framework for preclinical cancer drug response prediction and transfer to clinical response, with explicit unseen-drug generalization.
- THERAPI — Cell-line-to-patient transfer framework that aligns tumor transcriptomes with cancer cell lines and integrates perturbation and gene-level representations for patient drug response prediction.
- EXPRESSO — Supervised treatment-response framework using pretreatment tumor transcriptomics, drug targets, and context-specific biomarkers across multiple cancer types and therapies.
- PerturbRx — Treatment-conditioned representation learning framework that transfers drug-induced latent transitions learned from single-cell perturbation data to patient-level cancer treatment-response prediction.
Drug Perturbation¶
- CellOT — Neural optimal transport framework for predicting single-cell responses to drug and genetic perturbations.
- CMonge — Conditional optimal transport model for generalizable single-cell perturbation response prediction across drugs and doses.
- chemCPA — Compositional perturbation autoencoder for predicting single-cell transcriptional responses to unseen drug perturbations and dose combinations.
- cycleCDR — Interpretable cycle-consistency framework for modeling cellular responses to drug perturbations.
- PRNet — Deep generative model for predicting transcriptional responses to novel chemical perturbations for drug discovery.
- Bison — Cross-dataset model for globally unseen-compound response prediction using a shared gene representation, discrete diffusion models, and matched drug-contrast supervision.
- biolord — Deep generative model that disentangles known and unknown attributes for conditional generation of single-cell states.
- PerturbNet — Conditional generative model for predicting distributions of single-cell states under unseen chemical and genetic perturbations.
- CellFlow — Conditional flow-matching framework for modeling and predicting cellular phenotypes under chemical, genetic, and other perturbations.
- Prophet — Transformer model for predicting cellular phenotypes under unseen chemical or genetic perturbations across heterogeneous assays and contexts.
- PrePR-CT — Graph-based model using cell-type-specific co-expression networks as inductive priors for small-data chemical perturbation response prediction.
- XPert — Knowledge-informed dual-branch Transformer for drug-induced transcriptional perturbation prediction across dose, time, and cellular context.
- State — Transition model for predicting cellular perturbation responses across diverse contexts and sets of cells.
- TxPert — Knowledge-graph-informed latent-transfer model for transcriptomic perturbation prediction across unseen single perturbations, combinations, and cross-context settings.
- LPM — Large perturbation model that jointly learns heterogeneous perturbation experiments by disentangling perturbation, readout, and context representations.
Drug Repurposing¶
- DeepPurpose — Deep learning library for drug repurposing.
- TranSiGen — Dual-VAE architecture for ligand-based virtual screening, drug response prediction, and drug repurposing using chemical-induced transcriptional profiles.
Drug Target Interaction¶
- NeoDTI — Library for drug-target interaction prediction.
- DTINet — Network-based framework integrating heterogeneous biological data for DTI prediction.
- DeepDTA — Deep learning model using CNNs on protein sequences and drug SMILES.
- GraphDTA — Graph neural network–based DTI prediction using molecular graphs.
- MolTrans — Transformer-based DTI model leveraging molecular substructures.
- DrugBAN — Bilinear attention network for interpretable DTI prediction.
Compound-Protein Interaction¶
- MCPINN — Drug discovery via compound-protein interaction and machine learning.
- TransformerCPI — CPI prediction using Transformer.
Molecular Generation¶
- REINVENT — Reinforcement learning for de novo drug design.
- MolGPT — Transformer-based model for molecular generation.
- Molecular Transformer — Sequence-to-sequence model for retrosynthesis prediction.
- Matcha — Multi-stage Riemannian flow matching model for physically valid molecular docking with scoring, pose filtering, and benchmarks.
- TargetDiff — 3D equivariant diffusion model for structure-based drug design.
- DiffDock — Diffusion generative model for molecular docking, predicting the binding pose of small molecules to protein targets.
- JTVAE — Junction tree variational autoencoder for molecular graph generation that guarantees chemical validity via a hierarchical tree decomposition.
- DiffSBDD — Equivariant diffusion model for structure-based drug design that generates molecules and binding conformations for protein targets.
- ReLeaSE — Deep reinforcement learning framework for de novo drug design combining a generative and predictive model.
- PaccMannRL — Reinforcement learning-based generative model for de novo hit-like anticancer molecule design from transcriptomic data.
Protein Property Prediction¶
- NbBayesLM — Bayesian neural network integrating protein language model embeddings and physicochemical features to predict nanobody thermostability with uncertainty estimates. Paper
LLM for Biology¶
- AI4Chem/ChemLLM-7B-Chat — LLM for chemical & molecular science.
- BioGPT — LLM for biomedical text generation.
- GeneGPT — LLM for biomedical information, integrated with various APIs.
- GenePT — Foundation LLM for single-cell data.
- scPRINT — Pretrained on 50M cells for scRNA-seq denoising & zero imputation.
- BioMedLM — 2.7B parameter GPT-2-style language model trained exclusively on biomedical literature from PubMed for biomedical question answering and text generation.
- MolT5 — Language model for molecular tasks bridging text and SMILES, enabling molecule captioning and text-driven molecule generation.
- ChatDrug — LLM-based conversational pipeline for drug discovery, using natural language prompts for iterative drug editing and optimization.
Agentic AI for Biology¶
General Biomedical Agents¶
- Biomni — General-purpose biomedical AI agent integrating planning, code execution, specialized tools, databases, and software across diverse biomedical research tasks.
- ToolUniverse — Unified scientific tool ecosystem for building AI scientists that can discover, select, and execute biomedical tools and databases.
- ClawBio — Bioinformatics-native AI agent skill library with local-first pharmacogenomics, ancestry PCA, semantic similarity, nutrigenomics, and metagenomics skills.
- BioMedAgent — Self-evolving multi-agent framework for autonomous biomedical data analysis with tool discovery, workflow planning, code generation, execution, correction, and cross-omics analysis.
- BioMaster — Multi-agent system for automated and auditable bioinformatics workflows spanning RNA-seq, ChIP-seq, single-cell, spatial omics, Hi-C, long reads, metagenomics, and proteomics.
- BRAD — Retrieval-augmented bioinformatics assistant integrating scientific literature, databases, external tools, and executable workflows.
Therapeutics & Drug Discovery Agents¶
- TxAgent — Therapeutic reasoning agent using multi-step reasoning and a large scientific tool universe for drug interactions, contraindications, and personalized treatment analysis.
- Medea — Multi-agent therapeutic discovery system combining research planning, biological data analysis, literature reasoning, and multi-LLM deliberation across single-cell, cell-line, and patient contexts.
- DrugAgent — Multi-agent biomedical evidence synthesis framework for computational drug discovery with reliability-aware aggregation.
Bioinformatics & Omics Agents¶
- CellVoyager — Jupyter-based scientific agent that plans, executes, and evaluates single-cell transcriptomics analyses, accompanied by CellBench.
- CASSIA — Multi-agent LLM framework for reference-free and interpretable single-cell cell-type annotation with dedicated annotation, validation, scoring, and reporting agents.
- STELLA — Self-evolving biomedical research agent that expands its tool repertoire and supports literature reasoning, computational analysis, and laboratory-oriented scientific workflows.
- AutoBA — Automated multi-omics analysis agent that plans, executes, and repairs bioinformatics workflows from natural-language objectives.
- GenoMAS — Multi-agent framework for code-driven gene-expression analysis with planning, execution, debugging, backtracking, and GEO/TCGA-based scientific discovery.
- BIA — Bioinformatics agent for GEO search, sample metadata extraction, count-matrix processing, and pipeline extraction from papers.
- BioAgents — Multi-agent bioinformatics assistant using specialized language models and retrieval for genomics workflow development and troubleshooting.
- Genomi — Local-first genomics agent runtime that indexes personal variants, queries evidence, and generates evidence-grounded reports while keeping raw genome data on-device.
Multi-Agent Scientific Labs¶
- Virtual Lab — Human–AI collaborative research environment in which an LLM principal investigator coordinates specialized scientist agents for scientific discovery.
- Agent Laboratory — End-to-end multi-agent research workflow for literature review, experimentation, implementation, analysis, and report generation.
Paper & Workflow Agents¶
- Paper2Agent — Multi-agent system that transforms research papers and associated code into interactive, testable scientific agents and MCP tools.
Foundation Models¶
Transcriptomics Foundation Models¶
- Cell2Sentence / C2S-Scale — Single-cell language-model family representing gene-expression profiles as cell sentences for cellular interpretation, generation, and question answering.
- scFoundation — Large-scale foundation model for single-cell gene expression, enabling multiple downstream tasks.
- scGPT — Transformer-based foundation model pretrained on millions of single-cell profiles.
- Geneformer — Context-aware, attention-based deep learning model pretrained on a large corpus of single-cell transcriptomes.
- BulkFormer — Foundation model for bulk RNA-seq data; learns general transcriptomic representations.
- scBERT — BERT-based foundation model pretrained on large-scale scRNA-seq data for cell type annotation.
- CellPLM — Cell pre-trained language model with inter-cell transformer architecture for diverse single-cell analysis tasks.
- UCE — Universal Cell Embeddings: zero-shot single-cell embedding model trained on 36M cells across species, tissues, and assays without fine-tuning.
- GEARS — Graph-based model for predicting transcriptional responses to single and combinatorial genetic perturbations using biological priors.
- SATURN — Transformer-based model integrating gene expression and protein sequences via a protein language model to learn unified multi-species cell embeddings.
- CancerFoundation — Single-cell RNA-seq foundation model trained exclusively on a curated dataset of malignant cells to learn cancer-specific embeddings.
- CellFM — 800M-parameter single-cell foundation model pretrained on transcriptomics from 100 million human cells for annotation, integration, gene-function, and perturbation tasks.
- scPRINT-2 — Next-generation single-cell foundation model pretrained on 350M+ cells across 22K+ datasets and 16 species for embeddings, denoising, annotation, gene-network inference, and cross-species transfer.
Spatial Foundation Models¶
- GigaPath — Slide-level digital pathology foundation model pretrained on 1.3 billion pathology image tokens from whole-slide images.
- UNI — General-purpose self-supervised pathology foundation model trained on 100K+ whole-slide images for diverse computational pathology tasks.
- CONCH — Vision-language foundation model for computational pathology trained with contrastive captioning on pathology image–text pairs.
- Phikon — ViT-based pathology foundation model pretrained with iBOT self-supervision on TCGA whole-slide images.
- Nicheformer — Foundation model for single-cell and spatial omics using a transformer architecture with positional embeddings to encode spatial cell information.
- scGPT-spatial — Extension of scGPT for spatial transcriptomics with continual pretraining and a mixture-of-experts decoder for spatial gene expression analysis.
- DeepSpot — Deep learning model predicting spatial transcriptomics from H&E images at spot and single-cell resolution.
- DeepSpot2Cell — Predicts virtual single-cell spatial transcriptomics from H&E using spot-level supervision (NeurIPS 2025 Imageomics).
- DeepSpot-M — Multimodal foundation model for transcriptome-wide virtual spatial transcriptomics from histology.
- AESTETIK — Autoencoder for spatial transcriptomics representation learning using topology and histology image knowledge.
Pathology Foundation Models¶
- TITAN — Multimodal whole-slide pathology foundation model that combines image and language supervision for slide-level representation and zero-shot analysis.
- Virchow2 — 632M-parameter pathology vision transformer pretrained on 3.1M whole-slide images with mixed-magnification self-supervision.
- H-Optimus-0 — 1.1B-parameter histopathology foundation model trained with self-supervised learning on a large multi-center slide corpus.
- H-Optimus-1 — 1.1B-parameter pathology foundation model trained on billions of histology images from more than one million slides and 800K+ patients.
- UNI2-h — Billion-parameter histopathology vision foundation model for tile-level feature extraction and downstream computational pathology tasks.
- Phikon-v2 — Updated pathology foundation model for general-purpose histology feature extraction and transfer learning.
- GenBio-PathFM — 1.1B-parameter histopathology foundation model trained on public data using morphology-aware curation and dual-stage JEPA+DINO learning.
Multi-Omics Foundation Models¶
- scMulan — Single-cell multi-omic language model pretrained on ~10M cells spanning transcriptomics, epigenomics, and proteomics for cross-omics transfer tasks.
- totalVI — Probabilistic framework for joint analysis of paired scRNA-seq and protein (CITE-seq) data enabling multi-modal cell state representation across single-cell datasets.
- MultiVI — Multi-modal variational autoencoder for integrating paired and unpaired single-cell RNA-seq and ATAC-seq measurements into a unified latent space.
- MIRA — Probabilistic multimodal topic model jointly modeling single-cell transcriptomics and chromatin accessibility for regulatory network inference.
- GLUE — Graph-Linked Unified Embedding framework for unpaired single-cell multi-omics data integration across RNA, ATAC, methylation, and protein modalities.
- BABEL — Cross-modality translation model enabling prediction between scRNA-seq and scATAC-seq profiles without requiring paired single-cell measurements.
- Multigrate — Asymmetric multi-omics variational autoencoder for integrating single-cell data across RNA, ATAC, and protein modalities with missing-modality support.
- MOFA+ — Multi-Omics Factor Analysis framework identifying shared axes of variation across bulk and single-cell datasets including RNA, ATAC, proteomics, methylation, and copy number.
- GeneCompass — Large-scale foundation model integrating DNA regulatory sequences and single-cell transcriptomics from 120M+ cells across multiple species for gene regulation prediction.
- UnitedNet — Interpretable multi-task deep neural network for single-cell multi-omics integration spanning transcriptomics, chromatin accessibility, and proteomics.
- SpatialGlue — Graph attention network for spatial multi-omics integration jointly embedding spatial transcriptomics with chromatin accessibility or proteomics.
- MIDAS — Mosaic integration and differential accessibility model for single-cell multi-omics that handles arbitrary missing-modality combinations across transcriptomics, chromatin accessibility, and proteomics.
- Concerto — Contrastive self-supervised learning framework for single-cell multimodal data integration, batch correction, and reference-query mapping.
- scButterfly — Dual-aligned variational autoencoder for single-cell cross-modality translation between paired and unpaired multiomics data.
- JAMIE — Joint variational autoencoder for multimodal single-cell data imputation and embedding.
- scPair — Bidirectional feedforward network for single-cell multimodal analysis with cross-modality prediction leveraging single-cell atlases.
Domain Alignment¶
- scArches — Transfer learning framework for mapping new single-cell datasets onto pre-trained reference atlases across batches, conditions, and modalities.
- TOSICA — Transformer-based framework for one-stop interpretable cell-type annotation supporting cross-dataset and cross-species transfer.
Compound Embedding¶
- ChemBERTa-2 — RoBERTa-based molecular language model pretrained on SMILES for small-molecule representation learning.
- GROVER — Self-supervised graph transformer for large-scale molecular representation learning from unlabeled compounds.
- Mol2Vec — Unsupervised molecular embedding method inspired by Word2Vec for learning vector representations of chemical substructures.
- MolFormer — Linear attention transformer pretrained on millions of SMILES strings for efficient molecular embeddings.
- Uni-Mol — 3D molecular pretraining framework for universal representation learning on molecules and protein pockets.
- Uni-Mol2 — Scaled molecular pretraining model using atomic, graph, and 3D geometry features, with models up to 1.1B parameters pretrained on 800M conformations.
- ChemFM — 1B/3B-parameter chemical language model pretrained on 178M molecules for molecular representation, property prediction, generation, and synthesis tasks.
Pre-trained Embedding¶
- Evolutionary Scale Modeling (ESM) — Protein embeddings.
- ProtTrans — Suite of protein language models (ProtBERT, ProtT5, ProtXLNet) trained on billions of protein sequences from UniRef and BFD.
- ProGen2 — Protein language model trained on diverse protein families for sequence generation and fitness prediction.
- Ankh — Efficient protein language model optimized for downstream prediction tasks including secondary structure, localization, and function annotation.
- ESM Cambrian (ESM C) — Protein representation foundation-model family designed as an efficient next-generation successor to ESM2, spanning 300M to multi-billion-parameter models.
Protein Structure Prediction and Design¶
- AlphaFold3 — Predicts structures of proteins, nucleic acids, small molecules, and their complexes.
- Boltz-1 — Open-source all-atom biomolecular structure prediction model for proteins, nucleic acids, small molecules, and their complexes achieving AlphaFold3-level accuracy.
- Boltz-2 — Biomolecular foundation model jointly predicting complex structures and binding affinities for protein–ligand interaction modeling and virtual screening.
- Protenix — Trainable biomolecular structure-prediction framework for proteins, nucleic acids, ligands, and complexes with open training and inference pipelines.
- Chai-1 — Unified molecular structure prediction model covering proteins, nucleic acids, small molecules, and complexes.
- ESM3 — Multimodal protein language model that jointly reasons over sequence, structure, and function for generative protein design and engineering.
- ESMFold — Fast protein structure prediction using language model embeddings.
- RFdiffusion — Generative model for protein backbone design using diffusion.
- ProteinMPNN — Deep learning model for protein sequence design given backbone structure.
- OmegaFold — High-resolution de novo protein structure prediction from sequence.
- RoseTTAFold — Three-track neural network for protein structure prediction.
- OpenFold — Trainable, memory-efficient open-source reproduction of AlphaFold2 enabling custom protein structure prediction workflows.
- SaProt — Structure-aware protein language model using structure-aware tokens that encode both sequence and backbone geometry for improved function prediction.
- EvoDiff — Discrete diffusion framework for protein sequence generation trained on evolutionary-scale data, supporting unconditional generation, disordered region design, and functional motif scaffolding. [ paper-2023 ]
Multi-Modal Foundation Models¶
- TxGemma — Open Gemma-based therapeutic-development model family (2B, 9B, and 27B) for property prediction and interactive scientific reasoning.
- NVIDIA Nemotron for Science — Scientific-agent development stack based on Nemotron 3.5 Super VL, with post-training recipes and chemistry reasoning benchmark LitmusBench v0.2.
- CHIEF — Clinical Histopathology Imaging Evaluation Foundation model integrating histology images and clinical context for pan-cancer analysis.
- BiomedCLIP — CLIP-based vision-language foundation model for biomedical images and text trained on PubMed figure–caption pairs.
- PORPOISE — Pan-cancer integrative histology-genomic analysis framework using multimodal deep learning for patient stratification.
- PathomicFusion — Integrated framework fusing histopathology and genomic features via CNN, GNN, and attention gating for cancer diagnosis and prognosis.
- Virchow — Million-slide digital pathology foundation model using a vision transformer and self-supervised distillation for tile-level pathology image representation.
- TOAD — Tumor Origin Assessment via Deep-learning; weakly-supervised multi-task model predicting cancer primary origin from H&E whole-slide images.
- PLIP — Vision-language foundation model for pathology trained with contrastive learning on pathology image–text pairs for image classification and text-to-image retrieval.
- MUSK — Vision-language foundation model for precision oncology analyzing multimodal paired text and pathology image data for biomarker prediction and retrieval.
Cancer Genome Foundation Models¶
- TESSERA — Cancer-genome foundation model jointly pretrained on somatic SNVs and copy-number alterations from TCGA using masked reconstruction and cross-modal contrastive learning.
- MutationProjector — Pan-cancer genotype foundation model trained on mutations and copy-number alterations from >30K tumors for clinical representation learning.
Single-Cell Epigenomics Foundation Models¶
- EpiAgent — scATAC-seq foundation model pretrained on ~5M cells and >35B tokens for representation learning, annotation, imputation, perturbation prediction, and in-silico cCRE knockout.
- SCARF — Single-cell RNA+ATAC foundation model pretrained on >2.7M cells for multimodal representation, matching, cross-omics translation, and few-shot annotation.
- scDNAm-GPT — Foundation model for single-cell whole-genome bisulfite sequencing with whole-genome context modeling at single-CpG resolution.
- EpiFoundation — Foundation model for scATAC-seq using peak-to-gene aligned pretraining for cell representation, annotation, batch correction, and gene-expression prediction.
- ChromFound — Genome-aware scATAC-seq foundation model pretrained on 1.97M cells across tissues and disease contexts for zero-shot cell representations, annotation, and cross-omics prediction.
- Atacformer — Transformer foundation model for scATAC-seq that learns embeddings of cis-regulatory elements for clustering, annotation, and reference mapping.
- CLM-X — Multi-way Transformer foundation model jointly handling RNA-only, ATAC-only, and paired RNA–ATAC single-cell inputs for integration, translation, annotation, and perturbation prediction.
Other Omics Foundation Models¶
- CAPTAIN — Multimodal foundation model pretrained on co-assayed single-cell RNA and protein for joint representation learning and cross-modal downstream tasks.
- HiCFoundation — Hi-C foundation model pretrained on large-scale chromatin-contact maps for 3D-genome analysis, epigenomic prediction, and single-cell Hi-C adaptation.
- OmicsFM — Modality-agnostic molecular-expression foundation model with matched proteomics, bulk-transcriptomics, and single-cell-transcriptomics checkpoints.
- VirTues — Spatial-proteomics foundation model learning marker-aware representations across proteins, cells, niches, and tissues from multiplexed imaging.
- MethylGPT — Transformer foundation model for DNA methylation pretrained on >150K human methylomes across thousands of datasets, with 3M/7M/15M parameter variants.
- CpGPT — DNA methylation foundation model pretrained on >150K samples for zero-shot imputation, array conversion, reference mapping, and downstream phenotype prediction.
- Casanovo Foundation — Tandem mass-spectrometry proteomics foundation model that reuses a pretrained Casanovo spectrum encoder for spectrum quality, chimericity, and post-translational-modification prediction.
RNA Foundation Models¶
- RNA-FM — General-purpose RNA foundation model pretrained on large-scale RNA sequences for structural and functional representation learning.
- RiNALMo — RNA language-model family pretrained on tens of millions of RNA sequences for secondary-structure and functional prediction tasks.
Genomics Foundation Models¶
- Nucleotide Transformer — Foundation model for genomic sequences across multiple species.
- DNABERT — Pre-trained bidirectional encoder for DNA sequence analysis.
- DNABERT-2 — Improved genome foundation model with efficient tokenization.
- Enformer — Transformer model predicting gene expression from DNA sequence.
- Basenji — Sequential regulatory activity prediction from DNA sequences.
- Caduceus — Bidirectional equivariant long-range DNA sequence model based on Mamba.
- Evo — Long-context genomic foundation model (up to 1M tokens).
- Evo 2 — Genome foundation model trained on 9 trillion DNA base pairs across all domains of life with a 1M-token context window and single-nucleotide resolution.
- AlphaGenome — Long-context DNA model predicting multimodal regulatory outputs including expression, splicing, chromatin features, and contact maps at near base-pair resolution.
- modernGENA — ModernBERT-style DNA foundation-model family pretrained on hundreds of vertebrate genome assemblies for efficient long-sequence regulatory modeling.
- HyenaDNA — Long-range genomic foundation model handling sequences up to 1M tokens with sub-quadratic attention.
- Borzoi — Extended successor to Enformer for predicting RNA-seq coverage from long genomic sequence windows (524 kb) with improved resolution.
- DeepSEA — Deep learning framework for predicting chromatin effects of sequence alterations with single-nucleotide sensitivity across thousands of chromatin features.
- Sei — Sequence-to-function framework learning a genome-wide regulatory activity code from DNA sequences for variant effect prediction.
- GPN (Genomic Pre-trained Network) — Masked language model for DNA sequences enabling zero-shot variant effect prediction without requiring functional annotations.
Update & Link Rot Policy¶
- Link validity is monitored by the Link Check workflow.
Data Schema & Contribution Workflow¶
- Data schema reference:
docs/data/SCHEMA.md. - Contribution guide:
contributing.md.