Open-source tools across the cell-ag AI stack — media and cell-line modeling, bioprocess
scaling, metabolic and multi-omics analysis, mass spectrometry, workflow managers, AI agents,
and the data standards that connect them. Each entry links to its canonical home; the full
descriptions live in the canonical Software.md.
70 tools across 9 areas
Predicts the 3D structure of proteins from their amino acid sequence using deep learning. Can be used for design and engineering of cost-effective recombinant growth factors and signaling molecules. Predicting protein structure helps optimize stability, activity, and binding affinity, reducing the high cost associated with media components.
This is a collection of links to Protein Language Models, which are models trained to predict structure and function. Can be used for the design and engineering of cost-effective recombinant growth factors and signaling molecules.
Search large protein databases using monomers and multimers. Can be used to find known structures to infer function or guide engineering efforts of a new or target protein.
Open-source applied AI research code from the Alberta Machine Intelligence Institute (Amii) targeting cellular agriculture, hosted under the Amii-Applied-AI GitHub organization. The repository collects two Python subprojects: protein-thermostability-data-tools (code used in the development of a public protein thermostability dataset) and active-learning-for-cell-media (active-learning analysis applied to cell media optimization). MIT-licensed.
An open-source suite for Computational Fluid Dynamics (CFD) Used to model and a variety of environmental factors within bioreactors, to ensure viable conditions at scale.
A modeling environment for simulating cell-cell and cell-environment interactions. Enables the modeling of complex phenomena like cell proliferation, nutrient gradients within thick tissues, and the mechanics of cells interacting with scaffolds.
A physics-based framework specialized for multi-cell simulations. Can be used to model complex cell behaviors, including differentiation and nutrient limitations in tissue cultures.
The de-facto Python package for constraint-based modeling of metabolic networks, hosted by the openCOBRA consortium. Provides FBA, FVA, gene and reaction knockouts, flux sampling (OptGP, ACHR), gap-filling, production envelopes, dynamic FBA, and SBML I/O on a unified Model object that integrates cleanly with NumPy / pandas / SymPy. The reference implementation for FBA in the bioinformatics community (Ebrahim et al. 2013, BMC Systems Biology).
The MATLAB counterpart to COBRApy, also under the openCOBRA umbrella. Older but more feature-rich for advanced methods (thermodynamics-aware FBA via multiTFA, ME-models via COBRAme, elementary flux modes), and still the production tool of choice in many academic labs. For new cell-ag work, COBRApy is usually the better starting point unless a specific COBRA Toolbox feature is required. The current protocol is Heirendt et al. 2019 (Nature Protocols).
The genome-scale metabolic model test suite, providing automated QC and benchmarking for SBML GEMs. Computes >100 metrics covering annotation completeness, stoichiometric consistency, biomass formulation, growth simulation, and reproducibility, producing a versioned HTML report. Used by BiGG Models and the openCOBRA community as the de-facto curation standard; any cell-ag-specific GEM should be benchmarked with Memote before downstream use.
A web-based tool for building, sharing, and embedding visualizations of metabolic pathway maps, with bi-directional integration to COBRApy via JSON. Supports interactive overlay of flux distributions, reaction knockouts, and metabolite concentrations on hand-drawn or auto-generated pathway maps (King et al. 2015, PLOS Computational Biology).
Project page: https://escher.github.io.
A software application for simulation and analysis of biochemical networks and their dynamics, supporting deterministic ODE integration, stochastic Gillespie simulation, steady-state analysis, parameter estimation, sensitivity analysis, and metabolic control analysis. Cross-platform GUI plus Python bindings via basico; used as the simulation backend by Talk2Biomodels.
Project page: https://copasi.org.
A Python environment for modeling and simulating biological systems built around libRoadRunner (fast SBML simulator), Antimony (human-readable model language), and notebook-friendly workflows. Bridges between SBML, ODE simulation, parameter scans, and Python-native scripting. Maintained by the Sauro lab at University of Washington.
Project page: http://tellurium.analogmachine.org/.
A MATLAB toolbox for genome-scale metabolic model reconstruction, curation, and analysis, maintained by the Nielsen lab at Chalmers. Provides automated reconstruction from KEGG, MetaCyc, or template models, gap-filling, model curation tools, and integration with COBRA Toolbox; widely used to build species-specific GEMs in food, fermentation, and biopharma settings.
A Python package built on COBRApy that consolidates the strain-design MILP family — OptKnock, OptCouple, RobustKnock, Minimal Cut Sets (MCS / advanced MCS) — into a single API for in-silico engineering of metabolic networks. Useful beyond microbial work for identifying knockouts in cultivated mammalian cell GEMs that redirect flux toward biomass while suppressing lactate accumulation.
An integrated visual environment for metabolic modeling that wraps COBRApy with a GUI for FBA, FVA, Elementary Flux Modes, thermodynamic methods, Minimal Cut Sets, OptKnock, RobustKnock, OptCouple, and other strain-design techniques. Strong fit for teaching and for users who prefer interactive exploration over notebook-based work.
Docs: https://cnapy-org.github.io/CNApy-guide/.
Agent integration. Code-execution agents (Cursor, Claude Code, Biomni) can invoke any of these tools as Python; for COBRApy specifically, the cobrapy skill from K-Dense-AI's scientific-agent-skills (see the K-Dense-AI entry below) provides curated recipes that make this reliable in agent loops.
An ultra-efficient toolkit for molecular quantitative trait loci (molQTL) mapping across multi-omics data, from the Zhang group at China Agricultural University. The performance backbone of the FarmGTEx project family — eQTL, sQTL, mQTL, and other molQTL discovery at livestock-atlas scale, optimized for the throughput needed to handle the FarmGTEx tissue / sample matrices. Companion to Papers.md ref #143 (Teng et al. 2026, Nature Communications). Reproducibility deposits: ENA BioProject PRJEB58031, Zenodo 10.5281/zenodo.10072081 and 10.5281/zenodo.18280923.
An open-source C++ framework with Python bindings (pyOpenMS) for mass-spectrometry data analysis, from raw spectra processing through quantitative proteomics and metabolomics. Maintained by an international consortium led by the Kohlbacher lab (Tübingen); OpenMS 3 (Pfeuffer et al. 2024, Nat Methods) provides a modular toolkit (TOPP), Python bindings, KNIME nodes, Galaxy wrappers, and integrations with nf-core pipelines (e.g. nf-core/quantms). Widely used for flavor metabolomics, off-flavor characterization, and as a building block in larger workflow-manager pipelines.
Project page: https://openms.de/. pyOpenMS docs: https://pyopenms.readthedocs.io/.
A modular, open-source platform for LC-MS / GC-MS / IMS-MS data processing, maintained by the Pluskal lab (University of Münster) and the international MZmine consortium. Schmid et al. 2023 (Nat Biotech) describes the v3 release with multimodal MS support (LC-MS, IMS-MS, MS-imaging). Provides feature detection, alignment, gap-filling, MS/MS networking integrations (GNPS / FBMN, SIRIUS), and a CLI for batch processing. Standard preprocessor for flavor and natural-products metabolomics workflows.
Source: https://github.com/mzmine/mzmine. News and releases from mzio, the team behind MZmine: https://mzio.io/mzmine-news/.
A standalone Windows tool for DIA / DDA MS/MS spectral deconvolution and metabolite / lipid annotation, developed by Tsugawa et al. at RIKEN (Tsugawa et al. 2015, Nat Methods; the 2020 Nat Biotech lipidomics atlas paper extended MS-DIAL 4 to lipid identification). The de-facto standard for GC-MS deconvolution in flavor and food chemistry labs; ships with extensive built-in spectral libraries.
A C# tool for widely targeted metabolomics that processes multiple reaction monitoring (MRM) / selected reaction monitoring (SRM) data — plus SCAN and DIA MS/MS — developed by Tsugawa et al. (2013, Analytical Chemistry), same first author as the MS-DIAL entry above. Evaluates metabolite peaks by posterior probability and provides large-scale visualisation, data curation, and statistical analysis of widely-targeted metabolomics datasets — the targeted-quantitation complement to MS-DIAL's discovery-focused deconvolution.
Distributed via Zenodo: https://zenodo.org/records/11219831/latest.
The most-cited R / Bioconductor package for LC-MS / GC-MS metabolomics preprocessing, originally Smith et al. 2006 (Anal Chem) and continuously maintained since. Provides nonlinear retention-time alignment, peak picking, grouping, and gap-filling — the analytical workhorse of many academic metabolomics pipelines including the Galaxy-based Workflow4Metabolomics platform.
Bioconductor: https://bioconductor.org/packages/xcms/.
A cross-platform C++ library and command-line toolkit for mass-spectrometry data conversion and analysis, maintained by the Mallick lab at Stanford and an international community (Chambers et al. 2012, Nat Biotech). The msconvert utility is the universal first step in essentially every open MS pipeline — converting vendor-locked binary formats (.RAW, .D, .lcd, .wiff) to open standards (mzML, mzXML, MGF) so downstream tools can ingest the data.
Source: https://github.com/ProteoWizard/pwiz.
Java application for in-silico molecular formula determination and structure annotation from MS/MS spectra, maintained by the Böcker lab (University of Jena). SIRIUS 4 (Dührkop et al. 2019, Nat Methods) integrates fragmentation-tree-based formula prediction, CSI:FingerID for structure database search, CANOPUS for compound-class prediction, and COSMIC for confidence scoring. Standard tool for de-novo annotation in untargeted metabolomics and natural-products work.
In-silico fragmenter for MS/MS-based compound identification, originally Wolf et al. 2010 (BMC Bioinformatics) and substantially revised in MetFrag Relaunched (Ruttkies et al. 2016, J Cheminform). Scores candidate structures from compound databases (PubChem, ChemSpider, KEGG) against measured fragmentation spectra. Integrated into many metabolomics workflows including nf-core/metaboigniter and Workflow4Metabolomics.
A machine-learning-based mass-spectral analogue search tool from the iomega consortium (de Jonge et al. 2023, Nat Comms). Uses Spec2Vec and MS2DeepScore embeddings to find spectral analogues in reference libraries, including for compounds without exact matches — directly addressing the long-tail "unknown unknowns" problem in flavor and natural-products metabolomics.
A standalone tool for in-silico compound identification from MS/MS spectra, developed by Tsugawa et al. at RIKEN as a companion to MS-DIAL. Combines isotope pattern matching, formula prediction, in-silico fragmentation, and database search against curated reference libraries (HMDB, FooDB, ChEBI, PubChem, KEGG) to score candidate structures. Widely used in untargeted flavor metabolomics for annotating unknown compounds from GC-MS / LC-MS spectra.
The Global Natural Products Social Molecular Networking platform — a web-based MS/MS analysis platform from the Dorrestein lab at UCSD (Wang et al. 2016, Nat Biotech). Provides community-curated reference spectral libraries, Feature-Based Molecular Networking (FBMN), Ion Identity Molecular Networking (IIMN), and analog search via spectral similarity. Standard tool for compound annotation, dereplication, and pattern discovery in flavor and natural-products metabolomics workflows.
Also listed as a spectral-library reference in Databases.md / Mass Spectrometry Spectral Databases — dual-listed because the community-curated reference libraries are themselves a queryable database.
An R package on Bioconductor implementing PCA, PLS, OPLS, and OPLS-DA for chemometric analysis of metabolomics and other -omics data (Thévenot et al. 2015, J Proteome Res). Provides the multivariate engine in the Workflow4Metabolomics Galaxy platform and is widely used in flavor / sensory metabolomics for sensory-instrumental correlation, biomarker discovery, and quality-control modeling. See also ropls in the K-Dense-AI scientific-agent-skills collection for an agent-callable wrapper.
A comprehensive web-based and R-based platform for statistical, functional, and visual analysis of metabolomics data, maintained by the Xia lab at McGill in collaboration with the Wishart lab (current v6, Pang et al. 2024, Nucleic Acids Research). Provides modules for univariate / multivariate statistics, pathway enrichment, network analysis, biomarker discovery, time-series and dose-response analysis, plus a companion R package MetaboAnalystR for scripted workflows. The most-cited tool in food metabolomics applications; widely used for sensory-instrumental data analysis in flavor and off-flavor work.
An R / Bioconductor package for the integration and exploration of single- and multi-omics datasets, maintained by the Le Cao group (University of Melbourne) (Rohart et al. 2017, PLOS Computational Biology). Implements PCA, PLS, sparse PLS-DA, and the DIABLO method for multi-block supervised classification across heterogeneous data blocks. The methodological standard for fusing sensory panel scores, instrumental volatilome / non-volatilome data, and microbiome / -omics layers — directly applicable to multi-omics flavor and cell-ag quality-prediction work.
A food-specific R package for the analysis of sensory data, maintained by the Le and Husson group at Agrocampus Ouest (Lê & Husson 2008, J Sensory Studies). Implements QDA, napping, sorted napping, projective mapping, preference mapping, and panel performance diagnostics; built on top of FactoMineR. The de-facto open-source tool for descriptive-analysis sensory panel data in academic food science, including alt-protein flavor benchmarking.
A community-curated Nextflow pipeline for untargeted LC-MS metabolomics preprocessing, identification, and analysis within the nf-core consortium. Provides a containerized, parameterizable chain from raw MS data through peak picking (XCMS, CAMERA), alignment, annotation (MetFrag, CSI:FingerID), and statistical analysis. The most directly applicable existing reproducible Nextflow pipeline for flavor and off-flavor metabolomics work in cell-ag contexts.
A Galaxy-based collaborative research infrastructure for metabolomics, providing a large library of workflow modules for LC-MS, GC-MS, FIA-MS, and NMR preprocessing, identification, and statistical analysis (Giacomoni et al. 2015, Bioinformatics). Integrates XCMS, ropls, MetFrag, CAMERA, and many other established tools into a single web-based interface backed by French institutional compute (IFB), accessible without local installation. The closest "one-stop" reproducible food / flavor metabolomics platform; widely used in academic flavor metabolomics work.
A Snakemake workflow for untargeted LC-MS/MS metabolomics, maintained by the Biosustain group at DTU (Kontou et al. 2023, J Cheminform). Wraps MZmine 3, SIRIUS, GNPS / FBMN, and Ion Identity Molecular Networking (IIMN) into a containerized, parameterizable pipeline that produces feature tables, annotations, and molecular networks from raw MS data. Closest existing Snakemake pipeline that flavor-metabolomics work can build on.
A community-curated Nextflow pipeline for amplicon sequencing analysis (16S, 18S, ITS) within the nf-core consortium (Straub et al. 2020). Provides a full preprocessing + ASV-calling + taxonomic-assignment + diversity-analysis chain (Cutadapt → DADA2 → QIIME 2 → Picrust2) suitable for microbiome work in fermented foods, fermentation-based alt-protein, and the microbial communities relevant to cultivated-meat bioreactors. Routinely used for the microbiome arm of multi-omics flavor studies.
A community-curated Nextflow pipeline for shotgun metagenomic assembly and binning (Krakau et al. 2022, NAR Genomics and Bioinformatics). Produces metagenome-assembled genomes (MAGs) with quality control, taxonomic classification, and functional annotation — the analytical complement to ampliseq for higher-resolution microbial-community profiling. Relevant to cultivated-meat work involving complex bioreactor microbiomes, scaffold biofilms, or fermentation co-culture analysis.
An ecosystem for building AI scientists from any language or reasoning model, providing a unified interface across open- and closed-weight models and a shared tool/data/analysis environment. Provides infrastructure that bespoke biomedical AI agent projects have historically had to reinvent. Companion to the Gao et al. 2025 paper introducing ToolUniverse (Papers.md ref #41).
Project page: https://aiscientist.tools/. Docs: https://zitniklab.hms.harvard.edu/ToolUniverse/.
A general-purpose biomedical AI agent from Stanford's SNAP lab (Leskovec group) that performs autonomous multi-step research workflows across drug discovery, genomics, clinical analysis, and adjacent biomedical domains. Companion to the Huang et al. 2025 paper (Papers.md ref #49).
Project page: https://biomni.stanford.edu. AWS tutorial (Biomni + Bedrock AgentCore): https://aws.amazon.com/blogs/machine-learning/build-a-biomedical-research-agent-with-biomni-tools-and-amazon-bedrock-agentcore-gateway/.
An open-source AI-agent platform from VirtualPatientEngine targeting drug discovery and pharmaceutical R&D. Hosts a family of LLM-based agents — Talk2Biomodels (kinetic biological models), Talk2KnowledgeGraphs, Talk2Cells, and Talk2Scholars — that share infrastructure for tool use and reasoning over biomedical resources. Companion to the Wehling et al. 2025 Talk2Biomodels paper (Papers.md ref #50).
Docs: https://virtualpatientengine.github.io/AIAgents4Pharma/.
A bioinformatics-focused LLM agent (Bioinformatics Retrieval Augmented Digital Assistant) combining retrieval-augmented generation with bioinformatics tool orchestration. Targets automatic biomarker discovery, enrichment analysis, and general bioinformatics automation — patterns that map cleanly onto cell-ag tasks like cell-type marker discovery and pathway-enrichment analysis on cultivated-tissue scRNA-seq. Companion to Papers.md ref #94 (Pickard et al. 2025, Bioinformatics).
An agentic system for the design of virtual cell models from the Gerstein lab at Yale. Coordinates multiple LLM agents to plan, configure, and execute cell-modeling workflows — directly relevant to cell-ag for cell-type-specific virtual-cell pipelines that media-optimization, perturbation-response prediction, and bioprocess workflows can build on. Companion to Papers.md ref #93 (Tang et al. 2026).
Sakana AI's framework for fully automated open-ended scientific discovery via large language models, performing end-to-end idea generation, experimentation, and paper drafting. Version 2 (SakanaAI/AI-Scientist-v2) extends the system to workshop-level automated discovery via agentic tree search; the AI-Scientist-ICLR2025-Workshop-Experiment repository archives the run whose AI-generated paper passed peer review at an ICLR 2025 workshop. Companion to Lu et al. 2024 (Papers.md ref #45) and Yamada et al. 2025 (Papers.md ref #47).
Project home: https://sakana.ai/.
An open-source retrieval-augmented generative agent for answering questions over scientific literature with verified citations, from the FutureHouse research lab. Released as PaperQA in 2023 and substantially extended in PaperQA2 (2024), which achieves superhuman synthesis accuracy on literature questions. Companion to Papers.md ref #44 (PaperQA, Lála et al. 2023) and ref #46 (PaperQA2, Skarlinski et al. 2024).
FutureHouse cookbook (docs): https://futurehouse.gitbook.io/futurehouse-cookbook. Commercial spinout — Edison Scientific: https://edisonscientific.com/ (docs).
A gymnasium of language-agent environments for challenging scientific tasks, from the FutureHouse lab. Agents are framed as Language Decision Processes (LDP) and trained and evaluated against tasks spanning molecular cloning, scientific-literature QA, and protein stability — providing the reusable training-and-evaluation substrate beneath FutureHouse's task-specific agents. Companion to Papers.md ref #160 (Narayanan et al. 2024). Apache-2.0; built on the LDP framework, with Aviary docs at https://docs.edisonscientific.com/aviary.
An Aviary-based data-science agent that operates inside Jupyter notebooks, also from FutureHouse. It plans and executes notebook-based analyses, pairing the Aviary environment with the kind of exploratory bioinformatics work — single-cell exploration, media-response analysis — that cell-ag teams increasingly delegate to notebook agents. Apache-2.0; no companion paper at time of curation.
K-Dense-AI is an AI co-scientist ecosystem combining a commercial agent platform (K-Dense Web, which autonomously executes complex science / engineering / healthcare / finance tasks end-to-end) with a substantial open-source stack of agent infrastructure:
cobrapy (FBA / metabolic modeling — see Metabolic Modeling & Strain Design), pyopenms (mass spectrometry), scanpy / scvi-tools / anndata (single-cell analysis), rdkit / datamol / medchem (cheminformatics), biopython / bioservices / gget (bioinformatics utilities), cellxgene-census, pyzotero, and many others.SKILL.md / AGENTS.md file consumable by Claude Code or similar agents.Not an MCP server but a documentation-and-prompt-context layer that pairs well with code-execution agents like Biomni, Cursor, and Claude Code.
Project page: https://k-dense.ai.
An agentic skills framework and software-development methodology authored by Jesse "obra" Vincent (obra/superpowers) — one of the most-starred Claude Code skill collections. Provides domain-agnostic skills for planning, debugging, code review, and execution that compose cleanly with cell-ag-specific skill packs (e.g. K-Dense-AI's scientific-agent-skills) when assembling an agent stack for a cell-ag lab. Shell-based, MIT-licensed. K-Dense-AI's science-superpowers is a science-domain reimplementation of this methodology for data analysis, swapping test-driven development for pre-registration.
A meta-tool that converts documentation websites, GitHub repositories, and PDFs into Claude AI skills with automatic conflict detection. The closest existing automation for the pattern an AI-augmented cell-ag lab needs as it scales: turn a new wet-lab protocol PDF, a new bioinformatics package's docs, or a new GitHub library's README into a Claude Code / Cursor skill that the lab's agents can call directly — without hand-curating each integration. Python, MIT-licensed.
An open-source library of AI research and engineering skills covering vLLM, Megatron, GRPO, HuggingFace, and the broader LLM training and serving stack — maintained by Orchestra Research. Designed to package skills into Claude Code, Codex, or Gemini agents so they operate as fully-equipped AI research agents. Cell-ag teams building or fine-tuning their own biology foundation models (cf. TranscriptFormer) or running large-scale agentic workflows can pull from this library for the ML-training and serving infrastructure layer rather than reinventing it. MIT-licensed.
Seqera Cloud's AI assistant for bioinformatics workflows, providing an interactive co-scientist that helps users author and debug Nextflow pipelines, query workflow run data, and interpret results. Accessible through the Seqera CLI and Seqera Platform; no companion paper at time of curation.
A commercial lab-orchestration platform from Dotmatics that connects laboratory instruments, data systems, and AI assistance into a unified "connected digital lab" workflow — covering instrument integration, automated data flow, electronic-lab-notebook integration, and agent-assisted experimental planning. Marketed primarily to biopharma and biotech R&D groups; representative of the commercial-tooling layer that cell-ag startups increasingly evaluate as they scale beyond bench-scale workflows. See also the Dotmatics Luma webinar in Talks.md for an overview of the platform.
A family of generative foundation models for single-cell transcriptomics from the Chan Zuckerberg Initiative, trained on up to 112 million cells spanning 1.53 billion years of evolution across 12 species (Pearce et al. 2026, Science; see Papers.md ref #92). Provides state-of-the-art performance on cell-type classification and supports cross-species reasoning over transcriptomic data — directly relevant to cell-ag for translating biological knowledge between bovine, porcine, chicken, salmonid, and other livestock cells where annotated reference data is sparse (see the per-species pages in Datasets/ for the cell-ag-relevant data substrate). Distributed via CZI's Virtual Cells Platform with versioned releases.
Quickstart docs: https://virtualcellmodels.cziscience.com/quickstart/transcriptformer-quickstart. Announcement: https://chanzuckerberg.com/blog/transcriptformer-model-overview/.
A transformer-based foundation model for transfer learning in network biology from the Theodoris lab (Broad Institute / Gladstone), pretrained on ~30 million human single-cell transcriptomes via rank-encoded masked language modeling. Distributed exclusively through Hugging Face with tokenizer, pretrained weights, and example fine-tuning recipes for cell-type classification, gene-network inference, and in silico perturbation prediction; widely used as a single-cell-FM baseline. Companion to Papers.md ref #111 (Theodoris et al. 2023, Nature); pretraining corpus: Genecorpus-30M in Datasets/HumanReference.md.
A generative pretrained transformer for single-cell multi-omics from the Wang lab at the University Health Network (Toronto), trained on >33M cells spanning scRNA-seq, scATAC-seq, and CITE-seq. Provides downstream fine-tuning recipes for cell-type annotation, multi-batch integration, gene-regulatory-network inference, and perturbation prediction; one of the most-used baselines for newer single-cell foundation models. Companion to Papers.md ref #117 (Cui et al. 2024, Nature Methods). Documentation: https://scgpt.readthedocs.io/.
An early single-cell BERT-style foundation model for cell-type annotation from Tencent AI Lab Healthcare, treating individual genes as tokens with binned expression values. Released alongside Papers.md ref #112 (Yang et al. 2022, Nature Machine Intelligence); the independent re-evaluation in ref #113 (Boiarsky et al. 2024) and author reply in ref #114 (Yang et al. 2024) are core methodological reading for anyone benchmarking new single-cell FMs against existing baselines.
A large-scale foundation model on single-cell transcriptomics from BioMap Research, pretrained on ~50M cells with read-depth-aware encoding that explicitly handles the variable sequencing depths characteristic of public scRNA-seq corpora. Provides downstream applications spanning cell-type annotation, drug-response prediction, and perturbation-effect modeling. Companion to Papers.md ref #116 (Hao et al. 2024, Nature Methods).
Universal Cell Embeddings from Stanford's SNAP lab (Leskovec group) — a single-cell foundation model that represents each cell as an unordered set of expressed genes and each gene by its protein-language-model embedding, enabling zero-shot generalization to species and tissues never seen at training time. Releases include pretrained weights and zero-shot inference scripts for novel cell-type discovery across species. Companion to Papers.md ref #119 (Rosen et al. 2026, bioRxiv; Nature, in press at time of curation). The same lab's earlier SATURN method (ref #118, Rosen et al. 2024, Nature Methods) introduced the protein-LM-gene-embedding pattern that UCE generalizes — directly relevant to cell-ag where annotated livestock-species single-cell data is sparse and cross-species transfer is essential.
A generative pretraining model for single-cell deciphering, applying GPT-style autoregressive next-token prediction over gene-expression vocabularies. Smaller and earlier than scGPT or Geneformer, but methodologically important as one of the first demonstrations that next-token-prediction objectives (vs. masked-language-modeling) work for single-cell biology — the lineage that now includes Arc's State and Cell2Sentence. Companion to Papers.md ref #115 (Shen et al. 2023, iScience).
A framework for treating single-cell expression profiles as natural-language sentences — ordered lists of expressed gene symbols ranked by expression — enabling direct reuse of pretrained LLM architectures (and, in C2S-Scale, billion-parameter scaling) for single-cell biology. From the van Dijk lab at Yale. Companion to Papers.md ref #120 (Rizvi et al. 2026, bioRxiv). C2S-Scale project page: https://www.vandijklab.org/c2s-scale.
Graph-Enhanced gene-Activation Response Simulator — a graph neural network for predicting transcriptional outcomes of novel multi-gene CRISPR perturbations, including combinations never observed during training. From Stanford's SNAP lab (Leskovec group). Generalizes single-gene perturbation training data to combinatorial perturbation prediction via co-essentiality and gene-ontology graph priors. Companion to Papers.md ref #121 (Roohani et al. 2024, Nature Biotechnology).
Arc Institute's first-generation virtual cell model and companion evaluation framework, designed to predict stem-cell, cancer-cell, and immune-cell responses to drugs, cytokines, and genetic perturbations. Trained on ~170M observational and ~100M perturbational single-cell measurements across 70+ cell lines; uses a bidirectional transformer architecture with self-attention over cell sets and reportedly is the first model to consistently beat simple linear baselines on perturbation-response prediction. Released alongside cell-eval, the standardized evaluation framework for virtual-cell models. Companion to Papers.md ref #57 (Adduri et al. 2025, bioRxiv); see also the Arc Institute news article on State in OtherResources.md. The follow-on Stack model — companion to Papers.md ref #124 (Dong et al. 2026) — extends State with in-context learning, simulating cellular conditions via prompt engineering without further fine-tuning.
An LLM-based AI agent from Stanford's SNAP lab for designing genetic-perturbation experiments — including CRISPR-Cas9 single-gene and combinatorial knockouts — by reasoning over gene-function literature, prior screens, and experimental constraints. Demonstrates that an LLM agent with tool use can match or exceed specialized active-learning methods on hit-rate-driven experimental-design tasks. Companion to Papers.md ref #125 (Roohani et al. 2025, arXiv). Directly applicable to cell-ag as an off-the-shelf experimental-design layer for cell-line-engineering campaigns (selecting which TFs to overexpress for myogenic vs. adipogenic differentiation, or which media-pathway genes to knock down to test rate-limiting steps).
A large-scale benchmark for evaluating network-inference methods from single-cell perturbation data — including Perturb-seq, CROP-seq, and ECCITE-seq. Built around interventional ground truth from genome-scale CRISPR screens, providing standardized metrics, baselines, and dataset splits for ML methods that infer gene-regulatory networks. Companion to Papers.md ref #127 (Chevalley et al. 2025, Communications Biology).
A community hub for agentic biomedical systems — a registry of biomedical Model Context Protocol (MCP) servers plus a knowledgebase MCP server that exposes curated biomedical resources to LLM agents. Lets cell-ag teams plug standardized biomedical tools and data sources into agent stacks (Claude Code, Cursor, Biomni) without bespoke per-resource integration. Companion to Papers.md ref #133 (Kuehl et al. 2025, Nature Biotechnology). GitHub org: https://github.com/biocontext-ai.
A one-binary MCP server from GenomOncology unifying many biomedical knowledge sources — PubTator3, Europe PMC, ClinicalTrials.gov, MyVariant.info, cBioPortal, Reactome, Open Targets, MyDisease.info, MONDO, Monarch, DisGeNET — behind a single Model Context Protocol surface for LLM agents. MIT-licensed; the leanest existing MCP-native bridge between general biomedical literature, clinical-trial, and variant data and an agent stack. Sister project to BioContextAI, which catalogues biomedical MCP servers including BioMCP.
The de-facto XML-based standard for representing computational models of biological processes — metabolic networks, signaling pathways, gene-regulatory networks, and kinetic models. SBML is the interchange format for every genome-scale metabolic model catalogued in the per-species pages of the Datasets/ directory and the lingua franca of the constraint-based and kinetic modeling tools in Metabolic Modeling & Strain Design. Maintained by the SBML community with libSBML bindings across all major languages.
A schema language for authoring, validating, and transforming structured data models, with first-class support for ontology terms, code generation across languages, and export to JSON-Schema / SHACL / OWL. Increasingly used to define machine-readable metadata schemas for biological datasets and knowledge graphs — the structured backbone that agentic AI systems need in order to reason reliably over cell-ag data resources.
Project PISCES (Process Integration & Synthesis using Chemical Engineering Standards) standardizes process flowsheet data into a machine-readable Standard Flowsheet Format (SFF) for AI-powered knowledge extraction and analysis. For cellular agriculture, a standardized flowsheet format is the missing substrate for AI-assisted bioprocess design, scale-up modeling, and techno-economic analysis — letting agents reason over cultivated-meat process designs the way they reason over SBML metabolic models. SFF documentation: https://projectpisces.org/?page=sff-docs.
Process-flowsheet background (what SFF standardizes) — for readers approaching this from the AI / biology side, the LibreTexts Foundations of Chemical and Biological Engineering chapter on chemical processes and process diagrams and the ScienceDirect "Flowsheet" topic overview introduce the flowsheet concept and its notation.
Cell-ag application context — The Unjournal's cultivated-meat cost-modeling work, namely the unjournal/cm_pq_modeling repository and its techno-economic comparison of cultured-chicken cost models, is exactly the kind of bioprocess techno-economic analysis that a standardized flowsheet format like SFF is designed to make reproducible and machine-comparable.
An interlocking stack of MIT-licensed Python tools and registries supporting biomedical semantics and pragmatics. Each component is independently usable, and together they cover the full lifecycle of biomedical-entity identification, normalization, and cross-linking. Directly relevant to cell-ag agentic workflows that need to reason consistently across the livestock-genomics, metabolic-modeling, sensomics, and chemistry resources elsewhere in CAAIL.
GitHub org: https://github.com/biopragmatics. Core components: