Darwin: From Gene to Hypothesis — AI-Powered Research for Gene Regulation

Author: CAIRN Institute

Published: August 24, 2026

Read time: 10–12 minutes

#Genomics #Bioinformatics #GeneRegulation #SystemsBiology #OpenScience #AgenticAI

Explore Darwin on GitHub →

Darwin is powered by the CAIRN Institute GRN Atlas.

Quick Summary

Today we are publicly introducing Darwin — an AI-powered research environment for understanding gene regulation and developing testable biological hypotheses.

Darwin combines regulatory networks, expression, binding evidence, sequence context, pathways, traits, cross-species conservation, literature, perturbation analysis, and RNAi-oriented dsRNA design across human, mouse, Arabidopsis, tomato, petunia, pepper, potato, and rice. It includes both an interactive web UI and a structured skill layer — 100 documented research skills — for agent-driven workflows.

As of Saturday, August 22, 2026, the repository contains 100 documented Darwin research skills:

The current clean repository LLM results are:

Darwin is free for academic and non-commercial use. For commercial use, productization, deployment, or partnership discussions, contact CAIRN Institute.

Why We Built It

Researchers do not usually ask one-step questions.

A real workflow looks more like this:

Most biological software handles only one piece of that chain. Darwin was built to close that gap and give researchers a single place to move from question to hypothesis to next action.

What Darwin Is

Darwin is an AI-powered research environment for gene regulatory network analysis, built on the CAIRN Institute GRN Atlas. It combines:

The atlas currently supports eight species: human, mouse, Arabidopsis, tomato, petunia, pepper, potato, and rice.

Some layers are measured, some are projected, and some are computationally inferred or predicted. A core design rule is that these are never mixed without labeling.

What You Can Do With It

1. Explore a gene’s regulatory neighborhood

JAF13 regulatory network in Darwin showing activations, repressions, confidence levels, expression profile, and evidence sources for a petunia transcription factor.
Regulatory neighborhood of JAF13 (petunia) at 3-hop depth. Green edges are activations, red are repressions; line style encodes confidence. The right panel shows gene metadata, evidence sources, and per-tissue expression.

You can ask:

2. Interpret a gene set biologically

You can ask:

3. Compare across species

You can ask:

4. Predict interventions

You can ask:

5. Design and compare RNAi strategies

Side-by-side dsRNA comparison of EOBI and JAF13 with designed sequence and predicted downstream effects.
Side-by-side comparison of two RNAi targets (EOBI vs JAF13). Both achieve 100% specificity with zero off-targets. The atlas recommends JAF13 based on a smaller downstream program, then shows the designed 250 bp dsRNA sequence and predicted downstream gene effects with propagation paths.
Batch off-target screening of seven candidate genes ranked by dsRNA designability.
Batch screening of seven petunia anthocyanin genes ranked by off-target burden. Four candidates (JAF13, CHSB, AN2, EOBI) achieve fully specific dsRNA designs with zero off-targets in the best window.

For plant workflows, you can ask:

6. Turn analysis into decisions

Hit list first-pass interpretation and consensus ranking with overturn conditions.
A mapped hit list of petunia anthocyanin genes runs through first-pass interpretation (top candidates, upstream regulators, enrichment) and consensus ranking with explicit overturn conditions — what evidence would change the winner.

You can ask:

The Web UI

Darwin ships with a browser-based interface for interactive exploration and workflow-first analysis. The UI supports:

This release includes a working frontend for direct use by researchers as well as an API-backed analysis layer for more structured workflows.

The Skill Layer

Darwin also includes an AgentSkills-style skill library for structured tool use. The repository currently contains 100 documented skills:

99 callable skills across 7 research categories.
The 99 callable Darwin research skills organized across 7 research categories, from orientation and data readiness through perturbation design and collaborator handoff.

These skills are best understood by the high-level work they enable.

1. Orientation, search, provenance, and data readiness

These skills help a researcher identify the right genes, understand species support, clean inputs, and confirm what the atlas can defensibly answer before deeper analysis begins.

SkillWhat it does
grn-atlas-overviewOverview and workflow router for the atlas
grn-gene-searchSearch genes by symbol, alias, keyword, or partial text
grn-gene-infoRetrieve detailed metadata and TF status for a gene
grn-speciesReport per-species capability coverage
grn-statsReturn atlas-wide and per-species summary statistics
grn-provenanceExpose data sources, methods, and provenance metadata
grn-citationsExport BibTeX citations for atlas sources
grn-input-normalizationClean and normalize pasted biological input
grn-dataset-importParse and map gene lists or tabular inputs into atlas genes
grn-coverage-reportState whether a species is ready for a given analysis intent
grn-species-onboarding-planOutline the staged requirements for onboarding a new species

2. Core regulatory network structure and graph analysis

These skills answer the structural questions that sit at the heart of regulatory network analysis.

SkillWhat it does
grn-networkExplore upstream regulators and downstream targets
grn-pathfindingFind regulatory paths between genes
grn-subgraphExtract the induced subgraph for a gene set
grn-shared-regulatorsIdentify regulators shared by two or more genes
grn-regulonExtract a TF regulon at configurable depth
grn-regulon-compareCompare two TF regulatory programs
grn-upstreamRank upstream regulators for a gene set
grn-network-patternsDetect feed-forward loops, autoregulation, and bi-fan patterns
grn-centralityCompute centrality measures for genes in a network
grn-moduleDetect regulatory modules or communities
grn-motifQuery motif hits in gene promoters
grn-inferSurface expression-inferred regulatory edges
grn-diff-regulationCompare regulatory programs across tissues or conditions
grn-exportExport network edges with genomic and motif context

3. Expression, differential context, and support boundaries

These skills add biological context and also state clearly when deeper context layers are not yet available.

SkillWhat it does
grn-expressionRetrieve gene expression profiles
grn-coexpressionFind co-expressed partner genes
grn-differential-expressionCompare tissues, conditions, or imported DEG results
grn-celltype-regulationReport readiness and boundary conditions for cell-type analysis
grn-trajectory-regulationReport readiness and boundary conditions for trajectory analysis

4. Cross-species reasoning and conservation

These skills help researchers transfer hypotheses more carefully rather than assuming conservation without checking.

SkillWhat it does
grn-orthologyFind orthologs and inspect their local network context
grn-conservationCompare edge or network conservation across species
grn-transferabilityAssess whether a gene-level conclusion is likely to transfer

5. Perturbation, RNAi, promoter editing, and assay-oriented design

These skills move the platform beyond descriptive analysis and into intervention planning.

SkillWhat it does
grn-perturbationPredict downstream effects of silencing, knockout, or overexpression
grn-cascadeModel cascade effects from upstream interventions
grn-combinatorial-perturbationCompare pairwise and triple perturbation strategies
grn-dsrnaDesign or analyze dsRNA for RNAi silencing
grn-dsrna-screenScreen multiple genes for dsRNA designability and off-target burden
grn-variant-effectTest whether promoter-region variants overlap motif-supported sites
grn-promoter-edit-prioritizationRank promoter sites as editing targets
grn-crispr-designSuggest lightweight CRISPR guides
grn-primer-designSuggest lightweight primer pairs

6. Candidate ranking, uncertainty handling, and hypothesis comparison

These skills are designed for the part of research where a user needs to decide what to do next, not just inspect one more table.

SkillWhat it does
grn-candidate-triageRank a candidate gene list for a stated objective
grn-consensus-rankingCombine evidence layers into one weighted ranking
grn-counterfactual-analysisExplain what evidence would overturn the current winner
grn-confidence-boundaryState what is supported, unsupported, and ambiguous
grn-decision-boundaryProduce a decision-ready support and uncertainty summary
grn-hypothesis-compareCompare competing candidate hypotheses
grn-minimal-validationCompress a larger plan into the smallest defensible next step
grn-phenotype-targetingStart from a phenotype or design goal and ground candidates

7. Evidence synthesis, literature, planning, and collaborator handoff

These skills are the last mile between analysis and a usable research artifact.

SkillWhat it does
grn-evidence-auditAudit what evidence supports a gene or regulatory edge
grn-evidence-synthesisConvert evidence into a writing-ready synthesis
grn-literature-reviewRetrieve and classify recent external literature
grn-user-gene-set-analysisRun a first-pass interpretation over a user gene set
grn-experiment-prioritizationRecommend next analyses or experiments
grn-experiment-optimizerRe-rank follow-up options under budget, time, and assay constraints
grn-research-briefBuild a structured research brief
grn-validation-planBuild an execution-oriented validation plan
grn-study-packetAssemble a collaborator handoff packet
grn-study-reportProduce a collaborator-facing narrative report

Why the Skill Layer Matters

A useful research assistant is not just a model with access to endpoints. It needs a stable vocabulary of actions.

A question like:

Screen these RNAi candidates, pick the cleanest one, predict the perturbation effects, and summarize the affected biology.

is not one database call. It is a multi-step workflow.

In Darwin, that workflow can be expressed as:

Other realistic examples include:

That structure matters for reproducibility, testing, debugging, and failure analysis.

What the Full Skill Layer Enables for Researchers

Taken together, the skill system enables a broader class of work than a fixed dashboard or a raw database API. Researchers can use it to:

This matters because typical biological questions are rarely isolated. They are chained, conditional, and decision-oriented.

Testing With External LLMs

LLM agent testing results — GPT-5.4 and Nemotron-3-Ultra across single-skill and multi-skill orchestration.
Testing results across GPT-5.4 and Nemotron-3-Ultra for single-skill routing, multi-skill orchestration, and the full deterministic test suite.

To validate that the skill layer works not just in isolation but when driven by an external language model, we tested with two commercial LLMs — OpenAI’s GPT-5.4 and Nvidia’s Nemotron-3-Ultra (via OpenRouter) — neither fine-tuned on Darwin's skill definitions. The models received only the skill definitions and natural-language research questions, and had to select the correct tools, extract the right parameters, and chain multi-step workflows on their own.

The repository now has two useful LLM validation stories:

Current clean repository status: GPT-5.4

As of Saturday, August 22, 2026:

These current matrices are the best statement of the repository’s present skill-calling status.

Nemotron-3-Ultra comparison results

We also tested the skill system with Nvidia Nemotron-3-Ultra through OpenRouter to evaluate two distinct behaviors: single-skill selection and multi-skill orchestration.

Single-skill LLM testing

Earlier broad Nemotron single-skill reruns were useful primarily for frontmatter and routing hardening. They exposed which skill families were easy to confuse and helped tighten descriptions, overlap boundaries, and sequencing guidance.

Representative broad Nemotron rerun result:

Targeted Nemotron single-skill diagnostic subset on the later weak families:

Multi-skill orchestration testing

The latest August 22, 2026 paced Nemotron rerun did not complete the full matrices because the provider/model exited mid-run after partial completion.

Result:

These orchestration questions cover:

What Nemotron did well

Nemotron completed many substantial chained workflows successfully, including:

Where Nemotron remained weaker

The misses were concentrated in a few workflow families:

Failure shape in the latest August 22 Nemotron rerun:

That matters because Nemotron still demonstrates two distinct issues: reasoning/chaining quality gaps on some workflow families, and separate provider/model reliability issues on long runs.

What improved during testing

The Nemotron evaluation directly shaped the skill layer. We improved:

The result is not just a library of tools. It is a release that has been exercised with deterministic tests, UI tests, and external LLM-driven tool use.

Deterministic and Application-Level Testing

In addition to LLM validation, the repository has been exercised at the application and API layers.

Test tierCasesResultWhat it validates
Direct skill harness319319/319 PASSLocal execution and output validation for the legacy direct skill suite
HTTP skill harness8383/83 PASSREST-backed skill execution for the legacy HTTP suite
Integration4949/49 PASSCross-skill consistency, boundaries, performance, and idempotency
E2E Playwright2222/22 PASSBrowser workflows, navigation, and UI state
Backend API pytest165165/165 PASSAPI contracts, workflow endpoints, and science helpers
Frontend Vitest99/9 PASSFrontend regression coverage

Data, Evidence, and Trust

Darwin makes several distinctions explicit:

That makes the atlas useful for exploration without blurring the line between evidence classes.

Release Model

Darwin is being released publicly for academic and non-commercial use.

The repository is source-available under a non-commercial license. Academic research, education, and non-commercial experimentation are allowed under the repository terms. Commercial use, hosted productization, service deployment, or product integration requires separate permission.

The software license covers the code in the repository. Third-party data fetched into the atlas remains under the terms of the original upstream sources.

You can browse the code and documentation here:

For commercial use or potential partnerships, contact CAIRN Institute at info@cairninstitute.com.

What This Release Is For

This release is for researchers who want to:

It is also for teams interested in building agent-assisted biology workflows on top of a tested, skill-based analysis layer.

Frequently Asked Questions

Is this open source?

No, not in the OSI sense. It is source-available for non-commercial use. Academic research, education, and personal experimentation are allowed. Commercial use requires separate permission.

Does it separate curated and inferred results?

Yes. That distinction is a core design rule. Every inferred, predicted, or computationally derived result carries an explicit label distinct from measured data.

Can I use it through the web UI, the API, or agent tools?

All three. The UI supports direct interactive use, the FastAPI backend supports programmatic access, and the repository includes 100 documented skills in .agents/skills/ for LLM-driven workflows.

Further Reading

Questions or Feedback?

Contact us at info@cairninstitute.com.