Anthropic's life-science branch map needs one dataset first.

I am reading "gene branches in Anthropics" as a practical question about Anthropic's life-science product direction: which biological branches should a Claude-style research system cover, and which dataset creates the fastest path from broad AI synthesis to real, testable insight?

The answer:

Pull Open Targets Platform 26.06 first. It is the best first spine because it joins target-disease associations, genetic evidence, credible sets, locus-to-gene predictions, drugs, human-stage precedence, baseline expression, literature evidence, and downloadable open data into one agent-readable graph. Pull CZ CELLxGENE Census second when a hypothesis needs cell-type-level expression proof.

The branch map

Anthropic's recent life-science surfaces are not just a chat wrapper for papers. Claude Science is described as preconfigured for genomics, single-cell analysis, proteomics, structural biology, cheminformatics, and more, with specialist agents, reviewer behavior, and more than 60 scientific databases. Claude for Life Sciences adds scientific connectors, Agent Skills, prompt support, and use cases like literature review, protocol generation, bioinformatics, data analysis, and regulatory work.

That implies six practical branches for an operator building a demo or internal workbench:

Genomics and target genetics

Variants, GWAS studies, credible sets, gene burden, ClinVar-style evidence, and locus-to-gene models that decide which gene may sit behind a disease signal.

Single-cell and tissue context

Cell-type specificity, tissue distribution, disease-state expression, and the safety question: is the target where the biology needs it, or everywhere?

Proteomics and structure

Protein expression, interactions, structural annotations, subcellular location, confidence-aware structure reading, and assay-design consequences.

Chemistry and human-stage precedence

ChEMBL molecules, mechanisms of action, warnings, indications, trial reports, and whether any drug has already tested the target-disease pair.

Literature and protocol work

PubMed, bioRxiv, medRxiv, protocol QA, reviewer checks, citation grounding, and the mundane work of turning a clue into an auditable work packet.

Agent execution and remote data

MCP servers, APIs, open S3 buckets, local notebooks, reproducible pulls, and a route for the model to inspect data without pretending a paragraph is evidence.

Why Open Targets first

A groundbreaking life-science demo does not come from asking an LLM to brainstorm "novel targets" in a vacuum. It comes from making the model walk a real evidence graph, rank the branch that matters, and show what would change its mind. Open Targets is the rare public dataset that already has the right grain for that job.

The June 24, 2026 Open Targets 26.06 release added revamped baseline expression data, including single-cell RNA-seq and PRIDE mass spectrometry proteomics in the target prioritisation view. It also added GWAS updates, more credible sets and variants, stronger trial-report mining, trans-pQTL features in the locus-to-gene pipeline, and a stricter literature pipeline. That is exactly the bridge between Anthropic's branches: genetics, expression, proteomics, drugs, human-stage evidence, and literature.

The AWS Registry lists Open Targets as a no-subscription public dataset released every three months under CC0, and the Open Targets docs now expose the platform bucket at s3://open-targets-public-data-releases/platform/. I verified the bucket has 26.06/ live. The current output includes tables like association_overall_direct, evidence_gwas_credible_sets, credible_set, l2g_prediction, baseline_expression, trial-report tables, drug_molecule, target_prioritisation, and literature_vector.

The actual pull

The first pull should be narrow enough to run tonight and broad enough to avoid toy conclusions. Start with one disease family or indication area. Pull the direct associations, genetic evidence, credible sets, L2G predictions, target prioritisation, baseline expression, and trial reports. Then force the model to produce a ranked target-disease queue with visible confidence, novelty, cell-type context, and human-stage precedence.

Open Targets first: use it as the evidence spine and ranking table.
CELLxGENE second: validate cell-type or tissue claims in a dedicated single-cell corpus when Open Targets flags expression specificity as important.
AlphaFold/UniProt third: add structure and protein annotation only after a target survives causal, expression, and human-stage precedence checks.
ChEMBL fourth: expand chemistry once the target-disease pair is credible enough to ask about molecules, warnings, indications, and repositioning.

The insight pattern

The highest-value insight is not "gene X is associated with disease Y." Everyone can ask for that. The better insight is a contradiction the dataset makes visible:

a strong genetic signal with poor cell-type specificity,
a clean L2G prediction with no human-stage precedence,
a target with strong disease evidence but broad safety exposure,
a drugged pathway where the wrong indication has already tested the mechanism,
a literature-rich association that collapses when the evidence pipeline gets stricter.

That is where a Claude-style scientist agent earns its keep. It does not merely summarize the association. It finds the branch where the evidence changes the next experiment. It tells you whether to pull single-cell expression, structure, trial reports, or chemistry next. It names the fastest falsifier before naming the discovery.

Remote layer

Open Targets also has the remote affordance the demo needs. The docs list an official remote MCP endpoint at https://mcp.platform.opentargets.org/mcp, plus a local-server option and standard API/data-download paths. That matters because the demo should not be a static essay with vibes. It should be a repeatable agent route: ask a target question, inspect a current Open Targets table, cite the evidence, and show the pull command needed to reproduce the answer.

Demo

I built the public companion as the Anthropic Life Science ROI Console. It ranks the branch map, names Open Targets 26.06 as the default first pull, exposes the second dataset for each branch, and prints the local S3 pull plus remote MCP packet. The demo is deliberately inspectable: the ranking function, source stamp, remote endpoint, and pull command are visible in the page and testable from the browser.