virtual-biotech-agents

Multi-agent therapeutic-target-assessment orchestrator — a Chief-of-Staff briefing, decomposition + routing across four scientific divisions, a Scientific-Reviewer audit that can re-route to fill a gap, and a synthesized report; reproduces the loop from The Virtual Biotech (Zhang et al. 2026) over ClawBio skills. The skill makes no LLM call; reasoning is delegated to the driving agent.

For a therapeutic target (optionally + disease), query the live ClinicalTrials.gov API v2 (public, no key) and return a deduplicated list of registered trials, each carrying its NCT id and a deep link to the actual study record (clinicaltrials.gov/study/NCT…) — not the registry homepage. The clinical-precedent front-end. API-backed and reproducible, not page scraping.

For a therapeutic target (optionally + disease), run agentic web search via the Tavily Search API across three angles — recent literature/preprints, competitive/clinical landscape, and emerging safety signals — and return a cited, deduplicated snippet bundle. The current-evidence front-end. API-backed and reproducible, not page scraping.

Per-gene cell-type specificity profiler for single-cell atlases — computes the tau specificity index, expression bimodality coefficient, and ranked expressing cell types from an annotated .h5ad, with an optional clinical-trial-success prior. A pure analytic transform that chains downstream of scrna-embedding.

Fetch a gene × tissue slice of single-cell expression from the CZ CELLxGENE Census via its official API and write an annotated, log-normalized .h5ad — the data-acquisition front-end for scrna-embedding and celltype-specificity-profiler. API-backed and reproducible, not web scraping.

Given a gene and a tumour single-cell atlas (with a malignant-cell annotation), compute whether the target sits on the malignant compartment or on stroma/normal — expression in malignant vs non-malignant cells, malignant enrichment, and a tumour-cell-target call. Closes the "is the antigen on the cancer cells?" gap for tumour-targeting modalities (esp. ADCs).

For a drug, query openFDA and return a post-market safety snapshot — top FAERS adverse-event reaction terms with report counts, plus the label boxed warning if present. The FDA-Safety-Officer data front-end. API-backed and reproducible, not web scraping.

For a (target, disease) pair, fetch the Open Targets association score and its evidence breakdown by datatype — genetic association, somatic mutation, known drug, affected pathway, RNA expression, animal model, literature — in one GraphQL call. The target↔disease linkage-evidence axis, distinct from target-level prioritisation factors.

For one target, fetch the Open Targets prioritisation factors, tractability, and safety liabilities in a single GraphQL call and map them to the drug-discovery divisions they inform — DepMap essentiality (functional), tractability/clinical stage (modality), genetic constraint / mouse-KO / tissue specificity / safety liabilities (target safety). API-backed and reproducible.

For a gene, report its somatic mutation frequency across TCGA cancer types from the NCI Genomic Data Commons — fraction of each cohort carrying a simple somatic mutation in the gene, ranked. The somatic-driver axis (complements germline GWAS and expression). API-backed and reproducible.