The Virtual Biotech

A multi-agent AI system for pharmaceutical target identification and due diligence, built on the Claude Agent SDK. A Chief Scientific Officer (CSO) agent orchestrates a pool of specialist agents — each with access to curated biomedical databases via Model Context Protocol (MCP) servers — to answer arbitrary drug discovery and target biology queries.

---

Setup

1. Create the conda environment

conda env create -f environment.yml
conda activate vbt

The environment includes all dependencies for the CLI, MCP servers, and analysis scripts (Python 3.11, scanpy, anndata, CELLxGENE Census, R integration, etc.).

Live runs need the Claude Code CLI, which the pinned SDK provides. The claude-agent-sdk wheels bundle the CLI binary the SDK drives, and it is used ahead of anything on your PATH — so the environment above needs no separate install. If the SDK ever reports CLINotFoundError (pip having fallen back to the source distribution, say), install the CLI yourself — the native installer needs no Node.js:

curl -fsSL https://claude.ai/install.sh | bash

See Anthropic's setup guide for other install methods. (The audit tooling, tests, and verify need no CLI at all.)

2. Obtain the data

Open Targets Platform data (required)

Download the Open Targets Platform 25.09 data release — the exact archived version this system was built and tested on — in Parquet format. The loader expects a directory with one subdirectory per dataset, e.g.:

/path/to/open_targets/
├── target/
├── credible_set/
├── evidence/
├── drug_molecule/
└── ...

Tahoe-100M drug perturbation data (optional — enables the Tahoe functional-genomics tools)

The functional genomics MCP server's drug-perturbation tools use Tahoe-100M pseudobulk differential expression results. These are not distributed with this repository — obtain your own copy of the DE results from the Tahoe-100M dataset and point TAHOE_DATA_PATH at them. Without TAHOE_DATA_PATH set, the DepMap essentiality tools still work; only the Tahoe drug-perturbation tools are inactive. The loader expects the following layout under TAHOE_DATA_PATH:

tahoe/data/
├── tahoe_permissive_padj010.parquet       # permissive pseudobulk DE (padj < 0.10)
├── pseudobulk_de_significant/             # significant (padj < 0.05)
├── pseudobulk_de_high_quality/            # high quality (padj < 0.05, |FC| > 0.5)
└── metadata/
    ├── gene_metadata.parquet
    ├── drug_metadata.parquet
    ├── cell_line_metadata.parquet
    └── sample_metadata.parquet

3. Set environment variables

Environment variables can be set in a .env file in the project root (loaded automatically via python-dotenv) or exported in your shell.

Two kinds of paths — don't confuse them:

• Reference data (read-only; you download and provide it): OPEN_TARGETS_DATA_PATH — plus the optional TAHOE_DATA_PATH — point at large external datasets (see step 2 above). The app only ever reads these; it never writes to them.
• Working directory (writable; the app creates it): the optional MCP_OUTPUT_DIR is a directory the app makes and writes into. It starts empty; you only choose where it lives.

Required:

ANTHROPIC_API_KEY="sk-ant-..."
OPEN_TARGETS_DATA_PATH="/path/to/open_targets"   # READ-ONLY Open Targets dump you download (see step 2)

Optional — data:

TAHOE_DATA_PATH="/path/to/tahoe/data"            # optional; enables the Tahoe functional-genomics tools (see step 2)

Optional — output directory (defaults to data/ under the project root):

MCP_OUTPUT_DIR="/path/to/data"       # MCP server file outputs (parquet query results)

Optional — web app access password (only used by the Gradio web UI):

BIOTECH_APP_PASSWORD="choose-your-own"   # overrides the built-in default login password

The Gradio app (gradio_cso_app.py) is gated by a simple access password. It ships with a weak built-in default and is intended to bind to localhost; set BIOTECH_APP_PASSWORD to your own value before exposing it beyond your machine.

---

Apptainer (Linux cluster only)

vbt.def builds a self-contained Apptainer image with the conda environment baked in. Apptainer is Linux-only; for macOS or Windows, use the conda environment directly (see Setup — tested on macOS and Windows).

The source code is not embedded — bind-mount the project directory at runtime so you can update code without rebuilding.

Build

Run from the project root:

apptainer build vbt.sif vbt.def

Run

Use apptainer exec — not apptainer run — and set the working directory to the bind-mounted project with --pwd. Configure the MCP servers from inside the container first: setup_mcp.py records absolute interpreter and server paths, which differ between the container and a host conda environment.

apptainer exec \
  --bind /path/to/project:/workspace \
  --bind /path/to/open_targets:/data/open_targets \
  --pwd /workspace \
  vbt.sif python setup_mcp.py          # once per image

apptainer exec \
  --bind /path/to/project:/workspace \
  --bind /path/to/open_targets:/data/open_targets \
  --pwd /workspace \
  vbt.sif python run.py

mcp_config.json is environment-specific, so regenerate it whenever you switch between the conda environment and the container.

Environment variables (API keys, data paths, etc.) can be passed with --env or by bind-mounting a .env file into /workspace. Nothing else is needed for live runs: the CLI the SDK drives is bundled inside the image's Python environment, so it depends on neither PATH nor your host $HOME.

Test the image

apptainer test vbt.sif

---

Running the CLI

conda activate vbt
python3 setup_mcp.py     # once per environment — writes mcp_config.json
python3 run.py

setup_mcp.py records this environment's absolute interpreter and MCP server paths in mcp_config.json; run.py reads that file at startup and cannot run without it. Re-run it after moving the checkout or rebuilding the environment. source activate.sh and ./run.sh do this step for you.

By default the CSO and its specialist agents run on claude-sonnet-4-5-20250929. Pass --model with any current Anthropic API model ID to change it — this sets the CSO, and the specialists inherit it (the chief-of-staff and scientific-reviewer agents always run on Haiku):

python3 run.py --model claude-opus-4-6

Each run creates a new timestamped session directory under sessions/.

Interactive commands

Once the session starts and the CSO is ready:

You: Evaluate CD276 (B7-H3) as an immunotherapy target in NSCLC

# Multi-line input
You: """
Analyze OSMR as a target for ulcerative colitis:
1. Expression in disease-relevant cell types
2. Genetic evidence from GWAS
3. Existing drug modalities
"""

/summary   — print per-turn cost and agent breakdown
/done      — end session and write output files
/help      — show available commands
Ctrl+C     — graceful exit, writes output files

Session outputs

Each session creates a timestamped directory under sessions/:

sessions/
└── 20260524_103000/
    ├── session_report.json   # Full structured data: turns, costs, agent traces
    ├── transcript.md         # Human-readable conversation with reasoning traces
    ├── trace.jsonl           # Fine-grained event log (tool calls, sub-agent I/O)
    └── workspace/            # Files written by agents during the session

---

Architecture

Specialist agents

The CSO delegates to a flat pool of specialists, each with access to specific MCP servers:

Agent	Role	Data sources
genomics-analyst	GWAS, L2G, QTL colocalization	Open Targets Genetics
functional-genomics-analyst	CRISPR essentiality, DepMap (cancer only)	DepMap, Tahoe
single-cell-analyst	Cell-type expression, scRNA-seq	CELLxGENE Census
fda-safety-officer	Drug warnings, adverse events, mouse phenotypes	Open Targets
bio-pathways-ppi-analyst	Reactome pathways, GO, protein interactions	Reactome, STRING
clinical-trialist	Clinical trials, cancer genomics	ClinicalTrials.gov, cBioPortal
target-biologist	Druggability, protein structure, localization	Open Targets, GTEx
medchem-pharmacologist	Drug development, modality ranking	ChEMBL, Open Targets
chief-of-staff	Web intelligence, field overview	WebSearch, WebFetch
scientific-reviewer	Quality assurance, rigor review	(read-only)

MCP servers

Local MCP servers provide structured access to biomedical databases:

Server	Data source
expression	GTEx bulk RNA-seq expression
genetics	Open Targets Genetics (GWAS, L2G, QTL)
target	Open Targets target annotations and druggability
drug	Open Targets drug mechanisms and safety
disease	Open Targets disease ontology
association	Open Targets target-disease associations
single_cell	CELLxGENE Census
functional_genomics	DepMap CRISPR essentiality + Tahoe drug perturbations
pathway	Reactome pathways + Gene Ontology
interaction	Protein-protein interaction networks
clinicaltrials	ClinicalTrials.gov + cBioPortal cancer genomics

---

License

MIT License. See LICENSE for details.