Claude Skill MartBrowse skillsWhat is a Skill?
 
← Back to list

PyTDC (Therapeutics Data Commons)

Safely discover and use Therapeutics Data Commons datasets, splits, evaluators, benchmark groups, and molecular oracles via the PyTDC package.

Data & AnalyticsAdvanced★ 33,030⑂ 3,248AI score 7/10Last updated: Aug 9, 2026

What it does

Drives the full workflow for Harvard's Therapeutics Data Commons through the PyTDC package (pinned to 1.1.15).

  • Download-free discovery: discover_metadata.py lists tasks, datasets, benchmarks, evaluators, and oracles from installed package metadata
  • Task-aware splits: covers random, scaffold, cold_split, combination, and time, with explicit warnings that scaffold splitting does not prove absence of analog/temporal/label leakage
  • Exact evaluator names: PCC (not Pearson), ROC-AUC, micro-f1, kappa, plus notes on thresholds and metric direction
  • Benchmark groups: specialized classes such as admet_group, plus the rule that leaderboard aggregation needs at least five prediction dictionaries
  • Molecular oracles: bounded local scoring (e.g. QED) allowed; remote, docking, distribution, and composite oracles deliberately refused
  • Network/storage policy: enforces discover → plan → user approval → --execute, with cache_audit.py for read-only cache manifests

Who it's for

  • Researchers and grad students working on therapeutic ML (ADMET, DTI, DDI, molecular generation)
  • Teams that must review dataset licenses and download sizes before pulling data
  • Anyone who wants an AI agent to stop before downloading multi-hundred-megabyte archives or sending molecular structures to remote services

Example uses

  1. "Show me the ADME datasets without downloading anything" → runs discover_metadata.py --kind datasets --task ADME and returns bounded name/count summaries
  2. "Prepare Caco2_Wang with a scaffold split, seed 42" → prints the plan (license, size, cache dir, metric) first, then loads only after you approve with --execute
  3. "Score my predictions against the ADMET leaderboard" → validates the JSON prediction plan, asks before downloading the admet_group archive, then aggregates via evaluate_many

· · · Install guide · · ·

Install in the Claude app (no terminal)
  1. Download the ZIP with the button below.
  2. In Claude, open Settings → Capabilities and turn on 'Code execution and file creation'. (one time)
  3. Go to Customize → Skills → + → 'Upload a skill' and upload the ZIP.
↓ Download ZIP
Install in Claude Code

Let Claude do it — paste this into Claude Code

Install the skill I found on Claude Skill Mart.
Copy the skills/pytdc folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/pytdc/.
When it's done, tell me in one line what this skill can do.

Install with a command instead

git clone https://github.com/K-Dense-AI/scientific-agent-skills.git && mkdir -p ~/.claude/skills && cp -r scientific-agent-skills/skills/pytdc ~/.claude/skills/

⚠ This is a third-party skill. Check the source repository before installing.

  1. Open a terminal and clone the repo: git clone https://github.com/K-Dense-AI/scientific-agent-skills.git
  2. Create the skills folder and copy it in: mkdir -p ~/.claude/skills && cp -r scientific-agent-skills/skills/pytdc ~/.claude/skills/
  3. Install uv if you don't have it: curl -LsSf https://astral.sh/uv/install.sh | sh
  4. Create an isolated environment: uv venv --python 3.11 .venv-pytdc
  5. Preview the install size: uv pip install --dry-run --python .venv-pytdc/bin/python "setuptools==80.9.0" "PyTDC==1.1.15" (~120 packages, hundreds of MB)
  6. Confirm you have disk space, then rerun without --dry-run to install.
  7. Restart Claude Code and try: "use the pytdc skill to list ADME datasets".
  8. Only download real data after reviewing the plan and approving --execute (plus --download for MolGen corpora or oracle checkpoints).
View source on GitHub ↗License: MIT