# Build Therapeutic ML Workflows with PyTDC

Therapeutic machine learning requires reliable datasets, meaningful splits, and consistent evaluation. This skill guides PyTDC workflows from dataset selection through molecular optimization.

## Install

```bash
npx skillstore add davila7/pytdc
```

## Metadata

- Status: approved
- Slug: davila7-pytdc
- Skillstore revision: r2
- Version status: missing
- Tree hash: 4028b3712f4b4d043861cf1ef9a75afd5f9498585fe33acdc95db2d293ecb8cf
- Author: davila7
- GitHub username: davila7
- License: MIT
- Repository: https://github.com/davila7/claude-code-templates/tree/main/cli-tool/components/skills/scientific/pytdc
- Ref: c43861a65bb95efcae259cd161c9d6f4dc7eec6f
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: network, external\_commands
- Quality score: 82
- Quality tier: silver
- Public page: https://skillstore.pages.dev/skills/davila7-pytdc
- Manifest: https://skillstore.pages.dev/api/skills/davila7-pytdc/manifest

## Capabilities

- Selects PyTDC task classes and datasets for ADME, toxicity, interaction, and generation workflows.
- Explains scaffold, random, temporal, and cold split strategies for realistic model evaluation.
- Provides five-seed benchmark guidance with PyTDC evaluators and leaderboard-compatible summaries.
- Demonstrates molecular oracle scoring, batch evaluation, multi-objective weighting, and constraint checks.
- Documents molecule conversion, filtering, balancing, negative sampling, and biomedical identifier retrieval utilities.

## Use Cases

- Prepare ADMET Experiments: Choose property datasets, create scaffold splits, and select evaluation metrics for reproducible molecular prediction studies.
- Benchmark Interaction Models: Evaluate drug-target or drug-drug models with cold splits, multiple seeds, and standardized PyTDC metrics.
- Screen Generated Molecules: Score candidate molecules with property oracles and compare candidates against multi-objective constraints.

## Prompt Templates

### Load an ADME Dataset

```
Recommend a PyTDC ADME dataset for [property]. Explain its columns, task type, scaffold split, and suitable evaluation metric.
```

### Design a DTI Split

```
Plan a PyTDC drug-target interaction experiment using [dataset]. Compare random, cold-drug, and cold-target splits, then recommend one.
```

### Create a Benchmark Protocol

```
Design a five-seed PyTDC benchmark for [dataset group] and [model]. Include predictions, evaluation, aggregation, and reproducibility checks.
```

### Optimize Molecular Candidates

```
Plan a multi-objective molecular workflow using [oracles]. Define score normalization, weights, constraints, diversity checks, and experimental validation steps.
```

## Limitations

- Requires PyTDC and optional scientific dependencies that may need separate installation.
- Dataset loading and some oracles may download resources or require external API access.
- Example scripts use simulated predictions and do not provide a trained machine learning model.
- Oracle scores are computational estimates and cannot replace experimental or clinical validation.

## Best Practices

- Use scaffold or cold splits when random splits would overstate real-world generalization.
- Run multiple fixed seeds and report both mean performance and variability.
- Validate dataset names, metrics, oracle ranges, and optional dependencies before large experiments.

## Anti Patterns

- Do not treat simulated predictions in the example scripts as model baselines.
- Do not compare models that use different splits, seeds, preprocessing, or metrics.
- Do not interpret oracle scores as experimental evidence or clinical recommendations.

## Security Audit

- Audited at: 2026-07-23T13:29:01.191\+00:00
- Summary: All 90 static findings are false positives caused by Markdown backticks, dictionary keys, dataset partitions, and official resource links. No prompt injection, credential access, command execution, reconnaissance, or malicious network behavior was found.

## Stats

- Views: 528
- Downloads: 12
- Favorites: 0
- Popularity score: 0