DeepChem Molecular ML

What it does

Packages an end-to-end molecular machine learning workflow around DeepChem 2.8.0:

• Data loading: SMILES, SDF files and protein sequences into NumpyDataset / DiskDataset
• Featurization: circular fingerprints, graph-convolution inputs, RDKit descriptors
• Splitting: scaffold splitting as the honest default, plus stratified and Butina splitters
• Modeling: RF/XGBoost baselines → MultitaskRegressor → GCN, GAT, MPNN, AttentiveFP → ChemBERTa / GROVER / MolFormer fine-tuning
• Benchmarks: MoleculeNet datasets (Tox21, BBBP, Delaney) with published splits
• Three runnable scripts: predict_solubility.py, graph_neural_network.py, transfer_learning.py

It also documents real pitfalls: data leakage from random splits, GNNs underperforming fingerprints, small-data overfitting, lazy-import backend errors and the MKL 2025 / PyTorch conflict.

Who it's for

• Drug-discovery and chemistry researchers prototyping ADMET, toxicity or solubility models
• Graduate students reproducing MoleculeNet baselines
• ML engineers comparing molecular representations (fingerprints vs graphs vs pretrained embeddings)
• Teams extending into materials property prediction or protein/DNA sequence analysis

Example uses

1. Solubility baseline: "Train a regressor on my CSV (smiles, logS) and evaluate with a scaffold split" → CircularFingerprint + Random Forest with metric report.
2. Tox21 GNN comparison: "Train GCN and AttentiveFP on Tox21 and compare ROC-AUC" → multitask classification experiment built on the included script.
3. Small-data transfer learning: "I only have 300 actives and the model overfits" → ChemBERTa fine-tuning plus dropout and BalancingTransformer suggestions.

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