skills.rest
listentomi
listentomiCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

deepchem

Train and evaluate molecular machine learning models for property prediction and drug discovery.

Updated Aug 28, 2026
Search Tags:#drug discovery#featurization#smiles#deepchem#molecular machine learning#graph neural network#moluculenet
One-click install
npx skills add https://github.com/listentomi/Orcajack --skill deepchem-listentomi
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: deepchem
Source: https://github.com/listentomi/Orcajack/tree/main/skills/science/deepchem
Command: npx skills add https://github.com/listentomi/Orcajack --skill deepchem-listentomi

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires deepchem, scikit-learn, and includes scripts (resource) and references (resource) components.

What problem does it solve? Applying machine learning to chemistry and drug discovery requires specialized data loaders, molecular featurization, scaffold-aware splitting, and domain-specific models that general ML libraries do not provide. This Skill guides you through the complete DeepChem workflow so you can load molecular data, featurize it correctly, train appropriate models, and evaluate them on standard benchmarks without trial and error. ## Core Features & Use Cases - Molecular Data Loading & Featurization: Load SMILES, SDF, and FASTA data and convert molecules into fingerprints, descriptors, or graph representations with a decision tree for choosing the right featurizer. - Model Selection & Training: Covers Random Forest baselines, multitask deep networks, graph neural networks (GCN, GAT, AttentiveFP, DMPNN), and pretrained models (ChemBERTa, GROVER) with dataset-size-based recommendations. - MoleculeNet Benchmarks & Evaluation: Access 30+ curated datasets (Tox21, BBBP, Delaney) with scaffold splitting and standard metrics like ROC-AUC and R². - Use Case: Given a CSV of SMILES strings with solubility values, load it with CSVLoader, split with ScaffoldSplitter to avoid data leakage, train a MultitaskRegressor, and evaluate R² on a held-out test set. ## Quick Start Ask the AI to train a solubility prediction model on your CSV file of SMILES strings using DeepChem with scaffold splitting and report the test R² score.

Frequently Asked Questions about deepchem

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I predict molecular properties from SMILES strings in Python?▼

Use DeepChem's CSVLoader with a featurizer like CircularFingerprint to convert SMILES into model-ready features, split with ScaffoldSplitter, then train a model such as MultitaskRegressor or GCNModel. Evaluate with metrics like R² or ROC-AUC on the held-out test set.

Which featurizer should I use for molecular machine learning?▼

Use CircularFingerprint for traditional ML baselines, RDKitDescriptors for interpretable models, and MolGraphConvFeaturizer for graph neural networks. For transfer learning with pretrained models like GROVER, use the model-specific featurizer such as GroverFeaturizer.

Why should I use scaffold splitting instead of random splitting for molecules?▼

Random splitting lets structurally similar molecules appear in both training and test sets, causing data leakage and inflated scores. ScaffoldSplitter groups molecules by core structure so the test set contains genuinely novel scaffolds, giving a realistic estimate of generalization.

Does DeepChem support graph neural networks for molecules?▼

Yes, DeepChem includes GCNModel, GATModel, AttentiveFPModel, and DMPNNModel for molecular graph learning. These require graph featurizers like MolGraphConvFeaturizer and typically perform best on datasets larger than about 10K samples.

When should I use transfer learning with ChemBERTa or GROVER?▼

Use pretrained models when your dataset is small (under 1000 samples), contains novel scaffolds, or when you need rapid prototyping. Fine-tune with a lower learning rate such as 2e-5 for around 10 epochs rather than training from scratch.

Why is my graph neural network underperforming compared to fingerprints?▼

GNNs often underperform on small datasets below roughly 10K samples or with too few training epochs. Try training for 50-100 epochs, switching architectures to AttentiveFP or DMPNN, or using a pretrained model like GROVER.

Related Skills

Explore More Education & Research Skills→
eddieluengeddielueng
1

drug-discovery

Search ChEMBL for targets and bioactivities, estimate drug-likeness via PubChem, and check safety signals via OpenFDA.

Community
Intermediate
InternScienceInternScience
28

molclaw-mol-opt-physchem

Optimize drug molecule LogP, QED, and solubility through iterative SMILES edits.

Official
Advanced
Yezez9Yezez9

rowan

Run cloud-based quantum chemistry calculations and molecular simulations via a Python API.

Community
Advanced
SologaSologa
1

molfeat

Convert SMILES strings into over 100 molecular fingerprints and descriptors.

Community
Intermediate