DeepChem FAQ

These answers cover how to choose a DeepChem model architecture, train on your own data, judge whether a prediction is reliable, and how DeepChem compares to a dedicated ADMET panel.

Frequently asked questions

Which model architecture should I use?

For most molecular property prediction tasks, start with MPNN (Message Passing Neural Network). It provides strong general-purpose performance across a wide range of endpoints. If interpretability matters, try AttentiveFP which provides attention-based atom importance scores. For very large datasets (>50,000 molecules), DMPNN or ChemBERTa may offer improved performance. For small datasets (<200 molecules), classical models like Random Forest or XGBoost with ECFP fingerprints often outperform deep learning.

How does DeepChem compare to ADMET-AI?

DeepChem and ADMET-AI are complementary tools:

• DeepChem provides a flexible framework with multiple model architectures and the ability to train custom models on your own data. Best when you need custom endpoints or want control over the modeling pipeline.
• ADMET-AI provides curated pre-trained models specifically for ADMET endpoints with high-quality training data. Best for standard ADMET screening without custom model training.

For comprehensive compound assessment, run both and compare predictions.

Can I train on my own data?

Yes. Upload a CSV with SMILES and experimental labels. DeepChem supports both regression (continuous values like IC50) and classification (binary labels like active/inactive) tasks. Minimum recommended dataset size is 50 molecules, with 500+ providing more reliable models. The platform handles train/validation/test splitting, hyperparameter defaults, and early stopping automatically.

What is transfer learning and when should I use it?

Transfer learning uses a pre-trained model as a starting point and fine-tunes it on your specific dataset. Use it when:

• Your dataset is small (<500 molecules)
• The pre-trained model covers a related endpoint
• You want to leverage prior chemical knowledge

Transfer learning typically improves prediction accuracy by 10-30% compared to training from scratch on small datasets.

How do I know if a prediction is reliable?

Three indicators of prediction reliability:

1. Uncertainty estimate: Low uncertainty values indicate confident predictions. Enable ensemble or MC-dropout uncertainty to get quantitative confidence.
2. Applicability domain: Check whether the molecule is within the model's domain. Out-of-domain molecules have less reliable predictions.
3. Consensus across models: If predictions from MPNN, GCN, and AttentiveFP agree, the prediction is more likely reliable.

What is the applicability domain?

The applicability domain defines the chemical space where a model's predictions are expected to be accurate. It is determined by comparing each input molecule to the training data using molecular fingerprint distances. Molecules that are structurally dissimilar to any training compound are flagged as out-of-domain. Out-of-domain predictions should be treated as rough estimates, not quantitative values.

Can I predict multiple properties at once?

Yes. Multi-task models predict multiple endpoints simultaneously from a single SMILES input. This is more efficient than running separate models and can improve accuracy for related endpoints by sharing learned molecular representations. Select multiple endpoints in the configuration, or train a custom multi-task model on a dataset with multiple label columns.

How many molecules can I process in a batch?

Up to 10,000 molecules per prediction job. For larger libraries, split into multiple batches. Prediction time scales roughly linearly with batch size; 1,000 molecules typically take 5-15 seconds. Training on custom data supports datasets up to 100,000 molecules.

Read next