DeepChem quickstart

You want to predict a molecular property — solubility, toxicity, permeability — for one compound or a whole library. This walks you from entering a SMILES to interpreting the prediction in five steps: open the tool, add your molecules, pick a model and endpoint, submit, and review the results.

Quick molecular property prediction with .bloom* platform

DeepChem provides deep learning models for predicting molecular properties from SMILES strings. This guide walks through your first property prediction, from entering a molecule to interpreting the predicted properties.

Step-by-step prediction workflow

Common prediction scenarios

Single molecule profiling

1. Enter the SMILES of your compound
2. Select multiple endpoints (solubility, toxicity, permeability)
3. Review the comprehensive property profile
4. Use predictions to guide design decisions

Compound library screening

1. Upload a CSV with SMILES for your compound library (up to 10,000 molecules)
2. Select the target endpoints for screening
3. Sort results by predicted property values
4. Filter for compounds meeting your criteria

Post-generation screening

1. Run GenMol to generate novel molecules
2. Import generated SMILES into DeepChem
3. Predict ADMET properties for the generated set
4. Filter for drug-like compounds with favorable property profiles

Model comparison

1. Run predictions with multiple model architectures (MPNN, GCN, AttentiveFP)
2. Compare predictions across models for the same compounds
3. Use consensus predictions (average across models) for higher confidence
4. Identify compounds where models disagree for further investigation

Quality control checklist

Before submission

• SMILES strings are valid and parseable by RDKit
• Molecules are within the applicability domain of the selected model
• Selected endpoints are relevant to your research question
• Batch size is reasonable (under 10,000 molecules)

After prediction

• Predicted values are in reasonable ranges for the endpoint
• Uncertainty estimates are available and reasonable
• Out-of-domain warnings are noted for unusual molecules
• Predictions are consistent with known experimental data (spot checks)

Next steps after prediction

Compound triage:

• Filter compounds by predicted property thresholds
• Rank by multi-parameter optimization scores
• Feed top candidates into Compound Passport for comprehensive assessment

Iterative design:

• Identify property liabilities in your compound series
• Use DeepChem predictions to guide GenMol lead optimization
• Iterate between generation and prediction to converge on optimal candidates

Experimental validation:

• Prioritize compounds for experimental testing based on predictions
• Compare predicted vs. experimental values to assess model accuracy
• Use experimental data to fine-tune models for improved future predictions

Read next