MODEL LEARNING PATH
Chemprop / D-MPNN
Chemprop’s directed message-passing network predicts molecular properties from a chemical graph. Each bond message excludes its immediate reverse before atom states are pooled into a molecular representation.
Loading the interactive page…
Established directed-bond message passing as a practical learned molecular representation, with a documented path from property-prediction benchmarks to experimentally validated antibiotic-discovery studies.
Chapters
Encode the chemical graph
Prepared atom/bond descriptors become paired directed edges and initial bond states.
Represent each chemical bond as two directed edges and initialize their states.
Pass directed messages
Repeated nonbacktracking updates exclude the reverse bond, then incoming bond states produce atom embeddings.
Exclude immediate backtracking when updating directed bond messages.
Predict a molecular property
Pool atom embeddings and apply a task-specific feed-forward head.
Pool atom states into a graph representation and predict a molecular property.
Learn from labels
The supervised objective depends on whether the task is classification or regression.
Match the training objective to the prediction task.
All exercises in this guide
- Affine Projection
- Categorical and numeric molecular feature encoding
- Ragged segment sum and mean
- Directed molecular bonds and reverse-edge bookkeeping
- Affine projection and embedding lookup
- Nonbacktracking directed-bond recurrence and atom readout
- Masked Sum
- Masked Log-Sum-Exp
- Binary Cross-Entropy with Logits
- Masked Softmax
- Masked Mean
- Stable categorical and binary classification losses