GAT for Drug-Target Interaction Prediction
A dual-graph attention network that predicts binding affinity between cancer drug candidates and target proteins, benchmarked against a fingerprint baseline.
Overview
Drug-target interaction (DTI) prediction tries to answer a question that would otherwise take a wet lab months: given a candidate molecule and a target protein, are they likely to interact? Doing this experimentally is slow and expensive, so it’s a natural fit for learned models over molecular structure.
This project asks a specific version of that question — do graph attention networks actually beat the traditional fingerprint-based approach on DTI prediction? We implemented a GAT in PyTorch and benchmarked it head-to-head against a Morgan fingerprint + XGBoost baseline on the same splits.
A team project with three collaborators.
Data
We trained on CandidateDrug4Cancer: 29 cancer-related target proteins, 54,869 drug molecules, and 73,770 drug-protein interaction records. Molecules are represented as graphs with atoms as nodes and chemical bonds as edges. The DAVIS and KIBA benchmark datasets are wired up too.
Proteins get the same treatment rather than being flattened into a sequence embedding — we cache predicted structures from ColabFold/AlphaFold and build protein graphs from them, so both sides of the interaction are graphs.
Model
The core is a DualGraphAttentionNetwork that runs separate attention stacks over the drug graph
and the protein graph, then pools both into a joint representation for affinity regression.
- GPS-style layers combining local message passing with attention, carrying edge features through their own stream rather than discarding bond information.
- Multi-head attention — the final configuration uses 8 layers, 6 heads, and hidden size 96.
- LayerNorm at the feature-prep output, after the edge-stream residual, inside the feed-forward sub-layer, and on the graph-level pooled embeddings.
- Masked-language-model head over node categories, adapted from SSM-DTA, so unpaired molecules and proteins can contribute a semi-supervised pretraining signal.
Training uses Huber loss (β = 0.5), lr 3e-4 with weight decay 1e-3, ReduceLROnPlateau, batch size
64, and early stopping after 64 stagnant epochs against a 512-epoch cap, on an 80/10/10 split.
Model sizes are swappable through config files, which made ablations (no dropout, no pooling
dropout, SGD, lighter Huber) cheap to run.
Baseline
The comparison point is the standard cheminformatics pipeline: RDKit Morgan fingerprints for the molecule, protein embeddings for the target, and XGBoost regression over the concatenation — scored with the same accuracy-within-threshold metric as the GAT so the two are directly comparable.