The DAVIS dataset is a benchmark for drug-target affinity prediction. It records experimentally measured binding strengths between kinase inhibitors and protein kinases, supporting regression models that predict continuous affinity rather than only bind/no-bind labels. section 11.1 section 11.1.1
Use these properties when preparing or evaluating DAVIS data:
- Represents 68 unique compounds, 442 protein targets, and 30,056 measured drug-target pairs; a commonly curated version contains 433 kinases after removing nine records. section 11.1.1
- Focuses on kinase selectivity, where conserved ATP-binding sites make cross-reactivity and off-target effects important modeling concerns. section 11.1.1
- Stores affinity as pKd, calculated from Kd in molar units with pKd = -log10(Kd); higher pKd indicates stronger binding. section 11.1.1 section 11.5.5
- Creates repeated exposure because most compounds are tested against many proteins, making memorization and test-set leakage important evaluation risks. section 11.5.5
- Supports warm-start evaluation on unseen pairings, while cold-drug, cold-target, and cold-drug/cold-target splits measure generalization to unseen partners. section 11.5.5
Graph-based training record
A training example pairs a molecular graph with a protein representation and assigns the measured pKd as its target value. PyTorch Geometric stores graph features, edges, and the affinity label in a Data object, then batches variable-sized molecular and protein graphs with Batch.from_data_list. section 11.5.5
