The KIBA dataset supports drug-target affinity prediction by representing each drug-protein pair with a KIBA score: a normalized combination of Ki, Kd, and IC50 measurements. Unlike a bind/no-bind label, the score supports regression over continuous binding strength. section 11.1 section 11.5.5
A practical KIBA workflow uses the dataset as follows:
- Train a model on 118,254 measured interactions covering 2,111 compounds and 229 proteins. section 11.5.5
- Represent drugs and proteins as graph inputs when using graph neural networks, then predict one affinity score for each pair. section 11.5.5
- Interpret warm-start results carefully: test compounds and proteins may have appeared during training, although their tested combinations were held out. section 11.5.5
- Use cold-drug, cold-target, or cold-drug/cold-target splits to test generalization to unseen chemistry, targets, or both. section 11.5.5
Affinity prediction for a new pair
A model can learn from known drug-target interactions and estimate the KIBA score for a held-out combination, such as a known compound paired with a known protein. This warm-start task mirrors screening known drugs against new targets, but its results can overstate performance on genuinely new compounds or proteins. section 11.5.5
