Probability of improvement (PI) is an acquisition criterion that estimates how likely a candidate is to outperform the best observed binding-affinity result. It helps an active-learning system choose which molecule to evaluate next. section 9.2.5
Use PI within a drug-discovery active-learning loop:
- Train a surrogate model on molecules whose binding affinities have already been evaluated. section 9.4
- Score the unlabeled molecular pool by each candidate’s estimated probability of beating the current best result. section 9.2.5
- Evaluate the selected molecules with an expensive method such as docking, add the results to the training data, and repeat. section 9.4
Deep docking for VEGFR-2
PI was one of several acquisition functions used to prioritize compounds for docking against the VEGFR-2 kinase target (PDB ID: 4AG8). The broader active-learning experiment recovered 73 of the top 100 molecules after 10 iterations while docking only a small fraction of the library. section 9.4

