Method Profile
Combine separate three-dimensional kinase-pocket and molecule graphs, predict affinity with an eight-model ensemble, and calibrate uncertainty before retrospective selection experiments.
Prioritize kinase–compound binding experiments and compare exploration with conservative short-list selection.
Davis and KIBA experimental-affinity resources filtered for available structures; kinase-pocket and compound graphs; separate PDBbind complex-input comparison.
New-drug, new-protein, both-new and low-sequence-identity splits; five random split repetitions; eight independent network initializations; post-hoc uncertainty scaling on held-out validation data.
Inputs
Protein Data Bank kinase structures with KLIFS-defined 85-residue pockets, UniProt reference sequences and PubChem compound structures; published Davis dissociation constants or KIBA combined bioactivity scores.
Selection settings
Simulated KIBA active learning from 1% of training data; Davis 10,000-pair exploration pool; uncertainty-aware Bayesian acquisition functions for broad exploration or top-ten selection.
Calibrated uncertainty rescaling leaves the predicted mean unchanged. Distinguish retrospective revelation of known labels from prospective new binding experiments and keep Davis/KIBA assay metrics separate.
Compare structural-input ablations, baseline models, unfamiliar drug/protein splits and uncertainty-based selection policies.
Five random split repetitions; eight-model ensembles. KIBA acquisition starts at 1% of training data; Davis exploration evaluates 1,000 of 10,000 pairs against the top 500 known binders.
Sequence/molecule models without three-dimensional inputs, complex-input models, uncalibrated uncertainty and random/greedy acquisition.
KDBNet outperformed tested non-three-dimensional comparators and was comparable to slightly better complex-input models. KIBA active learning matched full-data performance at about 50% of training data; Davis exploration recovered 50% of the top 500 pairs after 1,000 selections.
A geometric neural network predicts binding from separate kinase and compound structures, then uses uncertainty to guide retrospective candidate selection. Benchmarks included unseen proteins and drugs; simulated active learning matched full-data performance with half the training data.
