KDBNet Combines Three-Dimensional Structures and Calibrated Uncertainty for Kinase–Drug Ranking

Separate Structures Supply a Binding-Affinity Prediction

Kinase inhibitors are an established drug class, but experimentally exploring every kinase–compound combination remains expensive. A 2023 Nature Machine Intelligence study introduced KDBNet, a neural-network model that combines the three-dimensional structure of a kinase binding pocket with a separate three-dimensional compound structure. It predicts binding affinity together with an estimate of model uncertainty.

The work was led by researchers at Georgia Institute of Technology and the University of Illinois Urbana-Champaign. Its graph neural networks represent protein residues and chemical atoms as nodes, with spatial relationships encoded between them. Learned protein and compound representations are then combined to estimate binding. A pre-existing structure of the protein bound to that particular ligand is not required.

The evaluation used the Davis and KIBA public affinity datasets. Davis contains measured dissociation constants, whereas KIBA integrates several bioactivity measurements into a combined score. Kinase-pocket structures came from the Protein Data Bank, using the KLIFS definition of an 85-residue binding pocket; compound structures came from PubChem. After filtering for available structures, nearly all compounds but only about 50–70% of proteins in the two resources remained. That selection defines the structural coverage of the benchmarks.

Testing Unfamiliar Drugs and Protein Targets

Four split settings tested unfamiliar compounds, unfamiliar proteins, both at once, and proteins with less than 50% sequence identity to the relevant training examples. Comparisons were repeated over five random train–test splits. These settings address different forms of transfer rather than relying only on new combinations of familiar drugs and targets.

KDBNet improved correlation and error measurements over the tested sequence- or two-dimensional-structure-based comparators. Removing the three-dimensional information from either the protein or compound reduced performance, supporting the contribution of both inputs. In a separate PDBbind comparison, methods given an actual protein–ligand complex retained a slight predictive advantage, while KDBNet achieved comparable performance using its less demanding separate-structure inputs.

The study therefore supports structure-aware affinity prediction within the evaluated datasets. It does not measure drug entry into cells, kinase inhibition in a cellular pathway or therapeutic response. Those downstream questions require evidence beyond the binding measurements used here.

Uncertainty Ranking and Calibration Are Different Checks

KDBNet estimated uncertainty from eight independently initialized versions of the same network. Their mean supplied the affinity prediction, and variation among their outputs supplied the uncertainty estimate. This captures uncertainty associated with the learned model; it was not designed to separately estimate every source of experimental measurement noise.

Predictions grouped by lower uncertainty generally had lower errors on the evaluated affinity data. The researchers also compared predicted interval coverage with the fraction of known measurements falling inside those intervals. A post-hoc calibration step rescaled uncertainty using a held-out validation set and improved that coverage relationship. The calibration changed interval widths while leaving the mean affinity prediction unchanged.

Retrospective Selection Tests the Value of That Uncertainty

The team next simulated active learning on KIBA. Starting with 1% of the training data, it repeatedly selected pairs with high model uncertainty, revealed their existing affinity labels and retrained. This exploration strategy reached performance comparable to full-data training after using about 50% of the training data. It improved more rapidly than random selection or a greedy strategy that selected the highest predicted affinities alone.

A second analysis combined predicted affinity and uncertainty in Bayesian optimization, a framework for sequential candidate selection. An exploration score favored promising but uncertain regions. In a Davis pool of 10,000 kinase–compound pairs, KDBNet recovered half of the 500 strongest-binding pairs after examining 1,000 pairs.

For a different objective, a conservative score favored high predicted affinity with lower uncertainty. The researchers compared the known measured affinities of the ten pairs prioritized by each method. This separates broad discovery of strong binders from choosing a short list intended to have fewer uncertain predictions.

All of these selection experiments reused existing affinity datasets. They demonstrate how uncertainty can change ranking and data acquisition under controlled retrospective conditions, with prospective biochemical or cellular confirmation still a separate step. The study’s useful distinction is between predicting a binding value, calibrating confidence in that value and choosing what to measure next.