Date

Thursday 17 Sep 2026

Time

15:15 - 15:45

Location

BE010

Supervisor

Anthe Janssen

2nd reviewer

Gerard van Westen

Jury

Francesco Buda

Protein kinases are among the most important classes of drug targets, yet structure-based affinity prediction across the kinome is constrained by the scarcity of experimentally determined ligand–protein complexes when compared to to the abundance of available bioactivity data. Molecular docking offers a bridge to this gap by generating 3-dimensional complexes suitable for structure-based machine learning. In this work we extend a docking-informed affinity-prediction framework with an environment-aware multimodal neural network ensemble designed by Xie et al. and apply it to a large-scale, kinase-focused dataset of 421,589 structure–ligand relations assembled from KLIFS co-crystal structures, Leiden Papyrus pChEMBL values, and QuickVina 2-GPU 2.1 docked poses. The model couples a protein-sequence encoder and a pocket–ligand cross-attention module with three structural graph modules, an equivariant pocket graph (EGNN), a heterogeneous complex graph (HGT), and an attentive ligand graph (AttentiveFP), feeding a multilayer perceptron that predicts binding affinity. Across all module combinations, trained in threefold under both a random and a ligand-balanced data split, the sequence modality accounted for nearly the full learnable signal (test R2 = 0.62 random, 0.28 balanced), whereas the structural graph modules neither performed well in isolation (R2 = 0.16 and 0.10) nor improved on the sequence-only baseline. The gap in performance between the random and balanced splits indicates that much of the apparent predictive power under random splitting reflects dataset redundancy rather than transferable interaction knowledge. Matching but not surpassing a simpler fingerprint-based baseline by Schifferstein et al., these results suggest that the limiting factor is not model capacity but the information content of the docked poses and the leakage-sensitivity of the data.