PLAIG: Protein–Ligand Binding Affinity Prediction Using a Novel Interaction-Based Graph Neural Network Framework

Rapid prediction of protein–ligand binding affinity
is important
in the drug discovery process. The advent of machine learning methods
has increased the speed of these predictions. Previous machine learning
models based on structural, sequence, and interaction-based approaches
have shown potential but often tend to memorize training data due
to incomplete feature representations that lead to poor generalization
on external complexes. To address this challenge, here, we developed
PLAIG, a Graph Neural Network (GNN)-based machine learning framework
for generalized binding affinity prediction. PLAIG represents binding
complexes as graphs, integrating protein–ligand interactions
and molecular topology to uniquely capture interaction and structural
features. To reduce overfitting, we tested principal component analysis
(PCA) and ensemble learning with a stacking regressor. During benchmarking,
PLAIG achieved a PCC of 0.78 on 4852 complexes from the PDBbind v.2019
refined set and 0.82 on 285 complexes from the v.2016 core set, outperforming
many existing models. External validation on the DUDE-Z data set demonstrated
its ability to differentiate active ligands from decoys, achieving
an average AUC of 0.69 and a maximum AUC of 0.89. To enrich de novo
prediction capabilities for subsequent model versions, PLAIG was hybridized
with sequence- and structure-based models. The hybrid models achieved
an average PCC of 0.88 on well-known drug–target complexes,
with the best reaching a PCC of 0.98. Future work will incorporate
an explicit inclusion of a docking methodology into PLAIG’s
pipeline and assess its performance on de novo ligands. PLAIG is freely
available at https://plaig-demo.streamlit.app/.