Ensembling methods for protein-ligand binding affinity prediction.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Accurate protein-ligand binding affinity prediction is crucial for computer-aided drug discovery.
• Existing deep learning methods often lack accuracy and generalization due to reliance on single models.

Purpose of the Study:

• To develop a novel deep learning approach, Ensemble Binding Affinity (EBA), for accurate protein-ligand binding affinity prediction.
• To enhance prediction accuracy and generalization by ensembling multiple deep learning models trained on diverse input features.

Main Methods:

• Trained 13 deep learning models using combinations of 5 input features, incorporating cross-attention and self-attention layers.
• Explored all possible model ensembles to identify optimal combinations for binding affinity prediction.
• Utilized 1D sequential and structural features of protein-ligand complexes, avoiding complex 3D structural data.

Main Results:

• Achieved a Pearson correlation coefficient (R) of 0.914 and a root mean square error (RMSE) of 0.957 on the CASF2016 benchmark dataset.
• Demonstrated over 15% improvement in R-value and 19% in RMSE on CSAR-HiQ datasets compared to the CAPLA predictor.
• Exhibited superior performance across all metrics on five benchmark datasets compared to state-of-the-art methods.

Conclusions:

• The EBA method significantly improves protein-ligand binding affinity prediction accuracy and robustness.
• The ensemble approach effectively leverages diverse models and features for enhanced predictive power.
• This work contributes to accelerating drug development by improving the success rate of potential drug candidates.