Prediction of protein-ligand binding affinity with deep learning

Area of Science:

• Computational chemistry and cheminformatics
• Pharmacology and drug discovery
• Artificial intelligence in medicine

Background:

• Accurate prediction of binding affinities is crucial for efficient drug design.
• Computer-aided methods, particularly deep learning, are increasingly vital in drug discovery pipelines.
• Various deep learning architectures, including CNNs, GNNs, and Transformers, have emerged for affinity prediction.

Purpose of the Study:

• To analyze and compare different deep learning methods for predicting protein-small molecule binding affinities.
• To evaluate the performance of selected deep learning models on the PDBbind v.2016 dataset.
• To explore strategies for improving affinity prediction accuracy by combining model strengths.

Main Methods:

• Classification of deep learning methods into CNNs, GNNs, and Transformers.
• Analysis of feature construction and model architectures for each deep learning approach.
• Experimental evaluation of four deep learning models on the PDBbind v.2016 core set, including statistical and visual analysis of predictions.

Main Results:

• Comparative analysis of the advantages and disadvantages of different deep learning models.
• Evaluation of prediction capabilities across various affinity intervals.
• Achieved a 1.6% improvement in Root Mean Square Error (RMSE) and a 2.9% increase in Pearson Correlation Coefficient (R) by combining model strengths, surpassing the state-of-the-art.

Conclusions:

• Deep learning methods show significant promise for accurate binding affinity prediction in drug discovery.
• Combining the strengths of diverse deep learning architectures can lead to enhanced prediction performance.
• Addressing current challenges in deep learning for affinity prediction is essential for future advancements.