Reproduce “Protein-Ligand Scoring with Convolutional Neural Networks” on OpenResearch

About this paper

Matthew Ragoza, Joshua Hochuli, Elisa Idrobo, Jocelyn Sunseri, David Ryan Koes

Researchers introduced a method applying convolutional neural networks directly to three-dimensional protein-ligand complex representations for binding prediction. Their approach outperformed traditional scoring functions like AutoDock Vina, achieving an AUC of 0.815 versus 0.645 for pose prediction and 0.85-0.86 versus 0.68 for virtual screening.

Key insights

• Models trained specifically for pose prediction or virtual screening performed best on their respective tasks, though a combined training strategy yielded a compromise model effective for both.
• The convolutional neural network (CNN) learned to extract relevant chemical information from simple atom representations, suggesting its ability to derive features from spatial patterns rather than relying heavily on pre-encoded chemical knowledge.
• Interpretability methods, such as systematic masking, revealed which atoms and protein residues contributed most to binding scores, aligning with experimental mutation data and validating the CNN's learned principles.

Summary by alphaXiv.