DeepFrag: a deep convolutional neural network for fragment-based lead optimization.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence

Background:

• Machine learning (ML) is advancing computer-aided drug discovery (CADD), particularly in binding-affinity prediction and virtual screening.
• Lead optimization, a crucial CADD stage for enhancing ligand binding, is underutilized by ML methods.

Purpose of the Study:

• To introduce DeepFrag, a deep convolutional neural network (CNN) designed for predicting chemical fragments to improve ligand binding affinity.
• To apply ML to the under-explored area of lead optimization in drug discovery.

Main Methods:

• Development of a deep convolutional neural network (DeepFrag) model.
• Training the model on receptor/ligand complex structures to predict fragment additions.
• Independent benchmarking using known ligands with deleted fragments.

Main Results:

• The DeepFrag model successfully identified the correct fragment from a set of over 6500 in 58% of cases.
• When the correct fragment was not selected, the top predicted fragment often exhibited chemical similarity, suggesting viable substitutions.
• The model demonstrates potential for guiding lead optimization strategies.

Conclusions:

• DeepFrag represents a novel ML application for lead optimization in CADD.
• The model shows promise in suggesting effective chemical fragments to enhance drug-target interactions.
• The trained model and software are publicly released to facilitate further research.