DrugForm-DTA: Towards real-world drug-target binding affinity model

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Bioinformatics and computational biology

Background:

• Drug-target affinity (DTA) prediction is essential for efficient drug discovery.
• Existing computational DTA methods often lack 3D protein structures, which are frequently unavailable.
• There is a need for accurate and accessible DTA prediction tools.

Purpose of the Study:

• To develop a novel computational model for predicting drug-target affinity (DTA).
• To utilize structure-less representations of proteins and ligands for DTA prediction.
• To achieve DTA prediction accuracy comparable to experimental measurements.

Main Methods:

• Developed DrugForm-DTA, a Transformer-based neural network.
• Employed ESM-2 for protein encoding and Chemformer for small molecule ligand encoding.
• Trained and evaluated the model on standard benchmarks (Davis, KIBA) and a filtered BindingDB dataset.

Main Results:

• DrugForm-DTA demonstrated superior performance on KIBA and Davis benchmarks.
• The model achieved DTA prediction confidence comparable to single in vitro experiments.
• DrugForm-DTA outperformed traditional molecular modeling methods in DTA prediction efficacy.

Conclusions:

• DrugForm-DTA offers a highly accurate computational method for DTA prediction.
• The model's performance rivals experimental measurement accuracy.
• The study provides a valuable, filtered BindingDB dataset and a robust DTA prediction tool.