PLANET: A Multi-objective Graph Neural Network Model for Protein-Ligand Binding Affinity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Predicting protein-ligand binding affinity is crucial for drug design.
• Existing deep learning models often require 3D complex structures and focus solely on affinity prediction.

Purpose of the Study:

• To develop an efficient graph neural network model, PLANET, for predicting protein-ligand binding affinity.
• To improve upon existing methods by incorporating multi-objective learning and diverse training data.

Main Methods:

• Developed PLANET, a graph neural network model.
• Input includes 3D binding pocket graphs and 2D ligand structures.
• Trained using a multi-objective approach with binding affinity, contact map, and distance matrix tasks.
• Incorporated PDBbind data and non-binder decoys for training.

Main Results:

• PLANET achieved scoring power comparable to top deep learning models on the CASF-2016 benchmark.
• Demonstrated superior performance in virtual screening on DUD-E compared to other models.
• Showed comparable accuracy to Glide on LIT-PCBA with significantly reduced computation time.

Conclusions:

• PLANET offers a balance of accuracy and efficiency for binding affinity prediction.
• The model shows potential as a valuable tool for large-scale virtual screening in drug discovery.
• Its multi-objective training and input flexibility contribute to its robust performance.