Multi-PLI: interpretable multi-task deep learning model for unifying protein-ligand interaction datasets.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Protein-ligand interactions are crucial for early-stage drug discovery.
• Computational methods, especially deep learning, show promise but face generalizability and data heterogeneity issues.
• Existing structure-based deep learning models struggle with limited data and dataset variations.

Purpose of the Study:

• To develop an interpretable multi-task model (Multi-PLI) for evaluating protein-ligand interactions.
• To address generalizability and data heterogeneity challenges in predicting protein-ligand binding and affinity.
• To provide biological interpretations of protein-ligand binding crucial for drug development.

Main Methods:

• An interpretable multi-task learning framework (Multi-PLI) was designed.
• The model unifies diverse datasets for concurrent classification (binding/non-binding) and regression (binding affinity) tasks.
• An occlusion algorithm was integrated for predicting key amino acids involved in binding.

Main Results:

• Multi-PLI demonstrated superior performance over traditional docking and machine learning methods.
• The model achieved competitive results against structure-based deep learning methods, even with similar training data sizes.
• The integrated occlusion algorithm successfully identified critical amino acids for protein-ligand binding.

Conclusions:

• The Multi-PLI model offers a robust and interpretable approach for assessing protein-ligand interactions.
• This method improves efficiency and accuracy in early-stage drug discovery.
• The model's ability to provide biological insights aids in understanding binding mechanisms and guiding drug design.