SSnet: A Deep Learning Approach for Protein-Ligand Interaction Prediction.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Protein-ligand interaction (PLI) prediction is vital for drug discovery, but current Deep Neural Network (DNN) models' performance depends heavily on feature selection.
• Understanding how protein features influence PLI in DNNs remains challenging.

Purpose of the Study:

• To develop a novel DNN framework, SSnet, for accurate PLI prediction.
• To leverage protein secondary structure information (curvature and torsion) for enhanced PLI prediction.
• To provide insights into the structural features driving ligand binding through model interpretability.

Main Methods:

• Developed SSnet, a DNN framework utilizing protein secondary structure features (backbone curvature and torsion).
• Compared SSnet's performance against established machine learning and non-machine learning models using various metrics.
• Visualized SSnet's intermediate layers to identify influential protein structural elements for ligand binding.

Main Results:

• SSnet accurately predicts PLI by learning from protein secondary structure.
• The model identifies key binding locations, including binding sites, allosteric sites, and cryptic sites, irrespective of protein conformation.
• SSnet is unbiased towards specific molecular interactions and captures essential protein fold information for PLI.

Conclusions:

• SSnet offers a powerful new approach for PLI prediction, enhancing drug discovery efficiency.
• The framework provides interpretability, revealing structural determinants of ligand binding.
• This work opens avenues for secondary structure-based Deep Learning (DL) in broader protein research and drug design.