Predicting protein-ligand binding residues with deep convolutional neural networks.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Protein-ligand interactions are crucial for biological processes.
• Identifying protein-ligand binding residues aids in understanding protein functions.
• Current computational methods are either sequence-based or 3D-structure-based, with 3D methods generally outperforming sequence-based ones.

Purpose of the Study:

• To develop a novel deep learning approach for *ab initio* protein-ligand binding residue prediction.
• To improve upon existing sequence-based methods by leveraging deep learning.
• To provide a powerful tool for predicting binding residues without relying on 3D structural information.

Main Methods:

• Proposed DeepCSeqSite, a sequence-based method utilizing a deep convolutional neural network.
• Stacked convolutional layers extract hierarchical features and capture long-distance residue dependencies.
• Features are combined via one-by-one convolution kernels and softmax for binding residue prediction.

Main Results:

• DeepCSeqSite demonstrated significant improvement over state-of-the-art methods, including COACH, with a minimum Matthews correlation coefficient (MCC) increase of 0.05.
• The method was evaluated on 151 nonredundant proteins and extended test sets.
• A data augmentation technique was explored, yielding slight performance enhancements.

Conclusions:

• DeepCSeqSite outperforms existing sequence-based and 3D-structure-based methods for protein-ligand binding residue prediction.
• The model achieves high accuracy without requiring 3D structural templates.
• The code, model, and datasets are publicly available for research use.