Sfcnn: a novel scoring function based on 3D convolutional neural network for accurate and stable protein-ligand...

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Accurate scoring functions are crucial for computer-aided drug design (CADD) but remain a bottleneck.
• Existing machine learning and deep learning models aim to improve binding affinity predictions.
• Novel featurization methods are needed to enhance the accuracy of protein-ligand binding affinity evaluation.

Purpose of the Study:

• To develop a novel scoring function for predicting protein-ligand binding affinities.
• To improve the accuracy and reliability of binding affinity predictions in CADD.
• To enhance the interpretability of deep learning models in drug discovery.

Main Methods:

• Development of a novel featurization method for protein-ligand complexes.
• Implementation of a deep 3D convolutional neural network (3D CNN) architecture.
• Utilizing Grad-CAM for interpretability of the 3D CNN model's intermediate layers.

Main Results:

• The 3D CNN model achieved high Pearson correlation coefficients (e.g., 0.7928 on CASF-2016).
• The model demonstrated accurate and stable performance across multiple independent datasets.
• The proposed featurization method simplified feature engineering and improved interpretability.

Conclusions:

• The developed 3D CNN scoring function is effective for predicting protein-ligand binding affinity.
• The model shows significant potential to improve virtual screening efficiency.
• This work accelerates the identification of potential drug candidates and novel biologically active compounds.