DecoyFinderNetAna: Application of Graph Convolution Neural Networks for Accurate Classification of True Small...

Area of Science:

• Computational biology
• Cheminformatics
• Machine learning in drug discovery

Background:

• Drug discovery faces challenges in differentiating true ligands from decoys during virtual screening.
• Deep learning and computational biology offer solutions to improve early-stage screening accuracy.
• DecoyFinderNetAna is introduced as a Graph Convolutional Neural Network (GCNN)-based approach for this purpose.

Purpose of the Study:

• To develop and evaluate DecoyFinderNetAna for distinguishing true ligands from decoys in compound libraries.
• To enhance the accuracy of early-stage virtual screening in the drug discovery pipeline.
• To provide a computationally efficient method for filtering compound libraries.

Main Methods:

• Compounds represented as molecular graphs with chemical features.
• A GCNN classifier trained on 85 protein targets from the DUD-E database.
• Case study on Mycobacterium tuberculosis Thymidylate kinase, validated with molecular docking and MD simulations.

Main Results:

• Achieved high performance metrics: average sensitivity 0.973, specificity 0.993, AUC 0.983 across 102 targets.
• Demonstrated strong precision and recall.
• Case study showed predicted true binder had significantly higher binding affinity (-28.01 kcal/mol) than decoy.

Conclusions:

• Machine learning, specifically GCNNs, can significantly enhance early-stage virtual screening.
• DecoyFinderNetAna effectively filters decoys, reducing costs and prioritizing compounds.
• This GCNN approach offers a scalable, time-efficient alternative to traditional methods, accelerating drug discovery.