Interaction prediction in structure-based virtual screening using deep learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Structure-based virtual screening (SBVS) is crucial for identifying drug candidates.
• Current machine learning (ML) methods for SBVS face challenges with existing benchmark datasets.
• Developing robust ML models requires high-quality, diverse datasets for accurate evaluation.

Purpose of the Study:

• To introduce a novel deep learning architecture for enhanced structure-based virtual screening.
• To generate effective fixed-size molecular and protein fingerprints for binding prediction.
• To address the limitations of current benchmark datasets in evaluating ML-based SBVS methods.

Main Methods:

• A deep learning architecture employing learnable atom convolution and softmax operations was developed.
• Fixed-size fingerprints for proteins and small molecules were generated independently.
• Non-linear transformation and inner product calculation were used for binding potential prediction.
• A new benchmark dataset was constructed using data from DUD-E, MUV, and PDBBind databases.

Main Results:

• The proposed deep learning architecture successfully generates molecular and protein fingerprints.
• The method demonstrates potential for predicting binding affinity in virtual screening.
• Analysis revealed that existing benchmark datasets may not be sufficient for rigorous ML-based SBVS evaluation.
• The newly constructed dataset offers a more comprehensive evaluation platform.

Conclusions:

• The novel deep learning approach offers a promising advancement in structure-based virtual screening.
• The developed molecular and protein fingerprints can improve binding prediction accuracy.
• The introduction of a new benchmark dataset is essential for the reliable assessment of future ML-based SBVS tools.