Advancing Ligand Docking through Deep Learning: Challenges and Prospects in Virtual Screening.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Molecular docking (MD) is crucial for predicting protein-ligand interactions in structure-based virtual screening (SBVS).
• Traditional MD methods often simplify algorithms for speed, potentially sacrificing accuracy.
• Deep learning (DL) has shown promise in conformation prediction, exemplified by AlphaFold2, suggesting its potential to revolutionize MD.

Purpose of the Study:

• To review the current state of DL in augmenting MD for SBVS.
• To highlight challenges and future prospects in DL-based MD (DLLD).
• To discuss contributions and scholarly insights in this rapidly evolving field.

Main Methods:

• Overview of virtual screening (VS) and molecular docking (MD).
• Introduction to deep learning (DL) paradigms and their deviation from traditional methods.
• Analysis of challenges in DL-based MD (DLLD), including evaluation metrics, application scenarios, and physical plausibility.

Main Results:

• DL offers enhanced accuracy and speed in binding pose prediction compared to traditional MD.
• DLLD models can achieve higher success rates but may produce physically implausible poses.
• A shift is observed from blind docking to identifying binding sites within DL models.

Conclusions:

• DLLD holds significant potential to advance SBVS.
• Future work should focus on improving generalization, balancing speed and accuracy, incorporating protein flexibility, and ensuring physical plausibility.
• Comparing generative and regression algorithms is essential for further DLLD development.