SurroDock: A Deep Learning Surrogate for Accelerated Pre-Docking Ligand Prioritization in Structure-Based Virtual...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Virtual screening is crucial for drug discovery but computationally intensive.
• Large chemical libraries pose challenges for exhaustive structure-based virtual screening.
• Existing methods struggle with the scale of make-on-demand and public chemical databases.

Purpose of the Study:

• To develop SurroDock, a deep-learning surrogate model for rapid approximation of AutoDock Vina docking scores.
• To serve as a pre-filter for structure-based virtual screening, enhancing efficiency.
• To evaluate SurroDock's performance on estrogen receptor alpha using agonist and antagonist-bound states.

Main Methods:

• Trained a deep-learning model on concatenated 2D molecular representations (fingerprints, descriptors).
• Utilized a dataset of ~334,000 compounds docked with AutoDock Vina against two estrogen receptor alpha conformations (1GWR, 3ERT).
• Evaluated model performance using R-squared values and retrospective screening enrichment factors.

Main Results:

• SurroDock achieved high R-squared values (0.88 for 1GWR, 0.93 for 3ERT) on held-out validation data.
• Demonstrated significant enrichment (EF@1% ~57-61 fold) in recovering top-ranked compounds.
• Showcased state-specific modeling effectiveness due to distinct docking-score distributions between receptor conformations.

Conclusions:

• 2D-based docking-score surrogate modeling offers a reproducible and retrainable strategy for large-scale virtual screening.
• SurroDock effectively concentrates docking resources on a smaller, enriched subset of compounds.
• SurroDock predictions are valuable prioritization aids, requiring complementary experimental validation.