Efficient Exploration of Chemical Space with Docking and Deep Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Commercial compound libraries have grown exponentially, exceeding one billion compounds.
• Exhaustive in silico screening of large libraries is becoming cost-prohibitive.
• Novel scaffolds and more potent compounds are discoverable in larger, diverse libraries.

Purpose of the Study:

• To introduce a machine learning-enhanced molecular docking protocol to increase virtual screening throughput.
• To develop a novel selection strategy balancing compound scoring and chemical space exploration.
• To reduce the computational cost of virtual screening while maintaining hit identification efficacy.

Main Methods:

• Active learning for machine learning-enhanced molecular docking.
• A novel compound selection protocol balancing scoring and diversity.
• Automated redocking of top-scoring compounds.
• Application to virtual screening campaigns against D4, AMPC, and MT1 targets.

Main Results:

• The protocol significantly increases throughput compared to traditional docking.
• It captures nearly all high-scoring scaffolds identified by exhaustive docking.
• More than 80% of experimentally confirmed hits were recovered with a 14-fold reduction in compute cost.
• Over 90% of hit scaffolds were identified within the top 5% of model predictions, preserving diversity.

Conclusions:

• Machine learning-enhanced molecular docking offers a cost-effective and efficient approach for virtual screening of large compound libraries.
• The developed protocol successfully identifies potent, diverse hit compounds for drug discovery.
• This method enables exploration of new chemical spaces and accelerates the identification of novel inhibitors.