Distillation enables scalable high-fidelity virtual screening across ultra-large chemical libraries

Area of Science:

• Computational chemistry and drug discovery.
• Development of novel computational frameworks for molecular interactions.

Background:

• Virtual screening of ultra-large chemical libraries is computationally intensive and often limited by predictive accuracy.
• Existing methods struggle with balancing speed, accuracy, and comprehensive exploration of chemical space.

Purpose of the Study:

• To present FastBindRank, a novel distillation-based framework for high-fidelity virtual screening of ultra-large chemical libraries.
• To enable efficient and accurate identification of potential drug candidates from massive compound datasets.

Main Methods:

• Developed FastBindRank, a sequence-based surrogate model using knowledge distillation from a structure-based model (Boltz-2).
• Trained FastBindRank on a significant subset (~1%) of the 122-million-compound PubChem library.
• Applied FastBindRank to screen for histone deacetylase 11 (HDAC11) inhibitors.

Main Results:

• FastBindRank achieved high-fidelity screening at scale, substantially enriching high-confidence binders.
• The model identified structural patterns linked to binding, offering insights into binding determinants.
• Compared to traditional subset screening, FastBindRank demonstrated a 74-fold increase in hit rate and over a 30-fold increase in discovery yield under similar computational budgets.
• Experimental validation confirmed the activity of two newly identified compounds.

Conclusions:

• Distillation is a practical and effective strategy for scalable, high-fidelity virtual screening of ultra-large chemical libraries.
• FastBindRank significantly enhances the efficiency and accuracy of drug discovery pipelines.
• The framework provides a powerful tool for exploring vast chemical spaces and identifying novel bioactive compounds.