Scaling Structure Aware Virtual Screening to Billions of Molecules with SPRINT

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Virtual screening accelerates drug discovery by predicting drug-target interactions (DTIs).
• Structure-based methods like molecular docking are computationally expensive for proteome-scale screening.
• Vector-based methods using protein language models (PLMs) offer a faster alternative by bypassing 3D structure modeling.

Purpose of the Study:

• To develop SPRINT, a novel vector-based approach for large-scale virtual screening of DTIs and identification of new mechanisms of action.
• To improve upon existing PLM-based methods by incorporating self-attention and structure-aware models for enhanced drug-target co-embedding.

Main Methods:

• Developed SPRINT, a self-attention based architecture utilizing structure-aware PLMs to create a co-embedding space for drugs and protein targets.
• Employed SPRINT for efficient binder prediction, search, and retrieval across large chemical libraries and proteomes.
• Evaluated SPRINT on DTI classification, binding affinity prediction benchmarks, and the LIT-PCBA dataset.

Main Results:

• SPRINT achieved state-of-the-art (SOTA) enrichment factors in virtual screening tasks.
• Demonstrated high accuracy and interpretability through residue-level attention maps.
• Achieved ultra-fast performance, screening 6.7 billion drugs against the human proteome in 16 minutes.

Conclusions:

• SPRINT enables accurate, interpretable, and ultra-fast virtual screening at an unprecedented scale.
• The approach facilitates in silico drug repurposing and development by identifying potential DTIs and novel mechanisms.
• SPRINT is accessible via the web as ColabScreen.