RECOVER identifies synergistic drug combinations in vitro through sequential model optimization.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Exhaustive screening of large small molecule libraries is infeasible for drug discovery.
• Deep learning models excel at predicting drug synergy scores in silico.
• Existing drug combination databases are biased, limiting generalizability.

Purpose of the Study:

• To develop and apply a sequential model optimization strategy using deep learning.
• To efficiently identify drug combinations with synergistic activity against cancer cell lines.
• To improve the selection process for potent drug combinations.

Main Methods:

• Employed a deep learning model for sequential model optimization over 5 experimental rounds.
• Focused on selecting drug combinations enriched for synergism and anti-cancer activity.
• Evaluated approximately 5% of the total chemical search space.

Main Results:

• Identified drug combinations with enhanced synergism and anti-cancer activity.
• Learned drug embeddings, derived from structural information, reflected biological mechanisms.
• In silico benchmarking showed a 5-10x enrichment for synergistic combinations compared to random selection.

Conclusions:

• Sequential model optimization is a highly efficient strategy for discovering synergistic drug combinations.
• Deep learning models can guide drug discovery by learning meaningful drug representations.
• This approach significantly accelerates the identification of effective anti-cancer drug combinations.