Active learning framework leveraging transcriptomics identifies modulators of disease phenotypes.

Area of Science:

• Computational biology
• Drug discovery
• Genomics

Background:

• Phenotypic drug screening is limited by chemical space and experimental scalability.
• Current computational methods often lack generalizability or optimization capabilities.
• Genomic proxies used in drug discovery are often heuristic and resist optimization.

Purpose of the Study:

• To develop a scalable and optimizable computational framework for identifying compounds that induce complex phenotypes.
• To leverage omics data within an active deep learning approach for drug discovery.
• To improve the efficiency and success rate of phenotypic drug screening.

Main Methods:

• Designed an active deep learning framework integrating omics data.
• Employed a lab-in-the-loop signature refinement strategy.
• Validated the algorithm on hematological discovery campaigns.

Main Results:

• The deep learning framework demonstrated superior performance over state-of-the-art models in recall.
• Achieved a 13- to 17-fold increase in phenotypic hit rate in discovery campaigns.
• A twofold increase in hit rate was observed when combined with lab-in-the-loop refinement, yielding molecular insights.

Conclusions:

• The developed framework enables efficient and scalable phenotypic hit identification.
• This approach has broad potential to accelerate the drug discovery pipeline.
• Integration with experimental refinement further enhances hit identification and provides mechanistic understanding.