Deep learning large-scale drug discovery and repurposing.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Drug discovery and repurposing are hindered by costly and low-throughput methods for identifying mechanisms of action (MOA).
• Accurate MOA identification is critical for efficient drug development and therapeutic application.

Purpose of the Study:

• To develop a cost-effective, high-throughput approach for MOA identification using mitochondrial phenotypes.
• To establish a deep learning model for analyzing time-resolved mitochondrial images to predict drug MOA.

Main Methods:

• A pipeline was created for time-resolved imaging of mitochondrial morphology and membrane potential.
• A dataset of 570,096 single-cell images from cells treated with 1,068 FDA-approved drugs was generated.
• A deep learning model, MitoReID, utilizing a re-identification (ReID) framework and an Inflated 3D ResNet backbone, was developed and trained.

Main Results:

• MitoReID achieved 76.32% Rank-1 and 65.92% mean average precision on the testing set.
• The model successfully identified MOAs for six previously untrained drugs based on mitochondrial phenotypes.
• MitoReID identified cyclooxygenase-2 inhibition as the MOA for epicatechin, which was validated in vitro.

Conclusions:

• The developed approach offers an automated and cost-effective alternative for MOA identification.
• This mitochondrial phenotype-based strategy can significantly accelerate large-scale drug discovery and repurposing efforts.
• The findings highlight the potential of deep learning in analyzing cellular imaging data for biological insights.