Confounder-aware foundation modeling for accurate phenotype profiling in cell imaging.

Area of Science:

• Computational Biology
• Drug Discovery
• Machine Learning

Background:

• Image-based profiling offers cellular insights but suffers from experimental variability, hindering accurate mechanism of action (MoA) and target identification.
• Current methods struggle to generalize to novel compounds, limiting exploration of new chemical spaces.

Purpose of the Study:

• To develop a confounder-aware foundation model integrating causal mechanisms and latent diffusion for robust biological effect estimation.
• To generate balanced synthetic datasets for improved prediction accuracy in drug discovery.

Main Methods:

• Developed a novel confounder-aware foundation model with integrated causal mechanisms within a latent diffusion model.
• Trained the model on over 13 million Cell Painting images and 107,000 compounds.
• Utilized the model to generate synthetic datasets for training and validation.

Main Results:

• Achieved state-of-the-art MoA and target prediction performance (0.66/0.65 ROC-AUC for seen, 0.65/0.73 ROC-AUC for unseen compounds).
• Demonstrated superior performance compared to real and batch-corrected data.
• Successfully mitigated confounder impact through robust cellular phenotype representation learning.

Conclusions:

• The developed framework provides robust biological effect estimations for novel compounds, accelerating hit expansion in drug discovery.
• Establishes a scalable and adaptable foundation model for cell imaging, advancing data-driven drug discovery.
• The model shows significant potential to become a cornerstone in computational drug discovery.