Predicting cell morphological responses to perturbations using generative modeling.

Area of Science:

• Cellular imaging
• Drug discovery
• Computational biology

Background:

• High-content microscopy generates rich phenotypic data for drug discovery.
• Analyzing large imaging datasets is challenging due to incomplete sampling and technical variations.

Purpose of the Study:

• To develop a generative model for predicting morphological changes in high-content imaging screens.
• To address challenges in analyzing complex screening data and account for technical variations.

Main Methods:

• Developed the IMage Perturbation Autoencoder (IMPA), a generative style-transfer model.
• Applied IMPA to predict morphological changes from genetic and chemical perturbations.
• Evaluated IMPA's performance on breast cancer and osteosarcoma cell lines, assessing its ability to handle batch effects.

Main Results:

• IMPA accurately captures morphological and population-level changes for both known and novel perturbations.
• The model demonstrates robustness across diverse experimental conditions by accounting for batch effects and technical variations.
• IMPA successfully models perturbations across various sources of technical variation.

Conclusions:

• IMPA is a robust tool for analyzing large-scale high-content imaging screens.
• The model facilitates efficient experimental design through in-silico perturbation prediction.
• IMPA is poised to enhance the analysis of microscopy data in drug discovery pipelines.