Triple-effect correction for Cell Painting data with contrastive and domain-adversarial learning

Area of Science:

• Cellular imaging
• High-throughput screening
• Computational biology

Background:

• Cell Painting (CP) is a high-throughput imaging technology providing morphological insights.
• CP data is susceptible to technical artifacts, including batch and well-position effects (triple effects).
• These artifacts obscure biological signals, necessitating robust correction methods for reliable analysis.

Purpose of the Study:

• To develop and validate cpDistiller, a novel method for correcting triple effects in CP data.
• To demonstrate cpDistiller's ability to preserve cellular heterogeneity while correcting technical variations.
• To showcase cpDistiller's utility in inferring gene functions, interactions, and identifying drug targets.

Main Methods:

• cpDistiller employs a pre-trained segmentation model.
• It utilizes a semi-supervised Gaussian mixture variational autoencoder with contrastive and domain-adversarial learning.
• The method was validated through extensive qualitative and quantitative experiments on diverse CP datasets.

Main Results:

• cpDistiller effectively corrects triple effects, particularly well-position effects, in CP data.
• The method successfully preserves crucial cellular heterogeneity.
• cpDistiller accurately captures system-level phenotypic responses to genetic perturbations.
• It reliably infers gene functions and interactions, even when integrated with scRNA-seq data.
• The tool shows capability in identifying gene and compound targets.

Conclusions:

• cpDistiller offers a powerful solution for mitigating technical artifacts in Cell Painting data.
• This method enhances the reliability of morphological profiling for biological discovery.
• cpDistiller holds significant potential for applications in drug discovery and systems biology research.