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

Abstract

Cell Painting (CP), as a high-throughput imaging technology, generates extensive cell-stained imaging data, providing unique morphological insights for biological research. However, CP data contains three types of technical effects, referred to as triple effects, including batch effects, gradient-influenced row and column effects (well position effects). The interaction of various technical effects can obscure true biological signals and complicate the characterization of CP data, making correction essential for reliable analysis. Here, we propose cpDistiller, a triple-effect correction method specially designed for CP data, which leverages a pre-trained segmentation model coupled with a semi-supervised Gaussian mixture variational autoencoder utilizing contrastive and domain-adversarial learning. Through extensive qualitative and quantitative experiments across various CP profiles, we demonstrate that cpDistiller effectively corrects triple effects, especially well position effects, a challenge that no current methods address, while preserving cellular heterogeneity. Moreover, cpDistiller effectively captures system-level phenotypic responses to genetic perturbations and reliably infers gene functions and interactions both when combined with scRNA-seq data and independently. cpDistiller also excels at identifying gene and compound targets, which is a critical step in drug discovery and broader biological research.