This Article presents a large-scale analysis of self-supervised learning methods for high-content morphological profiling.

We pretrained and benchmarked FMs for representation learning from multi-channel high-content images using the workflow summarized in Fig. 1. The models cover four SSL methods spanning three major learning mechanisms: masked-based generative (MAE, iTPN), contrastive-based (DINO) and hybrid (DINOv2). Unlike CellProfiler, our pipeline operates directly at the image level without requiring cell segmentation or feature engineering. The FMs were pretrained and benchmarked using a two-stage approach (Fig. 1). In the pretraining phase, all unannotated chemical perturbations from the JUMP–CP dataset were used for modeling optimization to learn general feature representations. In the evaluation phase, pretrained weights were frozen and used solely as feature extractors for two downstream tasks: zero-shot compound retrieval, where embeddings were directly used to assess reproducibility and biological relevance, and few-shot MoA classification, where a k-nearest neighbor (k-NN) classifier was trained on the extracted embeddings. Postprocessing strategies (sphering, plate-wise median absolute deviation (MAD) normalization and Harmony) were evaluated to mitigate batch effects and improve downstream performance (Methods). Postprocessing strategies for batch effect correction and improved SSL feature representation We conducted a detailed ablation of three postprocessing methods—sphering, MAD normalization and Harmony—examining their effects individually and in combination (Supplementary Fig. 1 and Methods). For SSL-derived features, postprocessing steps...

We pretrained and benchmarked FMs for representation learning from multi-channel high-content images using the workflow summarized in Fig. 1. The models cover four SSL methods spanning three major learning mechanisms: masked-based generative (MAE, iTPN), contrastive-based (DINO) and hybrid (DINOv2). Unlike CellProfiler, our pipeline operates directly at the image level without requiring cell segmentation or feature engineering. The FMs were pretrained and benchmarked using a two-stage approach (Fig. 1). In the pretraining phase, all unannotated chemical perturbations from the JUMP–CP dataset were used for modeling optimization to learn general feature representations. In the evaluation phase, pretrained weights were frozen and used solely as feature extractors for two downstream tasks: zero-shot compound retrieval, where embeddings were directly used to assess reproducibility and biological relevance, and few-shot MoA classification, where a k-nearest neighbor (k-NN) classifier was trained on the extracted embeddings. Postprocessing strategies (sphering, plate-wise median absolute deviation (MAD) normalization and Harmony) were evaluated to mitigate batch effects and improve downstream performance (Methods). Postprocessing strategies for batch effect correction and improved SSL feature representation We conducted a detailed ablation of three postprocessing methods—sphering, MAD normalization and Harmony—examining their effects individually and in combination (Supplementary Fig. 1 and Methods). For SSL-derived features, postprocessing steps...

Application for automatically renaming and reorganizing acquired images for 96-well microscope slides. - gregorymaubon/PlateArrangers

YouTube video by Ask Erin, Dear Beth

Anne Carpenter provides an overview of CellProfiler, a fr...

YouTube video by Ask Erin, Dear Beth

YouTube video by CellMorphoJSero

YouTube video by Ask Erin, Dear Beth