Advancing morphological profiling with self-supervised learning through comparative insights into representation learning of cellular perturbations

Self-supervised learning (SSL) enables scalable, label-free foundation models (FMs) that learn task-agnostic representations directly from images, bypassing cell segmentation and feature engineering inherent to classical morphological profiling. We present a large-scale evaluation of SSL FMs for image-based phenotypic profiling, covering contrastive (DINO), masked-based generative (MAE, iTPN) and hybrid (DINOv2) learning mechanisms, pretrained on all unannotated chemical perturbations from the Joint Undertaking for Morphological Profiling-Cell Painting (JUMP-CP) dataset (119,876 compounds, ~1.77 million images, 7 laboratories). Here we show that top-performing SSL models outperformed CellProfiler in compound-replicate retrieval on the held-out JUMP-Target2 set (302 annotated compounds, P ≤ 0.001). DINO and DINOv2 generalized across diverse perturbations and cell types, outperforming CellProfiler with 31-37% gains in site-averaged mean average precision for mechanism-of-action retrieval on EU-OPENSCREEN Bioactives (2,464 compounds), and consistently surpassing masked-based models across all benchmarks. This study provides a systematic assessment of SSL approaches for advancing morphological profiling and offers pretrained FMs to accelerate drug discovery.