Cell morphology profiling drops segmentation: self-supervised models retrieve about a third more mechanism classes than CellProfiler across labs

Cell morphology profiling can skip cell segmentation and handcrafted feature engineering by learning reusable representations directly from multi-channel high-content images. The dominant CellProfiler pipeline first segments cells, then extracts handcrafted morphological features and filters them; this dependence on segmentation and predefined features can introduce bias and limit scale. On the held-out JUMP–Target2 compound set of 302 compounds, every self-supervised model except MAE consistently outperformed CellProfiler on compound-replicate retrieval, with P ≤ 0.001.

Contrastive and hybrid self-supervision learn more robust representations than masked reconstruction and transfer better across cell types. Masked-based models trailed on several benchmarks; on held-out source 6, MAE retrieved only one gene target, while CellProfiler, iTPN, DINO and DINOv2 each retrieved six. On 2,464 EU-OPENSCREEN Bioactives compounds, DINO and DINOv2 outperformed CellProfiler and masked-based models at pathway, mechanism-of-action and protein-target annotation levels in both HepG2 and U2OS cells, with 31–37% gains in site-averaged mAP for mechanism-of-action retrieval.

With frozen weights and no fine-tuning, the models still transfer to cell types and a fluorescent channel absent from pretraining. Pretraining used only chemical perturbations in U2OS cells, while evaluation covered MCF-7, HUVEC, RPE and HepG2 cells. On BBBC021 MCF-7 data, where one channel was not present during pretraining, DINOv2 reached 91.3% overall NSC accuracy, above CellProfiler at 90%, DINO at 86.4%, iTPN at 79.6% and MAE at 68.9%.

Batch-effect correction matters for self-supervised features too, and the order of steps matters. Raw features carried severe batch effects that had to be corrected before similar perturbations clustered together. Applying sphering, then MAD normalization, then Harmony in that order gave the best results for self-supervised features; across 59,145 compounds in source 6, postprocessing consistently improved retrieval for all models, and DINO and DINOv2 scored highest on both retrieved-class percentage and mAP.

Pretraining used only chemical perturbation images in U2OS cells; evaluation covered the held-out JUMP–Target2 compound set, EU-OPENSCREEN Bioactives, BBBC021 and RxRx1, spanning MCF-7, HUVEC, RPE, HepG2 and U2OS cells. The pretrained models and embeddings are public, and the authors propose future work on fine-tuning strategies to reach additional cell types.

The authors note that sphering and Harmony cannot adapt precomputed transforms to new batches, so all data must be reprocessed; cross-source batch correction remains harder than within-source correction. Deep-learning features are less interpretable than engineered ones, more complex cell types such as neurons may require fine-tuning, and systematic evaluation on large-scale genetic perturbation data remains to be done.