The problem — A segmentation mask is not a measurement. The features stage is where the biology becomes computable: each object is reduced to a vector, and that vector determines which phenotypes are detectable at all. A feature space that cannot represent a particular morphological change means that perturbation is invisible to the entire downstream analysis — not weak, invisible.
What it is / how it works — Two families. The classical profiling family (CellProfiler lineage) computes hundreds of named morphology, texture, and intensity descriptors per object — the backbone of the Cell Painting assay, which yields ~1,500 features per cell across five channels (Bray et al. 2016; decade review: Seal et al. 2025). Each feature has a biological reading; the profile is interpretable and auditable. The learned embedding family trains an encoder to emit a compact representation, either supervised (foundation models for pathology: UNI, H-optimus-0, CONCH) or self-supervised (DeepProfiler, scDINO, cytoself). Self-supervised models — particularly those in the VQ-VAE or ViT lineage — can out-cluster classical features on organelle and complex membership, as cytoself demonstrated on protein localization. The decade review reports deep-learning features beating CellProfiler by up to +29 mAP on JUMP-scale benchmarks.
Where it breaks — Embeddings buy sensitivity and lose interpretability. A learned axis that separates two compounds tells you that they differ, not how — and it encodes acquisition nuisance (batch, scanner, plate position) just as readily as biology. Self-supervised models are notorious for batch-encoding; cytoself's authors flag residual batch effects. The features stage never validates itself: whatever vector is produced must clear the correction and reproducibility gates before any biological claim is credible. See From Masks to Morphology: Features & Self-Supervised Embeddings for the full trade-off.