The problem — Classical morphological profiling extracts hundreds of named features per cell: area, texture, intensity. Those features are auditable but bounded. A nucleus occupies a fixed coordinate space in the feature matrix, and phenotypes that don't express through the named descriptors are simply invisible. When a perturbation reorganizes the cytoskeleton in ways no hand-engineered filter captures, the assay misses it entirely.
What it is / how it works — DeepProfiler is a Broad Institute toolkit that replaces — or augments — hand-engineered extraction with a deep convolutional backbone. The current production checkpoint is the Cell Painting CNN v1 (EfficientNet B0), trained via weakly supervised learning on roughly 8.3 million single cells drawn from five experimental sources. It outputs 1,280-dimensional per-cell embeddings that encode morphology across the five Cell Painting channels (DNA, ER, RNA, AGP, Mito). The Moshkov et al. 2024 study, which established this architecture, showed improvements of up to 30% in downstream analysis relative to CellProfiler features, while also being more computationally efficient at scale. Outputs ship as .npz files wired directly into pycytominer, preserving the same normalization and sphering steps the classical pipeline uses.
The learned embeddings sit at the same stage as classical features in the From Masks to Morphology: Features & Self-Supervised Embeddings note: both are outputs of the features step, and both feed the same correction and readout layers.
Where it breaks — Embeddings encode everything the backbone attended to during training — including plate effects, scanner drift, and batch structure. A 1,280-dimensional vector that separates two compound treatments may be separating two imaging runs. The Moshkov study is explicit: batch correction is not optional after DeepProfiler; it is load-bearing. This is the same tension documented in Batch Correction Without Erasing Biology — the higher the embedding dimensionality, the more surface area for technical variance to hide in signal-shaped geometry. Applying Percent Replicating after extraction rather than before confirms whether the variance that survives correction is biological.