The problem — Most supervised classifiers in high-content screening require labelled examples of the thing you are looking for. Rare phenotypes, early cytotoxic responses, or novel mechanisms of action may produce only a handful of affected cells per plate — not enough to train on, easy to dilute into noise, and invisible to any metric that averages across wells. Treating outlier cells as noise rather than signal is how rare biology gets discarded before it reaches a scientist.
What it is / how it works — Anomaly detection reframes the problem: instead of learning a boundary between normal and abnormal from labelled pairs, the model learns a compact representation of normal morphology from unannotated control cells and then scores any image or cell-embedding by how far it sits from that learned distribution. Reconstruction-error methods (autoencoders, VAEs) flag cells whose appearance the model cannot reproduce faithfully; density-based methods (normalizing flows, one-class SVM) flag low-likelihood points in embedding space. Shpigler et al. (2024) applied this to Cell Painting profiles and showed it recovers biologically coherent outlier populations that matched-pairs approaches missed entirely. Because no labels are consumed in training, the signal budget is preserved for the rare events that matter.
In a production pipeline, anomaly scores serve two distinct roles: QC (flagging imaging failures — out-of-focus fields, edge artifacts, pipetting errors) and biology (surfacing cells with genuinely unusual morphology worth investigating). Conflating the two is a common failure mode; the scores must be interpreted in context of the upstream uncertainty-drift-and-failure-modes landscape before conclusions are drawn.
Where it breaks — The definition of "normal" is only as good as the reference set. A contaminated or phenotypically heterogeneous control population will produce a miscalibrated baseline, causing true anomalies to score as background and background to score as anomalies. Distribution shift across plates or batches compounds this: a cell that is anomalous relative to plate 1 controls may be entirely ordinary on plate 3. Without validation-without-ground-truth strategies — held-out positive controls, synthetic perturbations, rank-stability checks — anomaly scores are uninterpretable numbers rather than actionable signals.
The Fovea take — We treat anomaly detection as a first-pass triage layer, not a discovery endpoint. Scores drive attention — prioritising which wells or cells a scientist examines first — but every flagged population is validated against independent evidence before it enters a report. This is the same principle that governs our approach to rare-events-and-mutants: rare signal deserves more rigour, not less, precisely because it is easy to manufacture from noise.