← Back to the vault

Lab Note · updated 2026-06-20

The Challenge of Rare Events and Mutants

When the phenotype you care about appears in 10 cells out of 300,000, accuracy is meaningless, validation has almost no positives to learn from, and false discovery dominates. Rare-event detection is a class-imbalance and ground-truth-scarcity problem — best answered by adaptive acquisition that goes looking for the positives.

QCSegmentReadoutCell PaintingDigital PathologyLive-Cell Trackinganomaly detectionclass-weighted lossconformal predictionCellpose
← Back to constellation

The problem — The biologically interesting thing is often the rare thing: the one dividing cell entering mitosis, the rare mutant clone, the 10–20 malignant cells on a Pap smear of 300,000. Rarity breaks the standard machinery three ways at once. Accuracy collapses as a metric — a model that labels everything "normal" scores 99.99% while finding none of the positives. Validation starves — there are too few positives to train on, tune on, or test against, so confidence intervals are enormous and a single mislabeled example swings the result. And false discovery dominates — when the prior is one-in-a-million, even a tiny false-positive rate buries every true hit under noise. The positives you most want are exactly the ones the pipeline is least equipped to find or trust.

What it is / how it works — There is no single fix; there is a stack of partial ones. Reweight the loss so scarce classes count more — Hagos et al. regularize a detection loss by cell abundance and beat baselines specifically on rare cell types in imbalanced pathology. Reframe rarity as anomaly rather than classification — self-supervised representations flag morphological outliers in high-content profiling without needing labelled positives, learning "normal" and scoring deviation. Choose metrics that survive imbalance — Metrics Reloaded is explicit that under class imbalance, accuracy and even Dice mislead, and the metric must be chosen from the problem fingerprint. And crucially, go looking for the positives instead of scanning uniformly: event-driven systems like CelFDrive classify cells from cheap low-resolution frames and switch to high-resolution 3D only on a predicted rare event — an adaptive-acquisition approach that raises the hit rate by an order of magnitude.

Where it breaks — The general failure is reporting an aggregate that the rare class barely touches — and it recurs across every platform where the signal is sparse. In digital pathology, a slide-level accuracy is dominated by abundant benign cells while the malignant handful is missed. In Cell Painting, a rare mutant phenotype is averaged away in a well-level profile unless single-cell tails are modelled. In live-cell imaging, the rare transition is over in seconds and is simply not sampled under uniform acquisition. The connective tissue is ground-truth scarcity: with almost no positives, validation has no clean answer key, so honest practice means experiment-conditioned metrics (Seiffarth et al.), explicit false-discovery control, and acquisition that creates positives to validate against rather than hoping they appear. Anything else reports a number the rare event never entered.

Under extreme imbalance, a high accuracy is evidence of nothing — it can be achieved by ignoring every positive. Report recall and false-discovery on the rare class, validate against deliberately enriched positives, and prefer acquisition that hunts the event over scanning that averages it away.

References

References

Local graph