A split-control criterion for few-shot adaptation in single-cell perturbation prediction
Few-shot adaptation can help or harm single-cell perturbation prediction, but conventional evaluation shares a control estimate between calibration responses and held-out targets. We show how methods carry control error into their predictions and derive its contribution to score inflation. Nested control-depth experiments support inverse-depth scaling of inner-product inflation; correlation…
Few-shot adaptation can have both positive and negative effects on single-cell perturbation prediction, but current evaluation methods combine control estimates from both calibration responses and held-out targets. The researchers demonstrate how these methods propagate control errors into their predictions and quantify their impact on inflated scores.
By conducting nested control-depth experiments, they find that inner-product inflation scales inversely with depth, and correlation inflation depends on the control noise-to-signal ratio. When independent control scoring is used, the rank of 40 out of 52 configuration-dataset entries changes. To address this issue, the researchers introduce SafeDelta, a split-control criterion that utilizes calibration data to determine whether adaptation is necessary.
Across six datasets, mean gated gain remains non-negative under the original two-control evaluation. On a deep plate, fixed predictions and decisions maintain positive mean gains under three additional independent control references. The same reproducibility principle reveals a fibrosis programme that localizes to scar in human liver. These findings provide insights into adaptation decisions and control-depth planning.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
