Utilities for adapting feature distributions between measurement platforms before model transfer. This is designed for settings where standard frozen batch correction is not applicable because the source training set contains only one platform level.
Details
`CrossPlatformAdapter` implements lightweight platform adaptation strategies that can be applied before training on a source platform and predicting on a target platform:
- `"rank"`: within-sample percentile ranks across features - `"quantile"`: quantile normalization to the source reference distribution - `"zscore"`: within-sample z-scoring - `"reference"`: subtraction of per-sample reference miRNA summary - `"combat_pooled"`: pooled ComBat correction across source and target
The pooled ComBat strategy is semi-supervised because it needs some target labels to preserve biology while correcting platform effects.