Normalization Bias in Morpho-Transcriptomic Prediction

Swann Ruyter, Reuben Dorent, Daniel Racoceanu

Keywords: Spatial transcriptomics, diffusion models, normalization, H&E.

TL;DR: We show that transcriptomic target normalization strongly shapes reported morpho-transcriptomic performance because log-transformed targets preserve a more easily predicted cellularity- and depth-related signal than Pearson-residual targets.

Registration Requirement: Yes

Abstract: Spatial transcriptomics (ST) prediction from H\&E histology has attracted growing attention as a scalable approach to infer gene expression from tissue morphology. However, existing methods are typically benchmarked after target normalization, whose impact remains insufficiently characterized. This is critical because raw ST counts partly reflect local cellularity; consequently, normalization determines how much of this morphology-linked variation is retained in the prediction target. Using STEM as a controlled diffusion-based backbone, we compare its original log-normalization with a Pearson-residual-based normalization on a HER2-positive breast cancer ST cohort. We show that log-normalization preserves substantially stronger nuclei- and depth-related signals, and that the best-predicted genes under this regime are also the most strongly associated with nuclei-related features in the observed data. These findings suggest that part of the apparent performance gain in log-transformed space arises from the preservation of an easier-to-predict cellularity-related signal, with important implications for method comparison and for the interpretation of correlation-based benchmarks.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 102