Determine the causes of cohort-dependent exposure calibration shifts

Determine the relative contributions of mutation burden, cohort composition, variant-calling procedures, and the zero-inflated negative-binomial objective to the discrepancy between predicted and observed absolute mutational-signature exposures across TCGA and CPTAC cohorts.

Background

Hist2Sig transfers an image-to-genomics relationship learned from TCGA to CPTAC, but its absolute exposure predictions are imperfectly calibrated across cohorts. Predictions tend to be lower in CPTAC relative to observed exposures and overestimated in TCGA, especially for glioblastoma.

The paper identifies several possible explanations, including differences in mutation burden, cohort composition, variant-calling procedures, and the use of a zero-inflated negative-binomial loss for heterogeneous count data. However, it explicitly leaves unresolved how much each factor contributes to the observed calibration discrepancy.

References

Differences in mutation burden, cohort composition, and variant-calling procedures may all contribute to this discrepancy, as may the use of a zero-inflated negative binomial objective for highly heterogeneous count data. Their relative contributions remain to be determined.

— Predicting Mutational Signature Exposures from H&E Whole Slide Images: A Pan-Cancer Feasibility Study  (2609.30985 - Sartori et al., 25 Sep 2026) in Discussion, paragraph beginning “Generalization was nevertheless incomplete in terms of absolute exposure values”