Clinical utility of discordances between histomorphology and molecular biomarkers in precision oncology.

Abstract

3005 Background: Conventional molecular biomarkers guide treatment selection in precision oncology, yet therapeutic benefit remains variable even within biomarker-defined populations. Deep learning (DL) models can infer biomarker status from hematoxylin and eosin (H&E) histopathology, but systematic discordances arise between conventional biomarker assays and their histomorphology-based (HM) predictions. We hypothesized that such discordances reflect biologically meaningful tumor functional states rather than model error and may have clinical relevance. Methods: DL models were trained to predict 15 clinically relevant biomarkers from H&E images across six cancer types using public datasets. Discordant cases, defined as disagreement between assay-based biomarker labels and their DL prediction from H&E, were quantified and analyzed for alignment with functional gene expression states and clinical outcomes. Clinical utility was assessed in two landmark randomized trials: FINHER (HER2 prediction; N = 1010; trastuzumab randomized in HER2+ patients), and TAILORx (Oncotype DX recurrence score [RS] prediction; N = 10,273; chemotherapy randomization within RS 11–25). Adjusted Cox models tested added prognostic value and treatment beneft interactions. Primary endpoint was distant recurrence. Results: Models achieved strong performance for predicting the biomarkers (median AUC 0.78; range 0.65–0.87). Discordance rates ranged from 8% (MSI in colon) to 32% (KRAS in lung). In discordant cases, HM predictions correlated more strongly with functional gene expression states than the biomarker status they were trained to predict (mean Spearman r = 0.52 vs 0.31; p < 0.001). In FINHER, predicted HER2 status added prognostic value beyond HER2 status (p = 0.01) and predicted trastuzumab benefit in HER2+ patients (p-interaction = 0.026). In HER2+ patients, trastuzumab benefit was observed in those predicted as HER2+ (HR = 0.35, 95% CI: 0.15–0.83), but not in those predicted as HER2- (HR = 1.17, 95% CI: 0.45–3.02). In TAILORx, predicted RS added prognostic value over true RS in the no-chemo group (p = 0.005) and showed treatment interaction beyond RS within RS 11–25 (N = 1629; p-interaction = 0.021), as well as when restricted to postmenopausals (p = 0.029). Discordant patients (RS 11–25 predicted as RS > 26) derived chemotherapy benefit (HR = 0.30, 95% CI: 0.13–0.77), whereas those predicted as RS < 26 did not (HR = 0.95, 95% CI: 0.71–1.29). Conclusions: Discordances between molecular biomarkers and HM predictions are not random errors but reflect underlying tumor functional states. HM biomarkers may capture relevant information missed by standard assays that can improve therapeutic benefit prediction. This framework enables reinterpretation of DL pathology predictions, supports integrating H&E models to complement biomarkers, and enables derivation of functional information without outcome data.