Exploiting HER2 tumour heterogeneity with digital oncology

Intratumoural heterogeneity has emerged as an extremely relevant clinical marker, and in HER2-positive breast cancer, is a primary driver of resistance and an independent prognostic factor for poor outcomes (Cancers (Basel). 2025;17:2126; Cancers (Basel). 2023;15:2664). Conventional imaging assessments often fail to capture tumour heterogeneity and microenvironmental changes, and novel models are needed. In a presentation at ESMO TAT Asia 2026 (Hong Kong SAR, China, 12–14 June), a multi-regional fusion framework (HRP-Fusion) model was described, which attempts to improve the accuracy of pathological complete response (pCR) prediction in HER2-positive breast cancer (Abstract 23RO). The model combines radiomics – with integration of intratumoural and peritumoural imaging features – with deep-learning (DL) features and clinicopathological information.

The integrated HRP-Fusion model was compared with five other models (Clinical, Intratumoral, Peritumoral, Fusion Radiomics, DL-only) in a retrospective, multicentre study of 422 patients included in a training set (n=222) and two external validation sets (EVS1, n=108; EVS2, n=92). The HRP-Fusion model demonstrated superior discriminative power, achieving an area under the curve (AUC) of 0.915 (95% confidence interval 0.877–0.953) in the training set, significantly outperforming Clinical and DL-only models. Notably, high performance was maintained in external validation, with AUCs of 0.873 (EVS1) and 0.895 (EVS2). Intratumoural texture complexity and the tumour–stroma interface were the primary biological drivers of treatment response prediction.

Commenting on the abstract data, Dr Julien Vibert from Gustave Roussy, Villejuif, France, notes that unlike spatial transcriptomics, this ‘macroscopic’ imaging model is non-invasive, affordable and requires less bioinformatics expertise. “The findings suggest that we are potentially advancing towards a non-invasive marker that will allow us to better predict treatment response and, if prospectively validated in clinical trials across multiple centres, could support more personalised treatment adaptation, including escalation or de-escalation strategies. While costly spatial single-cell transcriptomics technologies are unlikely to be used routinely in the clinic, the biological insights they provide can be used to inform these types of more practical modalities,” he concludes.

In the future, he envisages artificial intelligence models being developed for routine use that are trained using insights from complex modalities coupled to more routine digital MRI or pathology. “This proof-of-concept study provides some insights for prospective, potentially practice-changing clinical studies and, ultimately, pave the way to significant advances in precision oncology,” Vibert concludes.