Prognostic RNA-splicing archetypes in breast cancer - identified by extended pre-training of histopathology foundation models

Sep 22, 2026

Upload by nayy☆

The study advances histopathology foundation models by extending pre-training on invasive breast tumor tissue, creating tumor-specialised models that capture richer semantic and biological information. It evaluates encoded biological concepts across multiple scales and identifies recurrent tumor archetypes with consistent morphological and molecular identities across patients. The dominant archetypes reveal aberrant RNA splicing-associated gene-expression programs and show prognostic value. These findings enable computationally efficient, biologically informed patient stratification using routine H&E slides.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

AccessFree to download

Overview of the strategy

Foundation models and extended pre-training

Tumor archetype discovery and molecular characterization

Prognostic RNA splicing-associated programs

Article [https://doi.org/10.1038/s41467-026-75217-z](https://doi.org/10.1038/s41467-026-75217-z) Prognostic RNA-splicing archetypesin breast cancer identiﬁed by extended pre-training of histopathology foundation models Received: 11 April 2025 | Accepted: 25 June 2026 | | --- | | | | Check for updates | Lisa Fournier1,2,3, Garance Haeﬂiger1,3, Albin Vernhes1,3, Vincent Jung1,4, Lena Loye5,6, Valentine Du Bois 7,8, Intidhar Labidi-Galy7,8, Pascal Frossard4, Igor Letovanec9,10, Cédric Vincent-Cuaz 4,5,6,11 & Raphaëlle Luisier 3,5,6,11 Recently, histopathology foundation models (hFM) have rapidly advanced in size and complexity, achieving excellent performance in cancer diagnosis and biomarker discovery. Here, we specialise pre-trained hFMs to invasive tumour tissue and present three key contributions. First, we systematically evaluate the biological concepts encoded in hFM representations across multiple biological scales. Second, we demonstrate that informed extended pre-training transforms generalist models into tumour-specialised ones encoding richer semantic information, enabling discovery of recurrent tumour archetypes with consistent morphological and molecular identities across patients. Third, we identify dominant tumour archetypes with aberrant gene-expression programs coexisting within tumours and recurring across heterogeneous epithelial cancers, including HER2-positive and triple-negative breast cancer. Crucially, these archetypes exhibit prognostic value, with RNA splicingassociated archetypes consistently predicting poorer outcomes. Our work shows that tumour-specialised hFMs unlock rich molecular and morphological information from routine H&E slides, providing computationally efﬁcient and biologically informed solutions for biological discovery and patient stratiﬁcation. Intratumor heterogeneity (ITH) is a hallmark of cancer1 and a major driver of therapy resistance, particularly in heterogeneous malignancies such as breast cancer2. Accurately quantifying and characterising ITH is essential for improving patient stratiﬁcation and predicting therapeutic outcomes. ITH has been extensively studied at the molecular level, particularly through single-cell RNA sequencing (scRNA-seq)3,4, revealing the coexistence of distinct tumour cell subpopulations within the same tumour, each deﬁned by unique molecular identities linked to tumour invasion and therapy resistance5–7. However, how these distinct molecular programs spatially co-exist within tumours, forming regions deﬁned by unique molecular proﬁles, tissue organisation, and composition, often referred to as “archetypes”, remains poorly understood. Recent advances in spatial transcriptomics (ST) and multi-omics 1Idiap Research Institute, Martigny, Switzerland. 2Center of Translational Research in Onco-Hematology, Faculty of Medicine, University of Geneva, Geneva, Switzerland. 3Swiss Institute of Bioinformatics, Lausanne, Switzerland. 4Signal Processing Laboratory (LTS4), School of Engineering, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland. 5Department for BioMedical Research, University of Bern, Bern, Switzerland. 6Department of Digital Medicine, University of Bern, Bern, Switzerland. 7Department of Oncology, Hôpitaux Universitaires de Genève, Geneva, Switzerland. 8Faculty of Medicine, Université de Genève, Geneva, Switzerland. 9Department of Histopathology, Central Institute, Valais Hospital, Sion, Switzerland. 10Institute of Pathology, Centre Hospitalier Universitaire Vaudois CHUV, Lausanne, Switzerland. 11These authors contributed equally: Cédric Vincent-Cuaz, Raphaëlle Luisier. e-mail: [cedric.vincent-cuaz@unibe.ch](cedric.vincent-cuaz@unibe.ch); [raphaelle.luisier@unibe.ch](raphaelle.luisier@unibe.ch) Fig. 1 | Overview of the strategy. a Schema of the publicly available datasets of breast tumors used in this study36,37, comprising H&E images, spatial transcriptomics data, and expert pathological annotations categorizing

FAQs

What is the main goal of extending pre-training for histopathology foundation models in this study?

To adapt generalist histopathology foundation models to invasive breast tumor tissue, producing tumor-specialised models with richer semantic and biological information.

How are tumor archetypes identified from histology images?

Embeddings from baseline hFMs and tumour-expert models are clustered in the image embedding space, and the resulting archetypes are characterised by integrating spatial transcriptomics data.

What does the study show about RNA-splicing-associated archetypes and prognosis?

RNA-splicing-associated tumour archetypes consistently predict poorer outcomes, linking aberrant gene-expression programs to prognostic value.

Sep 22, 2026

Upload by nayy☆