2O - Prostate cancer BRCA mutation prediction using multiple-instance learning and histopathology

Background

BRCA1/2 alterations are associated with poor prognosis in prostate cancer (PCa). However, tumours with genomic loss of BRCA1/2, show increased response to PARP inhibitors and platinum-based chemotherapy. Access to genomic testing remains limited in routine clinical practice, and technical challenges arise with next-generation sequencing of small tumour biopsies. Phenotypic biomarkers of BRCA1/2 status can support the deployment of precision medicine strategies in the clinic. Deep-learning architectures such as attention-based Multiple Instance Learning (att-MIL) applied to digital pathology have shown promise in oncology. Here, we study whether att-MIL can predict BRCA1/2 status in PCa from routine haematoxylin and eosin whole slide images (WSI).

Methods

Data from tumor specimens from PCa patients were collected from academic studies or public databases. BRCA1/2 status was determined by whole exome sequencing and/or targeted panel. We performed class-wise data augmentation to balance the training cohort. WSIs were tessellated into 224x224 tiles and a #tilesx2048 features embedding were extracted using cTransPath. Such embeddings fed a 5-fold cross-validated model following CLAM architecture. Model’s attention maps were extracted for predictions’ explainability. Point-Biseral correlation tested associations between predicted BRCA1/2 status and Gleason group.

Results

In total, 1075 tumor tissues from 864 patients were collected, including specimens from primary tumors (998) and metastasis (77). At diagnosis, 697 had localized and 167 stage IV disease. Mean Gleason group was 3.93 (95%CI 3.82-4.04) The prevalence of BRCA1/2 altered was 9.5%, increasing to 24.15% after data augmentation. The 5-fold cross-validation reached an area under the receptor operating and precision sensitivity curves of 0.87±0.04 and 0.71±0.08, respectively. Attention maps showed a lack of gland formation in high-scored tiles and clear well-formed glands in low-scored tiles. No correlation was found between predicted BRCA1/2 status and Gleason (r=0.04, p=0.65).

Conclusions

Digital pathology can provide phenotypic biomarkers associated with BRCA1/2 alterations PCa, aiding the deployment of precision medicine strategies in this disease.

Editorial acknowledgement

Clinical trial identification

Legal entity responsible for the study

The authors.

Funding

Vall d'Hebron Institute of Oncology.

Disclosure

D. Olmos: Financial Interests, Personal, Other, Honoraria: Janssen; Financial Interests, Institutional, Advisory Role: Janssen, AstraZeneca, Bayer; Financial Interests, Personal, Advisory Role: Clovis Oncology. P.G. Nuciforo: Financial Interests, Personal, Invited Speaker: Novartis; Financial Interests, Personal, Advisory Board: MSD Oncology, Bayer; Financial Interests, Personal, Other, Consultant: Targos Molecular Pathology GmbH. A. Vivancos: Financial Interests, Personal, Advisory Board: Roche, Bristol Meyers Squibb, Guardant Health, Bayer, Incyte; Financial Interests, Personal, Stocks/Shares: Reveal Genomics; Financial Interests, Institutional, Research Grant, Preclinical Research Grant: Bristol Meyers Squibb, Roche, Incyte. J. Carles Galceran: Financial Interests, Personal, Advisory Board: Astellas Pharma, AstraZeneca, Bayer, Johnson & Johnson, MSD Oncology, Novartis (AAA), Sanofi, Bristol Myers Squibb, Exelixis, Ipsen, Pfizer; Financial Interests, Institutional, Invited Speaker: Janssen-Cilag International NV, Lilly, S.A, MedImmune, Novartis Farmacéutica, S.A, Sanofi-Aventis, S.A; Other, Personal, Other, Member of the Commission: Catalan Program of Ambulatory Medication Commission (CAHMDA). E. Castro: Financial Interests, Personal, Advisory Board: Astellas, AstraZeneca, Janssen, MSD, Bayer, Pfizer, Daiichi Sankyo, Lilly, Medscape, Novartis; Financial Interests, Personal, Invited Speaker: Astellas, AstraZeneca, Janssen, Clovis, Pfizer, Pfizer, Medscape; Financial Interests, Institutional, Funding: AstraZeneca; Financial Interests, Institutional, Research Grant: Janssen, Bayer, Pfizer; Financial Interests, Institutional, Invited Speaker: Janssen, Pfizer, MSD. J. Mateo: Financial Interests, Personal, Advisory Board: AstraZeneca, Amgen, Janssen, Roche, Amunix, Pfizer; Financial Interests, Personal, Invited Speaker: AstraZeneca, MSD, Guardant Health; Financial Interests, Institutional, Advisory Board, Scientific Advisory Board Member for the company: Nuage Therapeutics; Financial Interests, Institutional, Research Grant: AstraZeneca, Pfizer Oncology, Amgen; Non-Financial Interests, Institutional, Product Samples, Access to drugs in early development for preclinical testing: AstraZeneca. R. Perez Lopez: Financial Interests, Personal, Full or part-time Employment, VHIO staff (team leader of the Radiomics Group): VHIO; Financial Interests, Institutional, Research Grant, Co-PI of 3 research studies: AstraZeneca; Financial Interests, Institutional, Research Grant, PI of a research study: Roche. All other authors have declared no conflicts of interest.