1202P - A deep learning approach using routine pathology images to guide precision medicine in metastatic CRC

Background

Survival outcomes of patients (pts) with metastatic colorectal cancer (CRC) have improved in recent decades, reflecting advances in targeted treatments, diagnostics, and biomarker (BM)-driven pt selection. Notably, microsatellite instability (MSI), RAS, and RAF have shown to be predictive of outcomes to specific targeted therapies, and the gene-expression based Consensus Molecular Subtypes (CMS) have demonstrated predictive value across stages of CRC. Nevertheless, factors such as socioeconomic disparities may preclude pt segments from access to molecular testing, while cost and technical challenges make RNA-based signatures difficult to implement in clinical practice. Alternative methods for BM detection, such as analysis of pathology images obtained via routine clinical care, can support subsequent clinical decision making. We developed a deep-learning (DL) algorithm to infer RAS/RAF/MSI and CMS classes using only H&E digital pathology images.

Methods

A DL model was pretrained using self-supervised learning on ∼55k unlabeled, digital H&E whole slide images (WSI) from multiple scanners, hospital systems, tumor / tissue types, etc. The model was then finetuned on 597 labeled H&E WSI from the TCGA-CRC dataset to detect alterations in BRAF, KRAS, NRAS, and MSI. A separate model was finetuned to derive CMS, where CMS labels were independently inferred using the RNA-Seq-based CMScaller algorithm. For all models, the data were split 80/20 for finetuning / testing, with the same BM prevalences in each split. Performance was evaluated on the test data using the Area Under ROC Curve (AUC).

Results

The models achieved AUCs of 0.85 - 0.87 for BRAF, KRAS, NRAS and MSI classification on the test data. The model achieved a macro averaged AUC of 0.82 for differentiating CMS classes, demonstrating robust performance across tasks.

Conclusions

If deployed as a pre-screening tool, each model could reduce subsequent confirmatory testing volumes by 58%, 72%, 67% and 42% for BRAF, KRAS, NRAS and MSI while retaining 90% sensitivity. Through accurate detection of RAS, RAF, MSI, and CMS from H&E slides, our DL algorithms could offer an alternative to genetic testing, streamlining BM assessment in clinical practice and broadening access to novel therapies.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Janssen Research and Development.

Funding

Janssen Research and Development.

Disclosure

C. Parmar: Financial Interests, Institutional, Stocks/Shares: Johnson and Johnson; Financial Interests, Institutional, Full or part-time Employment: Johnson and Johnson. A. Juan Ramon, S. Chowdhury, J.C. Curtin, M. Baig, R.W. Schnepp, X. Lyu, C. Moy: Financial Interests, Institutional, Full or part-time Employment: Johnson and Johnson; Financial Interests, Institutional, Stocks/Shares: Johnson and Johnson. P. Raciti: Financial Interests, Institutional, Stocks/Shares: Janssen R&D, Paige; Financial Interests, Institutional, Full or part-time Employment: Janssen R&D; Financial Interests, Institutional, Research Funding: Paige. K. Standish: Financial Interests, Institutional, Full or part-time Employment: Johnson and Johnson; Financial Interests, Institutional, Stocks/Shares: Johnson and Johnson, Merck; Financial Interests, Institutional, Research Funding: Johnson and Johnson. S. Shah: Financial Interests, Institutional, Full or part-time Employment: Johnson and Johnson; Financial Interests, Institutional, Research Funding: Johnson and Johnson, AstraZeneca; Financial Interests, Institutional, Stocks/Shares: AstraZeneca. F. Cruz-Guilloty: Financial Interests, Institutional, Full or part-time Employment: Johnson and Johnson; Financial Interests, Institutional, Stocks/Shares: Johnson and Johnson, Amgen. Z. Albertyn: Financial Interests, Institutional, Advisory Role: Johnson and Johnson. J. Greshock: Financial Interests, Personal, Full or part-time Employment: Johnson and Johnson; Financial Interests, Personal, Stocks/Shares: Johnson and Johnson; Non-Financial Interests, Leadership Role: Johnson and Johnson; Non-Financial Interests, Institutional, Proprietary Information: Johnson and Johnson. L. Demirdjian: Financial Interests, Institutional, Stocks or ownership: Janssen R&D.