% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Gustav:303950,
author = {M. Gustav and M. van Treeck and N. G. Reitsam and Z. I.
Carrero and C. M. L. Loeffler and A. Rabasco Meneghetti and
B. Märkl and L. A. Boardman and A. J. French and E. L.
Goode and A. Gsur and S. Brezina and M. J. Gunter and N.
Murphy and P. Hönscheid$^*$ and C. Sperling and S. Foersch
and R. Steinfelder and T. Harrison and U. Peters and A.
Phipps and J. N. Kather},
title = {{A}ssessing genotype-phenotype correlations in colorectal
cancer with deep learning: a multicentre cohort study.},
journal = {The lancet / Digital health},
volume = {7},
number = {8},
issn = {2589-7500},
address = {London},
publisher = {The Lancet},
reportid = {DKFZ-2025-01725},
pages = {100891},
year = {2025},
note = {2025 Aug;7(8):100891},
abstract = {Deep learning-based models enable the prediction of
molecular biomarkers from histopathology slides of
colorectal cancer stained with haematoxylin and eosin;
however, few studies have assessed prediction targets beyond
microsatellite instability (MSI), BRAF, and KRAS
systematically. We aimed to develop and validate a
multi-target model based on deep learning for the
simultaneous prediction of numerous genetic alterations and
their associated phenotypes in colorectal cancer.In this
multicentre cohort study, tissue samples from patients with
colorectal cancer were obtained by surgical resection and
stained with haematoxylin and eosin. These samples were then
digitised into whole-slide images and used to train and test
a transformer-based deep learning algorithm for biomarker
detection to simultaneously predict multiple genetic
alterations and provide heatmap explanations. The primary
dataset comprised 1376 patients from five cohorts who
underwent comprehensive panel sequencing, with an additional
536 patients from two public datasets for validation. We
compared the model's performance against conventional
single-target models and examined the co-occurrence of
alterations and shared morphology.The multi-target model was
able to predict numerous biomarkers from pathology slides,
matching and partly exceeding single-target transformers. In
the primary external validation cohorts, mean area under the
receiver operating characteristic curve (AUROC) for the
multi-target transformer was 0·78 (SD 0·01) for BRAF,
0·88 (0·01) for hypermutation, 0·93 (0·01) for MSI, and
0·86 (0·01) for RNF43; predictive performance was
consistent across metrics and supported by co-occurrence
analyses. However, biomarkers with high AUROCs largely
correlated with MSI, with model predictions depending
considerably on morphology associated with MSI at
pathological examination.By use of morphology associated
with MSI and more subtle biomarker-specific patterns within
a shared phenotype, the multi-target transformers
efficiently predicted biomarker status for diverse genetic
alterations in colorectal cancer from slides stained with
haematoxylin and eosin. These results highlight the
importance of considering mutational co-occurrence and
common morphology in biomarker research based on deep
learning. Our validated and scalable model could support
extension to other cancers and large, diverse cohorts,
potentially facilitating cost-effective pre-screening and
streamlined diagnostics in precision oncology.German Federal
Ministry of Health, Max-Eder-Programme of German Cancer Aid,
German Federal Ministry of Education and Research, German
Academic Exchange Service, and the EU.},
cin = {DD01},
ddc = {610},
cid = {I:(DE-He78)DD01-20160331},
pnm = {899 - ohne Topic (POF4-899)},
pid = {G:(DE-HGF)POF4-899},
typ = {PUB:(DE-HGF)16},
pubmed = {pmid:40829965},
doi = {10.1016/j.landig.2025.100891},
url = {https://inrepo02.dkfz.de/record/303950},
}