by Tim Heinemann1$, Christoph Kornauth2$, Yannik Severin1, Gregory I Vladimer3,7, Tea Pemovska3,4, Emir Hadzijusufovic4, Hermine Agis4, Maria-Theresa Krauth4, Wolfgang R Sperr4,5, Peter Valent4,5, Ulrich Jäger4, Ingrid Simonitsch-Klupp2, Giulio Superti-Furga3,6, Philipp B Staber4#, Berend Snijder1#
1: Department of Biology, Institute of Molecular Systems Biology, ETH Zurich, Zurich, Switzerland
2: Department of Pathology, Medical University of Vienna, Vienna, Austria
3: CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences, Vienna, Austria
4: Department of Medicine I, Div. of Hematology and Hemostaseology, Medical University of Vienna, Austria
5: Ludwig Boltzmann Institute for Hematology and Oncology, Medial University of Vienna, Austria
6: Center for Physiology and Pharmacology, Medical University of Vienna, Austria
7: Current address: Exscientia GmbH, Campus-Vienna-Biocenter 5, 1030 Vienna, Austria
$: These authors contributed equally to the work
#: Equal senior authorship
Correspondence: bsnijder@ethz.ch
Image-based drug testing in patient biopsies has recently been shown to identify potent treatments for patients suffering from relapsed or refractory hematological cancers. Here we investigate the use of weakly-supervised deep learning on cell morphologies (DML) to complement immunofluorescence (IF) in the classification of cancer and healthy cells in such drug testing. Across 390 biopsies from 289 patients with diverse blood cancers, DML-based drug responses show improved reproducibility and clustering of drugs with the same mode of action. DML does so by adapting to batch effects and by autonomously recognizing disease-associated cell morphologies. In a post-hoc analysis of 66 patients, DML-recommended treatments led to improved progression free survival compared to IF-based recommendations and physician’s choice-based treatments. Treatments recommended by both IF and DML enriched for patients achieving exceptional clinical responses. Thus, DML-enhanced ex vivo drug screening is a promising new tool in the identification of effective personalized treatments.
Please cite and see for more details Heinemann et al., Blood Cancer Discovery, 2022.
The Matlab code to perform 'deep morphology learning' using a pre-trained mCNN instance can be downloaded here as ZIP file: dml.zip (3,8MB). The ZIP file additionally contains example single-cell data, some required subfunctions, as well as a Matlab script that shows how to run mCNN.