Pan Pantziarka, Vidula Sukhatme, Lydie Meheus, Vikas Sukhatme, Gauthier Bouche

Background: Drug repurposing can speed up access to new therapeutic options for cancer patients. With more than 2000 drugs approved worldwide and 6 relevant targets per drug on average, the potential is quantitatively important. In this paper, we have attempted to quantify the number of non-cancer drugs supported by either preclinical or clinical cancer data. Methods: A PubMed search was performed to identify non-cancer drugs which could be repurposed in one or more cancer types. Drugs needed at least one peer-reviewed article showing an anticancer effect in vitro, in vivo or in humans. Results: A total of 235 eligible non-cancer drugs were identified (Table 1). Main characteristics of the drugs are summarized in Table 2. 67 (29%) are on the WHO list of essential medicines and 176 (75%) are off-patent. 133 (57%) had human data in cancer patient(s). Four were listed in clinical guidelines, namely thalidomide, all-trans retinoic acid, zoledronic acid and non-steroidal anti-inflammatory drugs (NSAID). Several drugs have shown a survival benefit in randomized trials such as cimetidine (colorectal cancer), progesterone (breast cancer) or itraconazole (lung cancer). Several other drugs induced responses in rare tumours, like clarithromycin, timolol or propranolol. Conclusion: We have found that the number of off-patent repurposing opportunities is large and increasing. Joint non-commercial clinical development (academics, governments, charities) may bring new therapeutic options to patients at low cost, especially in indications for which the industry has no incentive to invest in.

Jianhua Yin, Zhisheng Li, Chen Yan, Enhao Fang, Ting Wang, Hanlin Zhou, Weiwei Luo, Qing Zhou, Jingyu Zhang, Jintao Hu, Haoxuan Jin, Lei Wang, Xing Zhao, Jiguang Li, Xiaojuan Qi, Wenbin Zhou, Chen Huang, Chenyang He, Huanming Yang, Karsten Kristiansen, Yong Hou, Shida Zhu, Dongxian Zhou, Ling Wang, Michael Dean, Kui Wu, Hong Hu, Guibo Li

The tumor microenvironment is composed of numerous cell types, including tumor, immune and stromal cells. Cancer cells interact with the tumor microenvironment to suppress anticancer immunity. In this study, we molecularly dissected the tumor microenvironment of breast cancer by single-cell RNA-seq. We profiled the breast cancer tumor microenvironment by analyzing the single-cell transcriptomes of 52,163 cells from the tumor tissues of 15 breast cancer patients. The tumor cells and immune cells from individual patients were analyzed simultaneously at the single-cell level. This study explores the diversity of the cell types in the tumor microenvironment and provides information on the mechanisms of escape from clearance by immune cells in breast cancer.

Aurélie Kamoun, Aurélien de Reyniès, Yves Allory, Gottfrid Sjödahl, A. Gordon Robertson, Roland Seiler, Katherine A. Hoadley, Hikmat Al-Ahmadie, Woonyoung Choi, Clarice S. Groeneveld, Mauro A. A. Castro, Jacqueline Fontugne, Pontus Eriksson, Qianxing Mo, Jordan Kardos, Alexandre Zlotta, Arndt Hartmann, Colin P. Dinney, Joaquim Bellmunt, Thomas Powles, Núria Malats, Keith S. Chan, William Y. Kim, David J McConkey, Peter C. Black, Lars Dyrskjøt, Mattias Höglund, Seth P. Lerner, Francisco X Real, François Radvanyi, The Bladder Cancer Molecular Taxonomy Group

Muscle-Invasive Bladder Cancer (MIBC) is a molecularly diverse disease with heterogeneous clinical outcomes. Several molecular classifications have been proposed, yielding diverse sets of subtypes. This diversity hampers the clinical application of such knowledge. Here, we report the results of a large international effort to reach a consensus on MIBC molecular subtypes. Using 1750 MIBC transcriptomes and a network-based analysis of six independent MIBC classification systems, we identified a consensus set of six molecular classes: Luminal Papillary (24%), Luminal Non-Specified (8%), Luminal Unstable (15%), Stroma-rich (15%), Basal/Squamous (35%), and Neuroendocrine-like (3%). These consensus classes differ regarding underlying oncogenic mechanisms, infiltration by immune and stromal cells, and histological and clinical characteristics. This consensus system offers a robust framework that will enable testing and validating predictive biomarkers in future clinical trials.

Gabriela S Kinker, Alissa C Greenwald, Rotem Tal, Zhanna Orlova, Michael S Cuoco, James M. McFarland, Allison Warren, Christopher Rodman, Jennifer A Roth, Samantha A Bender, Bhavna Kumar, James W. Rocco, Pedro ACM Fernandes, Christopher C Mader, Hadas Keren-Shaul, Alexander Plotnikov, Haim Barr, Aviad Tsherniak, Orit Rozenblatt-Rosen, Valery Krizhanovsky, Sidharth V Puram, Aviv Regev, Itay Tirosh

Cultured cell lines are the workhorse of cancer research, but it is unclear to what extent they recapitulate the cellular heterogeneity observed among malignant cells in tumors, given the absence of a native tumor microenvironment. Here, we used multiplexed single cell RNA-Seq to profile ~200 cancer cell lines. We uncovered expression programs that are recurrently heterogeneous within many cancer cell lines and are largely independent of observed genetic diversity. These programs of heterogeneity are associated with diverse biological processes, including cell cycle, senescence, stress and interferon responses, epithelial-to-mesenchymal transition (EMT), and protein maturation and degradation. Notably, some of these recurrent programs recapitulate those seen in human tumors, suggesting a prominent role of intrinsic plasticity in generating intra-tumoral heterogeneity. Moreover, the data allowed us to prioritize specific cell lines as model systems of cellular plasticity. We used two such models to demonstrate the dynamics, regulation and vulnerabilities associated with a cancer senescence program observed both in cell lines and in human tumors. Our work describes the landscape of cellular heterogeneity in diverse cancer cell lines, and identifies recurrent patterns of expression heterogeneity that are shared between tumors and specific cell lines and can thus be further explored in follow up studies.

Geoff Macintyre, Teodora E. Goranova, Dilrini De Silva, Darren Ennis, Anna M. Piskorz, Matthew Eldridge, Daoud Sie, Liz-Anne Lewsley, Aishah Hanif, Cheryl Wilson, Suzanne Dowson, Rosalind M. Glasspool, M. Lockley, Elly Brockbank, Ana Montes, Axel Walther, Sudha Sundar, Richard Edmondson, Geoff D. Hall, Andrew Clamp, Charlie Gourley, Marcia Hall, Christina Fotopoulou, Hani Gabra, James Paul, Anna Supernat, David Millan, Aoisha Hoyle, Gareth Bryson, Craig Nourse, Laura Mincarelli, Luis Navarro Sanchez, Bauke Ylstra, Mercedes Jimenez-Linan, Luiza Moore, Oliver Hofmann, Florian Markowetz, Iain A. McNeish, James D. Brenton

Genomic complexity from profound copy-number aberration has prevented effective molecular stratification of ovarian and other cancers. Here we present a method for copy-number signature identification that decodes this complexity. We derived eight signatures using 117 shallow whole-genome sequenced high-grade serous ovarian cancer cases, which were validated on a further 497 cases. Mutational processes underlying the copy-number signatures were identified, including breakage-fusion-bridge cycles, homologous recombination deficiency and whole-genome duplication. We show that most tumours are heterogeneous and harbour multiple signature exposures. We also demonstrate that copy number signatures predict overall survival and changes in signature exposure observed in response to chemotherapy suggest potential treatment strategies.

Jessica N Spradlin, Xirui Hu, Carl C Ward, Scott M Brittain, Michael D. Jones, Lisha Ou, Milton To, Andrew Proudfoot, Elizabeth Ornelas, Mikias Woldegiorgis, James A Olzmann, Dirksen Bussiere, Jason R Thomas, John A Tallarico, Jeffrey M McKenna, Markus Schirle, Thomas J. Maimone, Daniel K. Nomura

Nimbolide, a terpenoid natural product derived from the Neem tree, impairs cancer pathogenicity across many types of human cancers; however, the direct targets and mechanisms by which nimbolide exerts its effects are poorly understood. Here, we used activity-based protein profiling (ABPP) chemoproteomic platforms to discover that nimbolide reacts with a novel functional cysteine crucial for substrate recognition in the E3 ubiquitin ligase RNF114. Nimbolide impairs breast cancer cell proliferation in-part by disrupting RNF114 substrate recognition, leading to inhibition of ubiquitination and degradation of the tumor-suppressors such as p21, resulting in their rapid stabilization. We further demonstrate that nimbolide can be harnessed to recruit RNF114 as an E3 ligase in targeted protein degradation applications and show that synthetically simpler scaffolds are also capable of accessing this unique reactive site. Our study highlights the utility of ABPP platforms in uncovering unique druggable modalities accessed by natural products for cancer therapy and targeted protein degradation applications.

Charles P. Couturier, Shamini Ayyadhury, Phuong U. Le, Jean Monlong, Gabriele Riva, Redouane Allache, Salma Baig, Xiaohua Yan, Mathieu Bourgey, Changseok Lee, Yu Chang David Wang, V. Wee Yong, Marie-Christine Guiot, Bratislav Mišić, Jack Antel, G Bourque, Jiannis Ragoussis, Kevin Petrecca

Cancer stem cells are critical for cancer initiation, development, and resistance to treatments. Our understanding of these processes, and how they relate to glioblastoma heterogeneity, is limited. To overcome these limitations, we performed single-cell RNA-sequencing on 38 296 glioblastoma cells and 22 637 normal human fetal brain cells. Using an unbiased approach, we mapped the lineage hierarchy of the developing human brain and compared the transcriptome of each cancer cell to this roadmap. We discovered a conserved neural trilineage cancer hierarchy with glial progenitor-like cells at the apex. We also found that this progenitor population contains the majority of cancer's cycling cells and is the origin of heterogeneity. Finally, we show that this hierarchal map can be used to identify therapeutic targets specific to progenitor cancer stem cells. Our analyses show that normal brain development reconciles glioblastoma development, unravels the origin of glioblastoma heterogeneity, and helps to identify cancer stem cell-specific targets.

Peter Ulz, Samantha Perakis, Qing Zhou, Tina Moser, Jelena Belic, Isaac Lazzeri, Albert Wölfler, Armin Zebisch, Armin Gerger, Gunda Pristauz, Edgar Petru, Brandon White, Charles E.S. Roberts, John St. John, Michael G Schimek, Jochen B. Geigl, Thomas Bauernhofer, Heinz Sill, Christoph Bock, Ellen Heitzer, Michael R. Speicher

Deregulation of transcription factors (TFs) is an important driver of tumorigenesis. We developed and validated a minimally invasive method for assessing TF activity based on cell-free DNA sequencing and nucleosome footprint analysis. We analyzed whole genome sequencing data for >1,000 cell-free DNA samples from cancer patients and healthy controls using a newly developed bioinformatics pipeline that infers accessibility of TF binding sites from cell-free DNA fragmentation patterns. We observed patient-specific as well as tumor-specific patterns, including accurate prediction of tumor subtypes in prostate cancer, with important clinical implications for the management of patients. Furthermore, we show that cell-free DNA TF profiling is capable of early detection of colorectal carcinomas. Our approach for mapping tumor-specific transcription factor binding in vivo based on blood samples makes a key part of the noncoding genome amenable to clinical analysis.

Wei Jiao, Gurnit Atwal, Paz Polak, Rosa Karlic, Edwin Cuppen, Alexandra Danyi, Jeroen de Ridder, Carla van Herpen, Martijn P. Lolkema, Neeltje Steeghs, Gad Getz, Quaid D. Morris, Lincoln D Stein, PCAWG Pathology & Clinical Correlates Working Grp, ICGC/TCGA Pan-cancer Analysis of Whole Genomes Net

In cancer, the primary tumour's organ of origin and histopathology are the strongest determinants of its clinical behaviour, but in 3% of the time a cancer patient presents with metastatic tumour and no obvious primary. Challenges also arise when distinguishing a metastatic recurrence of a previously treated cancer from the emergence of a new one. Here we train a deep learning classifier to predict cancer type based on patterns of somatic passenger mutations detected in whole genome sequencing (WGS) of 2606 tumours representing 24 common cancer types. Our classifier achieves an accuracy of 91% on held-out tumor samples and 82% and 85% respectively on independent primary and metastatic samples, roughly double the accuracy of trained pathologists when presented with a metastatic tumour without knowledge of the primary. Surprisingly, adding information on driver mutations reduced classifier accuracy. Our results have immediate clinical applicability, underscoring how patterns of somatic passenger mutations encode the state of the cell of origin, and can inform future strategies to detect the source of cell-free circulating tumour DNA.

Chengyan Dong, Li Huang, Sonia A. Melo, Paul Kurywchak, Qian Peng, Christoph Kahlert, Valerie LeBleu, Raghu Kalluri

Exosomes are man-sized vesicles shed by all cells, including cancer cells. Exosomes can serve as novel liquid biopsies for diagnosis of cancer with potential prognostic value. The exact mechanism/s associated with sorting or enrichment of cellular components into exosomes are still largely unknown. We reported Glypican-1 (GPC1) on the surface of cancer exosomes and provided evidence for the enrichment of GPC1 in exosomes from patients with pancreatic cancer. Several different laboratories have validated this novel conceptual advance and reproduced the original experiments using multiple antibodies from different sources. These include anti-GPC1 antibodies from ThermoFisher (PA5-28055 and PA-5-24972), Sigma (SAB270028), Abnova (MAB8351, monoclonal antibodies clone E9E), EMD Millipore (MAB2600-monoclonal antibodies), SantaCruz, and R&D Systems (BAF4519). This report complements such independent findings and report on the specific detection of Glypican-1 on the exosomes derived from the serum of pancreas cancer patients using multiple antibodies. Additionally, the specificity of the antibodies to GPC1 was determined by western blot and Protein Simple analyses of pancreatic cancer cells and their exosomes. Interestingly, our results highlight a specific enrichment of high molecular weight GPC1 on exosomes, potentially contributed by heparan sulfate and other glycosylation modifications.

Yotam E. Bar-Ephraim, Kai Kretzschmar, Priyanca Asra, Evelien de Jongh, Kim E Boonekamp, Jarno Drost, Joost van Gorp, Apollo Pronk, Niels Smakman, Inez J. Gan, Zsolt Sebestyén, Jürgen Kuball, Robert GJ Vries, Hans Clevers

Immune escape has been recognised as one of the hallmarks of cancer. Overcoming this immunomodulatory process by tumour cells has become a major therapeutic target. Here we utilize organoid technology to study immune-cancer interactions and assess immunomodulation by colorectal cancer (CRC). Transcriptional profiling and flow cytometry revealed that organoids maintain differential expression of immunomodulatory molecules present in primary tumours. Finally, we established a method to model antigen-specific epithelial cell killing and cancer immunomodulation in vitro using CRC organoids co-cultured with cytotoxic T cells. Our method may serve as a first step to rebuilding the tumor microenvironment in vitro.

Luiza Moore, Daniel Leongamornlert, Tim H. H. Coorens, Mathijs A Sanders, Peter Ellis, Kevin Dawson, Francesco Maura, Jyoti Nangalia, Patrick S Tarpey, Simon F Brunner, Henry Lee-Six, Raheleh Rahbari, Sarah Moody, Yvette Hooks, Krishnaa Mahbubani, Mercedes Jimenez-Linan, Jan J Brosens, Christine A. Iacobuzio-Donahue, Iñigo Martincorena, Kourosh Saeb-Parsy, Peter J. Campbell, Michael R. Stratton

All normal somatic cells are thought to acquire mutations. However, characterisation of the patterns and consequences of somatic mutation in normal tissues is limited. Uterine endometrium is a dynamic tissue that undergoes cyclical shedding and reconstitution and is lined by a gland-forming epithelium. Whole genome sequencing of normal endometrial glands showed that most are clonal cell populations derived from a recent common ancestor with mutation burdens differing from other normal cell types and manyfold lower than endometrial cancers. Mutational signatures found ubiquitously account for most mutations. Many, in some women potentially all, endometrial glands are colonised by cell clones carrying driver mutations in cancer genes, often with multiple drivers. Total and driver mutation burdens increase with age but are also influenced by other factors including body mass index and parity. Clones with drivers often originate during early decades of life. The somatic mutational landscapes of normal cells differ between cell types and are revealing the procession of neoplastic change leading to cancer.

Ron S. Gejman, Heather F. Jones, Martin G. Klatt, Aaron Y. Chang, Claire Y Oh, Smita S Chandran, Tatiana Korontsvit, Viktoriya Zakahleva, Tao Dao, Christopher A. Klebanoff, David A. Scheinberg

T cell receptor (TCR)-based therapeutic cells and agents have emerged as a new class of effective cancer therapeutics. These therapies work on cells that express intracellular cancer-associated proteins by targeting peptides displayed on major histocompatibility complex receptors. However, cross-reactivities of these agents to off-target cells and tissues have resulted in serious, sometimes fatal, adverse events. We have developed a high throughput genetic platform (termed “PresentER”) that encodes MHC-I peptide minigenes for functional immunological assays as well as for determining the reactivities of TCR-like therapeutic agents against large libraries of MHC-I ligands. In this report, we demonstrate that PresentER can be used to identify the on-and-off targets of T cells and TCR mimic antibodies using in vitro co-culture assays or binding assays. We find dozens of MHC-I ligands that are cross-reactive with two TCR mimic antibodies and two native TCRs and that are not easily predictable by other methods.

Marc J. Williams, Benjamin Werner, Christina Curtis, Chris Barnes, Andrea Sottoriva, Trevor A Graham

Recent studies have identified prevalent subclonal architectures within many cancer types. However, the temporal evolutionary dynamics that produce these subclonal architectures remain unknown. Here we measure evolutionary dynamics in primary human cancers using computational modelling of clonal selection applied to high throughput sequencing data. Our approach simultaneously determines the subclonal architecture of a tumour sample, and measures the mutation rate, the selective advantage, and the time of appearance of subclones. Simulations demonstrate the accuracy of the method, and revealed the degree to which evolutionary dynamics are recorded in the genome. Application of our method to high-depth sequencing data from gastric and lung cancers revealed that detectable subclones consistently emerged early during tumour growth and had considerably large fitness advantages (>20% growth advantage). Our quantitative platform provides new insight into the evolutionary history of cancers by facilitating the measurement of fundamental evolutionary parameters in individual patients.

Hamad Alshetaiwi, Nicholas Pervolarakis, Laura Lynn McIntyre, Dennis Ma, Quy Nguyen, Jan Akara Rath, Kevin Nee, Grace Hernandez, Katrina Evans, Leona Torosian, Anushka Silva, Craig Walsh, Kai Kessenbrock

Myeloid-derived suppressor cells (MDSCs) are innate immune cells that acquire the capacity to suppress adaptive immune responses during cancer. It remains elusive how MDSCs differ from their normal myeloid counterparts, which limits our ability to specifically detect and therapeutically target MDSCs during cancer. Here, we used single-cell RNAseq to compare MDSC-containing splenic myeloid cells from breast tumor-bearing mice to wildtype controls. Our computational analysis of 14,646 single-cell transcriptomes reveals that MDSCs emerge through a previously unrealized aberrant neutrophil maturation trajectory in the spleen giving rise to a unique chemokine-responsive, immunosuppressive cell state that strongly differs from normal myeloid cells. We establish the first MDSC-specific gene signature and identify novel surface markers for improved detection and enrichment of MDSCs in murine and human samples. Our study provides the first single-cell transcriptional map defining the development of MDSCs, which will ultimately enable us to specifically target these cells in cancer patients.

Alexandre Coudray, Anna M Battenhouse, Philipp Bucher, Vishwanath R. Iyer

To detect functional somatic mutations in tumor samples, whole-exome sequencing (WES) is often used for its reliability and relative low cost. RNA-seq, while generally used to measure gene expression, can potentially also be used for identification of somatic mutations. However there has been little systematic evaluation of the utility of RNA-seq for identifying somatic mutations. Here, we develop and evaluate a pipeline for processing RNA-seq data from glioblastoma multiforme (GBM) tumors in order to identify somatic mutations. The pipeline entails the use of the STAR aligner 2-pass procedure jointly with MuTect2 from GATK to detect somatic variants. Variants identified from RNA-seq data were evaluated by comparison against the COSMIC and dbSNP databases, and also compared to somatic variants identified by exome sequencing. We also estimated the putative functional impact of coding variants in the most frequently mutated genes in GBM. Interestingly, variants identified by RNA-seq alone showed better representation of GBM-related mutations cataloged by COSMIC. RNA-seq-only data substantially outperformed the ability of WES to reveal potentially new somatic mutations in known GBM-related pathways, and allowed us to build a high-quality set of somatic mutations common to exome and RNA-seq calls. Using RNA-seq data in parallel with WES data to detect somatic mutations in cancer genomes can thus broaden the scope of discoveries and lend additional support to somatic variants identified by exome sequencing alone.

Zachary Steinhart, Traver Hart, Megha Chandrashekhar, Zvezdan Pavlovic, Mélanie Robitaille, Xiaowei Wang, Jarrett Adams, James Pan, Sachdev Sidhu, Jason Moffat, Stephane Angers

CRISPR-Cas9 genome editing enables high-resolution detection of genetic vulnerabilities of cancer cells. We conducted a genome-wide CRISPR-Cas9 screen in RNF43 mutant pancreatic ductal adenocarcinoma (PDAC) cells, which rely on Wnt signaling for proliferation, and discovered a unique requirement for a WNT7B-FZD5 signaling circuit. Our results highlight an underappreciated level of functional specificity at the ligand-receptor level. We derived a panel of recombinant antibodies that reports the expression of nine out of ten human Frizzled receptors and confirm that WNT7B-FZD5 functional specificity cannot be explained by protein expression patterns. We developed two human antibodies that target FZD5 and robustly inhibited the growth of RNF43 mutant PDAC cells grown in vitro and as xenografts, providing strong orthogonal support for the functional specificity observed genetically. Proliferation of a patient-derived PDAC cell line harboring a RNF43 variant previously associated with PDAC was also selectively inhibited by the FZD5 antibodies, further demonstrating their use as a potential targeted therapy.

Edmond M. Chan, Tsukasa Shibue, James McFarland, Benjamin Gaeta, Justine S. McPartlan, Mahmoud Ghandi, Jie Bin Liu, Jean-Bernard Lazaro, Nancy Dumont, Alfredo Gonzalez, Annie Apffel, Syed O Ali, Lisa Leung, Emma A. Roberts, Elizaveta Reznichenko, Mirazul Islam, Maria Alimova, Monica Schenone, Yosef Maruvka, Yang Liu, Alan D’Andrea, David E Root, Jesse S. Boehm, Gad Getz, Todd R. Golub, Aviad Tsherniak, Francisca Vazquez, Adam J. Bass

Synthetic lethality, an interaction whereby the co-occurrence of two or more genetic events lead to cell death but one event alone does not, can be exploited to develop novel cancer therapeutics. DNA repair processes represent attractive synthetic lethal targets since many cancers exhibit an impaired DNA repair pathway, which can lead these cells to become dependent on specific repair proteins. The success of poly (ADP-ribose) polymerase 1 (PARP-1) inhibitors in homologous recombination-deficient cancers highlights the potential of this approach in clinical oncology. Hypothesizing that other DNA repair defects would give rise to alternative synthetic lethal relationships, we asked if there are specific dependencies in cancers with microsatellite instability (MSI), which results from impaired DNA mismatch repair (MMR). Here we analyzed data from large-scale CRISPR/Cas9 knockout and RNA interference (RNAi) silencing screens and found that the RecQ DNA helicase was selectively essential in MSI cell lines, yet dispensable in microsatellite stable (MSS) cell lines. WRN depletion induced double-strand DNA breaks and promoted apoptosis and cell cycle arrest selectively in MSI models. MSI cancer models specifically required the helicase activity, but not the exonuclease activity of WRN. These findings expose WRN as a synthetic lethal vulnerability and promising drug target in MSI cancers.

Nathan Wan, David Weinberg, Tzu-Yu Liu, Katherine Niehaus, Daniel Delubac, Ajay Kannan, Brandon White, Eric A. Ariazi, Mitch Bailey, Marvin Bertin, Nathan Boley, Derek Bowen, James Cregg, Adam M. Drake, Riley Ennis, Signe Fransen, Erik Gafni, Loren Hansen, Yaping Liu, Gabriel L Otte, Jennifer Pecson, Brandon Rice, Gabriel E Sanderson, Aarushi Sharma, John St. John, Catherina Tang, Abraham Tzou, Leilani Young, Girish Putcha, Imran S Haque

Background: Blood-based methods using cell-free DNA (cfDNA) are under development as an alternative to existing screening tests. However, early-stage detection of cancer using tumor-derived cfDNA has proven challenging because of the small proportion of cfDNA derived from tumor tissue in early-stage disease. A machine learning approach to discover signatures in cfDNA, potentially reflective of both tumor and non-tumor contributions, may represent a promising direction for the early detection of cancer. Methods: Whole-genome sequencing was performed on cfDNA extracted from plasma samples (N=546 colorectal cancer and 271 non-cancer controls). Reads aligning to protein-coding gene bodies were extracted, and read counts were normalized. cfDNA tumor fraction was estimated using IchorCNA. Machine learning models were trained using k-fold cross-validation and confounder-based cross-validation to assess generalization performance. Results: In a colorectal cancer cohort heavily weighted towards early-stage cancer (80% stage I/II), we achieved a mean AUC of 0.92 (95% CI 0.91-0.93) with a mean sensitivity of 85% (95% CI 83-86%) at 85% specificity. Sensitivity generally increased with tumor stage and increasing tumor fraction. Stratification by age, sequencing batch, and institution demonstrated the impact of these confounders and provided a more accurate assessment of generalization performance. Conclusions: A machine learning approach using cfDNA achieved high sensitivity and specificity in a large, predominantly early-stage, colorectal cancer cohort. The possibility of systematic technical and institution-specific biases warrants similar confounder analyses in other studies. Prospective validation of this machine learning method and evaluation of a multi-analyte approach are underway.

Korsuk Sirinukunwattana, Enric Domingo, Susan Richman, Keara L Redmond, Andrew Blake, Clare Verrill, Simon J Leedham, Aikaterini Chatzipli, Claire Hardy, Celina Whalley, Chieh-Hsi Wu, Andrew D Beggs, Ultan McDermott, Philip Dunne, Angela A Meade, Steven M Walker, Graeme I Murray, Leslie M Samuel, Matthew Seymour, Ian Tomlinson, Philip Quirke, Tim Maughan, Jens Rittscher, Viktor H Koelzer, on behalf of S:CORT consortium

Image analysis is a cost-effective tool to associate complex features of tissue organisation with molecular and outcome data. Here we predict consensus molecular subtypes (CMS) of colorectal cancer (CRC) from standard H&E sections using deep learning. Domain adversarial training of a neural classification network was performed using 1,553 tissue sections with comprehensive multi-omic data from three independent datasets. Image-based consensus molecular subtyping (imCMS) accurately classified CRC whole-slide images and preoperative biopsies, spatially resolved intratumoural heterogeneity and provided accurate secondary calls with higher discriminatory power than bioinformatic prediction. In all three cohorts imCMS established sensible classification in CMS unclassified samples, reproduced expected correlations with (epi)genomic alterations and effectively stratified patients into prognostic subgroups. Leveraging artificial intelligence for the development of novel biomarkers extracted from histological slides with molecular and biological interpretability has remarkable potential for clinical translation.