Search

CUP is a relatively common diagnosis that accounts for 3-9% of all cancers. The prognosis is poor with median survival of approximately 9 months. The identification of the primary tumor and therapy targets could improve the survival of these patients. We employed the BostonGene Tumor PortraitTM platform to interrogate 19 CUP cases. Trained on >19,000 samples and validated on 28,000 samples from independent datasets, the machine-learning based algorithm integrates whole-exome and RNA sequencing (WES and RNAseq) analysis to characterize cancer drivers, the tumor microenvironment, potential targets, tumor composition, molecular signatures, and site of origin (94% sensitivity and 99% specificity). The predicted tissue of origin was considered acceptable when it was compatible with tumor histopathology and immunoprofile, and included in the differential diagnosis of radiologic studies. The discovered biomarkers and possible treatments were discussed during our multidisciplinary precision oncology meeting. 16 out of 19 CUP cases (84%) had an acceptable predicted tissue of origin. Two cases lacked clinical evidence to support the predicted primary tumor, and 1 case failed RNAseq. Lung accounted for most of the sites of origin (31%) followed by gastrointestinal (15%) and breast cancers (8%). Other diagnoses included melanoma, uterine, bladder, and renal carcinomas, among others. Except for one case, a clinically significant biomarker or target was found. Those included relevant mutational signatures (e.g., homologous recombination deficiency, DNA mismatch repair), genetic characteristics (high tumor mutational burden (TMB), microsatellite instability), activating alterations (FGFR1, MYC, ERBB2 amplifications; NCOA4::RET fusion), loss of function in tumor suppressor genes (TP53, FANCA, ATM), and gene overexpression (ER). Further, microenvironment analysis characterized the tumor immune infiltrate and the level of RNA expression of PD-L1, PD-L2, and CTLA4. These oncology biomarkers and potential targets are of significant value independent of tissue of origin. In addition, mutations in NF1, KRAS, TP53, MSH2, BRCA1, and PTEN were found and validated by commercially available targeted NGS panels. Based on CUP tumor profiling by this platform, positive treatment response has been observed in 3 out of 4 CUP patients thus far, e.g., one patient with metastatic disease that showed high TMB and immune infiltrated microenvironment treated with Ipilimumab and Nivolumab had a sustained response. Therapy in oncology is often determined by the tissue origin, making CUP a therapeutic challenge. In this study, we demonstrate the application of an integrative WES and RNAseq platform to not only predict the site of origin, but also to identify relevant biomarkers and therapeutic targets in CUP. Citation Format: Majd Al Assaad, Michael Sigouros, Jyothi Manohar, Daniela Guevara, Zoia Antysheva, Daria Kiriy, Alexandra Boyko, Naira Samarina, Nara Shin, Viktor Svekolkin, Svetlana Podsvirova, Noel English, Alaina Villarreal, Brianna McKenna, Cagdas Tazearslan, Diana Shamsutdinova, Vladimir Kushnarev, Troy Kane, David Wilkes, Manish Shah, Barbara Ma, Scott T. Tagawa, David Nanus, Jones Nauseef, Olivier Elemento, Juan Miguel Mosquera, Cora N. Sternberg. Cancer of unknown Primary (CUP): Beyond the identification of the site of origin by an integrative genomic approach [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 2143.

The ability to locate a tumor primary site for patients with a cancer of unknown primary (CUP) is a major obstacle in providing personalized therapeutic options and access to clinical trials. Despite the recent use of molecular-based tools to identify the tumor tissue-of-origin (TOO), overall survival for CUP patients remains low. Here, we present an AI-based tool that predicts the TOO by using genomic and transcriptomic data to classify CUP tumors into hierarchically-organized molecular subgroups. The TOO predictor was composed of DNA, RNA, and Consensus classifiers that were hierarchically organized with respect to molecular diagnosis with upper level clusters based on common molecular features reflecting similar cell of origin, and lower levels containing specific diagnoses for further classification. The ML-based DNA classifier was trained on a dataset of publicly available genomic data generated from 8,000 samples, and independently validated using more than 5,500 samples. The ML-based RNA classifier was trained on a dataset of publicly available transcriptomic data created from more than 10,100 samples with tumor- and normal-specific features for each cancer type, and independently validated using 20,000 samples. The Consensus classifier, combining outputs from both DNA and RNA algorithms, was trained on a dataset of genomic and transcriptomic data from 1,000 samples, and validated on an independent dataset of 2,000 samples. Each classifier contained features selected based on data analysis according to the weighted F1-score, and the best hyperparameters for the final model. The 3-classifier algorithm predicts TOO for 33 cancer types and subtypes belonging to solid neoplasms, independently of sample source, sequencing methods, and cohort. Validation of the Consensus classifier showed a higher accuracy (95% f1-score) compared to the DNA and RNA classifiers (79% and 93% f1-scores, respectively), along with a high sensitivity (95%), specificity (99%), and precision (96%), as it takes into account both genomic events and expression patterns. The TOO predictor was prospectively validated on approximately 298 clinical samples with a known diagnosis using all classifiers. The diagnosis was identified in 90% of clinical cases (295/297) by the Consensus classifier with 90% sensitivity and 94% precision. The call rate for the DNA and RNA classifiers was above 95%. Of note, sensitivity of the top 4 predicted diagnoses was > 90% for all 3 classifiers, and the calculated rule-out accuracy of the Consensus classifier was 97%. In conclusion, an ML-based algorithm was developed that utilizes genomic and transcriptomic data to accurately predict the TOO for CUP tumors. Utilizing the Consensus classifier after DNA and RNA classifiers helps to identify the TOO of the tumor with high specificity, which can guide precision oncology therapeutic options. Citation Format: Zoia Antysheva, Daria Kiriy, Anton Sivkov, Alexander Sarachackov, Alexandra Boyko, Naira Samarina, Nara Shin, Jessica H. Brown, Ivan Kozlov, Viktor Svekolkin, Alexander Bagaev, Nathan Fowler, Nikita Kotlov. An ML-based tool for predicting tissue of origin for cancer of unknown primary (CUP) based on genomic and transcriptomic data. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5405.

With the wealth of next-generation sequencing (NGS) produced over the last decade, distinguishing between pathogenic and neutral single nucleotide missense mutations is essential to understanding disease pathogenesis and cancer treatment selection and optimization. Current experimental methods, such as overexpression assays, siRNA, knock-out models, and studies in cultured cells or model organisms, can help uncover the phenotype of a set of mutations in genes but are not optimal for the identification of pathogenic mutations in thousands of potential candidates. In this study, we present a new mutation annotation tool, MutAnt, for the prediction of mutation pathogenicity that achieves superior performance by integrating state-of-the-art machine learning techniques with feature selection and hyperparameter tuning. To develop MutAnt, a supervised machine learning algorithm was trained to classify pathogenic and neutral single nonsynonymous nucleotide substitutions. The training dataset was based on ClinVar v20190102, the CancerGenomeInterpreter database of validated pathogenic mutations, and benign mutations from the database of Cancer Passenger Mutations (dbCPM). The Boruta algorithm with Shapley values was used for feature selection, and the Bayesian optimization was applied for the model hyperparameter tuning. Diverse validation methods were utilized in the development of MutAnt, such as cross-validation; the Database of Curated Mutations (DoCM); holdout on new mutations documented in the latest released version of ClinVar database, which contains validated disease-associated single nucleotide polymorphisms (SNPs) not previously annotated in existing algorithms; mutations with high allele frequency; test dataset III from VariBench; and correlation with the mean function score in a saturation genome editing experiment for the BRCA gene. MutAnt exhibited a stronger performance than other methods on all of the tested validation strategies, which was demonstrated by its accuracy (high f1-score, 0.95 to 0.99) and sensitivity-specificity value (ROC-AUC value, 0.97 to 0.99) in classifying mutations in each validation dataset. Furthermore, the present approach allowed for the prediction of novel mutations that are not covered by extensive population studies with higher accuracy than other methods. This was demonstrated by its prediction of the clinically relevant BRCA variants in the saturation genome editing experiment. In conclusion, MutAnt outperformed all compared mutation annotation algorithms in this study for the classification of clinically relevant benign and pathogenic mutations in the ClinVar dataset. Importantly, this approach allows prediction of novel mutations, subsequently allowing for the selection of targeted therapies for a broader set of patients. Citation Format: Aleksandr Sarachakov, Viktor Svekolkin, Zoia Antysheva, Jessica H. Brown, Alexander Bagaev, Nathan Fowler. MutAnt: Mutation annotation machine learning algorithm for pathogenicity evaluation of single nonsynonymous nucleotide substitutions in cancer cells [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 192.