Deep Learning Model for Tumor Type Prediction using Targeted Clinical Genomic Sequencing Data

Tumor type guides clinical treatment decisions in cancer, but histology-based diagnosis remains challenging. Genomic alterations are highly diagnostic of tumor type, and tumor type classifiers trained on genomic features have been explored, but the most accurate methods are not clinically feasible, relying on features derived from whole genome sequencing (WGS), or predicting across limited cancer types. We use genomic features from a dataset of 39,787 solid tumors sequenced using a clinical targeted cancer gene panel to develop Genome-Derived-Diagnosis Ensemble (GDD-ENS): a hyperparameter ensemble for classifying tumor type using deep neural networks. GDD-ENS achieves 93% accuracy for high-confidence predictions across 38 cancer types, rivalling performance of WGS-based methods. GDD-ENS can also guide diagnoses on rare type and cancers of unknown primary, and incorporate patient-specific clinical information for improved predictions. Overall, integrating GDD-ENS into prospective clinical sequencing workflows could provide clinically-relevant tumor type predictions to guide treatment decisions in real time.

Deep Learning Model for Tumor Type Prediction using Targeted Clinical Genomic Sequencing Data

Tumor type guides clinical treatment decisions in cancer, but histology-based diagnosis remains challenging. Genomic alterations are highly diagnostic of tumor type, and tumor type classifiers trained on genomic features have been explored, but the most accurate methods are not clinically feasible, relying on features derived from whole genome sequencing (WGS), or predicting across limited cancer types. We use genomic features from a dataset of 39,787 solid tumors sequenced using a clinical targeted cancer gene panel to develop Genome-Derived-Diagnosis Ensemble (GDD-ENS): a hyperparameter ensemble for classifying tumor type using deep neural networks. GDD-ENS achieves 93% accuracy for high-confidence predictions across 38 cancer types, rivalling performance of WGS-based methods. GDD-ENS can also guide diagnoses on rare type and cancers of unknown primary, and incorporate patient-specific clinical information for improved predictions. Overall, integrating GDD-ENS into prospective clinical sequencing workflows has enabled clinically-relevant tumor type predictions to guide treatment decisions in real time.

Abstract 5440: Deep-learning model for tumor type classification enables enhanced clinical decision support in cancer diagnosis

Background: Knowledge of a patient’s tumor type is essential for guiding clinical treatment decisions in cancer, but histologically-based diagnosis remains challenging for a subset of cancers. Genomic alterations are highly indicative of tumor type and can be used to build classifiers that predict diagnoses, but most genomic-based classification methods use whole genome sequencing (WGS) data which is not feasible for widespread clinical implementation at present. Clinical sequencing is typically performed using cancer gene panels that target individual mutations, often drivers, but previous tumor type classifiers developed using driver-based features alone perform poorly. We hypothesize that a classifier developed using state-of-the-art deep-learning methods and a sufficiently large training cohort would be able to overcome previous accuracy limitations and support the development of a clinically-relevant tumor type prediction model.

Methods: We present Deep Genome-Derived Diagnosis (GDD-ENS), an ensemble-based deep-learning tumor type classification method trained using data from cancer gene panel sequencing. We specifically use data from MSK-IMPACT, an FDA-authorized clinical sequencing assay that reports genomic alterations including mutations, indels, copy number alterations, and gene fusions across 505 cancer-associated genes. We aggregated a discovery cohort of 35,372 patients with solid tumors profiled with MSK-IMPACT across 38 common cancer types and used this set to generate 4,487 somatic mutation features for development.

Results: GDD-ENS achieves 78.8% accuracy on a held out validation cohort of 6971 patients. For the 71.9% of predictions assigned a high confidence by the model, accuracy increases to 92.7%, rivaling WGS-based models. We use Shapley Values to report prediction-specific feature importance, and aggregate them across cancer types to show GDD-ENS identifies known cancer type-genomic alteration trends. GDD-ENS also, with high accuracy, identifies patients with cancer types not included in the 38 common types using metrics derived from ensemble statistics. For patients where non-genomic information could further guide predictions, we implement a customizable prediction-specific adaptive prior distribution and report improved accuracy after adjusting predictions to account for features such as metastatic sample biopsy site. Finally, we apply GDD-ENS to a set of 1,123 patients with Cancers of Unknown Primary (CUP) and return high confidence predictions for 49% of these patients, in some cases matching predictions on CUP samples with diagnoses that were later confirmed through additional sampling and disease progression.

Conclusions: Integrating GDD-ENS into prospective clinical sequencing workflows will enable clinically-relevant tumor type predictions that can guide treatment decisions in real-time.

Citation Format: Madison Darmofal, Shalabh Suman, Gurnit Atwal, Jie-Fu Chen, Anna Varghese, Jason C. Chang, Anoop Balakrishnan Rema, Aijazuddin Syed, Quaid Morris, Michael Berger. Deep-learning model for tumor type classification enables enhanced clinical decision support in cancer diagnosis. [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 5440.

Abstract 5046: DeepTumour: Identify tumor origin from whole genome sequences

The DeepTumour algorithm predicts the tissue of origin of a tumor based on the pattern of passenger mutations identified by whole genome sequencing. "Passengers" are incidental mutations that accrue in the genome over time due to random mutational processes, and are functionally distinct from the "driver" mutations that are responsible for the cancer's malignant behavior. In adult cancers, passenger mutations typically outnumber drivers by a hundred or thousand-fold; critically, the vast majority of passengers arise in the normal cell lineage that precedes the malignant transformation event and hence reflects mutational processes existing in the cancer's precursor cell and its ancestors.

Passenger mutations are not uniformly distributed across the genome, but are concentrated in areas of the genome that have a locally high mutation rate. Mutation rates are highest at places in the genome where chromatin is tightly packed and less accessible to the DNA repair machinery. Each distinct cell type has a different pattern of chromatin packing due to epigenetic modifications. DeepTumour takes advantage of this to infer the chromatin state in the cell of origin from the distribution of passenger mutations in the tumor.

Another characteristic of passenger mutations is that the probability of a particular type of mutation occurring (e.g. replacement of C by T) depends on the mutational processes that were active in the cell of origin and its ancestors. Because certain cancers are associated with distinct mutational exposures (e.g. lung cancer and smoking), DeepTumour uses the tumor's distribution of passenger mutation type as well as position on the genome.

The DeepTumour algorithm itself is a fully connected, feed-forward neural network which we trained using 28 cohorts representing different tumor types from the Pan-Cancer Analysis of Whole Genomes project. When applied to independent sets of tumors, the algorithm is able to achieve an overall accuracy of 88% on primary tumors and 83% on metastatic tumors for distinguishing the 28 cancer types. Furthermore, DeepTumour provides estimates of the models uncertainty, allowing it to automatically detect rare cancer samples with an accuracy of 93%.

The DeepTumour algorithm is now available as a fast, convenient and secure web-based service at https://deeptumour.oicr.on.ca. It accepts uploads of VCF files containing somatic mutations from tumor whole genome sequencing, and returns a ranked list of tumor type matches and their relative probabilities. It can be used to provide leads when evaluating a cancer of uncertain primary, to assist in resolving diagnostic ambiguities, and as a research tool for understanding tumors of intermediate histology.

Citation Format: Lincoln David Stein, Wei Jiao, Gurnit Atwal, Quaid Morris. DeepTumour: Identify tumor origin from whole genome sequences [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5046.

A deep learning system accurately classifies primary and metastatic cancers using passenger mutation patterns

In cancer, the primary tumour's organ of origin and histopathology are the strongest determinants of its clinical behaviour, but in 3% of cases a patient presents with a metastatic tumour and no obvious primary. Here, as part of the ICGC/TCGA Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium, 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 produced by the PCAWG Consortium. Our classifier achieves an accuracy of 91% on held-out tumor samples and 88% and 83% 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 accuracy. Our results have clinical applicability, underscore how patterns of somatic passenger mutations encode the state of the cell of origin, and can inform future strategies to detect the source of circulating tumour DNA.

Pan-cancer analysis of whole genomes

Cancer is driven by genetic change, and the advent of massively parallel sequencing has enabled systematic documentation of this variation at the whole-genome scale1-3. Here we report the integrative analysis of 2,658 whole-cancer genomes and their matching normal tissues across 38 tumour types from the Pan-Cancer Analysis of Whole Genomes (PCAWG) Consortium of the International Cancer Genome Consortium (ICGC) and The Cancer Genome Atlas (TCGA). We describe the generation of the PCAWG resource, facilitated by international data sharing using compute clouds. On average, cancer genomes contained 4-5 driver mutations when combining coding and non-coding genomic elements; however, in around 5% of cases no drivers were identified, suggesting that cancer driver discovery is not yet complete. Chromothripsis, in which many clustered structural variants arise in a single catastrophic event, is frequently an early event in tumour evolution; in acral melanoma, for example, these events precede most somatic point mutations and affect several cancer-associated genes simultaneously. Cancers with abnormal telomere maintenance often originate from tissues with low replicative activity and show several mechanisms of preventing telomere attrition to critical levels. Common and rare germline variants affect patterns of somatic mutation, including point mutations, structural variants and somatic retrotransposition. A collection of papers from the PCAWG Consortium describes non-coding mutations that drive cancer beyond those in the TERT promoter4; identifies new signatures of mutational processes that cause base substitutions, small insertions and deletions and structural variation5,6; analyses timings and patterns of tumour evolution7; describes the diverse transcriptional consequences of somatic mutation on splicing, expression levels, fusion genes and promoter activity8,9; and evaluates a range of more-specialized features of cancer genomes8,10-18.

Genome Gerrymandering: optimal division of the genome into regions with cancer type specific differences in mutation rates

The activity of mutational processes differs across the genome, and is influenced by chromatin state and spatial genome organization. At the scale of one megabase-pair (Mb), regional mutation density correlate strongly with chromatin features and mutation density at this scale can be used to accurately identify cancer type. Here, we explore the relationship between genomic region and mutation rate by developing an information theory driven, dynamic programming algorithm for dividing the genome into regions with differing relative mutation rates between cancer types. Our algorithm improves mutual information when compared to the naive approach, effectively reducing the average number of mutations required to identify cancer type. Our approach provides an efficient method for associating regional mutation density with mutation labels, and has future applications in exploring the role of somatic mutations in a number of diseases.

ASCL1 Reorganizes Chromatin to Direct Neuronal Fate and Suppress Tumorigenicity of Glioblastoma Stem Cells

Glioblastomas exhibit a hierarchical cellular organization, suggesting that they are driven by neoplastic stem cells that retain partial yet abnormal differentiation potential. Here, we show that a large subset of patient-derived glioblastoma stem cells (GSCs) express high levels of Achaete-scute homolog 1 (ASCL1), a proneural transcription factor involved in normal neurogenesis. ASCL1^hi GSCs exhibit a latent capacity for terminal neuronal differentiation in response to inhibition of Notch signaling, whereas ASCL1^lo GSCs do not. Increasing ASCL1 levels in ASCL1^lo GSCs restores neuronal lineage potential, promotes terminal differentiation, and attenuates tumorigenicity. ASCL1 mediates these effects by functioning as a pioneer factor at closed chromatin, opening new sites to activate a neurogenic gene expression program. Directing GSCs toward terminal differentiation may provide therapeutic applications for a subset of GBM patients and strongly supports efforts to restore differentiation potential in GBM and other cancers.