Performance of somatic structural variant calling in lung cancer using Oxford Nanopore sequencing technology.

BACKGROUND: Lung cancer is a heterogeneous disease and the primary cause of cancer-related mortality worldwide. Somatic mutations, including large structural variants, are important biomarkers in lung cancer for selecting targeted therapy. Genomic studies in lung cancer have been conducted using short-read sequencing. Emerging long-read sequencing technologies are a promising alternative to study somatic structural variants, however there is no current consensus on how to process data and call somatic events. In this study, we preformed whole genome sequencing of lung cancer and matched non-tumour samples using long and short read sequencing to comprehensively benchmark three sequence aligners and seven structural variant callers comprised of generic callers (SVIM, Sniffles2, DELLY in generic mode and cuteSV) and somatic callers (Severus, SAVANA, nanomonsv and DELLY in somatic modes). RESULTS: Different combinations of aligners and variant callers influenced somatic structural variant detection. The choice of caller had a significant influence on somatic structural variant detection in terms of variant type, size, sensitivity, and accuracy. The performance of each variant caller was assessed by comparing to somatic structural variants identified by short-read sequencing. When compared to somatic structural variants detected with short-read sequencing, more events were detected with long-read sequencing. The mean recall of somatic variant events identified by long-read sequencing was higher for the somatic callers (72%) than generic callers (53%). Among the somatic callers when using the minimap2 aligner, SAVANA and Severus achieved the highest recall at 79.5% and 79.25% respectively, followed by nanomonsv with a recall of 72.5%. CONCLUSION: Long-read sequencing can identify somatic structural variants in clincal samples. The longer reads have the potential to improve our understanding of cancer development and inform personalized cancer treatment.

Evaluation of Endobronchial Ultrasound-Guided Transbronchial Needle Aspiration (EBUS-TBNA) Samples from Advanced Non-Small Cell Lung Cancer for Whole Genome, Whole Exome and Comprehensive Panel Sequencing.

Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) is often the only source of tumor tissue from patients with advanced, inoperable lung cancer. EBUS-TBNA aspirates are used for the diagnosis, staging, and genomic testing to inform therapy options. Here we extracted DNA and RNA from 220 EBUS-TBNA aspirates to evaluate their suitability for whole genome (WGS), whole exome (WES), and comprehensive panel sequencing. For a subset of 40 cases, the same nucleic acid extraction was sequenced using WGS, WES, and the TruSight Oncology 500 assay. Genomic features were compared between sequencing platforms and compared with those reported by clinical testing. A total of 204 aspirates (92.7%) had sufficient DNA (100 ng) for comprehensive panel sequencing, and 109 aspirates (49.5%) had sufficient material for WGS. Comprehensive sequencing platforms detected all seven clinically reported tier 1 actionable mutations, an additional three (7%) tier 1 mutations, six (15%) tier 2-3 mutations, and biomarkers of potential immunotherapy benefit (tumor mutation burden and microsatellite instability). As expected, WGS was more suited for the detection and discovery of emerging novel biomarkers of treatment response. WGS could be performed in half of all EBUS-TBNA aspirates, which points to the enormous potential of EBUS-TBNA as source material for large, well-curated discovery-based studies for novel and more effective predictors of treatment response. Comprehensive panel sequencing is possible in the vast majority of fresh EBUS-TBNA aspirates and enhances the detection of actionable mutations over current clinical testing.

Discrepancies in tumor mutation burden reporting from sequential endobronchial ultrasound transbronchial needle aspiration samples within single lymph node stations - brief report.

Introduction: Tumour Mutation Burden (TMB) is a potential biomarker for immune cancer therapies. Here we investigated parameters that might affect TMB using duplicate cytology smears obtained from endobronchial ultrasound transbronchial needle aspiration (EBUS TBNA)-sampled malignant lymph nodes. Methods: Individual Diff-Quik cytology smears were prepared for each needle pass. DNA extracted from each smear underwent sequencing using large gene panel (TruSight Oncology 500 (TSO500 - Illumina)). TMB was estimated using the TSO500 Local App v. 2.0 (Illumina). Results: Twenty patients had two or more Diff-Quik smears (total 45 smears) which passed sequencing quality control. Average smear TMB was 8.7 ± 5.0 mutations per megabase (Mb). Sixteen of the 20 patients had paired samples with minimal differences in TMB score (average difference 1.3 ± 0.85). Paired samples from 13 patients had concordant TMB (scores below or above a threshold of 10 mutations/Mb). Markedly discrepant TMB was observed in four cases, with an average difference of 11.3 ± 2.7 mutations/Mb. Factors affecting TMB calling included sample tumour content, the amount of DNA used in sequencing, and bone fide heterogeneity of node tumour between paired samples. Conclusion: TMB assessment is feasible from EBUS-TBNA smears from a single needle pass. Repeated samples of a lymph node station have minimal variation in TMB in most cases. However, this novel data shows how tumour content and minor change in site of node sampling can impact TMB. Further study is needed on whether all node aspirates should be combined in 1 sample, or whether testing independent nodes using smears is needed.

Whole Genome Sequencing in Advanced Lung Cancer can be Performed Using Diff-Quik Cytology Smears Derived from Endobronchial Ultrasound, Transbronchial Needle Aspiration (EBUS TBNA).

INTRODUCTION: Maximising alternative sample types for genomics in advanced lung cancer is important because bronchoscopic samples may sometimes be insufficient for this purpose. Further, the clinical applications of comprehensive molecular analysis such as whole genome sequencing (WGS) are rapidly developing. Diff-Quik cytology smears from EBUS TBNA is an alternative source of DNA, but its feasibility for WGS has not been previously demonstrated. METHODS: Diff-Quik smears were collected along with research cell pellets. RESULTS: Tumour content of smears were compared to research cell pellets from 42 patients, which showed good correlation (Spearman correlation 0.85, P < 0.0001). A subset of eight smears underwent WGS, which presented similar mutation profiles to WGS of the matched cell pellet. DNA yield was predicted using a regression equation of the smears cytology features, which correctly predicted DNA yield > 1500 ng in 7 out of 8 smears. CONCLUSIONS: WGS of commonly collected Diff-Quik slides is feasible and their DNA yield can be predicted.

Multi-omic features of oesophageal adenocarcinoma in patients treated with preoperative neoadjuvant therapy.

Oesophageal adenocarcinoma is a poor prognosis cancer and the molecular features underpinning response to treatment remain unclear. We investigate whole genome, transcriptomic and methylation data from 115 oesophageal adenocarcinoma patients mostly from the DOCTOR phase II clinical trial (Australian New Zealand Clinical Trials Registry-ACTRN12609000665235), with exploratory analysis pre-specified in the study protocol of the trial. We report genomic features associated with poorer overall survival, such as the APOBEC mutational and RS3-like rearrangement signatures. We also show that positron emission tomography non-responders have more sub-clonal genomic copy number alterations. Transcriptomic analysis categorises patients into four immune clusters correlated with survival. The immune suppressed cluster is associated with worse survival, enriched with myeloid-derived cells, and an epithelial-mesenchymal transition signature. The immune hot cluster is associated with better survival, enriched with lymphocytes, myeloid-derived cells, and an immune signature including CCL5, CD8A, and NKG7. The immune clusters highlight patients who may respond to immunotherapy and thus may guide future clinical trials.

Evaluating Diff-Quik cytology smears for large-panel mutation testing in lung cancer-Predicting DNA content and success with low-malignant-cellularity samples.

BACKGROUND: Cytology smears are commonly collected during endobronchial ultrasound-guided transbronchial needle aspiration (EBUS TBNA) procedures but are rarely used for molecular testing. Studies are needed to demonstrate their great potential, in particular for the prediction of malignant cell DNA content and for utility in molecular diagnostics using large gene panels. METHODS: A prospective study was performed on samples from 66 patients with malignant lymph nodes who underwent EBUS TBNA. All patients had air-dried, Diff-Quik cytology smears and formalin-fixed, paraffin-embedded cell blocks collected for cytopathology and molecular testing. One hundred eighty-five smears were evaluated by microscopy to estimate malignant cell percentage and abundance and to calculate smear size and were subjected to DNA extraction. DNA from 56 smears from 27 patients was sequenced with the TruSight Oncology 500 assay (Illumina). RESULTS: Each microscopy parameter had a significant effect on the DNA yield. An algorithm was developed that predicted a >50-ng DNA yield of a smear with an area under the curve of 0.86. Fifty DNA samples (89%) with varying malignant yields were successfully sequenced. Low-malignant-cell content (<25%) and smear area (<15%) were the main reasons for failure. All standard-of-care mutations were detected in replicate smears from individual patients, regardless of malignant cell content. Tier 1/2 mutations were discovered in two cases where standard-of-care specimens were inadequate for sequencing. Smears were scored for tumor mutation burden. CONCLUSIONS: Microscopy of Diff-Quik smears can triage samples for comprehensive panel sequencing, which highlights smears as an excellent alternative to traditional testing with cell blocks.

Australian Genomics: Outcomes of a 5-year national program to accelerate the integration of genomics in healthcare.

Australian Genomics is a national collaborative partnership of more than 100 organizations piloting a whole-of-system approach to integrating genomics into healthcare, based on federation principles. In the first five years of operation, Australian Genomics has evaluated the outcomes of genomic testing in more than 5,200 individuals across 19 rare disease and cancer flagship studies. Comprehensive analyses of the health economic, policy, ethical, legal, implementation and workforce implications of incorporating genomics in the Australian context have informed evidence-based change in policy and practice, resulting in national government funding and equity of access for a range of genomic tests. Simultaneously, Australian Genomics has built national skills, infrastructure, policy, and data resources to enable effective data sharing to drive discovery research and support improvements in clinical genomic delivery.

Shariant platform: Enabling evidence sharing across Australian clinical genetic-testing laboratories to support variant interpretation.

Sharing genomic variant interpretations across laboratories promotes consistency in variant assertions. A landscape analysis of Australian clinical genetic-testing laboratories in 2017 identified that, despite the national-accreditation-body recommendations encouraging laboratories to submit genotypic data to clinical databases, fewer than 300 variants had been shared to the ClinVar public database. Consultations with Australian laboratories identified resource constraints limiting routine application of manual processes, consent issues, and differences in interpretation systems as barriers to sharing. This information was used to define key needs and solutions required to enable national sharing of variant interpretations. The Shariant platform, using both the GRCh37 and GRCh38 genome builds, was developed to enable ongoing sharing of variant interpretations and associated evidence between Australian clinical genetic-testing laboratories. Where possible, two-way automated sharing was implemented so that disruption to laboratory workflows would be minimized. Terms of use were developed through consultation and currently restrict access to Australian clinical genetic-testing laboratories. Shariant was designed to store and compare structured evidence, to promote and record resolution of inter-laboratory classification discrepancies, and to streamline the submission of variant assertions to ClinVar. As of December 2021, more than 14,000 largely prospectively curated variant records from 11 participating laboratories have been shared. Discrepant classifications have been identified for 11% (28/260) of variants submitted by more than one laboratory. We have demonstrated that co-design with clinical laboratories is vital to developing and implementing a national variant-interpretation sharing effort. This approach has improved inter-laboratory concordance and enabled opportunities to standardize interpretation practices.

Prospective Optimization of Endobronchial Ultrasound-Guided Transbronchial Needle Aspiration Lymph Node Assessment for Lung Cancer: Three Needle Agitations Are Noninferior to 10 Agitations for Adequate Tumor Cell and DNA Yield.

Introduction: Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS TBNA) is an important means of obtaining a tissue for advanced lung cancer. Optimizing the EBUS TBNA needling technique is important to maintain procedural simplicity and maximize sample quality for emerging molecular diagnostics. Methods: We prospectively explored three versus 10 agitations of the needle in sequential passes into the lymph node using separate needles. Resulting Diff-Quik cytology smears were quantitatively assessed using microscopic (tumor cell cellularity, abundance scores, erythrocyte contamination) and DNA yields. Microscopy was reported by two cytopathologists, and an inter-rater assessment was made by four additional cytopathologists. Results: In 86 patients confirmed as having malignant disease by EBUS TBNA (45 males, 41 females), a mean of 5.3 smears were made per patient with a total of 459 smears scored by pathologists and 168 paired smears extracted for DNA. There was no significant difference between three versus 10 agitations for smear cellularity (p = 0.44), DNA yield (p = 0.84), or DNA integrity (p = 0.20), but there was significantly less contamination by erythrocytes from three agitations (chi-square p = 0.008). There was significantly more DNA in the first pass into the node using three agitations than with other passes and with 10 agitations (pass × agitations interaction, p = 0.031). Reviewing pathologists correctly classified smears as more than or equal to 25% cellularity 86.3% of the time (κ = 0.63 [95% confidence interval: 0.55-0.71]). Conclusions: Three agitations are noninferior to 10 agitations for overall abundance of malignant cells and DNA content on smears. A smear with adequate DNA for panel sequencing could almost always be made with the first needle pass using three agitations.

Comparative Genomics Provides Etiologic and Biological Insight into Melanoma Subtypes.

Melanoma is a cancer of melanocytes, with multiple subtypes based on body site location. Cutaneous melanoma is associated with skin exposed to ultraviolet radiation; uveal melanoma occurs in the eyes; mucosal melanoma occurs in internal mucous membranes; and acral melanoma occurs on the palms, soles, and nail beds. Here, we present the largest whole-genome sequencing study of melanoma to date, with 570 tumors profiled, as well as methylation and RNA sequencing for subsets of tumors. Uveal melanoma is genomically distinct from other melanoma subtypes, harboring the lowest tumor mutation burden and with significantly mutated genes in the G-protein signaling pathway. Most cutaneous, acral, and mucosal melanomas share alterations in components of the MAPK, PI3K, p53, p16, and telomere pathways. However, the mechanism by which these pathways are activated or inactivated varies between melanoma subtypes. Additionally, we identify potential novel germline predisposition genes for some of the less common melanoma subtypes. SIGNIFICANCE: This is the largest whole-genome analysis of melanoma to date, comprehensively comparing the genomics of the four major melanoma subtypes. This study highlights both similarities and differences between the subtypes, providing insights into the etiology and biology of melanoma. This article is highlighted in the In This Issue feature, p. 2711.

Comprehensive genomic and tumour immune profiling reveals potential therapeutic targets in malignant pleural mesothelioma.

BACKGROUND: Malignant pleural mesothelioma (MPM) has a poor overall survival with few treatment options. Whole genome sequencing (WGS) combined with the immune features of MPM offers the prospect of identifying changes that could inform future clinical trials. METHODS: We analysed somatic mutations from 229 MPM samples, including previously published data and 58 samples that had undergone WGS within this study. This was combined with RNA-seq analysis to characterize the tumour immune environment. RESULTS: The comprehensive genome analysis identified 12 driver genes, including new candidate genes. Whole genome doubling was a frequent event that correlated with shorter survival. Mutational signature analysis revealed SBS5/40 were dominant in 93% of samples, and defects in homologous recombination repair were infrequent in our cohort. The tumour immune environment contained high M2 macrophage infiltrate linked with MMP2, MMP14, TGFB1 and CCL2 expression, representing an immune suppressive environment. The expression of TGFB1 was associated with overall survival. A small subset of samples (less than 10%) had a higher proportion of CD8 T cells and a high cytolytic score, suggesting a 'hot' immune environment independent of the somatic mutations. CONCLUSIONS: We propose accounting for genomic and immune microenvironment status may influence therapeutic planning in the future.

Whole genome deep sequencing analysis of cell-free DNA in samples with low tumour content.

BACKGROUND: Circulating cell-free DNA (cfDNA) in the plasma of cancer patients contains cell-free tumour DNA (ctDNA) derived from tumour cells and it has been widely recognized as a non-invasive source of tumour DNA for diagnosis and prognosis of cancer. Molecular profiling of ctDNA is often performed using targeted sequencing or low-coverage whole genome sequencing (WGS) to identify tumour specific somatic mutations or somatic copy number aberrations (sCNAs). However, these approaches cannot efficiently detect all tumour-derived genomic changes in ctDNA. METHODS: We performed WGS analysis of cfDNA from 4 breast cancer patients and 2 patients with benign tumours. We sequenced matched germline DNA for all 6 patients and tumour samples from the breast cancer patients. All samples were sequenced on Illumina HiSeqXTen sequencing platform and achieved approximately 30x, 60x and 100x coverage on germline, tumour and plasma DNA samples, respectively. RESULTS: The mutational burden of the plasma samples (1.44 somatic mutations/Mb of genome) was higher than the matched tumour samples. However, 90% of high confidence somatic cfDNA variants were not detected in matched tumour samples and were found to comprise two background plasma mutational signatures. In contrast, cfDNA from the di-nucleosome fraction (300 bp-350 bp) had much higher proportion (30%) of variants shared with tumour. Despite high coverage sequencing we were unable to detect sCNAs in plasma samples. CONCLUSIONS: Deep sequencing analysis of plasma samples revealed higher fraction of unique somatic mutations in plasma samples, which were not detected in matched tumour samples. Sequencing of di-nucleosome bound cfDNA fragments may increase recovery of tumour mutations from plasma.

Patient-derived xenograft models capture genomic heterogeneity in endometrial cancer.

BACKGROUND: Endometrial cancer (EC) is a major gynecological cancer with increasing incidence. It comprises four molecular subtypes with differing etiology, prognoses, and responses to chemotherapy. In the future, clinical trials testing new single agents or combination therapies will be targeted to the molecular subtype most likely to respond. As pre-clinical models that faithfully represent the molecular subtypes of EC are urgently needed, we sought to develop and characterize a panel of novel EC patient-derived xenograft (PDX) models. METHODS: Here, we report whole exome or whole genome sequencing of 11 PDX models and their matched primary tumor. Analysis of multiple PDX lineages and passages was performed to study tumor heterogeneity across lineages and/or passages. Based on recent reports of frequent defects in the homologous recombination (HR) pathway in EC, we assessed mutational signatures and HR deficiency scores and correlated these with in vivo responses to the PARP inhibitor (PARPi) talazoparib in six PDXs representing the copy number high/p53-mutant and mismatch-repair deficient molecular subtypes of EC. RESULTS: PDX models were successfully generated from grade 2/3 tumors, including three uterine carcinosarcomas. The models showed similar histomorphology to the primary tumors and represented all four molecular subtypes of EC, including five mismatch-repair deficient models. The different PDX lineages showed a wide range of inter-tumor and intra-tumor heterogeneity. However, for most PDX models, one arm recapitulated the molecular landscape of the primary tumor without major genomic drift. An in vivo response to talazoparib was detected in four copy number high models. Two models (carcinosarcomas) showed a response consistent with stable disease and two models (one copy number high serous EC and another carcinosarcoma) showed significant tumor growth inhibition, albeit one consistent with progressive disease; however, all lacked the HR deficiency genomic signature. CONCLUSIONS: EC PDX models represent the four molecular subtypes of disease and can capture intra-tumor heterogeneity of the original primary tumor. PDXs of the copy number high molecular subtype showed sensitivity to PARPi; however, deeper and more durable responses will likely require combination of PARPi with other agents.

Multiomic profiling of checkpoint inhibitor-treated melanoma: Identifying predictors of response and resistance, and markers of biological discordance.

We concurrently examine the whole genome, transcriptome, methylome, and immune cell infiltrates in baseline tumors from 77 patients with advanced cutaneous melanoma treated with anti-PD-1 with or without anti-CTLA-4. We show that high tumor mutation burden (TMB), neoantigen load, expression of IFNγ-related genes, programmed death ligand expression, low PSMB8 methylation (therefore high expression), and T cells in the tumor microenvironment are associated with response to immunotherapy. No specific mutation correlates with therapy response. A multivariable model combining the TMB and IFNγ-related gene expression robustly predicts response (89% sensitivity, 53% specificity, area under the curve [AUC], 0.84); tumors with high TMB and a high IFNγ signature show the best response to immunotherapy. This model validates in an independent cohort (80% sensitivity, 59% specificity, AUC, 0.79). Except for a JAK3 loss-of-function mutation, for patients who did not respond as predicted there is no obvious biological mechanism that clearly explained their outlier status, consistent with intratumor and intertumor heterogeneity in response to immunotherapy.

qmotif: determination of telomere content from whole-genome sequence data.

Motivation: Changes in telomere length have been observed in cancer and can be indicative of mechanisms involved in carcinogenesis. Most methods used to estimate telomere length require laboratory analysis of DNA samples. Here, we present qmotif, a fast and easy tool that determines telomeric repeat sequences content as an estimate of telomere length directly from whole-genome sequencing. Results: qmotif shows similar results to quantitative PCR, the standard method for high-throughput clinical telomere length quantification. qmotif output correlates strongly with the output of other tools for determining telomere sequence content, TelSeq and TelomereHunter, but can run in a fraction of the time-usually under a minute. Availability and implementation: qmotif is implemented in Java and source code is available at https://github.com/AdamaJava/adamajava, with instructions on how to build and use the application available from https://adamajava.readthedocs.io/en/latest/. Supplementary information: Supplementary data are available at Bioinformatics Advances online.

Correction: McCart Reed et al. The Genomic Landscape of Lobular Breast Cancer. Cancers 2021, 13, 1950.

The authors wish to make the following corrections to this paper [...].

Acquired RAD51C Promoter Methylation Loss Causes PARP Inhibitor Resistance in High-Grade Serous Ovarian Carcinoma.

In high-grade serous ovarian carcinoma (HGSC), deleterious mutations in DNA repair gene RAD51C are established drivers of defective homologous recombination and are emerging biomarkers of PARP inhibitor (PARPi) sensitivity. RAD51C promoter methylation (meRAD51C) is detected at similar frequencies to mutations, yet its effects on PARPi responses remain unresolved.In this study, three HGSC patient-derived xenograft (PDX) models with methylation at most or all examined CpG sites in the RAD51C promoter show responses to PARPi. Both complete and heterogeneous methylation patterns were associated with RAD51C gene silencing and homologous recombination deficiency (HRD). PDX models lost meRAD51C following treatment with PARPi rucaparib or niraparib, where a single unmethylated copy of RAD51C was sufficient to drive PARPi resistance. Genomic copy number profiling of one of the PDX models using SNP arrays revealed that this resistance was acquired independently in two genetically distinct lineages.In a cohort of 12 patients with RAD51C-methylated HGSC, various patterns of meRAD51C were associated with genomic "scarring," indicative of HRD history, but exhibited no clear correlations with clinical outcome. Differences in methylation stability under treatment pressure were also observed between patients, where one HGSC was found to maintain meRAD51C after six lines of therapy (four platinum-based), whereas another HGSC sample was found to have heterozygous meRAD51C and elevated RAD51C gene expression (relative to homozygous meRAD51C controls) after only neoadjuvant chemotherapy.As meRAD51C loss in a single gene copy was sufficient to cause PARPi resistance in PDX, methylation zygosity should be carefully assessed in previously treated patients when considering PARPi therapy. SIGNIFICANCE: Homozygous RAD51C methylation is a positive predictive biomarker for sensitivity to PARP inhibitors, whereas a single unmethylated gene copy is sufficient to confer resistance.

The Genomic Landscape of Lobular Breast Cancer.

Invasive lobular carcinoma (ILC) is the second most common breast cancer histologic subtype, accounting for approximately 15% of all breast cancers. It is only recently that its unique biology has been assessed in high resolution. Here, we present a meta-analysis of ILC sequencing datasets, to provide a long-awaited ILC-specific resource, and to confirm the prognostic value and strength of association between a number of clinico-pathology features and genomics in this special tumour type. We consider panel (n = 684), whole exome (n = 215) and whole genome sequencing data (n = 48), and review histology of The Cancer Genome Atlas cases to assign grades and determine whether the ILC is of classic type or a variant, such as pleomorphic, prior to performing statistical analyses. We demonstrate evidence of considerable genomic heterogeneity underlying a broadly homogeneous tumour type (typically grade 2, estrogen receptor (ER)-positive); with genomes exhibiting few somatic mutations or structural alterations, genomes with a hypermutator phenotype, and tumours with highly rearranged genomes. We show that while CDH1 (E-cadherin) and PIK3CA mutations do not significantly impact survival, overall survival is significantly poorer for patients with a higher tumour mutation burden; this is also true for grade 3 tumours, and those carrying a somatic TP53 mutation (and these cases were more likely to be ER-negative). Taken together, we have compiled a meta-dataset of ILC with molecular profiling, and our analyses show that the genomic landscape significantly impacts the tumour's variable natural history and overall survival of ILC patients.

DNA methylation patterns identify subgroups of pancreatic neuroendocrine tumors with clinical association.

Here we report the DNA methylation profile of 84 sporadic pancreatic neuroendocrine tumors (PanNETs) with associated clinical and genomic information. We identified three subgroups of PanNETs, termed T1, T2 and T3, with distinct patterns of methylation. The T1 subgroup was enriched for functional tumors and ATRX, DAXX and MEN1 wild-type genotypes. The T2 subgroup contained tumors with mutations in ATRX, DAXX and MEN1 and recurrent patterns of chromosomal losses in half of the genome with no association between regions with recurrent loss and methylation levels. T2 tumors were larger and had lower methylation in the MGMT gene body, which showed positive correlation with gene expression. The T3 subgroup harboured mutations in MEN1 with recurrent loss of chromosome 11, was enriched for grade G1 tumors and showed histological parameters associated with better prognosis. Our results suggest a role for methylation in both driving tumorigenesis and potentially stratifying prognosis in PanNETs.

Verifying explainability of a deep learning tissue classifier trained on RNA-seq data.

For complex machine learning (ML) algorithms to gain widespread acceptance in decision making, we must be able to identify the features driving the predictions. Explainability models allow transparency of ML algorithms, however their reliability within high-dimensional data is unclear. To test the reliability of the explainability model SHapley Additive exPlanations (SHAP), we developed a convolutional neural network to predict tissue classification from Genotype-Tissue Expression (GTEx) RNA-seq data representing 16,651 samples from 47 tissues. Our classifier achieved an average F1 score of 96.1% on held-out GTEx samples. Using SHAP values, we identified the 2423 most discriminatory genes, of which 98.6% were also identified by differential expression analysis across all tissues. The SHAP genes reflected expected biological processes involved in tissue differentiation and function. Moreover, SHAP genes clustered tissue types with superior performance when compared to all genes, genes detected by differential expression analysis, or random genes. We demonstrate the utility and reliability of SHAP to explain a deep learning model and highlight the strengths of applying ML to transcriptome data.