Dataset Information

Deep learning for explainable prediction of HPV-status in head and neck cancer using transcriptome data organized as pathway treemaps

ABSTRACT: We propose a convolutional neural network (CNN) model predicting HPV-status in head and neck cancer from transcriptome data allowing identification of molecular pathways driving individual classifier decisions. The CNN was trained on transcriptome data from 264 (13% HPV-positive) and tested on 85 (25% HPV-positive) patients after transformation into 2D-treemaps representing molecular pathways. Model stability was assessed by shuffling pathways within 2D-images. Grad-CAM saliency was used to quantify pathways contribution to CNN decisions. For comparison, a logistic regression model was generated. The CNN achieved ROC-AUC/PR-AUC of 0.96/0.90 for all treemap variants. Saliency heatmaps consistently found KRAS-, spermatogenesis-, bile acid metabolism- and inflammatory-signaling as most informative for classification of HPV-positive-, and MYC-targets-, epithelial-mesenchymal transition and protein-secretion pathways for HPV-negative patients. The regression-based 18-gene model achieved a ROC-AUC/PR-AUC of 0.97/0.97. We present a high-performance explainable CNN-model from transcriptome data of a typically sized oncological cohort providing molecular pathway information at the individual level.

ORGANISM(S): Homo sapiens

PROVIDER: GSE205308 | GEO | 2023/04/24

REPOSITORIES: GEO

ACCESS DATA

Similar Datasets

Project description:The effect of preventive vaccination against Human Papillomavirus (HPV) in reducing the burden of cervical cancer (CC) will not be seen before 30 years. It is still necessary to improve the procedures of screening and treatment against CC. Current methods for screening have low sensitivity (Pap test) or specificity (HPV tests) to detect high-grade cervical intraepithelial neoplasias (CIN) and CC (CIN2+). The objective of this study was the identification and characterization of cellular targets present in most CC and absent in normal cervical tissue that can be considered as potential markers for screening or therapeutic targets. Methods and Findings. A pyramidal strategy was used. Initially the expression of 8,638 genes was compared between 43 HPV16-positive CC and 12 healthy cervical epitheliums using the HG-Focus microarray. The expression intensity of 8,370 genes was validated in 24 samples with a second microarray (HG-ST1.0). A total of 997 genes were deregulated in CC and the 21 best ranked were validated and confirmed by qRT-PCR in 67 CC and 25 controls. According to the ROC analysis and the fold change (FC), the best (AUC M-bM-^IM-% 0.97 and FC M-bM-^IM-% 10) six genes (CCNB2, CDC20, PRC1, SYCP2, NUSAP1, CDKN3) belong to the mitosis pathway. They were further explored by qRT-PCR in 29 CIN1 and 21 CIN2/3 lesions to investigate whether they could differentiate CIN2+ from CIN1 and healthy cervical epitheliums (CIN1-). Three of these genes were associated exclusively with invasive tumors (CCNB2, PRC1, SYCP2) and three (CDC20, NUSAP1, CDKN3) also with CIN2/3. The sensitivity and specificity of CDKN3 and NUSAP1, along with CDKN2A, to detect CIN2+ was around 90%. The protein codified by these genes was confirmed by immunohistochemistry in CC. The effect of these markers on the survival was investigated in 40 CC patients followed by 42 months. Only high-expression level of CDKN3 was associated with a poor survival (20%; p = 2 x 10-4, Long-rank test) and it was independent from FIGO staging. Conclusions. CDKN3 and NUSAP1, together with CDKN2A, may be potential targets for development of screening methods. However, further studies are needed from a screening population to define the optimal trade-off between sensitivity and specificity for detection of CIN2+. CDKN3 may be also considered as a survival marker. Inhibition of mitosis is a well-known strategy to combat cancers. Therefore, CDKN3 may be a potential therapeutic target in CC. However, itM-BM-4s still necessary to demonstrate whether itM-BM-4s indispensable for tumor growth. cervical cancer vs. healthy cervix

Project description:Differentially methylated oral squamous cell carcinoma (OSCC) biomarkers, identified in-vitro and validated in well-characterized surgical specimens, have shown poor clinical correlation in cohorts with different risk profiles. To overcome this lack of relevance we used the HumanMethylation27 BeadChip, publicly available methylation and expression array data, and Quantitative Methylation Specific PCR to uncover differential methylation in OSCC clinical samples with heterogeneous risk profiles. A two stage-design consisting of Discovery and Prevalence screens was used to identify differential promoter methylation and deregulated pathways in patients diagnosed with OSCC and head and neck squamous cell carcinoma. Promoter methylation of KIF1A (k=0.64), HOXA9 (k=0.60), NID2 (k=0.60), and EDNRB (k=0.60) had a moderate to substantial agreement with clinical diagnosis in the Discovery screen. HOXA9 had 68% sensitivity, 100% specificity and a 0.81 AUC. NID2 had 71% sensitivity, 100% specificity and a 0.79 AUC. In the Prevalence screen HOXA9 (k=0.82) and NID2 (k=0.80) had an almost perfect agreement with histologic diagnosis. HOXA9 had 85% sensitivity, 97% specificity and a 0.95 AUC. NID2 had 87% sensitivity, 95% specificity and a 0.91 AUC. A HOXA9 and NID2 gene panel had 94% sensitivity, 97% specificity and a 0.97 AUC. In saliva from OSCC cases and controls HOXA9 had 75% sensitivity, 53% specificity and a 0.75 AUC. NID2 had 87% sensitivity, 21% specificity and a 0.73 AUC. This Phase I Biomarker Development Trial identified a panel of differentially methylated genes in normal and OSCC clinical samples from patients with heterogeneous risk profiles. This panel may be useful for early detection and cancer prevention studies. The 12 samples analyzed comprise equal numbers of three tissue-types: Primary oral squamous cell carcinoma, Oral leukoplakia and Normal oral mucosa. Gender, age and smoking-status were approximately equally represented in these goups. Normal oral mucosa served as a control.

Dataset Information

Deep learning for explainable prediction of HPV-status in head and neck cancer using transcriptome data organized as pathway treemaps

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure