Predicting clinical outcomes using cancer progression associated signatures

Our plain-English summary of Predicting clinical outcomes using cancer progression associated signatures, by J. Mamrot, N.E. Hall, R.A. Lindley (Oncotarget, 2021).

The study

• Authors: J. Mamrot, N.E. Hall, R.A. Lindley
• Journal: Oncotarget
• Published: 2021;12(8):845-858
• Cited by: 2 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Somatic mutation signatures are an informative facet of cancer aetiology, however they are rarely useful for predicting patient outcome. The aim of this study is to evaluate the utility of a panel of 142 mutation-signature-associated metrics (P142) for predicting cancer progression in patients from a 'TCGA PanCancer Atlas' cohort. The P142 metrics are comprised of AID/APOBEC and ADAR deaminase associated SNVs analyzed for codon context, strand bias, and transitions/transversions. TCGA tumor-normal mutation data was obtained for 10,437 patients, representing 31 of the most prevalent forms of cancer. Stratified random sampling was used to split patients into training, tuning and validation cohorts for each cancer type. Cancer specific machine learning (XGBoost) models were built using the output from the P142 panel to predict patient Progression Free Survival (PFS) status as either "High PFS" or "Low PFS". Predictive performance of each model was evaluated using the validation cohort. Models accurately predicted PFS status for several cancer types, including adrenocortical carcinoma, glioma, mesothelioma, and sarcoma. In conclusion, the P142 panel of metrics successfully predicted cancer progression status in patients with some, but not all cancer types analyzed. These results pave the way for future studies on cancer progression associated signatures.

Copyright: © 2021 Mamrot et al.

In plain English

This study used existing tumor genetic data from The Cancer Genome Atlas (TCGA) PanCancer Atlas. It tested whether a panel of 142 mutation-pattern measures linked to deaminase enzymes, called P142, could predict cancer progression. The panel was applied to 5,903 eligible patients, including 64 with mesothelioma. Patients were labeled "High PFS" or "Low PFS" by whether observed progression-free survival was above or below a set threshold. The threshold was 12 months for mesothelioma. Two cancer types with too few patients were excluded, leaving 28. Models were trained to predict the labels and tested on held-out validation patients over 20 rounds. Average accuracy across types was 60% (plus or minus 1.2%), versus 50% by chance. The best models for adrenocortical carcinoma, mesothelioma and cholangiocarcinoma reached 100% accuracy on their validation patients. Accuracy was low for 8 of the 28 types. Mesothelioma patients with Low PFS had significantly more mutations than those with High PFS. The study did not examine asbestos exposure. The authors concluded that outcomes correlated with these metrics in some, but not all, cancer types. They noted limits of the TCGA data.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Retrospective analysis of existing tumor-normal exome sequencing data from TCGA. Machine learning (XGBoost) models were trained to classify patients' progression-free survival status, with 20 rounds of cross validation per cancer type.
• Who was studied: 5,903 eligible patients from the TCGA PanCancer Atlas, grouped by cancer type. Patients were labeled High or Low PFS from their observed PFS relative to a threshold. The mesothelioma group had 64 patients: 16 above and 48 below the 12-month PFS threshold. The median PFS for this group was 12.00 months.
• Exposure: No asbestos or other environmental exposure was studied. The predictors were 142 metrics derived from somatic tumor mutations associated with AID/APOBEC and ADAR deaminase activity (codon context, strand bias, transitions/transversions).
• Main results:
  • Two cancer-type groups (DLBC and KICH) had too few patients in the tuning or validation cohorts and were excluded from further analysis, leaving 28 cancer types for modeling.
  • Across all cancer types, average predictive accuracy of the models in validation cohorts was 60% plus or minus 1.2% (mean plus or minus SD; 20 rounds of cross validation per cancer type). This compared with 50% expected by random chance.
  • The highest predictive accuracy was seen in models for adrenocortical carcinoma, mesothelioma and cholangiocarcinoma. These models predicted the PFS category of validation patients with 100% accuracy. The discussion describes this as accuracy of up to 100%, not average performance.
  • In lower grade glioma, mesothelioma and adrenocortical carcinoma, patients with Low PFS had significantly more mutations than patients with High PFS (p less than 0.05). For most cancer types, no significant correlation between total mutation burden and PFS status was found.
  • For mesothelioma, the PFS threshold used was 12 months, with 64 patients included: 16 above and 48 below the threshold.
  • In 20 of 28 cancer types, one or more models achieved more than 70% predictive accuracy.
  • Predictive accuracy was low for 8 of the 28 cancer types.
  • Individual metrics in the panel were weak predictors of PFS status on their own.
• The authors' conclusion: The authors concluded that they had identified a correlation between cancer patient outcomes and deaminase-associated metrics in some, but not all, cancer types investigated. They said the study provides a basis for further development of such biomarker panels.
• Limitations the authors state:
  • TCGA data had sparse personal and clinical history, missing metadata, predominantly Caucasian samples and a single timepoint for almost every patient. The authors said this may have reduced predictive power and may limit translation of the results.
  • The authors suggested that the accumulation of mutations without a discernable source may reduce the ability to resolve deaminase-mediated mutation signatures.

Topics

Medical terms in this study, explained in our Lexicon: Somatic Mutation, Carcinoma, Mesothelioma, Sarcoma.

Related studies

• Two Different Serum MiRNA Signatures Correlate with the Clinical Outcome and Histological Subtype in Pleural Malignant Mesothelioma Patients — M. Lamberti, R. Capasso, et al. (2015)
• Identification of a Five-Gene Signature for Predicting Survival in Malignant Pleural Mesothelioma Patients — Y. Bai, X. Wang, et al. (2020)
• Identification of glycolysis genes signature for predicting prognosis in malignant pleural mesothelioma by bioinformatics and machine learning — Y. Xiao, W. Huang, et al. (2022)
• miR-625-3p and lncRNA GAS5 in Liquid Biopsies for Predicting the Outcome of Malignant Pleural Mesothelioma Patients Treated with Neo-Adjuvant Chemotherapy and Surgery — J. Kresoja-Rakic, A. Szpechcinski, et al. (2019)
• MiR-score: a novel 6-microRNA signature that predicts survival outcomes in patients with malignant pleural mesothelioma — M.B. Kirschner, Y.Y. Cheng, et al. (2015)
• Predicting Malignant and Paramalignant Pleural Effusions by Combining Clinical, Radiological and Pleural Fluid Analytical Parameters — S. Herrera Lara, E. Fernández-Fabrellas, et al. (2017)

Browse more research in Oncotarget, or return to medical research.