Abstract 3510: A radiogenomic approach for triple-negative breast cancer risk stratification
Humaira Noor,Yuanning Zheng,Adam Mantz,Ryle Zhou,Andrew Kozlov,Wendy B. DeMartini,Shu-tian Chen,Satoko Okamoto,Debra Ikeda,Sarah Mattonen,Sandy Napel,Melinda L. Telli,George Sledge,Allison Kurian,Mina Satoyoshi,Olivier Gevaert,Haruka Itakura
DOI: https://doi.org/10.1158/1538-7445.am2024-3510
IF: 11.2
2024-03-22
Cancer Research
Abstract:Abstract Background: Triple-negative breast cancer (TNBC) is an aggressive disease that accounts for 15-20% of all breast cancers. Expressions of ER, PR and HER2 receptors are lacking in this disease, and thus targeted therapies are not effective. TNBC has a shorter relapse-free survival, higher metastasis rate and decreased overall survival compared with other breast cancers. However, when undergoing standard treatment, some patients respond well, while others have poor outcome, suggesting TNBC heterogeneity. Early stratification of patients with long versus short survival could identify the subgroup of patients who would not benefit from exposure to toxicity of chemotherapy treatment. Here, we developed a non-invasive radiogenomic approach for TNBC risk stratification. Methods: A transcriptomic-based prognostic gene signature was previously developed using the TCGA-BRCA cohort (n=860). Briefly, LASSO Cox regression model analysis with the ‘glmnet’ R package was used to identify the transcriptomic signature gene-set consisting of 50 genes. We tested this signature to prognosticate overall survival in a Stanford cohort (n=63) and a previously published SCANB cohort (n=604). The patients were stratified into high- and low-risk groups based on the median risk-score. Next, we developed a machine learning model that identified a radiomic feature set to predict the prognostic transcriptomic risk-groups. Radiomic features were extracted from pre-treatment breast MRI. Radiomics features were extracted using PyRadiomics. The model utilized Decision Tree Classifier and LeaveOneOut method was used for cross-validation. Results: The transcriptomic signature low-risk group was significantly associated with improved overall survival in the two TNBC cohorts, with hazard ratios of 0.11 [95% CI: 0.01-0.88] for the Stanford cohort and 0.71 [95% CI: 0.52-0.97] for the SCANB cohort (log-rank p-values p=0.012 and p=0.032, respectively). Including this transcriptomic signature in a multivariate analysis, which adjusted for clinical features (patient age, grade, stage and Ki67%), the transcriptomic prognostic signature remained a significant prognostic factor (p<0.05). The radiomic feature set (consisting of 20 features) predicted the high- and low-risk transcriptomic groups with a mean accuracy of 72.2% and a mean AUROC of 71%. The precision, F1 and recall scores were 67%, 74% and 82%, respectively. In an independent dataset consisting of 116 Stanford TNBC patients, we used this model to predict risk groups based on the MRI radiomics features, and evaluated the prognostic effects of predicted risk groups. The overall survival of the predicted high-risk group was significantly poorer than the predicted low-risk group (p=0.013). Conclusions: We present a prognostic model that can non-invasively stratify TNBC patients for low versus high mortality risk using radiomic features derived from pre-treatment patient MRI data. Citation Format: Humaira Noor, Yuanning Zheng, Adam Mantz, Ryle Zhou, Andrew Kozlov, Wendy B. DeMartini, Shu-tian Chen, Satoko Okamoto, Debra Ikeda, Sarah Mattonen, Sandy Napel, Melinda L. Telli, George Sledge, Allison Kurian, Mina Satoyoshi, Olivier Gevaert, Haruka Itakura. A radiogenomic approach for triple-negative breast cancer risk stratification [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3510.
oncology