A novel machine learning approach to breast cancer screening using non-invasive mRNA collection.

Abstract

e13609 Background: Current breast cancer diagnostic methods are often highly-invasive or painful, deterring many from screening. While DNA-based methods such as BRCA1/2 testing are available, they can miss cancer indicators due to epigenetic interactions. mRNA-based screening can capture dynamic signals, incorporating information from both DNA and environmental factors. We present a machine learning system for identifying breast cancer patients based on next-generation sequencing (NGS)-processed, non-invasively collected mRNA samples. Methods: We collected buccal cheek swab samples from breast cancer patients (n = 69) and non-cancer controls (n = 38). A custom microarray chip for quantifying 48 mRNA biomarkers was created by selecting cytokines associated with inflammatory pathways and cancer progression. Samples were sequenced using Thermo Fisher Genestudio S-5. We evaluated multiple classifiers (SVMs, decision trees, random forests, logistic regression, multi-layer perceptron) using 5x2 k-fold cross-validation. To optimize the biomarker panel to include a minimally-predictive set of features, we performed exhaustive searches over five- and six-biomarker combinations (1.7 and 12.3 million combinations, respectively). Results: Classifiers using the full 48-biomarker panel achieved similar performance (mean F-scores 0.78-0.83). We selected linear SVM for subsequent experiments due to its capacity for efficient inference. While automated feature selection methods degraded performance, our exhaustive search revealed synergistic biomarker combinations. The top-performing lower-dimensional panels are shown in Table 1, with our best classifier using six biomarkers and achieving an F1-score of 0.9. Conclusions: Our results demonstrate the efficacy of a machine learning system for identifying breast cancer patients from NGS of non-invasively-collected cheek swabs. The high-performing lower-dimensional biomarker panels show promise for both screening applications and research into breast cancer etiology. Table summarizing results of our top-performing five- and six-biomarker machine learning systems. "NA" in the BM6 column indicates only five biomarkers were used for that particular system. Our overall top-performing system used six biomarkers and achieved an F1-score of 0.9. BM1 BM2 BM3 BM4 BM5 BM6 P R AUC F1 ABCA1 ADAMTS4 CASP9 IL10 STK11 NA 0.763 0.929 0.823 0.837 ABCA1 CACNA1C CCL27 RGS2 SLC40A1 NA 0.734 0.970 0.726 0.834 CASP9 IL1B IL6 RGS2 SLC40A1 STK11 0.892 0.913 0.897 0.900 HAMP IL10 RGS2 SLC22A4 STK11 TGFB1 0.888 0.902 0.849 0.892