Genomic risk prediction of aromatase inhibitor-related arthralgia in patients with breast cancer using a novel machine-learning algorithm

Abstract

Many breast cancer (BC) patients treated with aromatase inhibitors (AIs) develop aromatase inhibitor-related arthralgia (AIA). Candidate gene studies to identify AIA risk are limited in scope. We evaluated the potential of a novel analytic algorithm (NAA) to predict AIA using germline single nucleotide polymorphisms (SNP) data obtained before treatment initiation. Systematic chart review of 700 AI-treated patients with stage I-III BC identified asymptomatic patients (n = 39) and those with clinically significant AIA resulting in AI termination or therapy switch (n = 123). Germline DNA was obtained and SNP genotyping performed using the Affymetrix UK BioBank Axiom Array to yield 695,277 SNPs. SNP clusters that most closely defined AIA risk were discovered using an NAA that sequentially combined statistical filtering and a machine-learning algorithm. NCBI PhenGenI and Ensemble databases defined gene attribution of the most discriminating SNPs. Phenotype, pathway, and ontologic analyses assessed functional and mechanistic validity. Demographics were similar in cases and controls. A cluster of 70 SNPs, correlating to 57 genes, was identified. This SNP group predicted AIA occurrence with a maximum accuracy of 75.93%. Strong associations with arthralgia, breast cancer, and estrogen phenotypes were seen in 19/57 genes (33%) and were functionally consistent. Using a NAA, we identified a 70 SNP cluster that predicted AIA risk with fair accuracy. Phenotype, functional, and pathway analysis of attributed genes was consistent with clinical phenotypes. This study is the first to link a specific SNP/gene cluster to AIA risk independent of candidate gene bias.

Keywords: SNP; Aromatase; arthralgia; breast cancer; estrogen; joint pain.

Figure 1

Flowchart of the analytical steps used to identify those SNPs most predictive of AIA risk in the study population.

Figure 2

Stability analysis of the small‐scale signature of the 70 SNPs which were identified as being most predictive of AIA risk. The figure shows the cumulative distribution function of the predictive accuracy obtained after 5000 random 75/25 hold out simulations. It can be observed that the median accuracy (percentile 50) is around 75%, being the lower and upper‐quartiles 71% and 78%. The minimum and maximum accuracy achieved was 54% and 100%.

Figure 3

Color visualization delineating the relationship of the 70 predictive SNPs and their functional relevance. (A) Correlation tree of the most discriminatory SNPs. This tree is built using the minimum spanning tree algorithm using the Pearson correlation coefficient. The algorithm looks for the maximum absolute values of the Pearson correlation coefficient (positive and negative correlations) within the set of most discriminatory SNPs. This hierarchical figure describes the strength of relationships between SNPs and how each SNP relates to the others in the cluster.(B) Associated phenotypes include similar phenotypes (RA, pain, inflammation and those associated with the tumor diagnosis).