Comparison of the Metastasis Predictive Potential of mRNA and Long Non-Coding RNA Profiling in Systemically Untreated Breast Cancer

SIMPLE SUMMARY: To support health care providers in clinical decision-making for breast cancer (BC) patients, profiles of gene activity patterns have previously been developed, where the majority have been based on messenger RNAs (mRNAs), molecules coding for proteins. However, we and others have recently developed profiles based on functional molecules that do not code for proteins—e.g., long non-coding RNAs (lncRNAs)—demonstrating great prognostic potential. Unfortunately, studies comparing such profiles for predicting relapse in BC patients are very scarce. Therefore, we aimed to compare these two types of molecules (mRNAs and lncRNAs) to forecast relapse in low-risk BC patients using advanced machine learning methods with two different approaches. Regardless of approach, our data suggested that profiles based on lncRNAs improved prediction of relapse and demonstrated potential advantages for future profile development. ABSTRACT: Several gene expression signatures based on mRNAs and a few based on long non-coding RNAs (lncRNAs) have been developed to provide prognostic information beyond clinical evaluation in breast cancer (BC). However, the comparison of such signatures for predicting recurrence is very scarce. Therefore, we compared the prognostic utility of mRNAs and lncRNAs in low-risk BC patients using two different classification strategies. Frozen primary tumor samples from 160 lymph node negative and systemically untreated BC patients were included; 80 developed recurrence—i.e., regional or distant metastasis while 80 remained recurrence-free (mean follow-up of 20.9 years). Patients were pairwise matched for clinicopathological characteristics. Classification based on differential mRNA or lncRNA expression using seven individual machine learning methods and a voting scheme classified patients into risk-subgroups. Classification by the seven methods with a fixed sensitivity of ≥90% resulted in specificities ranging from 16–40% for mRNA and 38–58% for lncRNA, and after voting, specificities of 38% and 60% respectively. Classifier performance based on an alternative classification approach of balanced accuracy optimization also provided higher specificities for lncRNA than mRNA at comparable sensitivities. Thus, our results suggested that classification followed by voting improved prognostic power using lncRNAs compared to mRNAs regardless of classification strategy.