The Importance of Development and Application of Subtyping Models for Breast Cancer on the Same Platform

Abstract

Abstract Background Intrinsic subtypes have played an important role in breast cancer research. The accuracy of breast cancer research findings depends on the accuracy of intrinsic subtype classification of breast cancer samples. Popular intrinsic subtype models, such as PAM50 and AIMS, were mainly developed on Microarray but are widely used in other platforms. The transferability of these models to RNA-seq and other platforms has rarely been studied. We aim to assess the effectiveness of popular intrinsic subtype models on RNA-seq data and improve the accuracy of breast cancer subtyping on this platform.Methods Assuming that one breast cancer sample only belongs to one subtype irrespective of the expression measuring platform, we assessed the consistency of subtype predictions of PAM50 and AIMS for TCGA Microarray and RNA-seq data from the same samples using Kappa statistic. We also built 12 models using common and intrinsic genes on both Microarray and RNA-seq data, and evaluated their performance under the same assumption.Results Both PAM50 and AIMS failed to produce consistent predictions for Microarray and RNA-seq data from the same samples, with Kappa values of 0.33 and 0.21, respectively. Data normalization improved prediction consistency but introduced false classification. Microarray-developed models showed low prediction consistency when applied to both Microarray and RNA-seq profiles of the same samples. However, prediction consistency between predictions of RNA-seq developed model predicting RNA-seq data and Microarray developed model predicting Microarray data was high with most Kappa values above 0.85.Conclusion Great caution should be taken when using PAM50 and AIMS for the subtyping of breast cancer RNA-seq data. Breast cancer subtyping models should be developed and applied on the same platform, using unnormalized expression data for accurate subtype prediction purpose.