Machine learning models based on immunological genes to predict the response to neoadjuvant therapy in breast cancer patients

Abstract

Breast cancer (BC) is the most common malignancy worldwide and neoadjuvant therapy (NAT) plays an important role in the treatment of patients with early BC. However, only a subset of BC patients can achieve pathological complete response (pCR) and benefit from NAT. It is therefore necessary to predict the responses to NAT. Although many models to predict the response to NAT based on gene expression determined by the microarray platform have been proposed, their applications in clinical practice are limited due to the data normalization methods during model building and the disadvantages of the microarray platform compared with the RNA-seq platform. In this study, we first reconfirmed the correlation between immune profiles and pCR in an RNA-seq dataset. Then, we employed multiple machine learning algorithms and a model stacking strategy to build an immunological gene based model (Ipredictor model) and an immunological gene and receptor status based model ICpredictor model) in the RNA-seq dataset. The areas under the receiver operator characteristic curves for the Ipredictor model and ICpredictor models were 0.745 and 0.769 in an independent external test set based on the RNA-seq platform, and were 0.716 and 0.752 in another independent external test set based on the microarray platform. Furthermore, we found that the predictive score of the Ipredictor model was correlated with immune microenvironment and genomic aberration markers. These results demonstrated that the models can accurately predict the response to NAT for BC patients and will contribute to individualized therapy.

Keywords: breast cancer; immunological gene; machine learning; neoadjuvant therapy; pathological complete response.

Figure 1

Association between immune profiles and pCR. (A) Comparisons of immune-related biological process enrichment scores between pCR and non-pCR patients. (B) Comparisons of pCR rates among patients with various immune subtypes. (C) Comparison of the level of stromal cell present, immune cell infiltration, and tumor purity between pCR and non-pCR patients. (D) Comparisons of infiltrating immune cell subsets between pCR and non-pCR patients.

Figure 2

Model performance evaluation in the training and test sets. Comparisons of the PSs in pCR patients and non-pCR patients in the training set (A) and test set (B). Brier scores of the models and clinicopathological characteristics in the training set (C) and test set (D). AUPRCs of the models and clinicopathological characteristics in the training set (E) and test set (F). AUROCs of the models and clinicopathological characteristics in the training set (G) and test set (H). SENs, SPEs, PPVs, and NPVs for the Ipredictor (I) and ICpredictor (J) models in the training set. SENs, SPEs, PPVs, and NPVs for the Ipredictor (K) and ICpredictor (L) models in the test set.

Figure 3

Model performance evaluation in the patient subgroups of the test set and GSE20271 dataset. AUROCs of the models for ER+/HER2- patients in the training (A) and test (B) set. AUROCs of the models for HER2+ patients in the training (C) and test (D) set. AUROCs of the models for ER-/HER2- patients in the training (E) and test (F) set. AUROCs of the models for all patients (G) and ER+/HER- patients (H) in the GSE20271 dataset.

Figure 4

Clinicopathological and biological implications for the PS. The correlations between the PS and ER status (A), PR status (B), HER2 status (C), ER/HER2 status (D), Ki67 status (E), age (F), clinical T stage (G), clinical N stage (H), and clinical stage (I). (J) Comparisons of the hallmark gene set enrichment scores between the patients with high and low PSs.

Figure 5

Relationship between the PS and immune microenvironment and genomic aberrations. The PS was associated with the immune state (A), tumor purity (B), and immune cell infiltration (C), but was not with the stromal cell infiltration (D). The PS was positively correlated to the TIL density (E) and cytolytic activity (F). Association between the PS and the abundance of immune cell subsets (G) and genomic aberration markers (H). ^∗P < 0.05,^∗∗∗P < 0.001.