A deep-learning model based on histological images predicts survival and pathological response to neoadjuvant chemotherapy and chemoimmunotherapy in breast cancer

Abstract Background Accurate biomarkers for prognosis and selection of patients most likely to benefit from neoadjuvant chemoimmunotherapy in early breast cancer (eBC) are urgently needed. The Immune Constant of Rejection (ICR) gene expression signature, a 20-gene signature reflecting cytotoxic immune activation, is associated with favorable prognosis and pathological complete response (pCR) to neoadjuvant chemotherapy (NACT) in eBC. However, ICR requires transcriptomic profiling, restricting its routine clinical implementation. We hypothesized that the biological features captured by ICR could be inferred directly from routine hematoxylin-eosin (H&E) whole-slide images (WSIs) using deep learning. Methods We developed an artificial intelligence (AI)-based deep-learning model to predict ICR class (AI-ICR model) from WSIs and tested its predictive value for survival and pCR to NACT and neoadjuvant chemoimmunotherapy (NACTIT). Our model was trained using WSIs from 2,881 TCGA non-BC multicancer samples paired with RNA-sequencing-derived ICR classification. It was externally validated in 1,087 TCGA BC samples and 104 pretreatment biopsies from two BC prospective neoadjuvant trials. We assessed its association with overall survival (OS) in TCGA and with pCR to NACT/NACTIT in the neoadjuvant cohorts (PELICAN+NEO-R-IPC). Results AI-ICR model showed a 0.917 average cross-validated AUC in the mutlicancer internal test set, and performed well in external eBC series, with 0.937 AUC in TCGA ( p = 4·13E-45) and 0.739 in PELICAN+NEO-R-IPC ( p = 5·00E-03), thus demonstrating robustness across tumor types, tissue preparations, and acquisition platforms. AI-ICR-high tumors displayed longer OS than AI-ICR-low tumors with Hazards-Ratio for death of 0.29 [95%CI 0.15–0.55] in multivariate analysis ( p = 1·69E-04). They also displayed larger pCR rate than AI-ICR-low tumors, with an Odds-Ratio for pCR of 4.33 [95%CI 1.68–11.2] in multivariate analysis ( p = 2·47E-03), and this difference was stronger in the NACTIT arm (81% versus 36%) than in the NACT arm (65% versus 30%). Transcriptomics showed enrichment of immune genes and activated immune cell types in AI-ICR-high tumors. Conclusions AI-ICR enables accurate inference of a clinically validated immune gene-expression program directly from routine H&E slides and is independently associated with survival and pCR to NACT/NACTIT in eBC. After further validation of both ICR inference and clinical outcome prediction in larger series, it could represent a simple, inexpensive and scalable alternative to transcriptomic profiling for precision management of eBC, including routine AI-assisted immuno-oncology.