2274P Deep learning-based prediction of pathologic complete response to neoadjuvant therapy in breast cancer using H&E images and RNA-Seq in the IMMUcan study

IMMUcan (SPECTA NCT02834884) is an European public-private effort to generate molecular and cellular profiling data of the human tumor microenvironment from up to 3000 cancer patients. Predicting pathologic complete response (pCR), which has been associated with better outcome after neoadjuvant treatment in breast cancer (BC), could help refining treatment strategies. Here, we aim to integrate multiple data layers using different Deep Learning (DL) approaches to predict pCR from baseline tumor samples in the context of the prospective IMMUcan Triple-Negative Breast Cancer (TNBC) and HER2-positive (HER2+) BC neoadjuvant cohorts. At the cut-off date of June 29th, 2022, we identified a first cohort of 132 and 149 patients diagnosed with TNBC and HER2+ BC, respectively, for preliminary analyses. To predict pCR at the patient level, benchmark models using RNA-Seq, image DL were trained on Whole Slide Images (WSIs) and RNA-Seq data. The image models included two main components: a tiling algorithm pre-trained on TCGA WSI to extract a spatialized representation of the WSI and a classification part for the pCR prediction. Baseline RNA-Seq data were available for 109 and 115 patients with TNBC and HER2+ BC, respectively, pCR status was available for 130 TNBCs and 117 HER2+ BCs, while a baseline H&E-stained WSI was available for all patients. Among the models applied independently to each data type, the best performance was obtained using RNA-Seq in HER2+ BC (ROC AUC = 0.61, std = 0.04), and WSI in TNBC (ROC AUC = 0.63, std = 0.03). These preliminary results show the potential of DL applied to WSI and RNA-Seq in predicting pCR for TNBC and HER2+ BC. Using DL models able to predict pCR provide the opportunity to better select patients and tailor neoadjuvant therapies in BC. Multimodal models combining RNASeq and WSI are currently being tested out by the team to improve performance.