Early prediction of lethal phenotypes in triple negative breast cancer using multiscale, multi-modality platforms

Triple negative breast cancer (TNBC) is the most aggressive breast cancer subtype and poses a clinical challenge due to the lack of tools to predict patients who are at risk of death due to disease. TNBCs subtypes were presumed to be a step forward, however, recent randomized clinical trials have failed to document differential impact of therapy within TNBC subtypes. Immunotherapies will benefit a small subgroup (~7 out of 100 women) of TNBC patients, but are associated with 15-30% increase in grade 3 immune toxicities. Currently, all patients with TNBCs are offered chemotherapy-based regimens either after surgery (70%) or less often (~30%) before surgery (neo-adjuvant). For the 30% of TNBC patients prescribed neo-adjuvant chemotherapy in TNBC patients presence or absence of residual disease has been used for optimizing escalation or de-escalation strategies. Studies such as the SWOG 2212 (SCARLET) are being proposed to de-escalate neo-adjuvant chemotherapy in TNBC based on TILs with residual disease as the endpoint. However, for the 70% of TNBC patients treated with adjuvant chemotherapy, there are no established prognostic tools, apart for TILs, for deescalating therapies. Better tools are necessary to identify potentially lethal TNBC cases with high accuracy. The overall objective of this grant is to develop novel tools for prediction of risk of lethality in TNBC using a multimodal approach. The central hypothesis is that accuracy of lethality prediction can be improved by a multi-scale approach comprising novel artificial intelligence (AI)-based tools of mammographic images and pathology including spatial immune analysis. The following specific aims are proposed: Aim 1) To generate a novel model of mammographic radiomic features to determine probability of TNBC lethality; Aim 2) To establish the role of generative adversarial network (GAN) augmented deep learning (DL) model from digital H&E images for predicting probability of lethality; Aim 3) To profile immune cell topology and spatial relationships and their contribution to risk of lethality; Aim 4) To integrate multimodal data models for identification of risk of lethality in TNBCs. A late fusion model will integrate the clinicopathological features with probabilities obtained from pathomics and radiomics and immune data to develop the CaPRI prediction score for assessing lethality of TNBC that will stratify patients based on their risk. The score will provide evidence-based rationale for escalating or deescalating therapy in TNBC. The developed tool, CaPRI score, will be based on multi-modal analysis and integrate them will clinicopathological features such as age, race and tumor size. The continuous CaPRI score could be used either as a continuous measure of risk or in a categorical manner (high or low risk). High-risk patients, like patients with residual cancer after neoadjuvant chemotherapy, would be candidates for escalation of therapy. Similarly, patients with low risk could be candidates for de-escalation.