Abstract PS3-04-04: Ai predicts response to neoadjuvant therapy in breast cancer across diverse cohorts

Abstract Background: Neoadjuvant therapy (NAT) is standard of care for many early-stage breast cancer patients. Traditional clinical characteristics such as ER and HER2 receptor status provide useful prognostic information, but lack sufficient precision when used in isolation. Basic histological features, such as tumor infiltrating lymphocytes, are also associated with NAT response. With the advent of computational pathology and machine learning, it is now possible to extract complex spatial and morphological predictive patterns from digitized slides that are not apparent to the human eye. Integrating histopathologic features with clinical variables holds promise for building robust predictive models that can guide personalized treatment planning, reduce overtreatment, and improve survival outcomes in breast cancer patients undergoing neoadjuvant therapy. Ataraxis Breast Neo (ATX-N) is a multi-modal artificial intelligence test that integrates morphological features extracted from standard H&E-stained slides with baseline clinical data (ER/PR/HER2 status, staging, age, IDC/ILC presence). Here we validate the accuracy of ATX-N in diverse patient populations. Methods: Whole slide images of H&E-stained biopsy specimens and corresponding clinical variables were collected from 563 patients who have undergone neoadjuvant therapy across UChicago (n=257), Providence Health (n=65), Jagiellonian University (n=115), and public IMPRESS dataset (n=126) cohorts. Across the cohorts, there are 186 TNBC patients, 201 HER2+ patients and 164 HR+HER2- patients. The remaining 12 patients had HER2 equivocal status or the information was missing. ATX-N scores were generated using a locked model. Performance of ATX-N was assessed for pathologic complete response (pCR) AUC. Reported AUCs are pooled using random effects models. Results: The overall pCR rate was 40.1%. ATX-N achieved a pooled AUC of 0.721 (95% CI: 0.646-0.796) across four independent cohorts, with cohort-level AUCs ranging from 0.644 to 0.804 (Table 1 Table 1: Breakdown of evaluated cohorts and model performance in each of the datasets. Dataset N total patients in the cohort pCR rate, N (%) AUC (95% CI) UChicago 257 94/257 (36.6%) 0.644 (0.577-0.716) Providence 65 30/65 (46.2%) 0.734 (0.598-0.846) Jagiellonian 115 37/115 (32.2%) 0.804 (0.717-0.881) IMPRESS 126 65/126 (51.6%) 0.714 (0.624-0.797) Pooled - All 563 226/563 (40.1%) 0.721 (0.646-0.796) Pooled - TNBC 190 84/190 (44.2%) 0.675 (0.597-0.752) Pooled - HER2+ 197 106/197 (53.8%) 0.633 (0.488-0.778) Pooled - HR+HER2- 164 34/164 (20.7%) 0.689 (0.578-0.799) ). Specifically, ATX-N achieved an AUC of 0.644 (CI: 0.577-0.716) on UChicago cohort, 0.734 (CI: 0.598-0.846) on Providence cohort, 0.804 (CI: 0.717-0.881) on Jagiellonian cohort, and 0.714 (CI: 0.624-0.797) on IMPRESS cohort. Notably, the confidence intervals for all cohorts did not overlap with 0.5, indicating that the model’s discriminative performance was better than random classification. In stratified analyses across molecular subtypes, pooled AUCs for TNBC, HER2+, and HR+HER2- patients are 0.675 (CI: 0.597-0.752), 0.633 (CI: 0.488-0.778), and 0.689 (CI: 0.578-0.799), respectively. When stratifying all patients based on ATX-N predictions, the bottom quarter had 15.6% pCR rate whereas the top quarter had 63.8% pCR rate. Conclusions: ATX-N demonstrates promising performance in prediction of pCR in breast cancer across diverse datasets. Larger studies are needed to confirm its efficacy and utility. Citation Format: J. Park, F. Howard, P. Wysocki, A. Przywara, K. Zeng, J. Cappadona, B. Machura, J. Szpor, C. Bifulco, B. Piening, V. Reid, J. Witowski, F. Esteva, K. J. Geras. Ai predicts response to neoadjuvant therapy in breast cancer across diverse cohorts [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-04.