Abstract P2-01-23: A 140-gene Machine Learning Classifier Predicts Survival and Response to Chemotherapy and Immunotherapy in 2500 TNBC

Abstract Background: TNBC is molecularly heterogeneous, showing varied responses to therapeutic agents (Txt). We developed a 140-gene classifier (TNBC-ICR) based on random-forest (RF) to cateogorize TNBC into four biological subgroups. It was trained and tested on 551 TNBC (Zhu ESMO Congress 2023), and had better precision in predicting time to recurrence with anthracycline-based (A) txt compared to TNBC-Baylor subtypes (Burstein Clin Cancer Res, 2015) in TACT2 trial (NCT00301925). Here, we assessed the genomic characteristics and clinical value of TNBC-ICR classifier to predict response for standard of care and emerging Txt. Methods: TNBC-ICR, consisting of 1000 decision trees, assigned each case to immune-enriched (IM), luminal-AR (LAR), mesenchymal-like (MES) or highly proliferative basal-like (BL-Prolif) based on RF probabilities linked to these subgroup features. Association with genomic characteristics and prognosis were tested in The Cancer Genome Atlas (n=155, TCGA Nature, 2012), METABRIC (n=269, Curtis Nature, 2012), and SCAN-B (n=571, Saal Genome Med, 2015). Association with pathological complete response (pCR) was tested across 4 neoadjuvant (NAT) clinical studies: 1) a combined cohort of MDACC, ISPY-1, LBJ/IN/GEI and USO (MDACC, n=188, Hatzis JAMA 2011), 2) CALGB 40603 (n=389, NCT00861705), 3) BrighTNess (n=482, NCT02032277) and 4) ISPY-2 (n=364, NCT01042379). These studies, all including Paclitaxel (P), assessed the additional benefits of Bevacizumab (Bev), Carboplatin (Carbo), PARP inhibitor (V) or Pembrolizumab (Pem), followed by A and Cyclophosphamide (C). Gene expression data (GE) was available for all studies, with batch correction applied to reduce platform effects before using the TNBC-ICR classifier. Statisical analyses included multivariable Cox regression models to assess the hazard ratio (HR), chi-squared tests to compare genomic characteristics and pCR rates among the subgroups and logistic regression models to assess the significance and odds ratio (OR) of tumor probabilities to IM, LAR, MES and BL-Prolif related to pCR. Results: Using whole exome sequencing data from TCGA and METATRIC (n = 306), TP53 was the most frequently mutated gene, observed in IM (88%), MES (76%), BL-Prolif (85%) and LAR (68%, p = 0.008), with PIK3CA mutations in LAR (41%) and the others (7%, p < 0.001). The pCR rates for IM, LAR, MES and BL-Prolif were as follows: MDACC (P+AC), 33% (18/54), 10% (3/30), 25% (9/36) and 48% (29/60), p = 0.002; CALGB 40603 (P+Carbo+AC+Bev), 67% (76/103), 56% (30/54), 36% (31/86) and 54% (73/136), p < 0.001; BrighTNess P+Carbo+V arm: 69% (51/74), 29% (8/28), 62% (34/55) and 43% (34/80), p < 0.001; P+Carbo arm: 73% (29/40), 37% (7/19), 62% (13/21) and 48% (20/42), p = 0.03; in P arm: 43% (16/37), 17% (3/18), 27% (8/30) and 34% (13/38), p = 0.21; ISPY-2 P arm: 22% (6/27), 50% (5/10), 12% (3/26) and 14% (3/22), p = 0.06; P+Pem arm: 100% (8/8), 60% (3/5), 40% (2/5) and 55% (6/11), p = 0.098; In P arms, pCR rate was generally low but addition of either Carbo or V increased pCR rates in IM and MES subtypes and addition of Pem over P increased pCR rates for IM and BL. There was no consistent association of TNBC-Baylor subtypes to pCR in these studies. Statistical modelling of TNBC-ICR classifier on txt response showed that probability to IM-subgroup predicted pCR (OR = 3.52, p < 0.001) in NAT and better recurrence-free survival in adjuvant chemotherapy-treated subgroups of SCAN-B (HR = 0.38, p = 0.04) and METABRIC (HR = 0.3, p = 0.04). Conclusion: Tumors with higher IM-features had increased pCR rates after NAT. We validated the reproducibility and clinical validity of TNBC-ICR classifier in predicting differential response to therapies, including taxane and immunotherapy, demonstrating its potential as a practical clinically relevant integrative biology-driven machine learning algorithm. Citation Format: Xixuan Zhu, Orsolya Sipos, Katherine A. Hoadley, Jane Bayani, Lucy Kilburn, John M.S. Bartlett, Judith Bliss, David Cameron, Andrew Tutt, Maggie Chon U Cheang. A 140-gene Machine Learning Classifier Predicts Survival and Response to Chemotherapy and Immunotherapy in 2500 TNBC [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-01-23.