Abstract PS1-13-06: Personalised Therapy in Triple Negative breast cancer (TNBC), evaluating predictive performance of Bayesian AI Digital Twins

Abstract Background: Despite multiple therapeutic strategies, approximately 40% of women with stage 2-3 Triple Negative breast (TNBC) cancer die. Biomarkers are sparse and treatment related toxicities can be significant, especially when drugs fail. Objective: Evaluate FarrSight®-Twin, a predictive tool for treatment response in TNBC based on integrating multi-modal clinico-pathological and genomic data, and assess its ability to prospectively identify patients unlikely to achieve a pathological complete response (pCR) and benefit less from neoadjuvant chemotherapy. Methods: VISION, an observational retrospective clinical study (n=200 RMH, UK) analysed the performance of the predictive algorithm FarrSight®-Twin. Study population included stage 2-3 TNBC, treated with neoadjuvant chemotherapy+/- immunotherapy (IO). Participants were recruited into Arm A (pCR) or Arm B if they had residual disease (non-pCR). Digital or virtual twins of individuals were created using their clinical data, cancer stage and tumour molecular data (WES (n=66), RNA seq n=(43), WES + RNAseq (n=42)) from an FFPE diagnostic biopsy. Treatment response and overall survival were predicted for each individual. Performance and accuracy were assessed by comparing predictions against real outcomes. Results: In this pre-planned analysis, 34 women were recruited into Arm A and 60 into Arm B with a median age of 51 (24-77yrs). 87% stage 2 (n=82) and 13% stage 3 (n=12).Neoadjuvant treatments administered were: anthracycline-taxane (33%), anthracycline-taxane-platinum (34%) or anthracycline-taxane-platinum-IO (21%) or Other. pCR was used to assess treatment response. pCR rates were highest for anthracycline-taxane-platinum-IO (50%). Similar pCR rates were reported for anthracycline-taxane (35%) and anthracycline-taxane-platinum regimens (31%). “Other” regimens consisting of de-escalated regimens following on treat toxicities and these women had pCR rates of 27%.Overall, 95% of women were alive at the 1 year time point (1 death; n=4 not reached), which dropped to 73% (7 deaths; n=19 not reached) alive for the study population.In the first interim analyses we present FarrSight®-Twin accuracy for predicting pCR using the leave one out approach. Following inputs were used: age, tumour type, clinical tumour size (cT), clinical nodal stage (cN) and molecular data from a diagnostic FFPE breast biopsy. Conclusion: FarrSight®-Twin integrates molecular data from routine diagnostic FFPE samples with limited clinical information, making it both scalable and suitable for incorporation into standard care pathways. In stage II-III TNBC, it accurately predicts an individual patient’s probability of non-response or suboptimal response (i.e., non-pCR) to neoadjuvant chemotherapy, an outcome observed in 50-70% of this population. Citation Format: N. Somaiah, R. Kow, S. Anbalagan, I. Roxanis, R. Chauhan, S. Kannambath, L. Gothard, S. Nimalasena, J. Noble, A. Johns, E. Misirlioglu, B. Plummer, A. Biankin, I. Babina, M. Griffiths, U. S. Asghar. Personalised Therapy in Triple Negative breast cancer (TNBC), evaluating predictive performance of Bayesian AI Digital Twins. 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 PS1-13-06