Prospective prediction of alisertib-induced Aurora-A kinase resistance mutations in patients with metastatic ER+ breast cancer using an innovative artificial intelligence approach.

3022 Background: Pharmacological blockade of Aurora-A mitotic kinase (AURKA) oncogenic pathway represents a promising therapeutic strategy to inhibit breast cancer plasticity and progression. However, treatment with the selective AURKA inhibitor alisertib is followed by acquired on-target mutations in a subset of patients that are linked to alisertib resistance and tumor progression. Current resistance mechanisms are typically identified retrospectively after therapeutic failure. We evaluated whether Dynaptive, a structure- and evolution-aware artificial intelligence (AI) platform, could prospectively predict treatment-induced AURKA resistance mutations prior to experimental observation. Methods: We identified acquired on-target mutations in patients from the alisertib Phase-II clinical trial. These mutations were cross-matched against Dynaptive predictions that had been generated independently of any experimental resistance data. Dynaptive integrates protein conformational ensembles and evolutionary sequence constraints to predict mutational hotspots and assess their functional and drug-binding impact. Identified mutations were evaluated against large-scale population datasets to determine natural occurrence. Results: Several AURKA mutations observed in patients that showed resistance to alisertib were prospectively predicted by Dynaptive despite being entirely absent from prior population datasets. These mutations were not observed previously in either healthy or cancer cohorts described in the public domain. Significantly, the most frequently occurring AURKA resistance mutation occurred at a residue ranked second among predicted mutational hotspots and were predicted to reduce alisertib binding affinity by ≥10-fold. Across all variants, absence from population datasets and emergence only following treatment support a treatment-induced resistance mechanism. Conclusions: A structure-aware AI platform prospectively identified clinically relevant, treatment-induced AURKA resistance mutations prior to experimental detection. These findings demonstrate the feasibility of predictive modeling to anticipate resistance mechanisms and support the development of next-generation AURKA inhibitors optimized to overcome mutations-driven resistance and to improve therapeutic efficacy, safety and durability.