Machine Learning-Guided Investigation of Electrophilic and Nucleophilic [77Br]Br Radiobromination Reactions
Shelbie J. Cingoranelli, Samantha Boisvert, Casey J. McCarthy, Annika E. Tharp, Jason Witek, Hong Beom Lee, Taylor R Johnson, E. William Webb, Allen F. Brooks, Jonathan W. Engle, Melanie S. Sanford, Paul A. Ellison, Peter J.H. Scott.
Objectives: Bromine-77 and Bromine-76 constitute a promising theranostic radionuclide pair and are increasingly being explored for nuclear medicine applications. [77Br]Br is well-suited for targeted radiotherapy applications through Auger electron and gamma emission and its 57-hour half-life enables sufficient time for target localization.1 Despite this potential, radiobromination remains comparatively underdeveloped relative to other radiometal and radiohalogen systems. Here, we investigate electrophilic and nucleophilic radiobromination reactions using 77Br and apply machine learning approaches to identify reaction features associated with successful radiolabeling and to guide future reaction optimization.
Methods: An initial screen looking at order of addition, solvent, oxidant, and additives identified favorable reaction parameters as a baseline for a substrate scope investigation. A total of 212 77Br-radiolabeling reactions were evaluated, specifically 124 electrophilic and 88 nucleophilic reactions. Automated machine learning workflows using the H2O AutoML2 framework were used to evaluate multiple supervised machine learning models, including tree-based, support vector machine, and k-nearest neighbor approaches for prediction of favorable radiochemical conversion (RCC). Model performance was evaluated using a repeated nested 5-fold cross-validation, with inner cross-validation for hyperparameter tuning and repeated outer folds to estimate predictive performance. Unsupervised machine learning methods were applied to investigate underlying data structure and feature relationships. Model predictions were compared with experimental results for selected bioactive substrates.
Results: Machine learning analyses identified reaction features and conditions associated with successful electrophilic and nucleophilic radiobromination. Preliminary model performance was modest and feature analysis further identified several properties associated with improved performance including fraction Csp3, localized lipophilic surface area, the third Kier shape index, branching topology (Kier-Hall connectivity index) and number of carbonyls. Across multiple model architectures, copper-mediated nucleophilic radiobromination consistently demonstrated higher predicted likelihoods of successful RCC compared to electrophilic reactions. Additional experimental conditions were identified using Bayesian optimization strategies.
Conclusions: This work establishes an integrated experimental and computational machine learning framework for investigation of radiobromination and provides insights into factors influencing successful 77Br-radiolabeling. These results posit nucleophilic radiobromination as a promising strategy for future radiobromine theranostic development and support continued radiobromine reaction optimization.
References: [1] Kassis, A. I.; Adelstein, S. J.; Haydock, C.; Sastry, K. S. R.; McElvany, K. D.; Welch, M. J. Lethality of Auger Electrons from the Decay of Bromine-77 in the DNA of Mammalian Cells. Radiation Research 1982, 90 (2), 362. [2] LeDell, E; Poirier, S. H2O AutoML: Scalable Automatic Machine Learning. 7th ICML Workshop on Automated Machine Learning (AutoML), 2020.