TY - JOUR T1 - Machine Learning for Translation of Published Radiochemical Methodologies JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 13 LP - 13 VL - 62 IS - supplement 1 AU - Eric Webb AU - Jay Wright AU - Liam Sharninghausen AU - Kevin Cheng AU - Allen Brooks AU - Melanie Sanford AU - Peter Scott Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/13.abstract N2 - 13Objectives: The development of Cu-mediated radiofluorination and radiocyanation of aryl organo-metal reagents (e.g. aryl boron and stannanes) has expanded chemical space compatible with 18F and 11C (1-7). However, it remains a significant challenge to predict reaction efficacy in the radiosynthesis of novel imaging agents with these methodologies using chemical structure and existing literature precedent alone. To maximize radiochemical yield of a new radiotracer, each available set of conditions (and associated precursor) needs to be screened, resulting in challenges and costs with respect to time and materials, slowing the development of novel radiotracers. This is particularly cumbersome and cost-prohibitive when developing automated radiochemical processes with multi-Curie quantities of radionuclide. We present a predictive machine learning model for Cu-mediated radiofluorination and radiocyanation that allows for prediction of radiochemical outcome for a given substrate and set of literature conditions. This serves as the first component of a radio-labeling platform capable of identifying optimal conditions and outcome for a variety of 18F and 11C labeling methodologies. Methods: Literature substrates were identified along with conditions for radiofluorination or radiocyanation. The aryl transmetallating reagents were decomposed into feature components by DFT calculation of the substrates. Structures were drawn in Gaussview6 to generate Gaussian09 input structures.(8) These structures were geometry optimized and frequency calculations were performed to verify optimization minima (B3LYP-LANL2DZ). 1H, 13C and heteronuclear NMR spectra were also calculated for the optimized structure. Calculated electronic features were supplemented with steric characterization by buried volume calculations using SambVca 2.1.(9) A home-written Python script was used to automatically parse and extract the calculated features into a CSV. Random shuffling and splitting of the dataset into training (70%) and test set (30%) was implemented in Python using Sci-Kit Learn.(10) Standard scaling was used to reduce variation in values with respect to units. Regression models were generated using several linear, neighbor-based, decision tree, ensemble, and neural network algorithms. Feature importance was determined by permutation and retraining on the test set. Results: Although linear and nearest neighbor models were poorly predictive for radiochemical incorporation (Test R2 < 0.50, RMSE: > 22.0%, MAE: > 17%), while simple decision trees resulted in a modest boost in predictive capacity (R2=0.44, RMSE: 20.7%, MAE: 15.9%), a significant boost was observed using ensemble methods. Random forest (R2=0.61, RMSE: 17.3%, MAE: 13.5%) and gradient boosted regression algorithms (R2=0.77, RMSE: 15.3%, MAE: 10.6%) proved the most effective at predicting radiochemical incorporation. Permutation importance identified n-butanol as a critical additive followed by the energy of the HOMO and the ipso-carbon NMR shift as most predictive of outcome for a given substrate. Conclusions: We described a machine learning regression model for the prediction of radiolabeling reactions. We anticipate this model will speed the development of new radiotracers by facile identification of conditions and prediction of efficacy. Efforts are underway to extend this approach to SNAr radiofluorination and other labeling methods, as as well as for prediction of those strategies for complex, late-stage substrates. Acks: NIH (R01EB021155) Refs: 1. Mossine et al. Org Lett 2015,17:5780; 2. Tredwell et al. ACIE 2014,53:7751; 3. Taylor et al. JACS 2017,139:8267; 4. Zischler et al. Chem Eur J 2017,23:3251; 5. Makaravage et al. Org Lett 2016,18:5440; 6. Makaravage et al., Org Lett 2018,20:1530; 7. Ma et al. Chem. Comm. 2017,53:6597; 8. Gaussian 16, Revision A.03, Gaussian, Inc., Wallingford CT; 9. Falivene et al. Nat Chem 2019,11:872; 10. Pedregosa et al. J Mach Learn Res 2011,12:2825. ER -