Abstract
Ensuring the safety of patients using pharmaceuticals and medical devices through chemical characterization requires accurate estimation of extractables and leachables to ensure tolerable risk from unintentional exposure to these chemicals. However, this is complicated by variability in chemical responses, particularly with mass spectrometry methods. High-quality relative response factor predictions provide the opportunity for both expedited and more accurate quantitation in extractables and leachables analyses. In-silico prediction models were developed to test a wide variety of compounds, which were selected utilizing a physicochemical property coverage approach. Three neural network models and their associated sub-models were applied to both positive and negative ionization modes for liquid chromatography mass spectrometry and gas chromatography mass spectrometry methods. Mean absolute errors across all methods were 0.47 for training data and 0.65 for out-of-sample data, indicating high predictability of response factors for the compounds chosen by the model and a low likelihood that the data were overfit. Model performance was evaluated using a series of chemicals outside the training data set. When tested, predictive accuracy was greater than 60% of known values for 44 of the 49 chemicals tested (90%), and direct comparison to an expert judgment approach that uses surrogates shows the predictions outperform that approach. This proof-of-concept work shows that sophisticated neural network modeling of response factor data is a potential solution for response factor variation.