Application of machine learning to predict monomer retention of therapeutic proteins after long term storage.

An important aspect of initial developability assessments as well formulation development and selection of therapeutic proteins is the evaluation of data obtained under accelerated stress condition, i.e. at elevated temperatures. We propose the application of artificial neural networks (ANNs) to predict long term stability in real storage condition from accelerated stability studies and other high-throughput biophysical properties e.g. the first apparent temperature of unfolding (Tm). Our models have been trained on therapeutic relevant proteins, including monoclonal antibodies, in various pharmaceutically relevant formulations. Further, we developed network architectures with good prediction power using the least amount of input features, i.e. experimental effort to train the network. This provides an empiric means to highlight the most important parameters in the prediction of real-time protein stability. Further, several models were developed by a different validation means (i.e. leave-one-protein-out cross-validation) to test the robustness and the limitations of our approach. Finally, we apply surrogate machine learning algorithms (e.g. linear regression) to build trust in the ANNs decision making procedure and to highlight the connection between the leading inputs and the outputs.