Recurrent Neural Networks Predict Future Peptide Aggregation for Drug Development.

Abstract

Physical stability of an active pharmaceutical ingredient (API) is a key consideration in the development of a pharmaceutical drug. Solution conditions such as pH, excipient concentrations, and storage temperatures can impact the physical stability of a therapeutic peptide in formulation. Optimizing these conditions is a critical activity in achieving a higher stability of a therapeutic peptide product. A Thioflavin T (ThioT) fluorescent reporter assay is widely used to measure the aggregation of peptide products. ThioT kinetic assays are used to predict the propensity of fibril formation by using ThioT curves for a peptide stored in a solution. However, there is no analytical relationship that can be used to relate the physical stability for different formulation conditions, resulting in execution of large-scale stability assays that require significant resources for pharmaceutical companies. Therefore, there is a need to develop new artificial intelligence (AI) methods to predict future ThioT curves in a fast and cost-effective manner. Here, we combined an experimental measure of time-varying conformational states from ThioT assays with AI models to predict peptide aggregation in different formulation conditions during drug development. We formulated the peptide aggregation problem as "language translation" in natural language processing, wherein the sequence of aggregation states at earlier time points was used to predict (or "translate") the aggregation states for future time points. We developed a new sequence-to-sequence long short-term memory (LSTM)-based recurrent neural network (RNN) model to predict entire ThioT curves at future time points (6 and 12 months) using data sets from initial and 1 month ThioT curves for different conditions. We achieved an excellent average mean absolute error (MAE) of 2.04 for the model, which was used to predict and experimentally validate ThioT curves for a 6 month time point. In contrast to the LSTM, the multilayer perceptron (MLP) baseline model showed a higher MAE of 5.17. However, at the 12 month time point, with limited training data, both models achieved comparable results with average MAEs of 4.25 and 4.45 for LSTM and MLP, respectively. Therefore, we conclude that LSTM models can be used to predict future ThioT curves only using the initial and 1 month ThioT curves as input. We believe that the use of recurrent neural network models will benefit the pharmaceutical industry to predict and explore the formulation landscape for future physical stability measurements of APIs based on short-term stability data.