Recurrent Neural Networks Predict Future Peptide Aggregation for Drug Development

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Biotechnology

Background:

• Physical stability of active pharmaceutical ingredients (APIs) is crucial for drug development.
• Solution conditions significantly impact therapeutic peptide stability.
• Thioflavin T (ThioT) assays measure peptide aggregation but large-scale studies are resource-intensive.

Purpose of the Study:

• To develop artificial intelligence (AI) methods for predicting peptide aggregation and ThioT curves.
• To enable fast and cost-effective prediction of peptide physical stability.
• To reduce the need for extensive, resource-heavy stability assays in pharmaceutical development.

Main Methods:

• Formulated peptide aggregation prediction as a natural language processing "language translation" problem.
• Developed a sequence-to-sequence long short-term memory (LSTM)-based recurrent neural network (RNN) model.
• Used initial and 1-month ThioT assay data to predict future (6 and 12 months) ThioT curves.

Main Results:

• The LSTM model achieved an excellent average Mean Absolute Error (MAE) of 2.04 for predicting 6-month ThioT curves.
• The LSTM model's predictions were experimentally validated.
• Both LSTM and Multilayer Perceptron (MLP) models showed comparable MAEs at the 12-month time point with limited data.

Conclusions:

• LSTM models can accurately predict future ThioT curves using short-term stability data (initial and 1 month).
• Recurrent neural network models offer a valuable tool for the pharmaceutical industry.
• These AI models can accelerate the exploration of formulation landscapes for API physical stability.