Recurrent Neural Network Model for Constructive Peptide Design.

Area of Science:

• Biochemistry
• Computational Biology
• Artificial Intelligence

Background:

• Recurrent Neural Networks (RNNs) excel at identifying patterns in sequential data like amino acid sequences.
• Generative models can learn from existing data to create novel instances, offering potential for bespoke peptide library design.

Purpose of the Study:

• To develop and evaluate a generative long short-term memory (LSTM) recurrent neural network (RNN) for de novo peptide design.
• To assess the model's ability to generate novel peptide sequences with predicted antimicrobial activity.

Main Methods:

• Training RNNs with LSTM units on a dataset of helical antimicrobial peptide sequences.
• Utilizing the trained model to generate new amino acid sequences for de novo peptide design.

Main Results:

• The generative LSTM RNN model was employed for de novo sequence generation.
• 82% of generated sequences were predicted to be active antimicrobial peptides, outperforming random sampling (65%).
• Generated sequences demonstrated closer proximity to the training data distribution compared to manually designed helices.

Conclusions:

• LSTM RNNs are capable of constructing novel amino acid sequences within their defined applicability domain.
• This approach facilitates prospective applications in peptide and protein design, reducing the need for extensive library enumeration.