Representing, predicting, and generating simple and complex peptides
Massina Abderrahmane
Data Scientist, Iktos Inc
Peptides represent a growing therapeutic space between small molecules and biologics. More than 80 peptide drugs have reached the market for a wide range of
diseases, including diabetes, cancer, osteoporosis, multiple sclerosis, HIV infection and chronic pain.
Predicting peptide properties using machine learning methods has gained interest in recent years, with a particular focus on anticancer, antimicrobial or improving permeability properties, using publicly available datasets.
Since 2018, multiple generative AI approaches for peptides have also been proposed. Methods include three popular deep generative model frameworks: neural language models (NLMs), variational autoencoders (VAEs), and generative adversarial networks (GANs).
These predictive and generative approaches have shown interesting performances, but also limitations, especially on the type of amino acids considered, most often restricted to natural amino acids.
In this work, we have developed new representations of peptides, including graph representations at the amino-acid, backbone - side chain and pharmacophore level.
These representations are suitable for peptides constituted of natural residues, but also for complex peptides which contain modified amino acids and cross-links. We applied a circular algorithm to transform these graphs into a vectorial representation and evaluated them for predictive tasks, including permeability. Finally, we proposed generative models based on the new representations proposed here, able to design new peptides with optimized properties.