Researchers at the University of Toronto’s Donelly Centre have developed an advanced artificial intelligence (AI) model, known as PepFlow, that can foresee the variety of shapes that peptides can form accurately. The different shapes a peptide can take are vital as they determine how it interacts with different molecules in the body, influencing its biological function.
Predicting the structures of proteins and peptides has been a long-standing challenge due to the complicated mathematics involved, which makes it suitable for machine learning. Other AI models, like DeepMind’s AlphaFold2, utilize deep learning to predict a protein’s likely 3D structure based on its amino acid sequence. While successful for proteins, AlphaFold2 encounters limitations with highly flexible molecules such as peptides.
Unlike AlphaFold2 which predicts only one structure, PepFlow has the ability to predict the entire energy landscape of a peptide. The energy landscape signifies every potential shape a peptide can form and how it transitions between these different conformations. Documented in a study in Nature Machine Intelligence, PepFlow uses AI models influenced by Boltzmann generators to comprehend the physical principles dictating how a peptide’s chemical structure defines its possible shapes. This capability helps PepFlow predict the structures of peptides with unusual features accurately, which is crucial for understanding how peptides function in varying biological contexts.
The capability to accurately predict peptide structures is significant for the development of peptide-based therapeutics. Peptides are essential biological molecules and are naturally dynamic. PepFlow can aid drug development by designing peptides that act as binders. The study’s lead investigator, Philip M. Kim, highlights that peptides are beneficial as therapeutics, as seen by GLP1 analogues like Ozempic, used to treat diabetes and obesity.
Peptide drugs have several advantages over traditional small-molecule drugs and larger protein-based therapeutics. They’re more specific in their actions, have lower toxicity than small-molecule drugs, and are cheaper and easier to produce than larger protein drugs. With PepFlow, the discovery and development of new peptide-based medicines could potentially be accelerated by enabling the design of peptides with therapeutic properties.
This development follows the release of EvolutionaryScale ESM3, a frontier generative model focusing on proteins. Abdin, the study’s first author, emphasized that although it took over two years to develop PepFlow and a month to train it, it was a significant step to advance beyond models that only predict a single structure of a peptide.
