Researchers have developed a deep learning model that outperforms Google's artificial intelligence system for predicting peptide structure

Researchers at the University of Toronto have developed a deep learning model called PepFlow that can predict all possible forms of peptides - chains of amino acids that are shorter than proteins but have similar biological functions.

PepFlow combines machine learning and physics to model the different folding patterns a peptide can take based on its energy landscape. Peptides, unlike proteins, are very dynamic molecules that can adopt different conformations.

"Until now, we have not been able to model the full range of peptide conformations," said Osama Abdin, first author of the study and a recent graduate of the Department of Molecular Genetics at the Donnelly Center for Cellular and Biomolecular Research at the University of Southern California. "PepFlow uses deep learning to determine accurate and precise peptide conformations within minutes. This model could be the basis for drug development by creating peptides that act as binding agents."

The role of a peptide in the human body directly depends on how it folds, as its 3D structure determines the way it binds and interacts with other molecules. Peptides are known to be highly flexible, adopting a wide range of folding patterns, and are therefore involved in many biological processes of interest to researchers developing therapeutics.

"Peptides became the subject of the PepFlow model because they are very important biological molecules that are naturally very dynamic, so we need to model their different conformations to understand their functions," said Philip M. Kim, principal investigator of the study and a professor at the Donnelly Center. "They are also important as therapeutic agents, as seen with GLP1 analogs such as Ozempic used to treat diabetes and obesity."

Peptides are also cheaper to produce than their larger protein counterparts, said Kim, who is also a professor of computer science in T University's Faculty of Arts and Sciences.

The new model extends the capabilities of Google's leading artificial intelligence system Deepmind for protein structure prediction, AlphaFold. PepFlow can outperform AlphaFold2 by generating a range of conformations for a given peptide, something AlphaFold2 was not designed to do.

What sets PepFlow apart are the technological innovations that underpin it. For example, it is a generalized model inspired by Boltzmann generators, which are highly sophisticated physics-based machine learning models.

PepFlow can also model peptide structures that take unusual shapes, such as a ring-like structure resulting from a process called macrocyclization. Peptide macrocycles are currently a very promising area for drug development.

Although PepFlow improves on AlphaFold2, it has its limitations because it is the first version of the model. The study authors noted a number of ways to improve PepFlow, including training the model with explicit data on the solvent atoms that dissolve the peptides to form a solution and constraints on the distance between atoms in ring-like structures.

PepFlow was built in such a way that it can be easily expanded to take into account additional considerations, new information, and potential uses. Even in its first version, PepFlow represents a comprehensive and efficient model with potential for further development of therapies that depend on peptide binding to activate or inhibit biological processes.

"Modeling with PepFlow allows us to understand the real energy landscape of peptides," says Abdin. "It took two and a half years to develop PepFlow and one month to train it, but it was worth it to move to the next frontier, beyond models that predict only one peptide structure."