Traditional protein design, which relies on physics-based methods like Rosetta can encounter difficulties in creating functional proteins due to parametric and symmetric constraints. Deep learning tools such as AlphaFold2 have revolutionized protein design by providing more accurate prediction abilities and the capacity to analyze large sequence spaces. With these advancements, more complex protein structures can be explored and synthetic proteins can be given new functional capabilities.
Despite these developments, designing large protein structures, particularly membrane proteins, remains a challenge. A team of researchers from institutes such as the Ecole Polytechnique Fédérale de Lausanne and the University of Washington set out to develop a deep learning pipeline to design complex protein folds and soluble analogs of membrane proteins. Through AlphaFold2 and ProteinMPNN, the team was able to produce stable protein structures and demonstrated the approach’s potential for protein design and drug discovery.
The researchers paid particular attention to the design of soluble membrane protein analogs due to their unique structural properties. The research team utilized their AlphaFold2 and ProteinMPNN pipeline to develop soluble versions of complex protein folds. Initially unsuccessful, the teams managed to create successful designs when training ProteinMPNN on soluble proteins. These newly developed proteins maintained their solubility and passed structural accuracy tests, illustrating their potential for biotechnological applications.
Throughout the research process, the researchers also explored the possibility of maintaining specific functional capabilities within the newly created solubilized proteins. Soluble versions of human proteins claudin-1 and claudin-4 were created, which still retained their ability to bind with the Clostridium perfringens enterotoxin, a trait present in their membrane-bound counterparts. This opens up possibilities for the development of therapeutic treatments.
The research as a whole demonstrated the potential of a deep learning approach to overcoming traditional difficulties in designing complex protein structures. The developed computational approach yielded high-quality protein backbones across different topologies without the need for retraining—significantly advancing the field of computational protein design and potentially accelerating drug discovery oriented toward membrane proteins. This substantial development broadens the scope of protein engineering.
For further information, readers are invited to check out the published research paper and follow their updates on Twitter, join their Telegram Channel and LinkedIn Group, subscribe to the newsletter, and join their subreddit for machine learning.
