Deep Learning in Protein Engineering: Designing Functional Soluble Proteins

Traditional protein design, often relying on physics-based methods like Rosetta, faces challenges in creating functional proteins with complex structures due to the need for parametric and symmetric restraints. Recent advances in deep learning, particularly with tools like AlphaFold2, have transformed protein design by enabling accurate prediction and exploration of vast sequence spaces. This has led to stable proteins with novel functions and intricate structures. However, designing large, complex protein folds, especially those mimicking membrane proteins in soluble forms, remains difficult. Understanding and expanding the fold space to include soluble analogs of membrane proteins could unlock new functional capabilities in synthetic proteins.

Researchers from several institutions, including the Ecole Polytechnique Fédérale de Lausanne and the University of Washington, have developed a deep learning pipeline to design complex protein folds and soluble analogs of membrane proteins. This approach uses AlphaFold2 and ProteinMPNN to create stable protein structures, including those mimicking membrane proteins like GPCRs, without parametric restraints or extensive experimental optimization. Biophysical analyses confirmed the designs’ high stability and experimental structures showed remarkable accuracy. This method expands the functional soluble fold space, enabling the incorporation of membrane protein functionalities, which could advance drug discovery and other applications.

Researchers have developed a deep learning-based pipeline that integrates AF2seq and ProteinMPNN to design complex protein folds, including soluble analogs of membrane proteins. AF2seq generates sequences to adopt target protein topologies, which ProteinMPNN optimizes for enhanced diversity and solubility. This approach successfully designed intricate structures like IGFs, β-barrels, and TIM-barrels without traditional parametric constraints. Experimental validation showed high stability and accurate structural alignment with the developed models. The pipeline’s success highlights its potential for exploring new protein topologies and integrating functionalities from membrane proteins, advancing drug discovery and protein engineering.

Researchers explored designing soluble analogs of membrane protein folds, which typically have unique structural features. Using the AF2seq-MPNN pipeline, they aimed to solubilize complex folds like claudin, rhomboid protease, and GPCRs. Initial attempts with standard methods failed, but retraining the ProteinMPNN on soluble proteins (MPNNsol) led to successful designs. They achieved soluble, thermally stable proteins with accurate structural alignments for these challenging folds. High-resolution X-ray crystallography confirmed the precision of their designs, showing that these membrane topologies could be converted to soluble forms, revealing their potential for diverse biotechnological applications.

The study extended the design of soluble analogs of membrane proteins to include functional capabilities. Researchers preserved specific functional motifs while solubilizing transmembrane segments, creating soluble versions of human claudin-1 and claudin-4 that retained their natural ability to bind Clostridium perfringens enterotoxin, mimicking their membrane-bound counterparts. They also designed chimeric soluble GPCR analogs incorporating functional domains from the ghrelin receptor and adenosine A2A receptor. These analogs could engage in specific protein interactions, demonstrating the preservation of critical functional sites. This approach holds the potential for designing functional proteins and advancing therapeutic discovery.

The study showcases a deep learning-based computational approach for designing complex protein folds, overcoming traditional challenges. It successfully generated high-quality protein backbones across various topologies without fold-specific retraining, achieving significant experimental success in producing soluble and properly folded designs. Structural validations confirmed precise modeling accuracy, which is crucial for functional protein design. Importantly, the method extended design capabilities to membrane protein analogs, including intricate folds like rhomboid protease and GPCR, demonstrating their solubility and monomeric state in solution. This breakthrough opens avenues for creating functional soluble proteins with native features, essential for accelerating drug discovery targeting membrane proteins, thus significantly broadening the scope of computational protein design.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter

Join our Telegram Channel and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our 46k+ ML SubReddit

🐝 Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...