Active-learning guided and high-throughput synthesis of membrane proteins

Introduction:: Membrane proteins underlie numerous human pathologies and are critical components of various drug-delivery vehicles, including nanoparticles and exosomes. However, the study and use of these proteins remain limited from our inability to synthesize them. Current methods rely on brute force synthesis using living cells, which often fail despite relentless optimization. To circumvent the shortcomings of cell-based synthesis, we create a novel high-throughput platform to rapidly synthesize and insert membrane proteins into liposomes. The system leverages cell-free protein synthesis, automation, microfluidics, and protein engineering. This method is highly adaptable, allowing us to vary lipid types, chemical environments, and chaperone proteins (Figure 1A). Leveraging our unique platform, we generate the first big data on membrane protein synthesis, with over 30,000 data points. Furthermore, based on the new data, we generate a novel active-learning model that enhances the speed of membrane-protein synthesis by at least 100fold and enables the synthesis of membrane proteins that have not been synthesized in the literature. Our new platform, data, and predictive model pave the way for new discoveries in protein pathologies and facilitate the development of novel membrane-protein therapeutics and nanotechnologies.




Materials and Methods:: To synthesize membrane proteins, we utilized a custom droplet printing robot capable of generating hundreds of cell-free reactions in a combinatorial array. Each reaction used an E. coli-based whole cell extract produced in house. The membrane proteins were synthesized outside of supplied liposomes to allow for co-translational outside-in insertion into the membrane. Protein insertion was determined by comparing the raw fluorescence of an engineered fluorescent folding reporter in the presence and absence of liposomes. We modulated various parameters, including lipid type, liposome concentration, small molecule concentrations, plasmid sequence, and chaperone proteins, generating a search space of over 16,000 unique conditions for each protein.


To transfer the knowledge gained from several membrane proteins to new ones, we constructed a membrane protein structure database using AlphaFold structures and Uniprot transmembrane helix annotations. These annotated structures were used to extract features from each membrane protein to capture how they interact with its environment. We combined these structural features with the reaction conditions to form the input layer of a deep neural network, which we trained using experimental data. We trained numerous networks with varying folds of the data and parameters, resulting in an ensemble of different models. We used these models to predict the results of numerous hypothetical reactions spanning the entire search space. We selected the best reaction conditions for the next round of experiments based on the average and standard deviation of these predictions.




Results, Conclusions, and Discussions:: The utility of the folding reporter was assessed using size exclusion chromatography and flow cytometry on numerous membrane proteins (Figure 1B). We used this reporter to evaluate the synthesis conditions for more than 30 membrane proteins, with the total number of reactions tested exceeding 30,000. The optimization of these proteins has resulted in the synthesis of numerous proteins at concentrations sufficient for structural investigation using cryo-electron microscopy. The membrane proteins include a calcium modulator and a protein lifeguard, which are associated with several diseases yet have no known structures, as well as the cell adhesion protein CD47 useful for targeted therapeutics. The optimal conditions for each membrane protein differ substantially, supporting the need for a protein-specific statistical model. The ensemble of neural networks captures the response of each membrane protein to given synthesis conditions (Figure 1C). The ensemble also predicts higher-performing reaction conditions for most membrane proteins tested in subsequent experiments (Figure 1D). The model incorporates more than 400 protein structural features extracted from the constructed database. These features allow for transfer learning between membrane proteins, as the model is trained on all membrane proteins. When these features are removed from the model, the prediction accuracy drops substantially, indicating their utility in relating membrane proteins to their response to specific reaction conditions.

This dataset constitutes the largest systematic evaluation of numerous membrane proteins in a wide array of reaction conditions to date. We have discovered numerous novel insights into the interactions between membrane proteins and their environment. In addition, several membrane proteins have been successfully synthesized in our work despite other groups failing to do so, and many of the proteins have no experimentally derived structures. Our work has transformed membrane-protein synthesis from an “art” to a high-throughput design-test-iterate workflow that leverages big dataset and active-learning approaches. Our work will accelerate the basic study of membrane proteins and enable the design of membrane-protein therapeutics for broad biomedical applications.


Acknowledgements (Optional): : We thank NIBIB and NIGMS for supporting the work. We also appreciate the advice and support of the members of Tan Lab.


References (Optional): : Contreras-Llano, L. E., Meyer, C., Liu, Y., Sarker, M., Lim, S., Longo, M. L., & Tan, C. (2020). Holistic engineering of cell-free systems through proteome-reprogramming synthetic circuits. Nature communications, 11(1), 3138.

Wang, J., Deng, K., Zhou, C., Fang, Z., Meyer, C., Deshpande, K. U. A., ... & Pan, T. (2019). Microfluidic cap-to-dispense (μCD): a universal microfluidic–robotic interface for automated pipette-free high-precision liquid handling. Lab on a Chip, 19(20), 3405-3415.