In a bold leap for medical and neuroscience research, MIT scientists have devised a computational technique that promises to improve the engineering of proteins, potentially leading to advancements in measuring brain electrical activity. Using this innovative method, the team successfully predicted mutations that enhanced proteins like the green fluorescent protein (GFP), commonly used in biological research, as well as a viral vector protein from adeno-associated virus (AAV), crucial in gene therapy delivery.
The research, backed by an array of commendable sources such as the U.S. National Science Foundation and the DARPA Accelerated Molecular Discovery program, was born out of a collaborative effort from MIT's brightest, including Ila Fiete, Regina Barzilay, and Tommi Jaakkola, among others. The main challenge, as Fiete explained in a statement obtained by MIT News, is the complex journey from DNA sequence to protein function. She compared the challenge to navigating mountainous terrain in search of a river basin — a journey often obscured by the rugged peaks en route.
The technique centers around a convolutional neural network (CNN), which learns from experimental data how to predict more efficient versions of proteins. The CNN created a 'fitness landscape,' a concept depicting the 'fitness' of proteins relative to their sequences. This helped the researchers identify which proteins might have higher functional potential. However, due to the unpredictability of evolutionary pathways, the CNN needed an extra edge to smooth out the fitness landscape's rougher patches, enabling it to find more fruitful mutations with greater ease.
Lead authors Andrew Kirjner and Jason Yim demonstrated this approach's efficacy not only with GFP but extended its application to the viability of AAV's capsid for DNA packaging. "Once we have this landscape that represents what the model thinks is nearby, we smooth it out and then we retrain the model on the smoother version of the landscape," Kirjner told MIT News. He further explained how this allowed for a progressive path from the starting point to the peak, a process otherwise hindered by the unsmoothed landscapes' irregularities.
The implications of this research stretch into neuroscience, where the team plans to refine proteins serving as voltage indicators, crucial for monitoring neuron activity sans electrodes. Success in this field has been limited despite decades of work, but with their computational model, the MIT researchers aim to rapidly accelerate progress. This could mark a new era for in vivo studies and electronic measurement adaptations, expanding the capabilities for understanding and treating neurological conditions. Shahar Bracha, an MIT postdoc working on related research, hopes the method can outperform two decades' worth of manual testing with just a small set of data feeding the ever-learning model.
