A Novel Computational Method Simplifies the Engineering of Beneficial Proteins

MIT researchers have unveiled a groundbreaking computational technique aimed at enhancing the engineering of proteins, marking a significant advance in the realm of synthetic biology. Traditionally, scientists have relied on a labor-intensive method involving numerous rounds of random mutations to refine natural proteins, such as those that emit fluorescent light, into more optimized forms. This traditional approach has seen success with proteins like the green fluorescent protein (GFP) but has encountered obstacles with others.

The innovative approach introduced by the MIT team simplifies the prediction of beneficial mutations, utilizing a minimal data set to forecast enhancements in protein functionality. By applying their model, the team successfully predicted mutations for GFP and a protein from adeno-associated virus (AAV) used in gene therapy, opening new avenues for neuroscience research and medical applications.

Ila Fiete, a prominent figure at MIT involved in the study, compares the challenge of protein design to navigating a mountainous terrain obscured by peaks, emphasizing the complexity of transitioning from DNA sequences to functional proteins. The study, a collaborative effort among MIT's brightest, including Regina Barzilay, Tommi Jaakkola, and graduate students Andrew Kirjner and Jason Yim, will be highlighted at the upcoming International Conference on Learning Representations.

The inception of this research stemmed from a desire to optimize proteins for use as voltage indicators in living cells, a venture that could revolutionize how neuronal activity is monitored. Despite extensive research efforts, the desired efficiency in these proteins has remained elusive. This prompted a collaboration between Edward Boyden's lab and Fiete's lab, merging computational prowess with biological inquiry.

The researchers tackled the immense challenge of predicting optimal protein sequences by developing a convolutional neural network (CNN) trained on GFP data. This model constructed a fitness landscape from a limited dataset, illustrating the potential of proteins and guiding the design towards more optimal forms. Through a process of smoothing this landscape, the model was refined to predict highly optimized protein variants, demonstrating significant improvements over the original.

This computational strategy was not only effective for GFP but also showed promise with AAV, optimizing its capsid for DNA delivery. The success with these well-documented proteins suggests the model's broad applicability across various protein engineering challenges.

Looking ahead, the researchers aim to apply their computational technique to optimize voltage indicator proteins, a quest that has confounded scientists for over two decades. With the potential to predict superior proteins from smaller datasets, this approach could revolutionize protein engineering.

The project received support from multiple sources, including the U.S. National Science Foundation and the Howard Hughes Medical Institute, underscoring the collaborative and interdisciplinary effort that underscores this significant scientific advancement.