BIO-AI: TransforMD: A Multiscale Machine Learning Platform for Explainable Protein Engineering with Molecular Dynamics Simulations

Protein engineering is an essential tool in a wide range of biotechnology applications, from manufacturing of valuable chemicals or materials, to unlocking new energy sources to improving the function of genes and proteins caused by mutations, to name a few. Because proteins are very complicated molecules, it can be difficult to predict how their properties or functions will change when they are mutated. The rise of artificial intelligence (AI) tools has recently begun to make powerful strides towards solving this problem, but existing AI models generally rely on highly simplified representations of proteins that don’t take into account their motions or flexibility. This is an important problem because scientists have long agreed that motion is essential for understanding how proteins work. This project develops a new tool that allows scientists to incorporate molecular motion into their protein AI models, which will improve their accuracy and make the models more useful for protein engineers. These molecular motions are learned from computer simulations of proteins that are based on physical laws, but we will develop a software tool that will put the knowledge learned from those simulations in the hands of researchers who aren’t experts in advanced computing. The software will also help researchers to better understand the reasons why a given change or mutation has a given effect on the protein, which will improve our understanding of how proteins function and could also be useful for teaching students about protein engineering. A key reason that molecular simulations have been neglected for training protein mutation effect prediction AI models is that it is difficult to perform simulations that accurately measure target protein properties such as catalytic rate constants, binding affinities, or fold stabilities. These complex properties are emergent from many individual motions over long timescales and need to be studied with computationally expensive methods. What’s more, the changes in these properties upon mutation are often too small to accurately measure with simulations. Here, we instead propose to use general data about the relationship between protein structure and dynamics to provide critically important context to AI models trained to predict mutation effects based on other types of data, thereby enabling the use of simple, high-throughput simulations to train models. Furthermore, this provides a means by which mutation effects observed on the benchtop can be traced back to specific molecular behaviors, facilitating testable mechanistic hypotheses to support basic and applied protein research. To facilitate the broad uptake of this approach, a core outcome of this project is a software platform that we call TransforMD, suitable for use by researchers who are not experts in computational methods, that handles the collection of MD simulation data, integration with other available data types into custom AI models, and interpretation and application of results. The project contributes significantly to the national priorities of artificial intelligence and biotechnology. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.