New AI Model Learns to Design Proteins Unlike Any in Nature

A protein's function comes from its structure, and its structure comes from its sequence of amino acids, the molecule's building blocks. Many methods for designing new proteins — including ones meant to bind disease-causing molecules — work in two steps: first choosing a structure, then using a machine-learning model to generate amino acid sequences that could fold into it. In nature, many different sequences can fold into the same structure, so the challenge for AI is learning to recognize that there is more than one useful answer, not just the one sequence evolution happened to select.

"For years, the field has measured success by asking whether a model can reproduce the protein sequence that evolution happened to select — our work shows that this isn't the best metric for protein design," said Amy Keating, head of MIT's Department of Biology and senior author of the study, published in PNAS. The new framework, called PottsMPNN, incorporates the physical principles that govern protein structure and stability, improving both sequence generation and the ability to predict how mutations will affect a protein's stability.

The most widely used protein design model today was released in 2022 and has not been surpassed since, said graduate student Foster Birnbaum, the study's lead author. His approach added deliberate variation, or "noise," to protein structures during training, which reduces the tendency of a model to overly mimic native sequences and increases the diversity of structures it can generate sequences for. PottsMPNN also uses a pairwise approach that accounts for physical interactions between all 20 possible amino acids at each pair of positions in a protein, which the researchers say is a key reason it more accurately models the relationship between a sequence and its stability.

"If we're thinking about a completely novel, designed structure, there would be no native sequence to compare it to," Birnbaum said. "What we actually care about is how likely the generated sequences are to fold into the desired structures, how well the model understands the sequence-energy landscape, and how well it can predict the effect of mutations on the stability of the protein." The team found that as the model relied less on native sequences, its predictions of structural compatibility and stability improved, including for entirely novel proteins.