One of the key challenges in developing therapeutic protein drugs is that novel proteins designed by AI often trigger immune responses once they enter the human body, thereby hindering their practical application. To tackle this issue, the research team led by Jiang Kaiyi has introduced a generative protein design model called MESSI. Built on the foundation of inverse folding, this model takes immune risk into account. By employing a discrete diffusion model and a forward-looking feedback mechanism, it steers the model to steer clear of options with high immune risk during sequence generation.

The team chose three types of proteins with distinct functions—luciferase, PD-1 binding protein, and urate oxidase—to validate the model. The results demonstrated that the variants designed by MESSI not only reduced the predicted immune risk but also maintained or even enhanced the original protein's function. Notably, some urate oxidase variants exhibited 2.7 times the activity of the original protein.

At present, this technology has not yet undergone immunogenicity validation in the human body and is still a considerable distance from clinical application. Looking ahead, it holds promise for the development of customized low-immunogenicity protein drugs tailored to different individuals' HLA types. This research builds upon the previous work of the Jiang Kaiyi team, which is firmly rooted in real-world needs and strives to make biological systems comprehensible, predictable, and designable.