Stanford University researchers are using machine learning algorithms to modify human proteins with an eye toward improving the efficacy and safety of targeted cell and gene therapies by reducing the risks of an adverse immune response. Details of the work are published in a new Cell Systems paper titled “Machine-guided dual-objective protein engineering for deimmunization and therapeutic functions.”
The research was led by Xiaojing Gao, PhD, an assistant professor of chemical engineering in the School of Engineering at Stanford and the senior author on the paper. According to the article, Gao’s team used three machine learning models to establish “a workflow for creating deimmunized zinc-finger arrays to target arbitrary DNA sequences.” Through their work, they hope to develop an alternative to nonhuman proteins in CAR-T and CRISPR-based therapies, which can trigger unwanted immune responses.
“We raise the question: Why not design treatments that avoid immune reactions from the start?” Gao explained. “With advances in computational tools, we’re now trying to predict which changes to a protein could trigger an immune response, and only move forward with designs that are less likely to be rejected by the body.”
One way to reduce the risk of an immune reaction is to start by modifying proteins that already exist in the human body. Gao and his colleagues chose zinc fingers, a readily available protein with a critical role in gene expression regulation, that binds naturally with human DNA, making it less likely to trigger an immune response. “The most significant part of our work is our progress in designing zinc finger DNA-binding domains that can target any genomic site we choose while maintaining a low predicted risk of triggering immune responses,” explained Eric Wolsberg, a PhD student in chemical engineering and the lead author of the paper.
Naturally occurring zinc fingers bind to specific sequences in the human genome. To repurpose them for a cell or gene therapy, Gao and his team used one algorithm to predict new DNA targets that could bind to combinations of zinc fingers. Since zinc fingers are typically linked together to recognize longer DNA stretches, the team assembled them into arrays, creating new junctions between the individual zinc finger units in the process.
However, these junctions created a complication. Because they do not occur naturally in the body, the researchers were concerned that the immune system might mark them as foreign and mount a response. To solve this issue, they deployed a second machine learning algorithm that was designed to make predictions about the immunogenicity of the zinc protein junctions in cancer vaccines. They used it to screen for junction designs that would avoid immune detection, as these would likely be safer.
Using both models resulted in functional zinc finger arrays with limited efficacy. To improve the functionality of the designs without compromising their lowered immunogenicity, the team applied a protein language model to make targeted mutations that would sharpen the zinc fingers’ performance. They used the second algorithm to test the edited arrays to ensure that the changes did not introduce new immunogenic properties. “We only moved forward with mutations that passed both tests—high functionality and low immunogenicity,” Gao said.
The scientists then compared the engineered zinc finger proteins with the originals using both computer-based predictions and lab-based testing. Their results showed that the original proteins increased the production of human genes by two- to six-fold, while the AI-enhanced proteins further increased production by two- to six-fold in the lab-based tests.
