Machine learning models permeate the structure of our lives, from curating playlists to explaining hard concepts in seconds. Beyond convenience, cutting-edge algorithms are finding their way into modern medicine as a powerful potential tool. With such advances, it was released on June 3rd Cell SystemStanford University researchers use machine learning to improve the efficacy and safety of target cell and gene therapy by potentially using our own proteins.
Most human diseases occur, either systematically or locally, due to malfunctioning proteins in the body. Naturally, it is ideal to introduce new therapeutic proteins to cure what is malfunctioning. Almost all therapeutic protein antibodies are designed to be fully human or human-looking, but similar approaches have yet to advance to other therapeutic proteins that work in cells, particularly those involved in CAR-T and CRISPR-based therapy. The latter still carries out the risk of triggering an immune response. To solve this problem, researchers at GAO Lab are now turning to machine learning models.
“In this paper, we raise a question: Don't design processes from the start to avoid an immune response? With advances in calculators, we are trying to predict which changes in proteins can trigger an immune response and only move forward with designs that are unlikely to be rejected by the body.
By combining three independent machine learning algorithms, the team has made great strides towards tools to avoid such immune response problems and efficiently design proteins that maintain function when introduced into the human body.
Design zinc fingers
One way to reduce the immune response to these therapeutic agents is to start with proteins already present in the human body. Therefore, the GAO team selected a small protein called the zinc finger, which is one of the most abundant proteins found in eukaryotes and is involved in regulating gene expression. Because of their ability to naturally bind to human DNA, they are good alternatives to existing technologies like CRISPR. This is more likely to trigger an immune response because it comes from bacteria.
“The most important part of our research is the developments in the design of zinc finger DNA binding domains that can target genomic sites of our choice, while maintaining the low predictive risk of triggering immune responses,” says Eric Ursberg, a doctoral student in chemical engineering and lead author of the paper.
Naturally occurring zinc fingers are bound to specific sequences in the human genome as a result of evolutionary processes. However, to reuse them for cells or gene therapy, the team used the initial algorithm to predict new DNA targets (such as disease-causing genes) that could bind to zinc finger combinations. Zinc fingers are usually linked to recognize long DNA stretches, so the team assembled them into sequences and created new junctions between the individual zinc finger units.
However, there were complications.
“These joints are unnatural and do not occur in our bodies,” GAO said. “This means that the immune system may recognize them as foreigners and respond. ”
The team then utilized the second machine learning algorithm, Maria. Maria was developed by paper co-author Binbin Chen, a graduate student at Stanford, and Ash Alizadeh, professor of medicine in the Moghadam family. Maria was designed to predict the immunogenicity of these zinc protein conjugation to design cancer vaccines.
My preference is that the vaccine is highly immunogenic. This means that the team will use Maria in reverse to screen for conjugation or mutations that avoid immunodetection. The design was considered safer if Maria predicted that the designed zinc fingers were not likely to be seen in the immune system. The combination of these two models produced functional zinc finger arrays, but their effectiveness is limited, perhaps due to the limitations of the algorithm used to predict zinc finger binding sequences.
To further improve the function of further engineered zinc fingers to preserve reduced immunogenicity, the team applied a third algorithm: a powerful protein language model called ESM-IF1.
The machine leads to strengthening
Drawing on training from millions of natural protein sequences, ESM-IF1 functions like a veteran editor, suggesting success rates where single letter genetic adjustment sharpens zinc finger performance.
“In the past, researchers have tried random mutations to improve zinc fingers, which were slow and inefficient. They were also incompatible with Maria's filtering,” Gao said. “This language model allows you to focus on smart, targeted changes.” After proposing a potential mutation in ESM-IF1, the team again ran a modified sequence via Maria to prevent the changes from introducing new immunogenic properties.
“We only made progress with mutations that passed both tests (high functionality and low immunogenicity),” GAO said.
The team compared the original zinc finger protein with a version with an AI attached mutation by assessing both computer-based prediction and lab-based testing to see improvements in performance. The original protein increased the production of human genes two to four times, but increased the production of AI by a further six-fold increase in lab-based testing.
“We took zinc finger engineering to places we had never visited before, conserving its functionality and reducing immunogenicity,” GAO said. Researchers plan to build on this method, aiming to one day end-to-end algorithms that will help design zinc finger gene therapy in humans.
For more information
Alizadeh is also a member of Stanford Bio-X, the Institute of Stem Cell Biology and Regenerative Medicine, the Mathern & Child Health Research Institute (MCHRI), and the Stanford Cancer Institute. GAO is a member of Bio-X, Wu Tsai Human Performance Alliance, Stanford Cancer Institute, Wu Tsai Neurosciences Institute and a faculty member of Sarafan Chem-H.
Other Stanford authors include Josh Tyco, Ph.D. '22. Binbin Chen, PhD '19, MD '21; Michael Bassik, Associate Professor of Genetics in the School of Medicine, Member of Bio-X, Faculty of Stanford Cancer Institute, Wu Tsai Neurosciences Institute, and Sarafan Chem-H. Lacramioara Bintu, assistant professor of bioengineering at the Engineering and Medicine school, is a member of Bio-X.
This study was funded by the National Institutes of Health, Longevity Grants, Stanford ChemSeed Grants, and the Stanford Bio-X Interdisciplinary Initiative Seed Grant Program.
Sign up every other week to read all Stanford Science stories Stanford Science Digest.
