AI Tool Reveals a Way to Predict Cancer Survivability Using MHC-1 Protein Patterns

By analyzing the nearly 6,000 Major Histocompatibility Complex-1(MHC-1) proteins in the human body, researchers from Arizona State University (ASU) were able to classify them into 11 different types according to their electrostatic signatures. Their new human leukocyte antigen (HLA)-Inception tool, created by AI and machine learning, revealed that individuals with a wider array of MHC-1 proteins were more likely to survive cancer. The work was published in Cell Systems.

diagram of HLA Inception Tool
The HLA Inception Tool categorizes MCH-1 proteins according to electrostatic properties to predict immune responses.

The mix of MHC-1 proteins is unique for every individual, with preferences for the types of protein fragments they interact with. The team wanted to see if they could devise a way using MHC-1 patterns to predict which peptides will bind to which MHC-1 molecules and observe whether a person’s immune defenses may recognize pieces of threatening viruses and cancers.

“Starting from the known sequence and 55 available protein structures of MHC-1 molecules we wanted to figure out the relationships for all 6,000 MHC-1 molecules from this small data set,” explains ASU professor Abhishek Singharoy, senior author of the study. “We came up with a biophysical neural network from sequences, electrostatic potentials, and structural properties based on the 55 structures that were known.” Electrostatic charges relate to the varied charges on the surface of the MHC-1 proteins.

With HLA-Inception, inputting an MHC-1 sequence generates its corresponding electrostatic property. “Once we did that, then we could predict the electrostatic properties of all the 6,000 MHC-1 complexes for which the structural properties were not known,” Singharoy adds.

Further, the HLA Inception tool revealed that there are 11 different types of MCH-1 complexes among the 6,000 total proteins. Singharoy explains that each of the 6.6 billion humans in the world could then be classified into one of the 11 different MCH-1 categories. “What this means is that one can use this for personalized medicine prediction,” he says. “As soon as I place that human being in one of these level classes, I will be able to know what the immune repertoire of this person is, and what the epitopes are to which this human being is more or less susceptible to now.” This information can then be used to predict whether the protein fragments, or peptides, MHC-1 are monitoring are self or foreign invaders (non-self).

The study revealed that most people have a mix of four to six MHC-1 categories based on electrostatic properties. Their theory was that the immune system will be more resilient against disease, the larger the mix of MHC-1 categories.

To assess that theory, the team used patient-derived data of 314 individuals treated with immune checkpoint inhibitors for melanoma and non-small cell lung cancer. “We found that people who were surviving that medication better were the same people whose electrostatic properties are more diverse than the others who are not,” says Singharoy.

Looking ahead, the team is investigating machine learning with another physics-based lens to stratify individual immune responses. “We are looking at catch bonds—where the bond is strengthened under force—at immune synapses,” Singharoy adds. “Perhaps we can profile human beings based on the behavior of the catch bonds which possibly will be complementary to electrostatics to learn more about the spaces our eyes cannot see.”

Also of Interest