Machine learning algorithms decode hidden data in the immune system for disease detection

Your immune system is worth lifelong worth regarding the threats you encountered – the Biological Rorodex of the Villain. Often, the perpetrator is the virus and bacteria you have conquered. Others are vaccine-like undercover agents administered to trigger a protective immune response in the form of healthy tissues caught in immunological crossfires.

Now, researchers at Stanford Medicine have mined this rich internal database to devise a method to diagnose diseases as diverse as diabetes's COVID-19 response to influenza vaccines. They envision this approach as a way to screen multiple diseases simultaneously, but they can also optimize machine learning-based methods to detect complex, difficult-to-diagnose autoimmune diseases such as lupus.

In a study of about 600 people, some studies of health, COVID-19 or other infectious diseases including autoimmune diseases including lupus and type 1 diabetes, the researchers developed algorithms for immunological diagnosis It was a remarkable success, called MAL-ID for machine learning. Those who had something based solely on the sequence and structure of B and T cell receptors.

“The diagnostic toolkit we use today doesn't make much use of the internal records of the immune system that it encounters,” said postdoctoral scholar Dr. Maxim Zaslavsky. “But our immune system is constantly monitoring our bodies with B. T cellsfunctions like a molecular threat sensor. Combining information from the two main arms of the immune system gives you a more complete understanding of the immune system's response to diseases and the pathways to autoimmunity and vaccine responses. ”

Zaslavsky and Erin Craig are the lead authors of the study, published February 20th. Science. Professor of Pathology Scott Boyd, MD, PhD, and Associate Professor of Genetics and Computer Science, Dr. Ansul Kundage, PhD, is a senior author of his research.

In addition to supporting the diagnosis of tricky diseases, MAL-ID can track responses to cancer immunotherapy and subcategorize disease states in ways that help guide clinical decision-making. Those believe.

“Some of the conditions we were seeing can vary significantly on a biological or molecular level, but we'll explain them in a wide range of terms that don't necessarily explain the specific immune system responses.” said Boyd, who co-directs Sean. N. Parker Center for Allergy and Asthma Research. “MAL-IDs can help you identify subcategories of specific conditions and can provide clues as to what treatments are most useful for someone's condition.”

Deciphering protein language

In the following dot approach, scientists used machine learning techniques based on large-scale language models. We modeled what underlies what relies on the threat recognition receptors of immune cells called T cells and the business end of antibodies (also known as receptors). It was made by another type of immune cell called B cells. These language models look for patterns in large datasets, such as text in books and websites. With sufficient training, these patterns can be used to predict the next word in a sentence.

In this study, scientists apply large language models trained with proteins, supplying millions of sequences to the model from B and T cell receptors, and using them to determine by the model We used it to group receptors that share important properties – which may suggest similar binding preferences. In doing so, Trigger gives you a glimpse into why a person's immune system mobilizes – an army of T cells, B cells, and other immune cells equipped to attack actual perceived threats Run around.

“The sequences of these immune receptors are very diverse,” Zaslavsky said. “This variability helps the immune system detect virtually anything, but it also makes it difficult to interpret what these immune cells are targeting. In this study, this vastly diverse interpretation has been achieved. We asked if we could decipher the record of encounters with these diseases. Information with some new machine learning techniques.

B cells and T cells represent two separate arms of the immune system, but the way in which they create proteins that recognize the infectious agents or cells that need to be eliminated are similar. In short, certain segments of DNA within the cell's genome are randomly mixed and matched – sometimes, with additional mutations to lift things up, trillion uniqueness Create a coding region where you can generate antibodies. (for B cells) or cell surface receptors (for T cells).

The randomness of this process means that these antibodies or T-cell receptors are not tailored to recognize specific molecules on the surface of the invader. But their eye-opening diversity ensures that at least a few people unite in almost every foreign structure. (Automatic immunity, or attacks by the immune system on the body's own tissues, are usually avoided by the conditioning processes T and B cells.

The act of binding stimulates the cells and carries out a full-scale attack by creating more of its own. The increased prevalence of cells using receptors consistent with similar three-dimensional structures then provides biological fingerprints of the disease or condition that the immune system is targeting.

To test their theory, researchers have included over 16 million B-cell receptor sequences and over 25 million T-cell receptor sequences from 593 people with any of six different immune states. We assembled the dataset: healthy controls, people infected with SARS-COV-2 (Covid-19) or viruses that cause HIV, people who recently received the flu vaccine, lupus or type 1 diabetes (both autoimmune diseases) ) people suffering from ). Zaslavsky and his colleagues later used a machine learning approach to find commonality among people in the same state.

“We compared other properties, including the frequency of segment use, the amino acid sequence of the resulting protein, and the way the model represents the “language” of the receptor,” Boyd said.

Together with T and B cells

Researchers provide the most relevant information about lupus and type 1 diabetes, and B-cell receptor sequences when identifying HIV or SARS-COV-2 infection or recent influenza vaccinations We found it to be the most beneficial. However, in all cases, combining T and B cell outcomes increased the ability of algorithms to accurately classify people by disease state, regardless of gender, age or race.

“Traditional approaches can sometimes have a hard time finding groups of receptors that appear different but recognize the same target,” Zaslavsky said. “But this is where large-scale language models are great. They can learn grammar and context-specific cues of the immune system, just as they have mastered the grammar and context of English. Thus, MAL-IDs can generate an internal understanding of these sequences. Give them insights you've never had before.”

Researchers have developed MAL-IDs in just six immunological conditions, but the algorithm can be adapted quickly to identify immunological signatures specific to many other diseases and conditions. I'm assuming it. They are particularly interested in autoimmune diseases like lupus.

“Patients can struggle for years before they get diagnosed. Still, the names that give these diseases are like umbrellas that overlook the biological diversity behind complex diseases.” Zaslavski said. “If we can use Mal-ID to unravel the heterogeneity behind lupus, or rheumatoid arthritis, it would be very clinically shocking.”

MAL-IDs may also help researchers identify new treatment goals for many conditions.

“The beauty of this approach is that it works even if the molecules and structures that the immune system is targeting are not completely clear at first,” Boyd said. “We can get information just by looking at similar patterns in the way people respond, and by digging deeper into these responses, we can uncover new directions for research and treatment. Maybe I will.”

Switzerland's Institute of Tropical Public Health, University of Basel, Oklahoma Medical Research Foundation, University of Pennsylvania, University of Cincinnati, Cincinnati Children's Hospital Medical Center, Duke University, Ikaan School of Medicine, Cincinnati Children's Hospital, Swedish Medical Center, University of Washington, Institute of Systems Biology , Harvard Chan School of Public Health, Beth Israel New York University's Deacones Medical Center and the American Lupus Foundation contributed to this work.

This study was funded by the National Institutes of Health (Grants R01AI130398, R01AI127877, U19AI057229, U54CA260518, U19AI167903, 5R01 EB001988-16, AI100645, UM1 AI144371, AI 101093 AI-086037, AI-48693, R01AI153133, R01AI137272, 3U19AI057229–17W1 COVID SUPP2, AR07375, UM1AI1444292, NIDDK P30DK116074, U54CA260518303030303030303030303030303030303030303175771-01, R01 CA264090-01, U19 AI057229 and 1U54CA26051), National Science Foundation, The Burroughs Wellcome Fund, The Sunshine Foundation, Henry Gustav Floren Trust, and Eva Grove.