How the immune system mimics machine learning to identify threats

The research, published by Assistant Professor Hannah Meyer and Associate Professor Saket Navlakha, addresses a long-standing immunological mystery: how T cells learn to avoid attacking healthy tissue when they only ever test against a tiny fraction of the body's proteins. By applying AI concepts to biological data, the team found that T cells interact with roughly 240 antigen-presenting cells out of a random sample of 2,000 during their training in the thymus. This limited exposure is sufficient because the immune system meets two specific conditions: the training data in the thymus mirrors the abundance of peptides found elsewhere in the body, and T-cell receptors are inherently cross-reactive.

This efficiency allows the immune system to correctly delete 90% of self-reactive T cells despite only encountering 10% of the body's self-peptides. Beyond basic biology, the team’s model accurately reproduced features of autoimmune polyendocrine syndrome type 1, suggesting that failures in this generalization process may drive autoimmune diseases. Navlakha refers to this interdisciplinary approach as ImmunoAI, framing the immune system not as a biological curiosity, but as a sophisticated processor solving fundamental machine learning problems.