How we discriminate between self and nonself is a central question in immunology and is particularly relevant in cancer immunology, where cells that start as self can acquire features visible to the immune system as they evolve.
When approaching this problem from AI/ML the impulse is to train a classifier on molecules that the immune system senses as nonself versus those from the self proteome and learn their underlying differences. In our work "How different are self and nonself?" with
@andimscience, our amazing Jonathan Levine,
@wbialek and great colleagues, we show viruses and self proteins are basically using the same language model, so the usual impulse is subverted.
As a result, the immune system can train on self and learn sequences close to it, a kind of "overfitting". We show that in antigen databases, consistently, peptides the immune system senses are often only one mutation from self. A cancer neoantigen that is just one mutation from self is therefore not a strange antigen at all.
Proud to publish this work in PRX Life with the American Physical Society. This project was a great deal of fun, opened my thinking up on the central question of our lab...
👇👇👇
@MSKCancerCenter
A statistical physics framework that models peptidomes across species shows that self and nonself peptides are nearly one and the same, implying that the immune system benefits by targeting antigens near those represented in the organism’s own proteome.
go.aps.org/40tAuIl