Â鶹´«Ã½

ONE of the most important features of the immune system is that it can learn and remember, a kind of biological intelligence that is rivalled only by the brain. And just like the brain, researchers are attempting to replicate this intelligence artificially, by creating an AI called Artificial Immunological Intelligence (AII) that will improve our understanding of the immune system and could be used in computing to prevent malicious invaders doing harm there too.

The brain’s intelligence is embodied in connections between fixed neurons, whereas the immune system’s is based on transient interactions between mobile immune cells. The principle is the same, however, and both can be replicated in a computer program.

For the past few years, Johannes Textor at Radboud University in the Netherlands and his colleagues have been building AIIs in the form of computer-based populations of simulated T-cells. Each simulated immune cell has unique properties and they collectively learn through reinforcement, with cells that make a positive contribution to eliminating simulated pathogens replicating, and those that don’t dying out.

Already, AII is helping immunologists make progress on some age-old problems. For example, Textor and his colleague Inge Wortel have built an AII to test a long-standing idea in immunology that explains how T-cells learn what is a threat.

Real-life T-cells have surface receptors that recognise proteins or fragments of proteins called peptides displayed on the surface of other cell types. These peptides are usually the breakdown products of old, malfunctioning proteins inside the cell, which are transported to the surface and displayed for T-cells to inspect. If the T-cells recognise the peptides as “self”, they allow the cell to go on its way. But if a cell is making non-self peptides, because it is infected with a virus or is cancerous, the T-cells attack.

One of the most important lessons the T-cell population has to learn is how to distinguish self from non-self in order to avoid launching autoimmune attacks. The textbook version of this is that young T-cells, called thymocytes, are dispatched to the thymus to face a loyalty test. After being presented with all possible self peptides, any of the cells that attack are eliminated.

There are problems with this, however. The number of self peptides in a human is around a million, whereas estimates of how many a thymocyte could encounter during its two or three-day residency in the thymus are much lower. And it has become clear that masses of rogue T-cells get around the test.

One possible explanation is that this is all a learning process, where thymocytes encounter a subset of self peptides and the system generalises from these. So Textor built an AII of the process and showed that, under the right circumstances, T-cells can generalise from a subset of self peptides and later respond much more strongly to non-self peptides, such as from the Ebola virus, HIV and other pathogens.

Although in its infancy, the use of AIIs goes beyond improving our understanding of the immune system. Textor says he can see applications in a field called adversarial information, where part of a data set, such as search engine results, has been maliciously manipulated and needs to be weeded out. Traditional AI struggles with this, he says, but AII – which, after all, is based on a system that evolved to spot and eliminate intruders – could probably manage it.