How Language Processing Algorithms Help Predict Coronavirus Mutations

Natural language processing (NLP) algorithms can now generate protein sequences and predict virus mutations, such as those changes that could help the coronavirus evade the immune system.

In a study recently published in the journal Science, a group of researchers, including experts from MIT, show how they use these algorithms to identify mutations in advance.

The basic idea is that the interpretation of a virus by an immune system is analogous to the interpretation that a human makes of a sentence.

It is a new method to identify and predict those mutations that can cause a virus to escape from the immune system or that does not respond adequately to vaccines and treatments already developed for the treatment of this pathogen.

To carry out this task, the researchers used the linguistic concepts of grammar and semantics. The genetic aptitude of a virus, that is, its characteristics to infect a person, can be interpreted in grammatical terms; whereas the mutations or variations of a virus can be interpreted in terms of semantics.

Thus, for example, changes in the proteins on its surface that make it undetectable by certain antibodies can be read, within this perspective, as an “alteration of its meaning”. In other words, a virus with different mutations of this type can have different meanings and, therefore, may require different antibodies capable of fighting it.

Anticipating mutations could help design treatments and vaccines (Stephanie Lecocq / Pool via REUTERS) 

To do this, the researchers trained the NLP model on thousands of genetic sequences taken from three different viruses: 45,000 unique sequences for a strain of influenza; 60,000 for an HIV strain and between 3,000 and 4,000 for a Sars-Cov-2 strain, as explained in the journal MIT Technology Review.

Knowing in advance which mutations, capable of eluding the responses of the immune system, could take place would help to take better health measures.

The researchers were applying this analysis model to new coronavirus variants, including the UK mutation, the Danish mutation, as well as the South African, Singaporean and Malaysian variants.

For now, they found that all of them have a high potential for immune escape, although this was not tested in the laboratory but arises from predictive analysis based on neural networks. But the model missed another change in the South African variant that has raised concern because it may allow it to not respond to vaccines.

The use of these algorithms helps to predict potential mutations immediately, which would help accelerate processes in the design of health strategies, study of new treatments and development of vaccines. Although it is something incipient, it shines a light on a path that will surely have to be explored.