AI Is Getting Better At Writing Genetic Code. Is Designed Life Closer Than We Think?

Getting your Trinity Audio player ready...

Researchers have developed an AI model that can predict the impact of genetic mutations and even write new genetic code.

The future of medicine is genetic. Most drugs work by binding to receptors on the surface of cells, which activates various signalling pathways triggering the increased or decreased expression of certain genes. This means that drugs generally have to target existing systems that already fulfill a biological purpose, and thus trigger many unwanted outcomes besides the intended effect, limiting the dose that can be given or simply making the drug completely unviable. Gene therapy, on the other hand, delivers genetic material that the cell uses to produce a specific protein, or shuts down the production of a disease protein. This bypasses the complex and messy molecular pathways that usually lie in-between a drug and its intended outcome by speaking to the cell in a language it understands – the genetic code.

The possibilities with gene therapy are endless, from fixing genetic mutations and personalised cancer treatments, to more sci-fi concepts like engineering new genes to enhance human health and lifespan. However, for the full potential of gene therapy to be achieved, we need a way to predict the impact of a given piece of genetic code, and to reverse engineer genetic code based on an intended outcome. This is where artificial intelligence is likely to play a significant role. The ability of machine learning algorithms to take large datasets, detect patterns beyond human understanding and make accurate predictions based on those patterns makes them ideal for this task.

Presenting Evo2

In this study published in preprint, researchers present Evo 2, a huge AI model trained on genetic data spanning all of evolution. Unlike their first model Evo 1, which was trained only on prokaryotes (organisms whose cells contain free-floating genetic material without a nucleus or organelles), Evo 2 is also trained on data from eukaryotes – organisms whose cells contain a nucleus. This includes human genetic data. In total, 9.3 trillion base pairs were included in its training set.

Predicting mutations and their impact

Despite being trained exclusively on genetic data and not on biological outcomes of different genetic sequences, Evo2 was able to predict whether a given genetic mutation would impact essential biological functions, which is a first for eukaryotes. It was also able to predict how dangerous some mutations would be. For example, it was able to predict whether mutations in the BRCA1 gene, which can cause breast cancer in humans, would be dangerous or not. This is even more remarkable considering that the training data only contained one human genome, meaning that these preditictions originated from a general model of how genetic code is supposed to look, not human data specifically. Evo2 appeared to ‘understand’ genetics from the ground up.

Building new sequences

The researchers also tested Evo2’s ability to build new genetic sequences from scratch. The researchers didn’t test the functionality of the generated sequences by creating physical versions: this would have been challenging, not to mention the ethical and safety considerations of letting a non-human intelligence have a go at designing new lifeforms! However, they did assess the viability of the code using other algorithms trained on natural sequences, and found that it had characteristics of natural and functional genomes. The researchers believe that, with further training that includes biological function, their AI could produce functional genetic code.

The implications

This kind of research brings us closer to making great strides forwards in medicine, but also brings with it immense threats. Let’s start with the good news. These AI models could significantly accelerate our understanding of the genetic basis of human diseases, including the ageing process. As we age, our cells accumulate countless genetic mutations that don’t directly cause cancer but do disrupt function in more subtle ways, contributing to age related deterioration. AI could help us to understand which of these mutations are most important, how exactly they lead to ageing, and potentially offer solutions. Further down the line, it may be possible for an AI to take genetic sequencing data and create personalised gene therapies for a specific individual. Indeed, it is hard to see how such personalised medicine could ever become widespread and affordable without the use of AI. The fact that we seem to be moving towards this is great news for humanity.

The dangers

What’s not such great news for humanity is the potential to use this technology for biological warfare or terrorism, for example by using AI to generate genetic code for new human viruses. The researchers were mindful of this risk, so they did not include any human pathogens in Evo2’s training data. However, since the model is open source, nothing is stopping someone else from training a new model that does include human pathogens. As was said for many years concerning a tangential taboo – the genetic modification of humans – someone somewhere is going to try this eventually. We’d be wise to focus our efforts on countermeasures – to which this very same technology may be essential – at the same time as upholding ethical principles when it comes to AI research.

This research is still in preprint, so it has not yet been peer-reviewed by other scientists. However, the research institutes involved in this research are reputable, and the fact that Evo 2 is open source means that the veracity of the claims should be relatively easy to confirm.