Noviota has published a research note examining a striking recent experiment in generative biology: When a Genome Model Produces a Working Virus, looking at what happens when a genome model moves from predicting biology to designing it.
From prediction to design
Much recent work applying artificial intelligence to biology has focused on prediction — given a protein sequence or a genome, can a model infer its structure or function? The experiment Noviota examines asks a different question: can a model propose an entirely new biological object? Researchers used genome-scale generative models to produce candidate genomes related to a small bacteriophage, ΦX174, then synthesised and experimentally tested them. Sixteen of the candidates proved to be viable bacteriophages capable of infecting E. coli.
As Noviota’s piece points out, the result matters less because a model “learned to manufacture life” and more because it exposes a sharper question: what does a computational model actually need to represent about a biological system in order to produce something that works?
Most candidates did not work — and that is the interesting part
A viral genome is not just a sequence of characters. Even a small one encodes an interconnected system whose components must together allow replication, expression, assembly, host interaction and propagation — meaning a generated sequence can look statistically plausible while being biologically useless. Of the hundreds of candidate genomes taken forward for testing, only a minority produced viable phages, which Noviota argues is at least as informative as the successes: generating sequences that resemble biology is not the same as generating a biological system capable of functioning.
Noviota Limited is the life-science portfolio of JDGS Holdings.
Questions & Answers
What experiment does Noviota’s research note examine?
Genome-scale generative models used to produce candidate genomes related to a small bacteriophage, ΦX174, which were then synthesised and experimentally tested.
How many of the candidates actually worked?
Sixteen of the candidates proved to be viable bacteriophages capable of infecting E. coli.
Why does the result matter, according to Noviota?
Less because a model “learned to manufacture life” and more because it exposes a sharper question: what does a computational model need to represent about a biological system to produce something that works?
Why is it significant that most candidates did not work?
A generated sequence can look statistically plausible while being biologically useless — only a minority of the hundreds of candidate genomes taken forward for testing produced viable phages.
What does this reveal about the difference between prediction and design?
Generating sequences that resemble biology is not the same as generating a biological system capable of functioning — the interesting shift is from predicting biology to actually designing it.
Read the full research note on Noviota →
Related project: Noviota