Scientists tested every possible single-letter mutation in a virus. The world's best biological AIs still got many consequences wrong.

Here's the full story, via Nature.

"Few biological systems have been studied as exhaustively as the phage ΦX174, or phi X." Its genome "became the first ever whole genome to be sequenced" in the 1970s, "the first to be chemically synthesized" in the 2000s, and "the first viral genomes designed by artificial intelligence were versions of, you guessed it, phi X." If any virus should be predictable by AI, it's this one.

"Lehner and his colleagues created more than 44,000 variants in the virus's genome by making all possible changes to individual nucleotides (three modifications each, one at a time) and all individual amino acids of its proteins (19 modifications to each)." To measure the effects, they "cultured thousands of viral variants together with Escherichia coli... for 80 minutes, enough time for two or three infection cycles," letting successful variants multiply and harmful ones die off.

"Half of the single-nucleotide mutations, and 60% of those altering amino acids, were harmful to the phage, a much higher proportion than he expected." And despite phi X being "generally thought to be completely optimized for laboratory conditions, a handful of mutations further improved its fitness."

"Cutting-edge AI systems for biology research that have shown promise in identifying harmful mutations struggled to predict the effects of the changes in the phage. The findings underscore the need for more, and better, experimental data to power these biological-AI tools."

"'Even in this super well-studied system, we can't explain why one-quarter of the mutations kill the virus,' says Ben Lehner." Of the harmful amino-acid mutations, "around half were suspected to have disrupted interactions with other proteins... One-quarter were in amino acids buried deep in a protein, possibly compromising structural integrity. The remaining ones were a mystery. 'There's hundreds of mutations in here where we haven't got a clue what they're doing,' says Lehner."

AI models are increasingly proposed for designing genes, proteins, and complete viral genomes. This study, using the single most exhaustively studied virus in biology, demonstrates that AI-generated biological plausibility is not the same as experimental truth, and that gap matters more, not less, as the stakes of AI-assisted biological design grow.

Source: Nature (Ewen Callaway): https://www.nature.com/articles/d41586-026-02609-y

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter