Columbia and Harvard researchers have engineered a strain of E. coli whose ribosomal small subunit runs on 19 amino acids instead of the 20 used by every known organism, replacing isoleucine across 20 of 21 genes in a 10,000-base stretch of the genome. The redesigned cells grow at about 70% the rate of normal E. coli when one stubborn gene is left untouched, and around 60% when the full set is swapped. The work, published April 30, 2026, leans heavily on AI protein-design software and AlphaFold 2 to suggest sequences that human biologists would not have written.
The genetic code — three DNA bases per amino acid, 20 amino acids total — is shared by essentially all life and is assumed to predate the last common ancestor. The Columbia and Harvard team set out to test whether life actually needs all 20, picking isoleucine because an analysis of the E. coli genome showed it was the amino acid most often swapped for a different one in related proteins from other species.
Isoleucine sits in a family of three branched, hydrophobic amino acids alongside leucine and valine, which made valine a natural substitute. The first test took 36 essential E. coli genes and replaced every isoleucine with valine. 22 of those swaps killed the cells, but 17 genes tolerated the change — including one where isoleucine was removed from 45 separate positions along the chain.
Key facts
- 01Columbia and Harvard researchers used AI protein-design tools to remove isoleucine from 20 of 21 ribosomal small-subunit genes in E. coli.
- 02The redesigned strain grew at about 60% the rate of unedited E. coli and held the changes across 400 generations with 20–30 incidental mutations.
- 03In a preliminary screen of 50 ribosomal genes, 18 tolerated an isoleucine-to-valine swap, 19 grew more slowly, and 13 swaps were lethal.
- 04Iterative runs across four protein-design packages fixed 25 of 32 problem proteins; structural redesign rescued 4 of the remaining 5.
- 05Outputs were validated with AlphaFold 2, the Nobel-winning protein structure model from Google DeepMind.
The ribosome, a protein-and-RNA complex that translates messenger RNA into amino acids, became the focused target. The team ran isoleucine-to-valine swaps on 50 individual ribosomal genes: 18 worked cleanly, 19 grew more slowly, and 13 were lethal. They then pointed deep-learning protein-design software at the 32 genes with reduced fitness, iterating across four different packages until 25 of them produced viable redesigns with no isoleucine.
“After redesigning 20 of the 21 small-subunit genes to drop isoleucine entirely, the engineered E. coli grew at roughly 60% the rate of an unedited cell across 400 generations.”— Jaeden Schafer
For the 5 holdouts, the researchers forced changes at the isoleucine site and let the software redesign neighboring amino acids in the 3D structure to compensate. That trick rescued 4 of the 5 problem proteins. AlphaFold 2 was used throughout to validate that the new sequences would still fold correctly.
Replacing the genes one by one was a stepwise march along the 10,000-base stretch encoding the 21 proteins of the small subunit. The first 10 swaps went in without trouble. By 17 of 21, growth had slowed; at 18 the cells died. Working back from the other end produced the same result, with both paths converging on a single problem gene called rplW.
Leaving rplW alone and replacing the other 20 genes produced cells that grew at roughly 70% the rate of unedited E. coli. The software, it turned out, had compensated for isoleucine changes in rplW by deleting short stretches of nearby amino acids — a fix that worked in isolation but clashed with the rest of the redesigned subunit.
The team then brute-forced rplW. They asked the design software for alternative amino acids at each of the 4 isoleucine positions and tested every combination — 16 designs in all. One of them produced a fully isoleucine-free small subunit. The resulting strain grew at about 60% the rate of the original and held its design across 400 generations, picking up 20–30 incidental mutations, none of which reintroduced isoleucine.
The authors argue that prior to the last common ancestor of life on Earth, organisms experimented with various genetic codes and probably used a mix of proteins and catalytic RNAs to run their metabolisms. Stripping a present-day organism back toward a 19-amino-acid code is a way to probe what chemistry an earlier code could actually support. They are explicit that none of this would have been tractable a few years ago: "It's not clear that any of this would have been possible without the heavy use of AI tools."
The caveat is what the AI cannot do. Different design packages produced very different sequences for the same protein, which the authors take to mean "they are exploring different regions of the space of possible sequences" — but they cannot confirm that, because the models do not explain themselves. In one case the software redesigned an entire alpha helix around a single isoleucine for reasons the researchers could not reconstruct. As they put it, "these software packages are tools: they let us do things that would otherwise not be possible, but they don't actually help us understand all that much."
That gap is becoming the defining tension in AI-for-science. AlphaFold 2 and its protein-design descendants are unblocking experiments that would have taken decades of rational engineering, and the Columbia and Harvard ribosome is a clean demonstration: 20 of 21 genes rewritten, isoleucine gone, cells still alive. But the same paper concedes the team is reverse-engineering its own results, guessing at why the model chose what it chose. The next frontier in interpretability — work like Goodfire's Silico, which we covered earlier this week — matters less for chatbots than for the biology labs already running on AI suggestions they cannot fully read.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
