Stanford Researcher Behind AI-Designed Viruses to Explain to MIT TR How a Model Writes Genomes

MIT Technology Review will host a subscriber-only online conversation on October 16 with Samuel King, a Stanford PhD student. In 2025, he used a generative model to propose genetic blueprints for viruses. It remains one of the most prominent examples of language-model architectures moving from text into biology. The announcement is worded cautiously: this is not yet AI-created life, but it may well be the next step. The publication's senior AI reporter will moderate the discussion.

What exactly King did

The announcement offers few details, so I am relying on the 2025 preprint prepared at Stanford and the Arc Institute. King worked in Brian Hie's group on models from the Evo family, which are trained on DNA sequences rather than words. The researchers chose bacteriophages as the target: viruses that infect bacteria and are not dangerous to humans. Their genomes are short, about 5,400 nucleotides, and encode just over a dozen genes.

The team fine-tuned the model on a family of related phages and asked it to generate new genomes in their entirety. The candidates then passed computational filters, were synthesized, and were tested on E. coli. According to the paper, of roughly 300 designs, 16 turned out to be functional. In lab tests, some of them competed with the natural prototype, and some outperformed it. The figures are worth double-checking against the original source, but the order of magnitude is a success rate of a few percent, which is not bad for de novo genome design.

Why this is more significant than protein generation

Protein models such as AlphaFold or RFdiffusion work at the level of a single molecule. Here, the model proposes a set of genes, their order, and regulatory regions that must work together inside a host cell. That is a system-level task.

Key parameters of the experiment:

• Target: bacteriophages, genome of about 5.4 thousand nucleotides
• Model: Evo in versions 1 and 2, fine-tuned on related genomes
• Output: several hundred generated genomes
• Result: 16 viable variants, according to the preprint
• Validation: DNA synthesis and infection of a bacterial culture

The limitations are also clear. A phage genome is orders of magnitude simpler than a bacterial one, and the model drew on close natural examples. So far, this is recombination and improvement of what is already known, not the invention of new biology. Several steps remain before a synthetic organism from scratch, and the main one is assembling a cell, not just its DNA.

Biosecurity as the central topic

Work like this always raises the dual-use question. The Evo 2 developers excluded eukaryotic virus genomes from the training data in advance, so that the model would perform worse on human pathogens. But data filtering is no guarantee: fine-tuning on the right family can restore some of those capabilities. The conversation with King is interesting precisely because the author will explain where he sees the risk himself and which barriers he considers effective.

What readers should do

For teams working with biological models, the practical takeaway is simple: look not at the demo, but at reproducibility in the wet lab. The share of viable variants and the speed of the generate–synthesize–test cycle matter more than polished sequence-likelihood metrics. For those following regulation, it is worth tracking how requirements for screening synthetic DNA orders develop: the line between a digital design and a physical organism runs through synthesis. It makes sense to look for a recording or recap of the conversation after October 16, and to read the preprint by King and Hie beforehand, so you can ask the authors substantive questions.