[Introduction] AI-written "Life Code" has become a reality! Today, Stanford and the Arc Institute unveiled a breakthrough, using bacteriophage ΦX174 as a template to generate a genome using AI for the first time. Sixteen of these genes successfully killed E. coli and even knocked out drug-resistant bacteria, making this a truly remarkable "ChatGPT moment" in life science.
For the first time in human history, a fully functional genome was generated using AI!
In 1977, biochemist Frederick Sanger and others completed the first genome sequencing in history - bacteriophage ΦX174.
More than 40 years later, Stanford teamed up with the Arc Institute team to use AI to generate the phage genome for the first time, starting with ΦX174.
One of the phage genomes designed by AI looks like this:
Evo-Φ36
Simply put, bacteriophage ΦX174 is a virus that infects E. coli. It can accurately kill bacteria but is harmless to the human body.
In the past, designing a genome was not an easy task and required considering numerous factors, which limited the progress in the field of synthetic biology.
To this end, Stanford and other teams came up with a "secret weapon" -
Based on training on millions of genomes, the DNA language models Evo 1 and Evo 2 can learn the complex features of genomes at an unimaginable scale.
Its working principle is similar to ChatGPT, and it is specifically designed to process DNA.
Paper address: https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1
They synthesized 285 genomes using bacteriophage ΦX174 as a template.
It was finally shown that the 16 genomes can effectively inhibit host growth, not only accurately killing specific E. coli, but also not accidentally harming other strains.
Some AI-designed phages replicate faster and are more competitive than the original versions, and can even deal with drug-resistant bacteria that natural phages have difficulty dealing with.
What does the success of this experiment mean?
It marks a major breakthrough in the field of AI in "synthetic biology"——
For the first time, it was successfully verified that AI can completely generate a phage genome with biological functions.
This not only expands the boundaries of human design of life, but also provides new alternative therapies to address health challenges such as "antibiotic resistance".
For the first time in history! AI generates a "complete" genome
In their latest technical blog post, the core team breaks down in detail the secrets to successfully designing the first batch of AI-generated genomes.
Whether designing a single gene or an entire genome, it is an extremely challenging problem.
Calculated in terms of the history of genetic information storage systems, the genome has existed for approximately 4 billion years, and the DNA genome has existed for approximately 3.5 billion years.
In February of this year, the Arc Institute demonstrated that the Evo "family" of genome-based models can successfully generate single proteins or complex multi-component systems, such as CRISPR-Cas complexes.
But designing the entire genome is a whole new battlefield!
Because the core challenge of genome design lies in complexity: multiple genes interact with each other and maintain a delicate balance to ensure replication, host specificity and evolutionary adaptability.
These challenges simply do not exist in single protein design.
To address this challenge, the Stanford Arc Institute team developed a series of innovative technologies, including:
- a gene annotation pipeline tailored for overlapping reading frames;
- Systematic fine-tuning and cue-word engineering strategies for sampling from genomic language models;
- A novel screening solution designed for synthetic phage genomes
ΦX174, a relay race spanning half a century
To generate a synthetic genome, you also need a reliable starting point.
Bacteriophage ΦX174 - a tiny viral genome of only 5386 nucleotides encoding 11 genes.
Left: Micrograph of ΦX174 phage; Right: 3D structure of a single ΦX174 phage
Its size is just within the affordable range of current DNA synthesis costs, but it is also complex enough to test the capabilities of genome design.
However, the overlapping structure of the ΦX174 gene creates a harsh test case:
A single mutation may affect multiple proteins, requiring them to function properly under multiple constraints.
In addition, ΦX174 encodes multiple regulatory elements and recognition sequences that work together precisely to ensure that the phage can be correctly packaged and replicated in the host cell.
The ΦX174 genome is a relay race spanning half a century.
In 1977, Fred Sanger and his team's research made it the first completely sequenced human genome.
In 2003, Craig Venter and his team synthesized it completely for the first time using chemical methods, proving that the genome can be constructed from scratch.
Now, in 2025, the team has used ΦX174 as a template to create the first AI-generated genomes.
This evolution marks the core capabilities that define modern genomics: first learning to read (sequencing), then writing (synthesis), and now designing (AI generation).
ΦX174 genome
AI "Genome Factory" solves the overlap puzzle
As mentioned above, ΦX174 overlaps genes, making it difficult to detect with standard tools, as it can only identify 7 of the 11 genes.
To this end, researchers created a dedicated annotation process:
By combining open reading frame (ORF) searches and homology comparisons with phage protein databases, we were able to successfully identify all genes and even predict some A* genes.
The tool proved effective when evaluating thousands of AI-generated sequences.
The researchers set a bottom line: the generated genome had to predict at least seven matches for the natural ΦX174 protein to ensure that the phage's "survival kit" was preserved.
Fine-tuning Evo to make AI better understand bacteriophages
The original Evo model, trained based on massive phage data, can generate sequences but lacks precise control over ΦX174.
For this reason, supervision and fine-tuning become the only option.
The team then had Evo continue training on 14,466 carefully selected small phage sequences. After reducing redundancy, the model focused on ΦX174-related mutations.
After fine-tuning, through carefully designed prompt words and sampling parameters, Evo can generate sequences that are similar to the evolution of ΦX174 but also innovative.
It's like giving AI an inspiration template, allowing it to inject new ideas into the familiar.
Assessment and screening
After generating the sequences, the authors developed a multidimensional assessment system that can examine gene arrangement, host specificity, and evolutionary diversity.
The key is to ensure that the AI phage can infect the non-pathogenic strain used in the experiment - type C Escherichia coli.
They then required the sequence to include a spike protein similar to ΦX174, as this protein determines the host range of ΦX174.
Experiments have shown that all 16 functional phages are strictly targeted to type C Escherichia coli and type W Escherichia coli.
Moreover, it was ineffective against the other six strains tested.
This demonstrates that host specificity can be maintained while other regions of the genome evolve significantly.
A new bacteriophage was born after wiping out bacteria in just 2 hours.
Traditional phage research is slow and cumbersome, so researchers have innovated the screening process.
They used Gibson assembly to synthesize the genome, transformed it into competent type C Escherichia coli, and then monitored its growth inhibition in 96-well plates.
Successful infection will cause the bacterial density (OD₆₀₀) to plummet within 2-3 hours.
This approach allowed the team to rapidly test 285 designs, ultimately validating 16 functional phages and characterizing their adaptability and host range.
Experimental testing to evaluate AI-designed phages
These AI genomes carried 67-392 new mutations compared to their closest natural counterparts.
Among them, Evo-Φ2147 carries 392 mutations and has an average nucleotide identity of 93.0% with phage NC51.
By some taxonomic criteria, it's good enough to be recognized as a new species.
In addition, 13 genomes contained mutations not seen in nature, demonstrating that Evo can exploit sequence space that has never been explored by natural evolution.
A very interesting finding was that one of the synthetic phages, Evo-Φ36, incorporated the DNA packaging protein J protein of the distantly related phage G4 (25 vs 38 amino acids).
This has been an unsolved engineering problem in the past.
Using cryo-electron microscopy, the researchers saw that it was embedded in the capsid structure in a unique way, and AI cleverly coordinated compensatory mutations to allow the new protein combination to function normally.
Cross-generational pursuit of drug-resistant bacteria, five reversals
Bacterial resistance to antibiotics is one of the most pressing challenges facing modern medicine, claiming hundreds of thousands, if not more, of lives each year.
Bacteria can rapidly evolve resistance to traditional antibiotics, greatly limiting the effectiveness of treatment.
Phage therapy holds promise for reversing this, but natural phages often cannot keep up with bacterial evolution.
In the study, the research team induced three type C Escherichia coli strains that were resistant to ΦX174, in which the waa operon (responsible for modifying bacterial surface receptors) had mutations.
The results showed that the AI-generated phage "cocktails" conquered three drug-resistant strains within 1-5 passages.
However, ΦX174 alone is completely ineffective.
It is worth mentioning that these phages that have achieved breakthroughs are "chimeric genomes." They fuse multiple AI fragments, and the mutations are concentrated in the receptor interaction region.
Sequence analysis showed that successful phages combined 2-3 different AI-designed genetic elements.
In this way, humans do not need to rely on rare phages in nature, but can let AI directly generate diverse groups, forming a "multiple hits" to make it difficult for bacteria to develop comprehensive drug resistance.
In short, AI can quickly screen out effective gene sequences, which means that phage therapy is no longer a "trial and error" process based on luck, but a precise "design."
In the future, humans will be able to proactively design treatments that are one step ahead and always stay ahead of bacterial mutations.
Genetic Revolution 2.0: Writing the Code of Life
Today, phage therapy is becoming an increasingly effective weapon against multidrug-resistant bacteria.
Recently, medical therapeutic targets are mainly aimed at plant pathogens or large DNA phages.
The latest research has demonstrated that AI models can capture evolutionary constraints and bridge the gap between AI-generated sequences and biological reality through training, quality control, and high-quality verification.
As model iteration and synthesis costs decrease, whole-genome design will open up unexplored evolutionary space and open up new frontiers for biotechnology and basic research.
This shift, from reading to writing to design, marks a new chapter in humanity's ability to transform biology at the most fundamental level.
Core Author
Brian Hie
I am an Assistant Professor of Chemical Engineering at Stanford University and an Arc Institute Innovation Fellow, where I conduct research at the intersection of biology and artificial intelligence.
He received his PhD from MIT CSAIL and his undergraduate degree from Stanford University.
Samuel King
Samuel King is a PhD candidate at Stanford University, currently working at the Arc Institute at the intersection of synthetic biology and ML.
He graduated from the University of British Columbia (UBC) with a Bachelor of Science (Honors) in Biology.
References:
https://x.com/samuelhking/status/1968329299364376698 https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1
https://arcinstitute.org/news/hie-king-first-synthetic-phage
This article comes from the WeChat public account "Xinzhiyuan" , edited by Taozi, and published by 36Kr with authorization.
