← The Health AI Brief
Loading summary
Transcript3 lines
- [00:01]
A
Some have reported this news as an existential threat. Others see it as a potential cure for many diseases. Both views contain truth. This is the very real paradox surrounding the first synthetic viruses designed entirely by artificial intelligence. A team at Stanford published a paper in Science demonstrating the generation of a functional living virus using generative AI. To understand the impact, we need to look past the headlines. The technology relies on genome language models. Think of this like the predictive text on your smartphone. The phone guesses the next word. It analyzes massive patterns in human language. Similarly, these genomic models analyze trillions of chemical letters. They predict the next nucleotide in a sequence. The model is called evo. Let's walk through exactly what the researchers did. They started with two foundation Evo 1 and Evo 2. These models were pre trained on trillions of nucleotides across all domains of life. Then the team fine tuned these models on over 14,000 genomes from the microviridiviral family. This is a family of small viruses that infect bacteria or bacteriophages. They use a natural virus called Phi X174 as a design template. The virus has a tiny genome of only 5,400 base pairs.
- [01:15]
B
Its genes are highly compact and overlap heavily. To generate new genomes, the team prompted the fine tuned AI with short starting sequences just four to nine nucleotides long. The AI then auto completed the rest of the sequence, predicting all 5,400 base pairs. But generating sequences is only the first step. Most AI generated sequences are biological gibberish. To find the viable ones, the team applied three layers of digital filters. First was quality control. The computer filtered out sequences containing non nucleotide characters. It rejected sequences that were too long or too short or had incorrect GC ratios. They also built a custom gene prediction tool. This tool scanned the digital sequence and discarded any genome that failed to encode at least seven essential viral proteins. Second was trophism, which means host targeting. The researchers wanted the virus to target
- [02:12]
A
a safe lab strain of the E. Coli bacteria called E. Coli C. So the filter required that the generated spike protein shared at least 60% similarity with the natural Phi X174 spike proteins. Third, they wanted to ensure diversification. The researchers wanted to avoid simply copying what already exists in nature. They filtered out any sequence that was more than 95% identical to natural viruses. They also used a filter to enforce new genomic architectures changing where genes start and stop so they didn't replicate the exact spacing of the templates. This pipeline narrowed down the pool to 302 candidate genomes. The team chemically synthesized 285 of these digital designs in a laboratory. They then assembled them into circular double stranded DNA. Then they introduced this DNA to E. Coli host cells. This step is called rebooting. If the design is correct, the cell reads the DNA and produces living active viruses. What did the work show then? It yielded 16 viable functional phages. This is a success rate of about 5%. The these viruses were genetically distinct from anything pre existing in nature. One of them, named Evo Phi 36 proved something remarkable. The AI gave it a much shorter DNA packaging gene taken from a distant phage. When humans tried to swap this gene manually, the virus dies. But the AI altered the surrounding genomic context to make it work. The genes co adapted. This demonstrates that the AI can handle complex multi gene coordination. The work also showed massive therapeutic potential. The researchers evolved E. Coli strains that were completely resistant to the natural Phi X174 viruses. These resistant bacteria had mutated a gene which alters their outer membrane so the virus cannot bind. Then the team treated these resistant bacteria with a cocktail of 16 AI designed phages. The AI cocktail rapidly overcame the resistance and wiped out the bacteria. The natural virus cocktail failed to do this during the infection. The AI designed phages recombined and mutated to target the new bacterial receptors. So I think it's important to be clear what the work did not do. The researchers built upon a well characterized natural template rather than designing complex cellular life from scratch. The designed bacteriophages still belong to the microviridae family, retaining 40% nucleotide identity with existing genomes. Most importantly, it didn't prove that the method can be easily applied to complex human pathogens. Eukaryotic viruses, which infect humans, plants and animals, are vastly more complex than the viruses here. The bacteriophages that infect bacterial. The positive implications are clear. We're losing the war against antibiotic resistant superbugs at the moment. If we're able to custom design phage therapies, then this could allow us to rapidly generate adaptive treatments. We can design phages that specifically target patient specific bacterial strains, bypassing resistance almost instantly. But the negative implications are also potentially deeply concerning. Evo 2 is open source. Anyone can download it. The Stanford team deliberately withheld data on human infecting viruses during pre training. This safeguard is commendable, but it's fragile. A bad actor could theoretically fine tune these open models on dangerous human pathogens. Coding agents and cloud computing make this fine tuning cheap and accessible. Commercial gene synthesis is also widely available. Existing laws govern the possession of known pathogens but don't cover completely new AI designed synthetic constructs as is astutely pointed out in the editorial accompanying this new research. Ultimately it's a significant technical generative Genomics has transitioned from theory to physical reality. The benefits for medicine could be immense, particularly for overcoming superbugs and bacterial resistance, but we also need to build global governance and screening tools that can flag dangerous synthetic sequences before they're printed. There's an opportunity to steer a powerful technology safely rather than running from its risks. It's just one of many instances where we can see huge great potential from what's possible now with AI, but also huge potential risk if things aren't handled appropriately. If you want to keep abreast of as many of these updates as possible, we'll aim to be covering as much as possible on this channel, so don't forget to hit like and subscribe to keep updated.