Recording intelligenceAI-generated brief · check the source for context

AI-Designed Bacteriophages

1:42:23 recording · AUTO · 2 speakers

Listen to the original

Listen to the episode
Executive Summary AI
  • Vincent Racaniello and Nels Elde record Twevo 121 from MicroTV on Christmas Eve, 24 December 2025, with Nels in Salt Lake City and a live audience of about 52 people.0:00
  • They open on science funding under stress: immunologist Shruti Naik's NY Cures is trying to raise state money for New York labs, which train 7% of all US physicians across 11 medical schools, while NIH, the world's biggest single funder of biomedical research, is being dismantled.1:55
  • The paper is a preprint titled Generative Design of Novel Bacteriophages with Genome Language Models, funded largely by the ARC Institute in Palo Alto with NIH conspicuously absent, using the Evo 1 and Evo 2 models trained on roughly 2 million bacteriophage sequences.14:30
  • Of 302 proposed genomes based on Phi X174 (5,400 base pairs, 11 genes), 285 were synthesized and only 16 formed plaques on E. coli strain C, and even those are about 95% identical to the original after heavy supervised fine-tuning and a bespoke coding sequence predictor for overlapping reading frames.58:37
  • The stated motivation is phage therapy cocktails against antibiotic-resistant bacteria, but Jamie Fraser's public comment, which Vincent reads out, faults the work for lacking explicit baselines, and Nels closes with NIH grant cancellations over seasonal flu gain-of-function fears and an OSTP call for public input.1:10:40

Brief overview

AI-designed phages are mostly fine-tuned Phi X174 variants, impressive engineering but no clear advance over evolution.

  1. Modification, not creation from scratchThe 16 viable phages carry 67 to 392 mutations and stay about 93 to 95% identical to Phi X174.
  2. Human intelligence did the heavy liftingSupervised fine-tuning fixed genome size, GC content, receptor identity above 85%, and overlapping reading frames.
  3. Nature runs bigger experiments every dayInfluenza and gut phages test more sequence space at biosafety level zero than 302 designed genomes ever could.

Questions this recording answers

10 questions, each answered where it is said
Did the researchers really create a brand new virus with AI?

No. The news said they made a virus, but they modified an existing one. The final phages are about 95% identical to PhiX174, carrying roughly 60 to 300 mutations, so the sequences are novel in the sense of unseen, not built from whole cloth.

Answered around 18:52
How many of the AI-designed genomes produced a working phage?

They picked 300 designed sequences and managed to synthesize 285; the rest failed for technical DNA-assembly reasons. Of those 285 modified PhiX174 genomes, 16 formed plaques on E. coli strain C. None of those 16 could plaque on the closely related lab strain K-12.

Answered around 58:15
What is a genome language model trained on?

Vincent quotes ChatGPT: like a language model that learns patterns in words and sentences, a genomic LLM learns patterns in biological sequences, including nucleotide sequence, protein sequence, motifs, genes, regulatory regions and biological functions. The Evo 1 and Evo 2 models were trained on roughly 2 million bacteriophage sequences.

Answered around 31:26
How much human intervention did the model's designs require?

A lot. They ran an extensive supervised fine-tuning process with strict limits on genome size and GC content based on PhiX174, exclusion of homopolymer runs, and a bespoke coding sequence predictor added because overlapping reading frames kept breaking the model. Nels calls that a heavy human hand.

Answered around 38:04
What are the basic numbers on the PhiX174 genome?

PhiX174 is about 5,400 base pairs encoding 11 genes, seven regulatory elements and two recognition sequences. It is a single-stranded DNA bacteriophage, one of the workhorses of molecular biology for 50 years, and it was the first genome of any virus ever to be sequenced.

Answered around 27:01
Why did they use E. coli and a phage instead of a human virus?

The choice of virus and host was deliberate for safety: making a phage that kills E. coli faster has little implication for humans. They also told the model not to make a eukaryotic virus, and the work was probably done at BSL-2. Vincent's position is containment, not prohibition.

Answered around 45:00
Did the designed phages compete well against natural PhiX174?

They ran 17-way competition experiments, their 16 designed phages plus PhiX174 at equal multiplicity of infection, in three replicates. Three of the synthesized phages consistently appeared in the top five, while the natural PhiX174 was in the running but never the best. EvoPhi 2483 was the most lytic.

Answered around 1:07:12
Can the phage collection beat E. coli that is already phage resistant?

They first evolved E. coli resistant to the phages, then applied the collection. The resistant strains held at first, but after one, two or five passages, depending on the strain, the diversified phage population overcame the bacteria and lysed them, the same cocktail principle used to hold HIV in check.

Answered around 1:13:57
What practical application is this work aiming at?

Phage therapy: designing collections of phages that persistently lyse pathogenic, antibiotic-resistant E. coli and other bacteria instead of using antibiotics. Vincent notes about 100 phage therapy cases in Belgium, done as one-offs rather than clinical trials, where the combination of phage plus antibiotics beats either alone.

Answered around 1:10:40
What criticism has been raised against the preprint?

Jamie Fraser commented that the manuscript is a great demonstration of what is feasible with whole genome phage engineering, but misses the opportunity to contextualize the model's performance with explicit baselines, so the 16-out-of-300 success rate is hard to judge. The sequence space stays close to known natural sequences.

Answered around 1:27:37
Key Quote

“Now, if I take polio and change one codon, I now have a novel polio virus.”

— Vincent Racaniello, AI-Designed Bacteriophages, at 34:35▶ listen at 34:35
Key Quote

“It's that caveat combined with inference, time, steering, and filtering.”

— Nels Elde, AI-Designed Bacteriophages, at 1:03:38▶ listen at 1:03:38
Key Quote

“so basically now it's in the end ai is only as good as the information we provide it.”

— Vincent Racaniello, AI-Designed Bacteriophages, at 1:24:37▶ listen at 1:24:37
Key Quote

“And so it's hard to judge the sizzle from the steak.”

— Nels Elde, AI-Designed Bacteriophages, at 1:28:37▶ listen at 1:28:37

Cite this

“AI-Designed Bacteriophages.” https://traffic.libsyn.com/secure/twievo/TWiEVO121.mp3. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/ai-designed-bacteriophages.html.

Report this page
What is wrong with this page?