Niko:

On Rosie and Bureaucracy: The most viral health story this month concerned Paul Conyngham, a Sydney tech entrepreneur with no biology background, who used AI tools to design a personalized mRNA cancer vaccine for his dog, Rosie. Conyngham paid $3,000 to sequence Rosie’s tumor DNA, used AI tools to understand this sequence, and then worked with the university to manufacture and inject a custom mRNA vaccine. “The tennis ball-sized tumour on Rosie’s hock has shrunk in half” as a result, according to reporting in The Australian.

Making a one-off mRNA vaccine is not especially hard, as bioengineer Patrick Heizer has explained. Researchers cure mice of cancer routinely, and it’s much harder to prove a therapy is both safe and effective in human trials. Note that the $3,000 figure was also just the cost of sequencing. The true cost, including reagents and labor, was much higher.

Palli Thordarson, Director of the RNA Institute where the work was done, also clarified that not all of Rosie's tumors responded to the vaccine. The team is investigating whether those tumors mutated in unique ways. Thordarson also said the treatment required co-administration of a checkpoint inhibitor, anti-PD-1, which is itself an anticancer drug. [FWIW, I’m sorta skeptical we can learn from what they did because they administered both together, and the anti-PD1 drug seems pretty effective. (Except that it should be easier to do small clinical studies.)]

This story also says a lot about the ridiculous, bureaucratic hurdles required for clinical testing (even on dogs). “The red tape was actually harder than the vaccine creation, and I was trying to get an Australian ethics approval and run a dog trial on Rosie,” said Conyngham. “It took me three months, putting two hours aside every single night, just typing the 100 page document.”Similar bureaucracy pervades clinical trials in the US and UK. We should urgently find faster ways to design and test drugs in human patients.

The Brain Coaxes Tumors to Grow Faster: For a new study, researchers mapped how a tumor in a lung “talks with” the brain in mice. It turns out that lung tumors secrete proteins that attract sensory nerves from the vagus, which connects internal organs to the brain. Some of these nerves then send signals from the tumor through the brainstem. In response, the brainstem ramps up sympathetic nerve firing, coaxing noradrenaline to go into the tumor and shield it from attacks by the immune system. When the researchers destroyed this “circuit” — either by silencing the sensory neurons with chemicals, or deleting the proteins involved entirely — tumors grew more slowly, the immune system (specifically T cells) were more active, and the mice lived much longer. There is now a great deal of interest in finding these circuits in humans, and also devising therapeutics to disrupt it.

Screening DNA: Anyone can order custom DNA online. Companies build the DNA using chemistry and then ship it through the mail. To make sure the DNA doesn’t encode dangerous pathogens or toxins, the companies are supposed to do two things: Verify that the customer is a legitimate researcher, and screen their DNA order against databases of dangerous pathogens.

For a new paper, researchers tested five different AI models to see if they could automate that first part, legitimacy screening, which today is usually done by humans at a cost of about $14 per customer. The researchers began by building “profiles” of 41 real scientists, each paired with a simulated order for a dangerous DNA sequence. Each AI model was then prompted to confirm the scientists’ institution, verify it’s a legitimate place, check their email domain, cross-reference whether the scientist is on a sanctions list, and so on. Gemini 2.5 Pro, connected to four specialized APIs to help with Internet search and database access, did this best. It categorized customers correctly 90 percent of the time, versus 80 percent for the humans, for a tenth of the cost. This is good news, because many companies (especially outside the U.S.) don’t screen all DNA orders. Automation and lower costs are one way to increase adoption.

The second study checked whether AI could identify obfuscated DNA sequences encoding dangerous toxins or pathogens. Many new AI tools can “redesign” proteins, such that different DNA sequences all encode a protein with the same 3D shape. This means that would-be bioterrorists could design toxic proteins which look identical to those found in nature, but do so using entirely different DNA sequences that don’t appear in databases. Another way to get around DNA screening checks is to chop up the sequence into short snippets, order each snippet individually, and then reassemble the strands in the laboratory.

For this study, then, researchers took 5,505 “redesigned” proteins from toxins and pathogens (plus 1,036 negative controls made from benign proteins), converted them to DNA, and fragmented each one into pieces of 25, 50, 75, 100, 125, 150, 175, and 200 nucleotides, with short overlaps between adjacent fragments. This produced roughly 700,000 fragments from the dangerous set and 170,000 from the controls. The, they ran every fragment through four existing biosecurity tools used by DNA synthesis companies.

To score each tool, they used a metric called Matthew's correlation coefficient, which basically just asks: how often does the tool correctly flag dangerous sequences while ignoring harmless ones? A perfect score is 1.0 and random guessing is 0. At the "sequence level," where a protein counts as “caught” if any of its fragments gets flagged, two tools (Providers 1 and 2) performed well even on very short fragments. At 200 nucleotides, both scored above 0.9. At 50 nucleotides, they still scored between 0.5 and 0.8, meaning they caught most redesigned dangerous proteins even when broken into very small pieces. This is good news for biosecurity, but there’s still no guarantee that companies will actually adopt and use these tools.

Hypermutation of Living Cells: When a researcher wants to make a better protein, they usually just mutate the corresponding gene and see what happens. This is often done in a test tube. The genes are randomly mutated, and then the proteins are purified and tested, again and again until a useful variant emerges.

A much faster way to evolve genes is to use living cells. A single tube of bacteria contains billions of cells, each of which can “run” its own mutation experiment in parallel. For a new study, researchers made a tool to mutate genes six orders of magnitude faster than wildtype mutation rates. The nice thing about this method is that it only requires two parts: a DNA-copying enzyme from a virus which had been deliberately designed to be sloppy (making one mistake every 1,000 bases of DNA) and a DNA strand which is only recognized by that error-prone enzyme. All scientists have to do, then, is put their gene onto this DNA strand, place it into cells, and then wait for the error-prone enzyme to insert random mutations. The bacterium's own genome is left completely alone.

This tool could be used to evolve better drug molecules, for example, or to evolve thousands of proteins in parallel to help build datasets to train predictive models.

CRISPR-Carrying Phages: E. coli causes up to 30 percent of bloodstream infections in cancer patients undergoing stem-cell transplants. These patients are often given levofloxacin (an antibiotic) as a prophylaxis, but up to 65 percent of these E. coli strains are resistant to the drug.

Prior clinical trials have tried to solve this problem using bacteriophages, or viruses that infect and kill bacterial cells. In a 2005 study, for example, researchers gave healthy volunteers T4 phages, by mouth, and found the phages were safe but did not reduce E. coli counts.