🧬 DeepMind Predicted Every Possible Typo in Your DNA. All 9 Billion of Them.

This week an AI claimed to have read every possible typo in the human genome, a fruit fly's brain got a brokerage account, and Stanford handed most of a mouse's cortex over to human cells. Somewhere in between, a gene therapy finally did the thing gene therapies keep promising, for the kids who needed it most. Also, placenta went up a mouse's nose. Normal week, apparently.

RESEARCH & NEWS

Your genome is three billion letters long, and at every position nature could have typed one of three wrong ones. Most of those typos do nothing. A few wreck a childhood. Until now, working out which was which meant running variants through a model one at a time, on hardware you probably don't have.

Google DeepMind recently did everyone's homework. AlphaGenome Atlas is a 1-petabyte catalogue of predicted effects for all 9 billion possible single-letter changes in the human reference genome, free for academic use. Each variant comes with thousands of predicted consequences for splicing, gene expression and chromatin, plus a new single-number score (AVI) that fuses AlphaGenome with its protein-focused sibling AlphaMissense.

The interesting part is the non-coding 98% of the genome, where most trait-linked variants live and where older tools mostly shrug. Re-running solved rare-disease cases, AVI put the known culprit in the top 50 candidates 29.5% of the time, versus 12.5% for the standard tool CADD. It also flagged an overlooked splicing variant in DNM1 in a child with epileptic encephalopathy, which lab experiments then backed up.

Now the fine print. These are predictions, not measurements, made against a single reference genome, and the paper is a DeepMind-hosted PDF rather than a peer-reviewed publication. UBC genomicist Carl de Boer calls it a useful resource but warns the headline score is "probably going to be easily misinterpreted".

So yes, the AI has mapped every letter of the code, in the same way a weather service has mapped tomorrow. It's an extremely good forecast. You still take an umbrella to the lab.

NEWS

This month, researchers formally published the first finished connectome of an entire male fruit fly central nervous system, brain and nerve cord together, built by Janelia's FlyEM team, Cambridge and Google. It's a real neuroscience milestone. The internet immediately asked whether it runs Doom.

It does, sort of. Coinbase engineer Alex Wormuth wired it up so that each Doom frame stimulates sensory neurons and taking damage zaps two dopamine cells as reinforcement. At last check the fly was on round 6,385 and, in PC Gamer's gentle assessment, not very good. Others have since handed it Beat Saber and Super Mario 64.

Then came Stonkfly, because of course it did. It turns live crypto prices into a colored chart and pipes the pixels into the full MaleCNS graph: 166,700 neurons and 25.6 million connections. A fixed readout picks buy, sell or hold. Gains stimulate 15 dopamine cells, losses hit two aversive ones, and live trading is capped at $10 an order. The README states plainly that profitable learning has not been demonstrated, and Hackaday notes nobody has yet checked how good a fly is at trading.

To be fair to the fly, none of this is really a simulated fly. A connectome is a wiring diagram: it tells you who connects to whom and roughly how much, not how each neuron actually behaves from moment to moment. These projects are closer to bolting a joystick onto a subway map.

Still, it's the only trader in history with its entire brain published under a Creative Commons license. Most hedge funds can't say that.

RESEARCH

Growing human brain organoids inside rodents has always had a housing problem. The rodent brain is already occupied, matures faster than human tissue, and does not share. Sergiu Pașca's lab at Stanford solved it the way landlords do: they evicted the tenants.

The team bred immunodeficient "apallial" mice genetically missing most of their neocortex and hippocampus, roughly half their total brain tissue, then transplanted four human cortical organoids into the empty space in pups only a week or two old. Three months later, the human graft accounted for about 92% of the cortical tissue.

The human tissue didn't just sit there. It wired into the mouse brain, sent projections as far as the spinal cord, produced a deep-layer projection neuron type that dish-grown organoids essentially never make, and fired in coordinated waves while the mice were awake. Mice without a cortex failed a working-memory maze; mice with the human graft passed it, though the authors stress nobody knows yet whether the human cells are what made the difference.

Now the part the headlines skipped. The graft only matured to roughly mid-gestation levels, has no proper cortical layers, and sits on an otherwise entirely mouse brain. Monash neuroscientist Adeel Razi asks people not to call it a human brain in a mouse, but rather "a hybrid brain system containing developing human brain cells within an overwhelmingly mouse brain".

The authors themselves say anything more mature will need ethical guidance worked out first. So nobody is defying god just yet. They're still filling in the paperwork.

NEWS

Sanfilippo syndrome type A is the kind of disease that makes rare-disease drug development feel less like business and more like an obligation. Children develop normally, then lose speech, cognition and movement, which is why families often call it childhood dementia. Median life expectancy is 15 years.

On September 17, the FDA approved Ultragenyx's Fayuvi, the first treatment for the disease. It's a one-time intravenous AAV9 gene therapy carrying a working copy of SGSH, the gene for the enzyme these children lack. Without it, a sugar chain called heparan sulfate piles up inside cells and the brain takes the worst of it.

The efficacy data deserve a careful read. Seventeen treated children scored 23.5 points higher on a cognitive scale than 27 untreated children from a natural-history cohort. That's an external comparison, not a randomized trial. In a disease this rare and this fast, that's a reasonable trade-off, but still a trade-off.

Getting here was not smooth. The vector was built at Ohio State and Nationwide Children's more than a decade ago, licensed to Abeona, then handed to Ultragenyx when Abeona ran short of money. The FDA sent it back in July 2025 over manufacturing issues and asked for more long-term data before approving the resubmission. Ultragenyx also collects a priority review voucher on the way out.

For families who spent decades fundraising for research themselves, this is that rare biotech headline without a catch in the first sentence.

RESEARCH

The placenta spends nine months stopping a mother's immune system from attacking a half-foreign tenant, and then gets thrown away. Researchers at Università Cattolica in Rome wondered whether that talent for calming inflammation could be put to work on Alzheimer's instead.

The team grew stromal cells from donated placental amniotic membrane and harvested the extracellular vesicles they shed, tiny membrane parcels packed with microRNAs and signaling molecules. Female Alzheimer's-model mice got them up the nose twice a week from three to nine months of age, starting before symptoms appear. Each dose was 4 microlitres, so the "nasal spray" in the headlines is less a spray than a sniffle.

In mice, it worked. The vesicles reached the hippocampus and were taken up by both neurons and microglia. Treated mice did better on memory tests, hippocampal amyloid-beta dropped, and the brain's immune cells calmed down. Tau, the other famous Alzheimer's protein, didn't budge. In neurons grown from Alzheimer's patients' skin cells, the vesicles also prevented neurite shrinkage and restored synaptic proteins.

The press release says the vesicles can also counteract cognitive decline once it has already occurred. The paper is more modest: one month of treatment in already-impaired mice improved one memory test and not the other. Co-author Claudio Grassi, who heads the university's neuroscience department, is clear that these are preclinical results still needing validation in humans.

It joins a very long line of Alzheimer's treatments that work beautifully in mice. At this point, the mice should be dementia-free forever.

Look closely and almost every story this week is about a stand-in. The Atlas is a forecast of your genome, not your genome. The connectome is a subway map of a fly, not a fly. The Stanford mouse is a human cortex draft running on mouse hardware, and the placenta drops have so far only ever met a mouse. Useful stand-ins, all of them, but still stand-ins.

The exception is Fayuvi, where the patients are real children and the result is a real approval. That one earned its headline.

Which stand-in would you trust least with your money? Hit reply, we read everything. And if you know someone who would enjoy a fruit fly with a trading account, forward this along.

Keep questioning everything (especially anything trading crypto on 166,700 neurons),

Prateek & Jere

P.S. The fly has never lost money it didn't have, which already puts it ahead of a surprising number of crypto funds. And terveiset Suomesta, where the autumn equinox just landed and the fruit flies around the compost bin are, as far as we know, not yet trading.