🤖 Microsoft's AI doctors are dunking on humans (and babies are hoarding dementia proteins)

Well, well, well. Look who decided to show up to the "let's make humans obsolete" party. This week, Microsoft's AI doctors are diagnosing patients like they're speedrunning a medical degree (spoiler alert: they're winning), Google built an AI that reads DNA like your ex reads your texts (obsessively and with concerning accuracy), and babies are casually flexing Alzheimer's proteins like it's a weird baby superpower. Oh, and scientists taught brain cells to play Pong (again) because apparently that's how we test epilepsy drugs now? Meanwhile, we’re revisiting the fact that the FDA finally admitted mice are terrible at cosplaying humans. Grab your favorite caffeinated beverage and maybe a stress ball - this week's serving of science is spicier than your take on pineapple pizza.

Microsoft dropped an AI bombshell that makes ChatGPT look like a Magic 8-Ball. Their new Medical AI Diagnostic Orchestrator (MAI-DxO) achieved 85.5% diagnostic accuracy on the most challenging medical cases from the New England Journal of Medicine, while 21 experienced human physicians managed only 20%. That's not just winning, that's the diagnostic equivalent of dunking on someone from the free-throw line.

The secret sauce is pure Silicon Valley meets Grey's Anatomy: five AI agents acting like a virtual medical team. There's a Hypothesis Agent throwing out diagnostic possibilities, a Doctor Agent asking all the right questions, a Cost Analysis Agent keeping everyone financially honest, a Quality Control Agent fact-checking everything, and a Decision Panel orchestrating the whole show through what they call a "chain-of-debate" process. It's like having House, Cameron, Chase, Foreman, and Cuddy in your computer, minus the workplace drama and Vicodin addiction.

Also! Microsoft basically poached Google's entire health AI dream team to make this happen. Dominic King, former head of DeepMind's health division, now leads Microsoft AI Health alongside other strategic defectors from Google. It's the most expensive talent acquisition since the NBA, and the results speak for themselves: This system diagnosed patients 4x better than humans while cutting diagnostic costs by 20%.

The kicker? They tested this on NEJM's most challenging cases, the medical equivalent of Jeopardy! Tournament of Champions questions. One case involved alcohol poisoning, where baseline OpenAI o3 ordered $3,431 worth of brain imaging for the wrong diagnosis, while MAI-DxO asked about hand sanitizer ingestion, confirmed with targeted testing for $795, and nailed it. That's the difference between throwing darts blindfolded and using a GPS.

The elephant in the room: This hasn't been peer-reviewed yet, and they tested on cherry-picked complex cases rather than routine diagnoses. But with 7.4 million annual misdiagnoses in US emergency departments alone, even imperfect AI assistance could save lives and money. The path from research demo to your doctor's office remains long, but Microsoft just made the strongest case yet for AI medical superintelligence.

Remember when AlphaFold revolutionized protein folding and basically solved biology? Google DeepMind just said "hold my test tube" and created AlphaGenome - the AI equivalent of reading the universe's instruction manual. While AlphaFold tackled the 2% of our genome that codes for proteins, AlphaGenome is going after the remaining 98% that controls when and how genes actually turn on and off.

This thing processes up to 1 million DNA letters while maintaining single-base-pair precision. Imagine reading War and Peace while noticing every comma placement. It predicts thousands of molecular properties simultaneously: gene expression levels, splicing patterns, chromatin accessibility, transcription factor binding sites, and 3D chromosome folding. Basically, it's like having a crystal ball for DNA, except the crystal ball went to MIT and has a PhD in computational biology.

The technical breakthrough is genuinely impressive: AlphaGenome outperformed existing models on 22 of 24 sequence prediction tasks and matched or exceeded top performers on 24 of 26 variant effect predictions. It can analyze how a single DNA letter change affects all these properties in under one second, work that previously required separate models and weeks of lab experiments.

Dr. Caleb Lareau from Memorial Sloan Kettering called it "the most powerful tool to date to model genetic differences" and "a milestone for the field." Coming from cancer researchers who've seen every AI snake oil pitch, that's basically a standing ovation.

The reality check: AlphaGenome struggles with tissue-specific contexts and long-range genetic effects beyond 100,000 base pairs. It's trained only on human and mouse data, so don't expect it to decode your pet goldfish's genome anytime soon. Google's being refreshingly honest about calling this "AlphaFold 1—a big first step" rather than claiming they've solved genetics forever.

Still, for biotech companies working on rare disease diagnosis, cancer research, or synthetic biology, this is game-changing. The API is free for non-commercial research, and a full open-source release is planned after peer review. It's democratizing advanced genomic analysis in a way that would make the Human Genome Project coordinators weep tears of joy.

You can also find the code on GitHub if you fancy a look.

Plot twist: The agency that's been requiring mouse testing for decades just announced they're probably going to skip it. The FDA dropped new guidance a few months ago, allowing and encouraging non-animal alternatives, marking the biggest shift in drug development methodology since, well, ever. Turns out those 90% failure rates from animal studies to human trials weren't just bad luck. They were a feature, not a bug.

The new approach is called NAMs (New Approach Methodologies), which sounds like a trendy wellness brand but is an AI-based computational model, human cell lines, organoids, organ-on-a-chip systems, and 3D human tissue models. Instead of wondering if your drug will kill a mouse, you can now test directly on human-relevant systems that predict human responses.

But honestly, this is old news at this point. We have covered this before. But things are now starting to move, and something as important as this needs repeating.

The paradigm shift is already happening: no biotech company has historically conducted full mouse lifespan studies for therapeutics before human trials, even for chronic drugs. The FDA Modernization Act 2.0 first authorized non-animal alternatives in December 2022, and now they're actively encouraging it with fast-track reviews and expedited approvals for NAM-based submissions.

The industry reaction has been overwhelmingly positive, with 60% of life sciences professionals previously hesitant due to regulatory concerns now seeing a clear path forward. BioAge Labs and Insilico Medicine are pushing for standardized approaches, with the compelling argument that discovering geroprotective effects could dramatically expand addressable markets, potentially adding over $1 trillion in value from just one additional year of human health span. It’s a nice thought, and we finally seem to be on the cusp of having a robust standard for animal studies. Which feels interesting to say the least when we are trying to phase out animal experiments. We do see the value, though.

The catch: Only 8 NAMs are currently in the FDA's pilot program, with minimal advancement through qualification stages. This is classic FDA, announcing revolutionary changes while maintaining glacial implementation speed. But the momentum is real, and first-mover biotech companies are already positioning themselves for the post-animal testing world.

Scientists have officially crossed into "we were so busy asking if we could, we never stopped to ask if we should" territory. Cortical Labs created living brain computers from human neurons, taught them to play Pong, then discovered that epilepsy drugs don't just stop seizures—they turn neurons into better gamers. Yes, you read that correctly. We're testing seizure medications by checking if brain cells can improve their arcade game high scores.

The setup sounds like something a stoned grad student would pitch at 3 AM: grow human brain cells on silicon chips, hook them up to Pong, and watch them learn. These "DishBrain" systems (actual scientific name, not making this up) use 800,000 human neurons that self-organize into networks and start playing within five minutes. They learn faster than any AI system, which is simultaneously impressive and slightly concerning for our robot overlords.

Here's where it gets spicy: when researchers induced seizure-like activity in these neural networks, their Pong skills went to absolute trash. But add carbamazepine (a common epilepsy drug) at 200 µM concentration, and suddenly they're back in the game, literally. The drug didn't just stop the chaotic firing—it restored the networks' ability to think and learn. Meanwhile, other epilepsy drugs like phenytoin stopped the seizures but left the neurons playing like they'd had one too many at happy hour.

This matters because 30% of epilepsy patients don't respond to current meds, and traditional drug testing on mice is about as useful as asking your goldfish for stock tips. By testing on actual human neurons playing video games, we can finally see if treatments restore brain function or just make neurons shut up. The system is so sensitive it can detect cognitive improvements that traditional tests miss entirely.

The existential crisis bonus round: These brain-in-a-dish systems have sparked heated debates about consciousness. If neurons can learn to play Pong, feel frustrated when they miss, and improve with practice, are they... aware? The researchers diplomatically note they "cannot predict its full potential," which is science-speak for "we might have accidentally created sentient Petri dishes and we're not ready to process that emotionally."

For epilepsy patients, this could mean personalized treatments grown from their own cells. For humanity, it means we're literally growing biological computers that beat silicon chips at their own game. Nature: 1, NVIDIA: 0.

You can read the publication from Nature communications biology or more from here

Scientists just discovered that newborn babies have 3x higher levels of phosphorylated-tau217 than Alzheimer's patients, and somehow this is completely normal and healthy. It's like finding out that toddlers are walking around with nuclear reactors in their backpacks, except the reactors are powering their brain development instead of causing meltdowns. Nature really said "hold my beer" with this one.

The University of Gothenburg study analyzed 462 people across Sweden, Spain, and Australia, finding that premature infants show the highest p-tau217 concentrations of all groups tested. These levels decline sharply over the first 3-4 months of life, suggesting this protein isn't just tolerated in newborns—it's actively beneficial for brain development.

Here's where it gets wild: In Alzheimer's patients, p-tau217 elevation is driven by β-amyloid plaque aggregation, leading to neurofibrillary tangles that destroy neurons. In newborns, the same protein elevation supports microtubule dynamics, neuroplasticity, and rapid synaptic connections without any toxic aggregation. It's the same molecule playing completely opposite roles depending on the biological context.

Lead researcher Fernando Gonzalez-Ortiz put it perfectly: "If we can learn how the newborn brain keeps tau in check, we might one day mimic those processes to slow or stop Alzheimer's." Senior author Kaj Blennow added that understanding these protective mechanisms "could offer a roadmap for new treatments."

The implications are significant. This discovery suggests newborn brains possess protective mechanisms that prevent tau aggregation despite high phosphorylation levels - mechanisms apparently lost with aging. Instead of just reducing tau (current therapeutic approaches), we could potentially recreate nature's own solutions by reactivating these endogenous protective systems.

This fundamentally challenges the amyloid cascade hypothesis and opens unprecedented opportunities for biomimetic therapies. The finding validates p-tau217 as the premier Alzheimer's biomarker while revealing its dual nature as both an essential developmental factor and a pathological marker. For the FDA-approved p-tau217 blood test, this means interpretation must consider age and biological context: high levels in babies indicate normal development, not disease.

The research creates immediate opportunities for age-stratified biomarker development and protective mechanism research platforms, while pointing toward long-term therapeutic strategies that work with biology rather than against it. Sometimes the best solutions are hiding in plain sight, and apparently, this time they were hiding in the NICU.

Read more of the results from the published paper or from here

Phew. Another week, another reminder that we're living in the timeline where AI doctors are better at their jobs than actual doctors, brain cells have gaming addictions, and babies are biochemically trolling us. Science really said "normal is boring," and we're honestly here for it.

So, which story broke your brain this week? Are you ready to let Microsoft's AI squad diagnose your mysterious rash? Did the Pong-playing neurons give you an existential crisis about consciousness? Are you side-eyeing every baby you see now, knowing they're packing more tau proteins than a nursing home?

Drop us a line – we read every single reply (yes, even the ones where you correct our spelling). And if you enjoyed this week's chaos, share it with someone who needs their weekly dose of "wait, science did WHAT now?"

Still cruising at that sweet 70% open rate, which means most of you actually like this nonsense. You beautiful nerds make our week!

Stay curious, stay caffeinated, and maybe check if your baby is hoarding any other dementia proteins, Prateek & Jere

P.S. - If you're a mouse who lost your job at the FDA, we're hiring for taste testers. Must be willing to try experimental cereal. Previous experience being wrong about drug effects preferred.

And if you’re still not convinced, no worries! Just unsubscribe here, you can always check back on us later