💊 The Most Personalized Medicine Company in the World Just Ran Out of Time

Sometimes the future arrives exactly on schedule, and nobody can afford to wait for it. This week, a company built around making drugs for one patient at a time shut down just as the FDA started writing rules to help them. Meanwhile, brain cells are playing video games, AI is writing genomes, and someone finally froze a mouse brain and got it to do something useful afterward. Let's get into it.

NEWS

Picture this: you've built a company around the radical idea that every patient with an ultra-rare genetic disease deserves their own custom therapy. Not a therapy for their disease category. Their own therapy. One drug, one patient, manufactured from scratch. It's the kind of vision that makes biotech investors tear up at pitch meetings, and accountants reach for the whiskey.

EveryONE Medicines (EoMeds) was exactly that company. Co-founded by Julia Vitarello, the mother of Mila Makovec (the child who received milasen, the first-ever drug designed for a single patient), EoMeds was trying to industrialize individualized antisense oligonucleotide therapies for children with severe neurological conditions caused by unique genetic mutations. Backed by GV, Khosla Ventures, and Third Rock Ventures, the Boston-based startup had secured early agreements with European regulators and partnered with the UK's MHRA for a process-based approval pathway.

Then, in early March 2026, it shut down.

Here's the thing: just days earlier, the FDA had released draft guidance for a "Plausible Mechanism" framework designed to accelerate approval of exactly these kinds of therapies. The framework would allow developers to demonstrate a plausible biological mechanism rather than running traditional clinical trials for each individual drug. On paper, this was the regulatory breakthrough EoMeds had been waiting for.

But draft guidance is still draft guidance. The 60-day public comment period doesn't close until April 27, and for a startup burning cash with roughly 10 employees, "the rules may improve soon" is not the same as having a business now. An HHS spokesperson explicitly denied that the framework caused the closure, calling such suggestions "baseless."

The real villain? Probably reimbursement. Making a custom drug is technically heroic. Getting someone to pay for it, repeatedly, at scale, when your addressable market is sometimes literally one person, is a business problem that regulatory frameworks alone can't solve. To be fair, the n-of-1 movement isn't dead (academic centers and the N=1 Collaborative continue the work), but EoMeds' closure is a sharp reminder that science and sustainability don't always arrive on the same timeline.

NEWS

If you've been reading this newsletter for a while, you might remember when we covered neurons learning to play Pong. Well, the same company is back, and apparently, Pong was too easy. Melbourne-based Cortical Labs now says 200,000 human neurons grown on a chip can play Doom.

Before you picture sentient brain tissue fragging demons, though, let's talk about what's actually happening. The neurons sit on a multi-electrode array inside Cortical Labs' CL1 unit (roughly $35,000 a pop). A reinforcement learning algorithm processes game state, translates it into electrical stimulation, zaps the neurons, records their spike patterns, and runs those through a decoder that maps them to seven Doom actions. The neurons are part of the loop, but turns out, so is a PPO reinforcement learning algorithm, doing significant computational work.

Chief Scientific and Operational Officer Brett Kagan put performance in perspective: "The cells play a lot like a beginner who's never seen a computer." The system beats random play but still loses frequently. And unlike the original Pong experiment published in Neuron in 2022, the Doom demo has no peer-reviewed paper (it exists on YouTube and GitHub). One skeptical analysis even noted that the developer's own code acknowledges the AI decoder may learn to route around the neurons entirely.

Still, Cortical Labs isn't really trying to speedrun Doom. They're testing whether biological neurons can be integrated into computing architectures at all, which is a genuinely interesting question, even if the answer right now is "sort of, with a lot of help."

RESEARCH

A year ago, the Arc Institute dropped Evo 2 as a preprint, and the computational biology world collectively lost it. Recently, the paper was published in Nature, and Arc released a one-year retrospective showing what the model has actually done in the wild.

The numbers are absurd. Evo 2 is a 40-billion-parameter DNA language model trained on 9.3 trillion nucleotide tokens from 128,000+ whole genomes spanning all three domains of life: bacteria, archaea, and eukarya. It processes raw DNA sequences (A, T, C, G) at single-nucleotide resolution with a context window of up to one million tokens. For context, that's like reading an entire genome in one pass, which is exactly the point.

What makes Evo 2 matter beyond its size is that it's fully open-source under Apache 2.0: weights, training code, inference code, and the entire OpenGenome2 training dataset. Led by co-senior authors Patrick Hsu (Arc Institute co-founder, UC Berkeley), Brian Hie (Stanford), and Hani Goodarzi (UCSF/Arc), the team used a StripedHyena 2 architecture rather than a standard transformer.

The one-year stats are striking: 200+ citations, 100,000+ model downloads, 88,000+ GitHub downloads. More importantly, Arc reports that 16 of 285 AI-designed bacteriophage genomes successfully propagated and killed bacteria, making them apparently the first AI-designed and experimentally validated organisms. The model can also predict pathogenic BRCA1/BRCA2 mutations with >90% AUROC and design chromatin accessibility patterns with 0.92-0.95 concordance.

To be fair, the model has clear limits. AI-generated bacterial genomes are still "missing critical elements" and "would likely not function if synthesized," as Hie himself acknowledged. But as a foundation for others to build on (because, of course, it's open), Evo 2 is already doing things its creators didn't anticipate.

RESEARCH

Researchers at Friedrich-Alexander-Universität Erlangen-Nürnberg did something that sounds like a science fiction premise: they froze mouse brain tissue to −196°C, kept it deeply frozen for up to seven days for slices and 8 days for whole brains, thawed it out, and found that the neurons could still fire, synapses could still transmit signals, and the cellular machinery of learning (long-term potentiation) was still operational. Published in PNAS, it's an advanced demonstration of comprehensive functional electrophysiological recovery in adult mammalian brain tissue after cryopreservation.

Now for the asterisks. The primary experiments used 350-micrometer-thick slices of hippocampus, not whole brains. Lead author Alexander German and colleagues also vitrify whole mouse brains in situ via vascular perfusion, but with a lower success rate that required repeated protocol adjustments.

The method is called vitrification: replacing water with cryoprotectants (a cocktail including DMSO, ethylene glycol, formamide, and polyvinylpyrrolidone) that convert tissue to a glass-like state rather than letting ice crystals shred everything. Rewarming at 80°C per second prevents ice formation during thawing.

Before anyone books a cryonics appointment: the paper explicitly states results "should not be interpreted as directly transferable to the cryopreservation of large organs." Memory preservation was not tested. The more realistic near-term application is biobanking brain tissue for drug testing, which is less cinematic but potentially very useful.

NEWS

The Cavendish banana, which accounts for over 90% of global banana exports, is essentially a genetic clone. Every single one. Essentially, because there are variations, but you get the point. And a soil fungus called Panama Disease Tropical Race 4 (TR4) is slowly eating through them, with no cure, no effective fungicide, and spores that survive in soil for decades. If this sounds familiar, it's because the exact same thing already happened once (the Gros Michel banana, wiped out in the 1950s).

Norwich-based Tropic Biosciences just closed a $105 million Series C co-led by Forbion and Corteva Agriscience to scale their answer: gene-edited bananas engineered using their proprietary GEiGS platform, which uses CRISPR to redirect the plant's own RNA interference machinery. Crucially, no foreign DNA is introduced, sidestepping GMO classification in most markets.

Tropic already commercially launched two varieties in 2025: a non-browning banana (named one of TIME's Best Inventions of 2025) and an extended shelf-life variety that adds 12 extra days of green life. The TR4-resistant banana is targeted for 2027. Given that TR4 threatens an industry supporting $25 billion in trade and 400 million livelihoods, the math on this one is pretty straightforward.

This week's lineup is basically the biotech version of "the spirit is willing, but the flesh is weak." EoMeds had the vision but not the economics. Cortical Labs' neurons are trying their best, but still need software to carry the team. Evo 2 can write genomes, but they're not quite alive yet. And we can freeze a brain and get it working again, as long as you define "working" very carefully and don't ask about memories.

The good news? A banana company raised $105 million to save a fruit that feeds hundreds of millions of people, and they're actually shipping product. Sometimes the most boring-sounding stories are the ones that matter most.

If any of this made you think (or at least exhale through your nose), forward this to someone who could use a little more biotech absurdity in their inbox. And as always, hit reply with questions, corrections, or your hottest take on whether those neurons actually deserve credit for the Doom kills.

Keep questioning everything (especially headlines about frozen brains),

Prateek & Jere

P.S. Somewhere, a 200,000-neuron chip is reading this newsletter and losing a game of Doom at the same time. Multitasking, sort of.