The Lyceum: Biotech & Life Sciences Weekly — Mar 17, 2026

The FDA just killed the two-trial rule for drug approval, AI designed a million proteins and a startup tested every single one, and two completely different pills are converging on sleep apnea — a disease that currently requires you to strap a machine to your face. Biology is becoming a manufacturing platform, computation is becoming the engine of discovery, and the regulatory systems built for a slower era are scrambling to keep up.

If you've spent any time around drug development, you know the two-pivotal-trial rule like you know your own phone number. It has governed how drugs reach patients since 1998 — and it's gone.

The FDA announced that one adequate, well-controlled study combined with confirmatory evidence will now serve as the default basis for marketing authorization. The agency framed the shift plainly: the two-trial standard "was intended to provide credible causal evidence" in an era of more limited biological understanding. That era, they're saying, is over.

The implications cut both ways. For rare disease developers and gene therapy companies working with tiny patient populations where running two 500-person trials is simply impossible, this is potentially transformative. For skeptics, one rigorous trial with confirmatory evidence still leaves meaningful room for judgment calls about what "confirmatory" means — and that ambiguity will be litigated in advisory committee rooms for years.

Watch what the agency accepts as confirmatory evidence, because that's where the real policy is hiding. The FDA is also mandating AI use by staffers and offering a one-month assessment for drugs deemed to serve a national interest. This isn't a tweak — it's a philosophical rewrite of how America decides which medicines are safe enough to sell.

A design on a computer is just a prediction. This week, Boston-based Manifold Bio closed the loop.

NVIDIA used its generative model to design one million novel protein binders — molecules engineered to stick to specific disease targets, the first step in creating many new drugs. Manifold Bio then created and tested all of them simultaneously using its high-throughput platform, measuring over 100 million protein-protein interactions. The result: the AI successfully generated functional, specific binders for 68% of the 127 disease targets tested in the experiment.

That number deserves a moment. Traditional drug discovery tests candidates one at a time, maybe a few hundred in a good campaign. This is massively parallel, compute-driven engineering validated in the wet lab at a scale that simply didn't exist two years ago. It's the same playbook companies like Insilico Medicine are using to push AI-designed molecules into late-stage clinical trials — Insilico's TNIK inhibitor for idiopathic pulmonary fibrosis is now in Phase IIb/III, making it one of the furthest-along AI-origin drugs in the world.

The loop between computational design and experimental validation is now closed at meaningful scale. Expect partnerships, IP disputes, and a scramble to apply this pattern to enzymes, metabolic pathways, and cell-programming parts.

If you know anyone who's abandoned a CPAP mask in frustration, this one is for them.

Two independent pharmacological approaches to obstructive sleep apnea showed meaningful clinical results within weeks of each other. First: a European trial published in The Lancet found that sulthiame — an old epilepsy drug repurposed for sleep apnea — reduced breathing interruptions by up to 47% in the trial in patients with moderate to severe disease. Sulthiame tweaks carbonic anhydrase in the brain's respiratory centers, nudging the body to breathe more consistently during sleep. Its long safety history in epilepsy gives it a meaningful advantage in getting to Phase 3 quickly.

Second: Apnimed reported that its Phase 3 trial of AD109 — a combination of aroxybutynin and atomoxetine that works via a completely different neuromuscular mechanism — achieved a 46.8% reduction in the apnea-hypopnea index on the trial versus 6.8% for placebo on the trial, with the effect holding at 51 weeks. About 23% of participants hit complete disease control in the trial. Apnimed plans to file a New Drug Application with the FDA in early 2026 — which means a decision could come this year.

Sleep apnea affects an estimated 80 million adults in the US alone, and most are untreated because the current solution is a machine strapped to your face. Two independent mechanisms converging on pharmacotherapy simultaneously is a pattern that historically precedes rapid market formation. If follow-up studies confirm benefit and safety, sleep medicine may shift from hardware to pills.

The history of CRISPR is basically a story of learning to make ever-smaller, ever-more-targeted cuts to DNA. The next chapter might be about not cutting at all.

Researchers at UNSW Sydney developed a modified CRISPR system that switches genes on and off by removing chemical tags — methyl groups that act like molecular anchors silencing genes — without making a single cut to the DNA. The work also settles a long-running scientific debate: DNA methylation (those chemical tags) actively silences genes, rather than being a mere consequence of silencing.

The practical target is sickle cell disease. By switching the fetal globin gene back on after birth, the approach could provide a workaround for the faulty adult globin gene — no permanent DNA cuts required. This is still cell-culture work, not yet tested in animals or humans. But the engineering logic is elegant: if you can control gene expression by rewriting chemical tags rather than the DNA sequence itself, you get reversibility and no permanent cuts — two properties that regulators and patients both want.

In a related clinical signal, Wired reported that a small trial transplanted CRISPR-edited pancreatic cells into a Type 1 diabetes patient, with stable glucose for 30 days post-transplant. Different disease, same engineering thread: editing cells for function plus immune evasion, moving toward human studies.

Three signals this week tell the same story: biology is becoming a manufacturing platform.

Amyris opened its fourth production line at its Barra Bonita, Brazil plant — notable because the company was in serious financial distress in 2023-2024, restructured, and is now adding infrastructure. COO Adam Blaziak described the new line's automation as "unmatched," positioning Amyris as a molecule-agnostic platform that can pivot between ingredient categories depending on what customers actually buy. That's the biomanufacturing business model maturing from "we make one cool thing" to "we are a platform."

Meanwhile, New Culture announced a 50,000-liter facility producing casein for cheese at a reported cost near $3 per kilogram — a jump from 5,000-liter pilot runs and a price point approaching parity with traditional dairy. And Perfect Day reported price parity for animal-free whey protein at $3.50/kg at bulk scale. Both are company-reported figures that deserve skepticism, but they drive supply deals and retailer interest when believable.

Dutch startup Vivici also launched its precision-fermented lactoferrin — Vivitein LF — at Expo West 2026. Lactoferrin is one of the highest-value dairy proteins (traditionally $1,500+/kg), meaning precision fermentation doesn't need to match commodity whey prices to win — it just needs to beat an ingredient that's already expensive. That's a much easier bar, and it's precisely the kind of product that makes the economics work right now.

A million proteins designed by AI and tested in a single experiment, a 50,000-liter tank of cheese protein that never saw a cow, and an old epilepsy pill that might replace the machine your uncle refuses to wear to bed. The FDA just told drug developers they only need to be right once — which is either the most efficient regulatory reform in a generation or the setup for a very expensive lesson in confirmation bias. Stay skeptical, stay curious.

If someone you know is building, investing in, or just trying to understand where biology is headed — send them this.

The FTC declared "unfair" AI an enforcement target under Section 5 — biotech companies using AI in trials or diagnostics, take note. Read → FTC Draws a Line: "Unfair" AI Is Now an Enforcement Target

Gut microbes can directly tune memory circuits, making the microbiome a live lever for cognitive aging research. Read → Your Gut Is Running the Controls on Your Memory

ARPA-H committed $144 million to validate surrogate biomarkers and run human trials of candidate healthspan drugs. Read → Longevity

A five-day brain-stimulation protocol matched six weeks of standard treatment for depression. Read → Neuro

Self-driving labs are running closed-loop materials discovery campaigns that traditional R&D can't match in speed. Read → Catalyst