When people talk about AI discovering new drugs, they usually picture a super-brain on a screen calculating a miracle molecule from scratch. Yet the actual bottlenecks inside the pharmaceutical industry look almost the exact opposite. First-cycle Complete Response Letters issued by the US FDA over the past five years show that up to 74% of rejections stem from factory manufacturing and quality-control defects rather than the drug molecule itself. Inside pharma companies, analytical chemists spend roughly 75% of their working hours manually checking hundreds of pages of scanned, handwritten batch records just to verify that operators did not miss a signature and that scales were calibrated on schedule. Even in molecular design, scientists are routinely blocked by the lack of reliable physical simulators or by software bugs inside precision lab instruments.
That sobering engineering reality delivered a public lesson just a week ago in a scorpion-venom protein test disclosed by Anthropic itself. In that effort, a computational team used a large language model to drive protein-design tools and generated 318,000 candidate proteins in a few dozen hours; when tested at the wet bench, only a dozen or so even stuck to the toxin surface, and the number of functional antivenoms that actually neutralized toxicity was 0. Yet just days later, on October 27, Anthropic released a starkly different drug development progress report, disclosing deep pipeline collaborations with six organizations—Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group, and Bristol Myers Squibb—alongside wet-lab activity data surpassing current clinical standard-of-care drugs and the July acquisition that formed its wholly owned therapeutics subsidiary, Mithril Medicine. In a biological world where reliable simulators are still far from universal, how a large language model actually moves drug discovery forward is hidden right inside these unglamorous engineering trenches.
After AlphaFold won the Nobel Prize, many assumed that solving 3D protein structures would naturally let us compute new drugs. But existing 3D structure simulators have a built-in capability boundary: they only excel at molecules with rigid, fixed 3D shapes like Lego bricks. Once they encounter flexible structures that wiggle constantly inside a cell like wet noodles without a fixed shape, or dynamic physical processes like prying open a lock in slow motion, existing simulation software draws a complete blank.
Facing dead ends where simulators turn in blank sheets, brute-forcing 3D physical structures is a losing game. What Claude did inside two partner biotechs was to help scientists trade an uncomputable problem for one they could actually solve.
In Generate’s cancer immune-cell engineering program, the goal was to drastically shorten the long intracellular signaling tail that sounds the alarm inside an immune cell, freeing up cargo space inside the lipid nanoparticle delivery vehicle for other anti-cancer payloads. The catch is that this alarm tail is a flexible wet noodle with no fixed 3D shape, so physical simulators cannot compute how to trim it. Claude abandoned 3D geometric simulation altogether, read through more than two hundred immunology papers, compiled existing short signaling modules from different immune cells into a dictionary, and recombined them like Lego blocks while preserving the core alarm switch. Through this literature-driven problem reframing, the team cut the tail’s length by roughly 70%; in wet-lab tests, the immune cells not only retained their cancer-killing function, but boosted surface production of cancer-detecting receptors by roughly 10-fold.
The same reframing intuition powered Divergence’s severe food-allergy drug program. Existing allergy drugs can only mop up free-floating allergy molecules in the bloodstream and are helpless against allergy triggers already locked tightly onto immune-cell surfaces. To truly defuse the bomb, a new drug must forcibly pry those bound allergy molecules off the cell—yet simulation software can only score how tightly static shapes fit together, not how to pry open a lock in slow motion. After reading the structural literature, Claude turned this dynamic lock-picking problem into a static geometry problem. It noticed that the allergy molecule stays locked onto the cell only because a door-hinge-like joint in its middle is bent into a fully open position. Design a tiny wedge to jam beside that hinge and trap it in a half-closed, awkward angle, and the lock pops open on its own. Following that blueprint, Claude compressed the early discovery cycle from a traditional nine months down to three or four weeks within a single month-long campaign; in wet-lab tests, 20 designs successfully pried the allergy molecule off cell surfaces, and 3 outperformed the current clinical standard-of-care drug.
Even when off-the-shelf computational software exists—and even when robotic labs can automatically test a thousand proteins a week—research still routinely hits a wall. Blocking the way are two maddening engineering traps: search algorithms walking in circles, and laboratory instruments reporting falsified numbers.
Before Claude stepped in, the search algorithm used to optimize drug sequences behaved like a blindfolded hiker who only steps uphill, getting trapped circling the same small local hillock over and over and spitting out nearly identical molecules. Worse still, to game a high score, the algorithm would let designed molecules grab the wrong spot or leave a superglue-like greasy patch on the surface that caused them to clump together inside high-concentration liquid formulations. Rather than acting as a button-pushing operator, Claude dove straight into the sampling software’s underlying code and added an alternating warm-and-cool temperature schedule: whenever the search got stuck on a local hillock, it automatically raised the random exploration temperature to jump out, then cooled back down to lock in the structure once a new path opened up. At the same time, it added explicit penalty terms to the scoring formula for grabbing the wrong spot and for surface clumping, pulling the search software out of its dead end.
Even more insidious than a stuck algorithm is a lab instrument reporting bad books. In BigHat’s antibody thermostability campaign, the team used a UV laser heating instrument to ramp up the temperature on nearly a thousand antibody variants and measure their heat limits, feeding the melting temperatures exported by the instrument software into machine-learning models. After the first round, all seven predictive models failed completely, and the instrument’s built-in quality-control software flagged more than half of the experimental wells as garbage and threw them out.
Prompted by a human researcher, Claude bypassed the instrument’s summary spreadsheet and pulled out the raw heating curves for every single well to inspect them one by one—catching an absurd bug in the instrument software. The instrument normally estimates melting temperature by tracking fluorescence changes as a protein unfolds under heat, but some antibodies clumped and precipitated like boiled egg white before they finished unfolding, causing the laser reading to plummet just the same. The instrument’s built-in formula could not tell the two apart and lumped signal drops from high-temperature clumping together with normal thermal unfolding, scrambling the thermostability scores across the board. Combining light-scattering signals, Claude rewrote the physical scoring formula to separate high-temperature clumping from unfolding, rescued more than four hundred discarded wells, brought the predictive models back to life, and helped the team select an upgraded lead molecule that tolerated 18 degrees Celsius higher temperature in liquid formulation.
Computing a promising molecule at the bench is barely the first step in a new drug’s long march. The back half is a swamp of regulatory paperwork and cross-organization coordination. As noted at the outset, more than 70% of first-cycle FDA rejections trace back to manufacturing and quality-control flaws, and pharma analytical chemists spend three-quarters of their time visually checking hundreds of pages of scanned handwritten batch records.
Feed several hundred pages of handwritten batch-record scans straight into a general-purpose large model, and it will quietly skip dozens of pages halfway through—making it unusable inside a regulated pharmaceutical plant. Enveda’s batch-record review system puts Claude inside a strict compliance cage with a rigid exoskeleton: it feeds only forty high-resolution scanned pages at a time, uses external Python code to maintain a page-by-page coverage ledger that forces a rerun if a single page is missed, and restricts the system to read-only access so human experts make the final sign-off. In a blind test across more than a thousand pages of real clinical batch records, the system caught every issue found by human reviewers plus 7 additional real manufacturing deviations that humans had missed, compressing nearly twenty-six hours of manual review down to 2.8 hours.
More moving than saving time for pharma companies is how this project-management capability brings hope to children with ultra-rare diseases. Take congenital muscular dystrophy in children: with fewer than 5,000 patients across the US and Europe, affected children are born with severe muscle weakness and often do not survive to adulthood. Scientists had long known how to design a bridging molecule that acts like an artificial rivet to re-anchor muscle fibers, yet commercial drugmakers consistently walked away. The reason is brutal: pushing that single drug into the clinic requires simultaneously coordinating 11 specialized external contract research laboratories, from breeding mouse models and running toxicology to manufacturing clinical-grade material. For a traditional drugmaker, simply hiring a twenty- to thirty-person team of project managers, procurement specialists, legal counsel, and regulatory staff to wrangle those vendors creates a fixed overhead so high that a market of fewer than five thousand patients is guaranteed to lose money—killing the rare-disease drug by administrative overhead alone.
Mithril Medicine, the therapeutics company Anthropic acquired and formed this July, puts Claude to work as a super project manager. As documented in the Mithril rare-disease preprint, a single founding scientist teamed up with Claude to manage all 11 external vendors, handle cross-vendor experimental protocols and daily handoff emails, and read through more than three hundred pages of vendor manufacturing quality-control reports in a few minutes to triage scheduling. By slashing fixed management overhead, this rare-disease pipeline completed in-vivo validation in diseased mice in under eighteen months—restoring muscle strength in treated mice to nearly double that of untreated mice—and is preparing to file for human clinical trials next year.
String these stories together, and the most illuminating takeaway is how different they look from the popular imagination of AI doing science. Look back at Anthropic’s moves across science over the past few months—from launching desktop tools mid-year to help scientists clean databases and write supercomputer scheduling scripts, to doubling the runtime speed of more than thirty biomolecular computing packages in September without changing a single scientific formula, to embedding inside six biomedical organizations’ pipelines and building its own drug company today—and the path that actually works in the lab shares a unmistakable signature.
What large models are worst at is often the exact thing people hope they will do: stepping in for scientists to make major scientific decisions without real experimental feedback. In a blind test on engineering anti-cancer immune cells, for instance, researchers asked five frontier large models to score a set of candidate sequences before seeing any wet-lab data, testing whether they could pick out the winners zero-shot. Every model’s scores were barely better than flipping a coin. And in a separate control test, the moment human scientists removed a baseline rule requiring the core alarm switch to be preserved, the algorithm deleted the alarm switch altogether to boost cell survival and surface receptor yield—producing a “blind” immune cell that aced every surface metric on paper but refused to fire a single shot when facing cancer cells. Setting the direction, guarding the physical baseline, and having the judgment to suspect something is wrong and hit the brakes when data looks off remain irreplaceable human territory.
Conversely, once human scientists lock in the core direction and physical guardrails, large models are remarkably effective at doing the supporting legwork. Modern scientific research has long since moved past the romantic image of Newton sitting under an apple tree waiting for a flash of inspiration; more than eighty percent of researchers’ time goes into unglamorous engineering friction and organizational coordination. Where Claude actually delivered results across these pipelines was by reading hundreds of papers to reframe uncomputable physics problems into combinations of existing modules, diving into legacy software to patch sampling code and pull thousands of raw instrument curves to fix broken scoring formulas, and pairing with deterministic code ledgers to audit over a thousand pages of handwritten factory records while helping a single scientist coordinate eleven external labs. It does not need to be an all-knowing miracle doctor; simply crawling into the engineering blind spots to trade uncomputable problems for tractable ones, recalibrate distorted instruments, and flatten the crushing administrative overhead that prices life-saving rare-disease drugs out of existence is already enough to change the economics of scientific discovery.