Anthropic reports Claude Code cut a CAR-T component 70% and generated 35,000 antibody designs in a month
- Anthropic's life sciences team ran research partnerships with six organizations (BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, The Column Group) and started its own discovery programs, including rare childhood muscular dystrophy.
- Six Generate Biomedicines scientists used Claude Code to redesign part of the CAR T-cell genetic instructions; one design was about 70% shorter, let immune cells make roughly 10 times as many receptors, and killed tumor cells more efficiently in lab tests.
- For Divergence Labs, Claude turned food-allergy drug goals into a design plan and then largely autonomously generated roughly 35,000 antibody designs in about a month; early tests found 3 IgE-binding hits and 20 that dislodge IgE, 3 of which beat the standard medication.
- Anthropic argues the bottlenecks are distributed across the whole development process (target selection, molecule design, manufacturing, regulatory filings and trial design), so it is embedding Claude in real programs instead of attacking one problem.
- Other work covers lead optimization at BigHat Biosciences and manufacturing and filings at Enveda, and Anthropic says what it learned has already driven prioritized model improvements.
Hacker News 의견들
I have a rare disease with no cure, so I want this to work. But murine models still take time and trials are slow, and most biotechs lose money at exactly this stage.
That line about needing to integrate Claude more deeply into biotech development processes sounds invasive. If it isn't working, the fix is apparently to spend more on Anthropic's products.
Have you actually tried their best public model on your problem yet? That's a cheaper first step than waiting on the industry.
The giant chapter picker pinned at the top is funny when there are only two chapters and the second one is a tiny blob at the end.
Verification loops in biology are just much longer: wet lab, trials, regulatory steps. Sequential steps are slow, so I expect a lot of parallelization, and I don't know how regulators keep up.
Math is fully formalized by design and biology is the opposite, so AI will move much faster in the axiomatic fields first.
I'd bet robotics comes before biology. A human can look at a robot and tell whether it works, which is more verification than most bio results get.
The real limitation is humans: emotions, laws, liability. An agent driving robotic arms could iterate on real experiments, but AI already finishes a task before I have time to assign the next one.
'Generate scientist' as a job title, is that like prompt engineer?
It's six scientists from Generate Biomedicines, the company. Short for the company name, not a new profession.
The line between good and bad AI use is 'machine, do magic' versus 'machine, do x, y, z, avoid i, j, k.' Articulating what matters is the hard part.
The Enveda bit about manufacturing and filings is the unexpected part. Documentation volume has outgrown our ability to produce it reliably, and construction projects have the same problem.
A lot of the helping here is just filling paperwork. If that's the bottleneck, the bureaucracy is the problem, not the document.
Regulatory filings aren't lorem ipsum, real intelligence goes into them, which is why templates never fixed it. Going from two weeks of sequential review to a two hour session is a 25% speedup worth having.
Cautious optimism is right, and that line about needing many more biotech partners reads as self-serving.
Sure it's self-serving. I'm optimistic about AI in medicine generally, not about Anthropic specifically.
Drug discovery is heavily regulated for good reasons, so it'll be one of the last things AI actually disrupts.