🔥 Trending on HN

Claude Enters More Parts of Drug Development, but Medicines Remain Far Away

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
preclinical

Work done before a possible medicine is tested in people.

antibody

A protein designed to attach to a particular target.

CAR T-cell therapy

A treatment that uses changed immune cells to attack cancer cells.

Anthropic, the company behind Claude, published an early report on October 27. The report describes using Claude across several parts of drug development. Anthropic says it worked with six organizations: BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. It also started its own drug programs. On Hacker News, the item received 178 points and 160 comments. That measures community attention. It does not prove that the reported results are correct.

Background: Drug development has many bottlenecks

Drug development has many steps. Teams find a useful target, design a molecule, improve its properties, prepare manufacturing documents, and plan clinical trials. Anthropic says the delays are spread across this whole process. Its approach is therefore to place Claude inside real drug programs. The goal is to learn where Claude helps and where it needs improvement.

What the case studies report

At Generate Biomedicines, six scientists used Claude Code to propose a shorter genetic design for CAR T-cell therapy. One design was about 70% shorter than the standard version. In lab tests, it helped immune cells make about ten times more receptors. The cells killed tumor cells more efficiently and stayed active longer, according to the report.

At Divergence Labs, Claude generated about 35,000 antibody designs in one month. The project targets food allergies. Early tests found three designs that bound to IgE and 20 that could remove it from immune cells. Three of those 20 appeared stronger than the standard medicine in that test.

BigHat tested about 1,010 variants in about ten weeks. The best versions showed more than ten times tighter binding, an 18-degree-Celsius increase in solution stability, and more than 100 times slower clumping. At Enveda, Claude reviewed 1,061 pages of manufacturing documents. It found all 11 issues human reviewers had found, plus seven more. Review time fell 89%, from over 25 hours to under three.

Why this matters

The report presents Claude as a helper across a workflow, not as a machine that finishes a medicine alone. It can suggest candidates, reconsider experimental data, check paperwork, and help assess trial plans. If these results continue, researchers may test more ideas with less administrative work. That could change how small teams use their time.

However, these are company-reported examples. The page does not establish that every claim has been independently reproduced. The numbers also come from different projects, so they should not be treated as one controlled comparison.

What is confirmed

Anthropic says six partner organizations are involved. It gives concrete counts for designs, variants, documents, and review time. It also says it acquired Mithril Medicine, where a possible muscular-dystrophy treatment moved from an idea to promising mouse results in under 18 months. Bristol Myers Squibb is using Claude to analyze early clinical-trial plans, with an expansion planned for 2027.

What remains unknown

Anthropic explicitly says the work is still early. Most of it is preclinical, meaning the candidates have not reached testing in people. The report does not show how many candidates will become medicines, how safe they will be, or whether outside teams will reproduce the findings. A promising lab result is not the same as a proven treatment.

What to watch next

The next checkpoints are the combination of Divergence’s two antibody functions, further work on Mithril’s candidates, and the planned 2027 expansion with Bristol Myers Squibb. The important evidence will be human trials, independent checks, and public results that include failures as well as successes.

💬 AI may speed some drug work, but biology and clinical proof remain slow

Comments saw plausible gains in molecule search, data work, and paperwork, but stressed that experiments, animal studies, human trials, and regulatory review remain the slow proof loop. Early cases and personal anecdotes do not yet establish a large clinical benefit.

  • The proposed near-term upside is candidate generation and selection, mining older data, and handling documents. That may improve productivity, but it is not evidence that AI will produce many new cures.
  • A biotech-industry commenter’s self-report said results from two clients using ML/AI were mixed. An ML-identified drug failed animal testing, while another had better efficacy but a worse safety profile and required a new trial said to add two years and $10 million. These are anecdotal self-reported results.
  • A drug-discovery startup worker’s self-report said that, after nine years, eight clinical candidates had advanced and every one encountered a version of the problem of translating preclinical work into human trials. This is personal experience, not controlled evidence.
  • Human biology is a systems problem. Knowing how a drug binds to a protein often does not provide enough data to predict downstream effects across tissues, diseases, and patients.
  • Simple molecular screening misses factors such as solvent positions, molecular flexibility, electrostatics, entropy, coordinating ions, and time-dependent conformational changes. Several commenters argued that specialized models with system-specific data may be more suitable than a generic LLM.
  • Wet-lab chemistry, animal studies, clinical trials, and regulatory steps create long feedback loops. Robotics and parallelization could help, but they do not remove expensive experiments or the sequential parts of validation.
  • Some users self-report that a public Claude model might refuse biology questions, and that they experienced refusals or downgrades in security-related work. This is individual experience or suspicion, not a systematic benchmark.
  • Automating paperwork may reduce workload without removing the underlying bureaucracy. A commenter who said they have a rare disease welcomed possible help but remained cautious about slow experimental and clinical work, and viewed deeper integration into many biotech workflows as unsettling.

growing digest at 303 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Claude Helps With Many Steps of Drug Research

📰 Full story: Claude Enters More Parts of Drug Development, but Medicines Remain Far Away

Anthropic says Claude helped teams test ideas, study results, and check documents.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
preclinical

Testing done before people receive a possible medicine.

antibody

A protein that attaches to a particular target.

receptor

A small cell part that receives a signal.

💡 The gist

  • Claude helped with drug ideas, experiments, and paperwork.
  • Some early tests showed promising results.
  • No finished medicine has been proven yet.

What did Anthropic report?

Anthropic is the company that makes Claude. It shared early results on October 27. Six organizations worked with Anthropic. They included research and drug companies. Claude joined several parts of their work.

Drug research has many steps. Scientists choose a target. They design possible medicines. They test those designs. They prepare documents. They also plan studies with people. Anthropic says Claude can help across this whole path.

At Generate Biomedicines, Claude helped design instructions for immune cells. One design was about 70% shorter. The cells made about ten times more receptors. They also killed tumor cells more efficiently. The cells stayed active longer in lab tests.

At Divergence Labs, Claude made about 35,000 antibody designs. It made them in about one month. The designs targeted a food-allergy project. Early tests found several useful-looking designs. People still need to test them more.

At BigHat Biosciences, the team tested 1,010 versions. The work took about ten weeks. At Enveda, Claude checked 1,061 pages. It found 11 known problems. It found seven more problems, too. Human review time fell from over 25 hours to under three.

Why does this matter? Claude did not create a finished medicine. It helped people handle many smaller tasks. That may let teams test more ideas. It may also reduce paperwork.

The results have important limits. Anthropic made the report itself. Most work is preclinical. Preclinical means testing before people receive the medicine. The report does not prove safety or success in humans.

Hacker News listed 178 points and 160 comments. This shows community attention. It does not prove the results are correct.

Next, watch human studies and independent checks. Also watch the planned 2027 work with Bristol Myers Squibb. Those steps will show whether early results can continue.

💬 AI may help, but the hard part of drug development is still slow

The thread’s main view was that AI might speed up candidate searches and paperwork, while proving that a drug works safely in people remains a different and difficult problem.

  • AI might help find drug ideas, search older data, and prepare documents. That could save time, but it is not proof of new cures.
  • Self-reports from people in drug development describe mixed results: an AI-found drug failed animal testing, and one startup worker said all 8 of 8 candidates had trouble moving toward human trials after 9 years. These are personal reports.
  • Bodies are complicated and the data needed to predict all effects is often missing. Matching a molecule to a target is not enough; specialized models may be needed.
  • Lab work, animal studies, clinical trials, and regulatory checks take a long time. Robots and parallel tests may help, but they cannot skip the proof.
  • Some users self-report that public Claude may refuse biology questions or has refused security-related work. Paperwork tools may help, but bureaucracy and worries about deeper company involvement remain.

growing digest at 303 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Claude Helps Make Medicine Ideas

📰 Full story: Claude Enters More Parts of Drug Development, but Medicines Remain Far Away

Claude is helping grown-ups try new medicine ideas.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
receptor

A small cell part that helps it notice something.

Hacker News

A website where people discuss technology news.

Anthropic is a company that makes Claude, an AI helper. Claude is helping people who make medicines. It gives them possible ideas. It also checks their work.

One team asked for food-allergy medicine ideas. Claude made about 35,000 designs. It made them in about one month. People tested some designs afterward.

Another team made a shorter cell instruction. It was about 70 percent shorter. The cells made about ten times more receptor parts. These parts help cells notice things.

Another team gave Claude 1,061 pages. Claude found 11 known problems. It found seven extra problems. Checking the pages became much faster.

These are early tests. Most tests have not used people yet. No safe medicine is proven.

Hacker News had 178 points and 160 comments. That means many people noticed the story. It does not prove the story is true.

People must do more tests. They must check safety. They must learn whether these ideas become medicines.

💬 AI can help with drugs, but people still have to test them

AI can think quickly. The human body is much harder to check.

  • AI can help think of drug ideas and sort papers. People still have to test the drugs.
  • People who work in this area self-reported that some AI-found drugs failed animal tests and that candidates can run into trouble before human trials.
  • The body is very complicated, so tests take a long time. Robots may help, but they cannot skip the tests.
  • Some users also self-reported that the public model refused biology questions. A person with a rare disease hoped for help but stayed cautious about deeper company involvement and paperwork rules.

growing digest at 303 comments (revision 3). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources