Claude Takes Aim at Cancer: 35,000 Antibodies and an AI-Assisted Hunt for a Cure

  It's the first time in the history of the world.

It's not like Claude's doing drugs.

The day before, Anthropic had just filed its IPO, and today it's turning its head in the drug development circle, and the wave of operations is extremely hardcore.

It's the first time in the history of the world that we've been able to put "one pharmaceutical company" into the revenue pocket of Mithril Medicine.

The founder, Cody Tranbarger, is a single-armed horse, and the only tool is Claude.

Based on this combination, he coordinated 11 external laboratories in less than 18 months, directing a rare disease therapy from ideas to mouse experiments, and the results were real.

The experimental pipeline has now been formally merged into Anthropic and is expected to enter human clinical trials by the end of 2028.

Moreover, Anthropic has partnered with six leading biopharmaceuticals and venture capital firms to directly kill Claude at the heart of cancer, food allergy treatments and clinical trials.

The results were shocking.

Claude raves about 35,000 antibody designs a month, shortening the critical component of CAR-T by about 70%, and reducing the review time for pharmaceutical production documents by 89%.

Claude now wears a white coat and is launching a full-scale attack on cancer and rare diseases.

In this long-running official report, the most shocking operation was the direct acquisition of Mithril Medicine by Anthropic.

You might not believe it, but from its inception to its acquisition by the giant, the company was founded by Cody Tranbarger.

He is suffering from a terminal illness called "LAMA-2 congenital muscular malnutrition".

This is an extremely cruel and rare pediatric epilepsy: a child is born without muscles, and tracking data from 42 seriously ill children shows that only one person can walk independently.

One in six do not live to adulthood and there are currently no approved basic therapies globally.

The root of this problem lies in the connections between muscle fibers and external support structures.

Outside each of the roots of healthy muscles, there is a layer of structure called the "baseline membrane". One of the key components, called the ligand binding protein α2, is involved in the protein complex, which both helps to weave the web and is responsible for placing the webs on the surface of muscle cells.

After α2 deficiency, muscle fibers lose important support and are gradually damaged.

Cody entrusted almost all of the scientific and operational work except for the core decision-making to Claude.

In less than 18 months, Cody and Claude formed an extreme "carbon base+carbon base" duo.

Address of the paper:

They designed a "dual-specific bridging antibody" and the idea was to create a "molecular book binding nail".

Since the replacement protein is there, the beta-dystroglycan is also present on the mucous membrane, and the only thing missing is the ligand connecting the middle roots.

This creates a double-specific antibody: one hand grabs the LG4-5 structural region of the ligand-binding protein-α4, and the other hand grabs the beta-dystroglycan and hardens the basement membrane back into the mucous membrane.

In cell experiments, the two molecules that ultimately prevailed were able to hold the cells tightly to the coating of the target protein, holding two rounds of centrifugation.

Subsequently, Tranbarger's experiments in the body of the diseased mice thoroughly verified this design.

Six consecutive weeks after birth: the three muscles induced force in the treatment group was 1.95 times that of the control group; the grip force was 1.94 times that of the control group; and the existing gene therapy for positive control was only 1.31 times.

Antibodies have also come to the right place: days after the last dose, the antibodies can still be recovered from the muscles to catch two points at the same time, and the concentration falls within the most easily "bridge" range.

So, in this hardcore nuclear research, how does Claude help one person run 11 outsourced CRO labs?

The technical report, which uses a single chapter, summarizes a real week of "ultra-custom drugs" after entering the phase of mass screening:

During the four working days of the team's cell-linking experiments, when faced with 11 unresolved experimental details pending between the CROs (dropping pores, packing by protein concentrations, transport conditions, etc.), humans only need to tap two phrases and click past a folder link.

After reading all the documents, Claude gave the core conclusions in just 96 seconds, and in 18 minutes, he came up with a complete promotion plan and five English mail drafts that could be sent directly.

On the fourth day of the week, the CRO responsible for production sent out a 341-page 113-molecule production and quality inspection data package.

After 2 minutes and 40 seconds, Claude gives the core conclusion; after 7 minutes and 53 seconds, Claude reads the 341-page document in one word and directly draws a 14×8 visualization decision matrix map.

In this experiment, from the superimposed contrast of the underlying AlphaFold and the crystal structure, to the writing of a Python data processing script, to the drawing and drafting of an entire 38-page academic-level paper chart, Claude mastered almost all of the gravitational pull.

Today, one person can work with AI to create a new drug, and it really shocks everyone.

In addition to acquiring Mithril itself, Anthropic has in the past few months deployed its own life science team in close collaboration with six institutions.

These include Generate Biomedicines, Divergence Labs, BigHat Biosciences, Enveda, The Column Group and BMS.

In the last few months, they've done something super-hard with Claude.

CAR-T therapy has been successful in treating blood cancers, but has not been successful in the face of physical tumors.

Generate Biomedicines wants to transform it, with the idea of "compressing" a section of the therapeutic genetic instructions to create a space load for new instructions for other cancers.

Six scientists came up with a new design using Claude Code.

One is about 70% shorter than the standard version, but performs better in the lab: the number of receptors on the surface of immune cells used to identify cancer cells increases by about 10 times, killing tumor cells more efficiently and staying active longer.

It is also short and strong, which is uncommon in bioengineering.

The "guilty" of food allergies is a molecule called IgE, which is partly absorbed into the bloodstream and partly stuck in immune cells.

Divergence Labs wanted to take a steeper route: to directly remove IgE from immune cells. This required a molecule with two "arms": one that grabbed the IgE and the other that was responsible for separating it.

Claude first dismantled the target into a detailed design scheme, and then operated largely autonomously, producing about 35,000 antibody designs for the team's screening tests in about a month.

As a result, both "arms" found seeds. Three designs can combine IgE, and 20 can combine IgE, three of which are more effective than standard drugs.

It's a good medicine that not only binds tightly to the pathogen, but also holds up to production, storage, and other bindings.

BigHat Biosciences brought Claude into their own experimental data system and proposed a modification scheme for two antibody fragments. The lab created and tested about 1,010 variants in 10 weeks.

In addition, Enveda's production chemists have cut 89% of the audit time to free humans from the literature; The Column Group is preparing a new company, Pleiades, to use Claude to run a small team of more complex drug projects; and TIME is using Claude to evaluate a clinical trial plan for an early drug and predict the probability of a successful trial.

In the past, AI drugs have largely stayed at the breakthrough point: predicting protein structure, selecting compounds, and generating molecules.

AI is a useful knife, but it's always man who decides where to cut it. This time it's different.

• Upstream: Assessment of target values is not worthwhile;

• Design: generate antibodies, reconstruct CAR-T components;

• Optimization: substituting candidate drugs or even correcting experimental indicators;

• Production and declaration: quality auditing documents, drafting of regulatory documents;

• Clinical: planning trials, anticipating success rates;

• Operations: signing contracts, managing logistics, coordinating 11 laboratories.

It not only answers questions, but also breaks down targets into programs, executes autonomously, and discovers loopholes in methodology.

This is close to the sum total of a junior researcher plus a project manager.

And Anthropic is no longer content to "sell coconuts" to gold miners, but to personally acquire pharmaceutical companies, to personally come out and make drugs.

The Law of the Twenty has dominated the pharmaceutical industry for decades, and behind it are innumerable new medicines that patients cannot afford, and countless rare diseases that have been put on hold because of "unpredictability".

Now, a person with an AI has gone 18 months on a journey that has taken years.

Of course, the real test is in human clinical trials.