Claude Science

24/7 continuous loop of AI trying to cure cancer.
8k6YidRtD893CYRRKeWvGp2ixwTVU2haf5ymrwwZpump

[Claude Science @ClaudeScience]

The volume is back and were not doing any good with it.

Introducing Claude Science a continuous loop using AI to accelerate cancer research.

Using $Science creator rewards, we can bring a real success story to @Pumpfun.

https://t.co/hXKIeUvOaz

[Claude Science @ClaudeScience]

Most interesting thing it's proposed so far:

deferasirox- an approved iron chelator, against VHL-null kidney cancer, through an iron/ferroptosis axis, alone and combined with GPX4 blockade.

Approved drug. Known safety profile. Testable by anyone who already has the cell lines.

Cheap to try, cheap to kill. That's the good kind of hypothesis.

[Claude Science @ClaudeScience]

Chordoma is rare enough that most people treat it as one disease.

It doesn't. Four copy-number clusters, each routed somewhere different — CDKN2A/chr9-loss tumours to CDK4/6 inhibition, chr2+7-gain tumours to GLI.

[Claude Science @ClaudeScience]

Germline TP53 breast cancer, from the archive:

ERBB2 amplification is the only recurrent co-driver, and the apoptotic arm — BAX, TP53I3 — is transcriptionally dead.

So: HER2-directed ADCs, and stop expecting anything routed through apoptosis to work.

Dead pathways are worth mapping.

[Claude Science @ClaudeScience]

Why CAR-T stalls in the pleural cavity, per the desk:

NF2 copy-number loss drives retinoic-acid- and chemokine-dependent recruitment of CRIg+ macrophages — and that's what makes the resistance cavity-restricted.

Not "the microenvironment is immunosuppressive." A named cell, a named axis, a named reason it's local.

[Claude Science @ClaudeScience]

Its sharpest clinical call so far:

stop gating PARP inhibitors in non-BRCA carriers by gene name. Gate them on function — RAD51 foci plus a genomic HRD scar — because PALB2/RAD51C/RAD51D and ATM/CHEK2 don't behave alike.

The gene is the label. The function is the target.

[Claude Science @ClaudeScience]

Give it a dead tumour suppressor and it goes hunting for the partner.

Ewing: EWSR1::FLI1 gets pulled to double-strand breaks, ATM goes quiet, ATR becomes collateral.
Osteosarcoma, TP53/RB1-null: KIF18A, selective for chromosomal instability.
TP53-mutant adrenocortical: WEE1 + RSK1.
You can't drug a loss. You drug what the loss depends on.

[Claude Science @ClaudeScience]

It told a drug to stay in its lane.

In MEN1-related neuroendocrine tumours, c-MET protein showed up in only ~20% — 3 of 43 strongly.
So the call wasn't "try cabozantinib." It was: test it prospectively only in the expressing minority.
Most of the value is in the word "only."

[Claude Science @ClaudeScience]

It bet against the trial that's further along, and said why.

In HPV+ head and neck cancer it ranked a high-avidity E6-directed TCR-T above the low-avidity E7 TCR currently advancing clinically — on the grounds that the E7 clone was lost in metastasis.

Disagreeing with the frontrunner only counts if you name the reason.

[Claude Science @ClaudeScience]

Eight childhood cancers in one rotation:
neuroblastoma, medulloblastoma, retinoblastoma, Wilms, osteosarcoma, Ewing, rhabdomyosarcoma, ATRT.
Each got its own dossier, its own leads, its own open questions.

Paediatric oncology is small markets and hard biology. A loop doesn't care about market size.

[Claude Science @ClaudeScience]

Systemic mastocytosis: CNS-sparing KIT D816V-selective inhibition, KIT VAF as the pharmacodynamic endpoint.
Paraganglioma: HIF2α antagonism for the pseudohypoxic cluster, plus ¹⁷⁷Lu → ²¹²Pb alpha escalation.
Renal cell: an iron chelator.
50 cancers in. It hasn't repeated itself yet.

[Claude Science @ClaudeScience]

Cholangiocarcinoma, sharpened.

Not "FGFR inhibitors work and then stop working."
Its claim: co-occurring CDKN2A/B or PTEN loss not kinase-domain mutation, drives the short pemigatinib responses. So the next move is FGFR + CDK4/6, not a better FGFR inhibitor.
Specific enough to be wrong.

[Claude Science @ClaudeScience]

Not every lead is a drug for people who are already sick.

For Lynch syndrome it ranked low-dose aspirin chemoprevention and organ-sparing checkpoint blockade in germline carriers, plus a specific question about why TGFBR2 frameshifts behave asymmetrically.

Prevention is a therapeutic strategy. It gets ranked like one.

[Claude Science @ClaudeScience]

Juvenile myelomonocytic leukemia. Rare, paediatric, brutal.

Its lead: a repurposable triplet, IL-17A neutralisation, celecoxib, NLRP3 inhibition, layered on MEK blockade, in primary patient samples.

Rare cancers don't get new chemistry. They get combinations of things already in humans.

[Claude Science @ClaudeScience]

Every visit ends by writing the questions for the next one.

Verbatim, from the CML dossier:
"What is the kinase-domain mutation spectrum at asciminib failure, specifically in the ~21 T315I patients from NCT02081378 who never reached MMR?"
That isn't "more research is needed." That's naming the missing table.

[Claude Science @ClaudeScience]

It read a failed trial correctly.

eRapa missed its primary endpoint in familial adenomatous polyposis. Most summaries stop there.

It logged the miss, isolated the duodenal-specific signal under intermittent dosing, and rebuilt the programme around that instead of the colorectal endpoint that failed.

Negative results are data.

[Claude Science @ClaudeScience]

Most interesting thing it's proposed so far:

deferasirox - an approved iron chelator, against VHL-null kidney cancer, through an iron/ferroptosis axis, alone and combined with GPX4 blockade.

Approved drug. Known safety profile. Testable by anyone who already has the cell lines.

Cheap to try, cheap to kill. That's the good kind of hypothesis.

[Claude Science @ClaudeScience]

State of the board:

50 cancers worked. 222 live hypotheses. 8 have survived independent corroboration. 1 retired.
That ratio is the whole point. A hypothesis generator that promotes everything is a random number generator with citations.

All findings get published in the archive https://t.co/c2eMtNVAr9

[Claude Science @ClaudeScience]

This is a loop, not a chatbot.
One cancer per visit. 72 on rotation. Memory is the dossier.

Every visit:
1. pull new papers and live trials
2. integrate them into what it already believes
3. strengthen, weaken or retire each lead
4. write its own queries for next time

One attempt. Nothing retried for a better answer

Watch Claude try to cure cancer live:
https://t.co/jSoJ92MpR3

[Claude Science @ClaudeScience]

Dev supply locked until end of year. Lets be the change we need to see onchain.
https://t.co/sNbTi3S6TA