Broad Institute · April 18, 2025

Inside a real FDA clinical hold

In April 2025, FDA put part of the Broad Institute's prion siRNA trial on clinical hold. We rebuilt the program's files from the public record — the filed dossier, the agency's letter, and the lab data behind them — and had AI models do the work the response required.

The best model passed most checks — and still authorized work FDA had frozen. See where it broke.

Assignments: 4
Model runs scored: 7
Models tested: Claude Opus & GPT-5.6

1What does this company do?

The Broad Institute is a nonprofit research institute; one of its labs is developing a medicine for prion disease.

Prion disease is rare, always fatal, and caused by a normal brain protein (the prion protein) folding into a toxic shape. The lab’s idea is to lower how much of that protein the brain makes. This is a real program, and its filings are public.

For experts

Sponsor: Broad Institute of MIT and Harvard. Target: PRNP / PrPᶜ. Regulator: FDA CDER, Office of Neuroscience, Division of Neurology 1.

2Where is the product on its way to patients?

Its drug, 2439-exNA, is ready for its first human studies, but the FDA has frozen one of them.

1Discovery
2Lab & animal studies
3Human trials
- 2439-exNA
4Filing & review
5Approved
6On the market

2439-exNA is a small piece of genetic material (an siRNA) injected into the spinal fluid that tells cells to make less prion protein. The lab asked the FDA to start two studies: one in people who already have symptoms and one in healthy people who carry the gene mutation. The FDA let the first go ahead and put the second on hold.

For experts

Divalent siRNA, intrathecal. IND 167326 (February 2025). Protocol 5.3.5.1.1 (symptomatic) may proceed; protocol 5.3.5.1.2 (presymptomatic carriers) on partial clinical hold under 21 CFR 312.42(b)(1)(iv).

3What is happening right now?

On 14 March 2025 the FDA put the healthy-carrier study on hold until a monkey safety study is done.

Healthy carriers may stay well for years, so the FDA wants near-certainty the drug is safe for them. Its letter asks for a longer safety study in monkeys and adds further, non-blocking comments on the lab test, the animal record and the manufacturing math. The lab must respond to all of it.

Next level: pick a job below to see who had to do what, the files, the model’s steps, and each check. Choose a job ↓

Company documents

Documents from the real public record.

FDA correspondence · PDF

The letter that froze a trial

FDA placed the presymptomatic study on partial clinical hold because the required animal safety study was never run. Until the agency says otherwise, no activity under that protocol may begin.

Two protocols, two different answers: the symptomatic study may proceed; the presymptomatic study may not be initiated. Every file in this record exists to answer this letter.

IND 167326 · Ref ID 5570921PDF

Page 1 of the real FDA partial clinical hold letter for IND 167326

Source excerpt · Letter ¶1–3; signature page, April 11, 2025The letter itself misspells the asset name — “2349-exNA” — a detail preserved verbatim from the public record.

1 / 7

The desk

The real filing, nothing simplified

Every file comes from the program's public record — sections of the filed IND, the agency's correspondence, the raw plate exports, the sequence tables. The same record a regulatory team would open.

Files in the record: 294
Data tables: 157 TSV / CSV
Filed documents: 103 PDFs
Sequences & figures: FASTA · 20 PNG

Browse the actual record

figure-1.png
figure-2.png
figure-3.png
figure-4.png
figure-5.png
figure-6.png
figure-7.png
figure-s01.png
figure-s02.png
figure-s03.png
figure-s04.png
figure-s05.png
figure-s06.png
figure-s07.png
figure-s08.png
figure-s09.png
figure-s10.png
figure-s11.png
figure-s12.png
stats_for_text.txt
supplement.xlsx
table-s01.tsv
table-s02.tsv
table-s03.tsv
table-s04.tsv
table-s05.tsv
table-s06.tsv
table-s07.tsv
table-s08.tsv
table-s09.tsv
table-s10.tsv
table-s11.tsv
table-s12.tsv
table-s13.tsv
table-s14.tsv
table-s15.tsv
table-s16.tsv
table-s17.tsv
table-s18.tsv
table-s19.tsv
table-s20.tsv
table-s21.tsv
table-s22.tsv
table-s23.tsv
table-s24.tsv
table-s25.tsv
table-s26.tsv
table-s27.tsv
table-s28.tsv
table-s29.tsv
table-s30.tsv
table-s31.tsv
table-s32.tsv
table-s33.tsv
table-s34.tsv
table-s35.tsv
table-s36.tsv
table-s37.tsv
table-s38.tsv
table-s39.tsv
table-s40.tsv
table-s41.tsv
table-s42.tsv
table-s43.tsv
table-s44.tsv
table-s45.tsv
table-s46.tsv
table-s47.tsv
table-s48.tsv
table-s49.tsv
table-s50.tsv
table-s51.tsv
table-s52.tsv
table-s53.tsv
table-s54.tsv
table-s55.tsv
table-s56.tsv
table-s57.tsv
table-s58.tsv
table-s59.tsv
table-s60.tsv
table-s61.tsv
table-s62.tsv
table-s63.tsv
table-s64.tsv
table-s65.tsv
table-s66.tsv
table-s67.tsv
table-s68.tsv
table-s69.tsv
table-s70.tsv
table-s71.tsv
table-s72.tsv
table-s73.tsv
LICENSE.md
README.md

What the benchmark measures

Built to be hard to fake

Real

Every file the model sees is the Broad Institute's actual public record — the filed IND dossier, FDA's hold letter, the ELISA plates, the sequence tables. Nothing synthetic.

Hard

The assignments are the work the hold response actually required: deciding which site activity the letter froze, defending the assay data the agency questioned, reconciling records that contradict each other. Frontier models get these wrong.

Verifiable

Each submission is scored against an expert-written checklist — the judgments a specialist would verify, not whether the writing sounds good. Every failed check links to a verbatim moment in the run, and the full transcripts are public.

Where it broke

The model authorized the work FDA froze

FDA · hold letter · verbatim

“Studies that are on clinical hold under this IND may not be initiated until you have submitted the required information and we notify you that you may initiate these clinical studies.”

IND 167326, April 2025

The model · submitted plan · verbatim

“IRB submission, IRB maintenance, and site contracting for the held protocol — PERMITTED”

Task 01, best-scoring run

FailedSite contracting under a frozen protocol, marked PERMITTED — in a document a compliance team would ship. See the run →

The assignments

Choose an assignment

Each page shows the assignment, the files on hand, the model's work, and every model's score — including the failures.

How submissions are graded

Examples of scoring checks

Does the model recompute the filing's numbers or just repeat them? Does it notice a dated extract that contradicts the governing record? Does it permit work the FDA froze? Each assignment page shows which checks each submission passed and which it failed.

From the recordThe FDA’s hold letter misspells the drug it regulates — “2349-exNA” — and the error is preserved verbatim in this record.

Talk to usRequest the full Biopharma Bench V0.1 — company environments and run receipts: team@raycaster.ai