On September 7 a group at nference posted a preprint counting BPC-157 users in the clinical notes of a 15.2 million patient network. On September 15 Reuters covered it. The paper is more careful than the coverage, and the coverage is where most readers will meet it.

This is a structured read of the paper itself. In July I covered the regulatory side of the same question, when an FDA advisory committee voted 8 to 6 to allow compounding of six unapproved peptides over the objection of the agency's own scientists (The FDA Cleared Six Peptides for Compounding. The Evidence Didn't.). That piece was about supply. This one is about who is already taking the stuff, and about how much weight the count can carry.

MAIA is the Metabolism Agentic Intelligence Atlas, also in Cambridge, carried by three of the authors alongside their nference affiliation. It signals where the group's other work sits, which is metabolic disease. Two of this paper's twenty-five references are the same group's preprints on retatrutide and on incretin trial participants, and a third is their published work on GLP-1 initiation and compounding ascertained by the same method.

There is no academic co-author, no clinician co-author from an unaffiliated institution, and no named site investigator. The physician who performed the accuracy validation is described only as "a clinician reviewer" at "one participating site."

The conflict statement, in full:

Funding is declared as none. That is consistent with a study run on data the company already holds using a platform it already built.

Two things follow, and they point in different directions.

First, the narrow one. The paper's co-therapy figure counts tirzepatide in 6.1% of users and semaglutide in 4.3%. Those are products of companies in the class the disclosure describes. The disclosure says no collaborator shaped the work. It does not say which collaborators exist, so a reader cannot check the overlap themselves.

Second, the structural one, which matters more. nference sells large language model curation of de-identified health records to pharmaceutical clients. A paper demonstrating that this method can find and characterize a patient population that is invisible to claims data, prescription fills, and structured records is a methodological contribution and a capability demonstration at the same time. The abstract's closing sentence calls for "facilitating structured EHR capture of broader gray-market peptide use," which is a recommendation that the world buy more of what the authors make.

That is a reason to read the methods closely. It is not a reason to discount the count. Commercial motivation to demonstrate a method is compatible with the method working, and this one largely does.

This is the part of the paper that is genuinely well built, so it deserves the detail.

A language model was given each patient's entire note history rather than a keyword window around the drug name. The stated reason is sound and is the kind of thing only someone who has read a lot of charts would think to say. Procurement, indication and response details turn up in social history, medication reconciliation and plan sections, often nowhere near the mention that triggered the search.

The model then sorted each patient into confirmed exposure, discussion without use, denial of use, or unclear documentation. Separate passes classified supply channel, indication, injury site, post-surgical region, route of administration, and concomitant interventions. A further pass classified the direction of any documented symptomatic response, who reported it, the symptom type, and the affected quality-of-life domain.

One design choice is worth crediting specifically. Co-use of another named peptide was counted only when that same patient was independently confirmed as a user of that peptide by the identical note-based method. An unverified mention was not enough. That is a stricter bar than most real-world evidence papers set for themselves.

A clinician at one site reviewed random samples of the model's extractions, stratified by the value the model had assigned, and marked each correct, incorrect or indeterminate. Agreement was computed within each predicted class and combined by class prevalence, with Wilson score confidence intervals. Extractions validated below 80% agreement are labelled exploratory and, the authors say, are not used to support mechanistic claims.

Proportions carry Wilson score 95% intervals. Mean age across initiation years was tested by one-way ANOVA, categorical distributions by chi-square, ordered strata by Cochran-Armitage trend test. Multiplicity was controlled by Benjamini-Hochberg within each prespecified comparison family, with q values reported alongside P. Endpoints, comparison families and small-cell suppression were prespecified before the outcome analyses ran. Any cell below 11 patients is reported as "<11." Reporting follows STROBE and RECORD.

De-identification was by HIPAA expert determination under 45 CFR §164.514(b)(1). Identifiers were replaced by one-way cryptographic hashes, ZIP codes truncated to two digits, ages 89 and over collapsed, BMI over 40 collapsed, and free text scrubbed by an ensemble system the authors report at greater than 99% recall.

One consequence of that method is worth noting against the trend analysis. All dates were shifted backward by patient-specific random offsets of 1 to 31 days. That preserves intervals within a patient and blurs the boundary between adjacent quarters across patients. It is a small effect against a 33-fold change and a real one against any claim tied to a specific month.

Neither the data nor the analysis code is available. The data cannot be shared under the terms of the expert determination. For the code, the paper says to contact the corresponding author.

Across initiation years the male share rose from 57% in 2020 to 70% in 2026 (Cochran-Armitage P < 0.001) and the race difference clears chi-square at P < 0.001, a 1.23-fold enrichment for White patients. The age decline from 55 to 49 is reported in the abstract and was repeated by Reuters. The figure carrying it reports its own one-way ANOVA across initiation years at P = 0.178.

Documented adverse events by organ system were neuropsychiatric 1.4%, dermatologic or injection site 1.4%, and gastrointestinal 1.3%. The authors state plainly that these are documented signals rather than prospectively ascertained incidence, and are likely under-reported. The 13 all-cause deaths are derived from structured records, are not attributed to the compound by the paper, and should not be read as a safety signal in a cohort with a mean age of 50 and no comparator.

Reuters supplied one figure the paper does not contain. The FDA's adverse event reporting system holds 10 reports involving BPC-157 since 2021, all of them classified serious, with 8 of the 10 filed in 2026. Ten spontaneous reports cannot support an incidence estimate. The concentration in a single year is a statement about reporting behavior and attention.

Limitation 6 is the one a reader would be least likely to catch and the authors volunteered it. The accuracy table above is the paper's main claim to credibility, and the authors are saying the reviewer could see what the model had already decided. That does not void the validation. It does mean the intervals are optimistic on top of being wide.

Four things, stated plainly.

A denominator. Before this, nobody had published a count of BPC-157 users derived from medical records. Now there is one, with a known method and a stated error rate.

Evidence that the supply migration preceded the regulatory decision. The July committee's harm-reduction argument was that reclassification would move users from the gray market into regulated pharmacies. In data ending before that vote, compounding pharmacies were already the largest documented channel, and 7.1% of the cohort obtained the compound on a prescription.

A measurement of how late clinicians find out. Among the 80 patients with an ascertainable start date, use preceded the first documenting note by a mean of 60 days. 58% of patients have exactly one note mentioning it, ever.

A number the paper did not intend to provide. The flat duration-response curve is the closest thing here to evidence about the compound itself, and it points away from an effect.

What it does not add is any evidence that BPC-157 works, or that it is safe, or that its use rose because of anything in particular. The paper says all three of those things about itself. The coverage said the first one anyway.