The Croatian lab: how one research group built the BPC-157 literature

On 20 October 2026 we searched PubMed for every record that mentions BPC 157, or its code names PL 14736 and bepecin, in the title or abstract, or that is indexed under the substance “BPC 157”.² We downloaded all records and checked the author lists and publication types with the National Library of Medicine’s own tools. Each record outside the Zagreb group we then classified by hand.

The search returned 235 records, published from 1992 to 2026. Sikirić is an author on 175 of them (74%). Before 2016 the share was higher: 85 of 96 records (89%).

Reviews and commentaries inflate any count, so we then set aside everything PubMed tags as a review, editorial, comment, letter or erratum. That left 177 records, mostly primary studies:

Counted that way, the Zagreb group produced about 140 of roughly 160 primary experimental papers, about 85%.

Independent work is also clustered. Five of the 17 independent experimental papers come from one laboratory in Taiwan, and five from one laboratory in Xi’an, China.^3,4 Collaborations add a further wrinkle. Papers by a gastroenterology group in South Korea, for example, include Sikirić as a co-author. We counted them as the group’s work, because shared authorship means shared methods.

PubMed is not the whole literature. Conference abstracts, Croatian-language journals and studies indexed only in other databases are missing, and so is the developer’s clinical programme, which survives only as abstracts.⁵ Author matching can miss name variants. Our split between “primary” and “review” relies partly on PubMed’s tags, and some of the group’s 140 records are review-like papers that PubMed did not tag. None of this changes the overall picture. It is why we give round numbers and describe the method.

Any single experiment can be wrong for reasons that cannot be seen from inside it. Chance can produce a striking result in a small sample. A batch of reagent can be contaminated, or an animal strain can respond oddly. An outcome scored by someone who knows which animals were treated can drift towards the expected answer. A choice between several reasonable ways to analyse the data can tip a borderline result over the line.

A second laboratory with different people, animals, suppliers and habits is unlikely to make the same mistake in the same direction. If the effect appears again there, the most likely explanation is that it is real. That is why independent replication is the main error-correcting mechanism of experimental science.

Within one group, studies share far more than a topic. They use the same injury models and surgical techniques, the same scoring systems and the same source of peptide. The same people are trained the same way, and they share statistical habits and assumptions about how the peptide should behave. If any of these introduces a bias, it introduces it into every study in the same direction.

A hundred such studies then do not test the effect a hundred times. They measure the bias a hundred times, very precisely. John Ioannidis put it bluntly in 2005: in many fields, “claimed research findings may often be simply accurate measures of the prevailing bias”.⁶

Incentives are shared too. A group whose reputation, funding and patents are tied to a compound has natural reasons to look hard for positive results and to publish them; Sikirić is named as an inventor on a 1998 international patent application for salts of the peptide.⁷ This is not a special failing. It is true of every drug developer, and it is exactly why regulators rely on registered, independently monitored trials before they accept a claim.

Positive results are easier to publish than null ones, everywhere. A review of 525 animal stroke publications found that only 2% reported no significant effect. It estimated that about one in seven experiments went unpublished, and that publication bias alone accounted for roughly a third of the apparent benefit.⁸ When scientists at a company tried to confirm 53 “landmark” preclinical cancer studies, they succeeded with 6.⁹ A systematic replication of 50 experiments from high-profile cancer-biology papers found that effects first reported as positive were, on median, 85% smaller than the originals.¹⁰ Many animal studies also do not report randomisation or blinded assessment, the basic protections against these biases.¹¹

None of this is specific to Zagreb. But it shows what to expect when a large animal literature has not been independently repeated: some real effects, some overstated ones, and no way to tell them apart from inside the literature.

The group reports benefits from BPC-157 in an unusually wide range of conditions, from stomach ulcers and fistulas to tendon tears, nerve injury and cardiovascular damage.⁵ That breadth can be read in two ways. It might reflect a basic protective mechanism. Or it might reflect a consistent research method that tends to produce positive results in any model it touches. The existing literature cannot tell these apart. Independent replication can.

Recent years add another trap. Between 2024 and 2026 most new BPC-157 records were not from Zagreb. But more than half of those were reviews and commentaries, largely summarising the Zagreb data. A review that cites a study is an echo of it, not a second test.

• Cell signalling. A laboratory in Taiwan found plausible pro-healing signals, mostly in cultured cells: more VEGF receptor 2, more nitric oxide, more activity in cell-migration proteins Cells(Evidence level: Shown in cells or tissue in the laboratory).^3,12,13 Two of the same effects have since been reported by other groups, both in 2026. The laboratory in Xi’an saw more vessel-like tubes form in cultured human blood-vessel lining cells Cells(Evidence level: Shown in cells or tissue in the laboratory).¹⁴ A group in Ankara, Turkey, found that rings of human artery, left over from bypass surgery in 12 men, relaxed in a dish, mostly through nitric oxide Cells(Evidence level: Shown in cells or tissue in the laboratory), as the Taiwanese rat arteries had.^15,13 We found no independent repeat of the specific signalling pathways the Taiwanese laboratory described.
• Toxicology and drug handling. A laboratory in Xi’an produced the only formal toxicology and pharmacokinetic studies.^16,4 These are useful, but they test safety and handling, not efficacy.
• Gut protection. An industry laboratory at Parke-Davis in the US reported in 1995 that the peptide reduced chemically induced colon damage in rats, with a dose–response Animals(Evidence level: Shown in animal studies). The abstract does not give the number of rats.¹⁷
• Tendon healing. The first independent test in living animals of the group’s flagship tendon claim came from Turkey in 2026. It used a related but different model: 32 rats in four groups (8 given BPC-157) whose cut Achilles tendons were surgically repaired, while the Zagreb studies left them unrepaired. Strength and overall tissue scores moved in the expected direction but were not statistically significant. One collagen measure was significantly higher Animals(Evidence level: Shown in animal studies), but each analysis used only 4 tendons per group.¹⁸
• Humans. The developer’s randomised trial in ulcerative colitis, with 53 patients, was reported only as a meeting abstract and never published in full. Three small uncontrolled reports come from one US clinician.^5,19 A fourth, on three dental patients given BPC-157 in a mixture with other drugs, came from a group of US dentists in 2026.²⁰

The group also engages with its critics in print. When an independent review raised concerns about angiogenesis and tumours, the group published a detailed reply.^24,7 Single-group dominance is not unusual for compounds that never reached late-stage development, and many medicines began in one laboratory. Nor is the missing human evidence simply the group’s fault: in public records, the industrial development of PL 14736 never progressed beyond a phase 2 trial in the 2000s.⁵

The fix is well known, and none of it requires new science:

• Independent, pre-registered animal studies of the key claims (tendon, ligament and gut healing). These should be randomised and blinded, use a sample size calculated in advance, and test several doses. Treatment should start after the injury is established, and the peptide should be bought and analysed independently. Ideally several laboratories would run the same study design in parallel.
• Full publication of the old phase 2 trial data, so the only controlled test of whether it works in people can be properly assessed.⁵
• Registered human trials reported in full whatever they find. A randomised, placebo-controlled trial in hamstring injury is recruiting, with results expected after February 2027.²⁵
• An identified and independently confirmed molecular target, which would make effects predictable and doses meaningful.
• Independent long-term toxicology, including the carcinogenicity studies that no one has yet published.⁵