Yak collagen peptides picked by a language model slow a skin enzyme and lengthen worm lives

A team funded by Jilin University has used a machine learning screen to pick yak collagen peptides from leftover bone and then tested three of them. In a test tube the short fragments slowed MMP 1, an enzyme that cuts collagen, and in tiny roundworms they extended lifespan under the conditions used. The paper appeared in npj Science of Food on 22 October 2026. The work stops at enzymes and worms, so it says nothing yet about skin or about people.

How a language model ranked bone fragments

The starting material was collagen from yak bone, a by product of yak meat processing that the authors want to turn into something of value. The authors built a screening model they call AgeAttnMLP on top of ESM 2, a protein language model trained on huge numbers of natural sequences to learn which patterns of amino acids tend to occur together. Their model was built to rank candidate peptides by their likely anti ageing activity, using that learned knowledge of sequences. On five fold cross validation it scored an area under the curve of 0.891, and 0.879 on a stricter test that kept similar sequences out of both training and test sets. Those scores describe how well the model sorted known examples, not how often its new picks would work in a laboratory.

To generate candidates the team digested the collagen sequence in silico, simulating cuts by Alcalase, a protease widely used to break food proteins into peptides. The model ranked the resulting fragments, and three were chosen as leads: TGPPGPSGI, GEPGAPGE and GPSGEPGTA. All three carry the glycine and proline rich repeats typical of collagen. Using a protease that industry already relies on means that a real digest could in principle contain similar fragments. The paper frames the work as a way to add value to bone waste. That framing is about food processing and should not be read as evidence for any collagen product on the market, none of which were tested here, and readers following other computer screened peptides will know that predicted activity often fails at the bench.

What the yak collagen peptides did in a dish and in worms

MMP 1 belongs to the matrix metalloproteinases, enzymes that break down the protein scaffolding between cells, and it is one of the main collagen cutting enzymes in skin. That is why it features in research on skin ageing. In a cell free enzyme assay, all three yak collagen peptides reduced its activity, and GEPGAPGE was the strongest. Its half maximal inhibitory concentration was 126 micrograms per millilitre, the amount needed in the mixture to cut enzyme activity by half. The enzyme test shows only that the fragment can slow purified MMP 1 in a controlled mixture. It does not show that the peptide survives digestion, reaches skin or reaches the enzyme inside living tissue, which are the steps any real effect would depend on.

Every one of the three peptides extended the lifespan of the worms under the tested conditions, and GPSGEPGTA gave the largest effect.

The second test used Caenorhabditis elegans, a millimetre long roundworm that lives for a few weeks and is a standard first model in ageing research. Every one of the three peptides extended the lifespan of the worms under the tested conditions, and GPSGEPGTA gave the largest effect. The abstract does not give the size of the gains in a form we could check, so we do not report a percentage. The ranking matters in itself. The best enzyme inhibitor was not the peptide that most extended worm life, which suggests the two results may not share one cause. The abstract does not name the worm pathways involved, so the route from peptide to longer life remains unknown and may have little to do with collagen at all.

What the simulations suggest

To explain how a short collagen fragment might slow MMP 1, the authors ran molecular dynamics simulations, which follow the predicted motion of the enzyme and peptide atom by atom, and estimated binding energies with a method called MM PBSA. The simulations suggested that peptide binding loosens a small beta sheet, around residues 76 to 78, close to the catalytic region where collagen is cut. That would be a plausible way to reduce activity without blocking the active site directly. It remains a model. No crystal structure, mutation study or binding measurement was reported to confirm that the peptides sit where the simulation placed them, and simulations of short flexible peptides are especially sensitive to starting assumptions.

Simulations of this kind are a standard way to turn a test tube result into a hypothesis that later experiments can check. If the loosening of that sheet is real, changing single amino acids in a peptide should shift both the predicted binding and the measured inhibition in step, and a solved structure of the enzyme with a peptide bound would show the contact directly. Neither experiment appears in the abstract or in the public summaries we could read. The table above sets out each step of the pipeline in order, and only two of its five rows involve a physical measurement. The rest are predictions, which is normal for an early screening paper but worth keeping in mind when reading headlines about anti ageing collagen.

What the study cannot show

The paper links two findings, enzyme inhibition and longer worm life, but does not show that one causes the other. It used three peptides, one enzyme assay and one invertebrate, with no cell culture of human skin, no mouse work and no human participants. It did not test whether the peptides survive stomach acid and gut enzymes, a barrier that stops most food derived peptides before they reach the blood, or whether they reach skin when applied to it. The only funding listed is a graduate innovation fund at Jilin University, which suggests an early academic project rather than a development programme with a product in view.

The result is still useful as a method. Pairing a protein language model with simulated digestion is a cheap way to sort thousands of possible fragments before anyone synthesises a single one, and a similar logic has been used in machine learning screens for antibiotic peptides. The yak collagen peptides now need independent repeats of the enzyme and worm data, tests in human skin cells and, much later, controlled studies in people before any claim about ageing could be made. Until then, the study is not a reason to change diet or skincare, and readers should be wary of products that cite it as evidence.

Conclusion

The yak collagen peptides study shows a sensible screening pipeline producing three fragments that slowed MMP 1 in a test tube and extended lifespan in worms. Those are early signals from simple systems. The mechanism rests on simulation, the two biological results may not be linked, and no skin, mammal or human data exist. It is a lead for laboratories, not evidence that collagen products slow ageing in people.

Source Liu Y, Zhang Y, Li Q, Zhuang J, Li W, Han W. Valorization of yak bone collagen byproducts: computational screening and experimental validation of anti-aging peptides targeting MMP-1. npj Science of Food, published 22 October 2026. DOI 10.1038/s41538-026-01223-6

Amino Weekly reports research. It is not medical advice and never gives doses.