AI designed nucleosome binding peptides hold up in a test tube

A team in Moscow and Philadelphia has reported nucleosome binding peptides that software drew and a laboratory assay then measured. The target was the acidic patch, a small negatively charged dent on the surface of the nucleosome where many chromatin proteins dock. Software produced about one hundred candidate sequences, ten were synthesised, and two bound roughly as tightly as the viral peptide the field uses as its benchmark. Five others showed no measurable binding at all. The brief report appeared in Frontiers in Bioengineering and Biotechnology on 22 October 2026.

How one hundred sequences became ten test tubes

The authors ran the peptide protocol of BindCraft, an open design program built on AlphaFold2, asking it for short sequences of roughly fifteen to thirty residues that would sit on the acidic patch. Because the underlying model cannot handle DNA, the design target was an isolated histone dimer taken from a published nucleosome structure rather than the full particle. The run produced about one hundred designs. The authors kept the forty with the best predicted binding energy and interface confidence scores, rescored those with a separate structure program, then picked ten by eye for chemical synthesis. Every design came back alpha helical, which narrows the shapes the experiment could test.

The patch they aimed at is small, well mapped and crowded with competition. Chromatin remodellers, the silencing machinery and several viral proteins all grip it, usually through a protruding arginine that slots into the acidic cavity. The best studied example is a short stretch from a herpesvirus protein known as LANA, which binds in the submicromolar range and has served for years as a measuring stick. Designing new nucleosome binding peptides for the same spot is therefore a fair test of a design program. Success means matching a sequence that evolution has already optimised, and failure is easy to detect because the benchmark peptide is available and well characterised.

What the FRET quenching assay measured

Rather than tagging each designed peptide, the group built a competition test. A fluorescent donor dye sat on the nucleosomal DNA, a second dye sat on the benchmark LANA peptide, and binding brought the two close enough to quench the donor signal. Adding an unlabelled design pushed LANA off the particle and restored the signal, and the concentration needed to do that was converted into a dissociation constant. Nucleosomes were present at roughly ten nanomolar. This FRET quenching assay reports displacement rather than direct occupancy, so it measures competition for one site instead of total affinity for the whole particle.

Half of the shortlist failed, and the ranking from the scoring program did not predict which half.

Two designs reached the benchmark. One matched LANA almost exactly, a second came within a factor of three, three more landed near or above one micromolar, and the remaining five did not shift the benchmark peptide at the concentrations tested. That is a hit rate of two in ten for strong binders and five in ten for anything measurable, from a shortlist already filtered by two scoring programs. The authors describe the assay as refined for this purpose, since the earlier approach of labelling every candidate would have cost more and risked changing the binding it was meant to measure. The numbers below are the values reported in the paper, with lower meaning tighter binding.

Where de novo peptide design still guesses wrong

Half of the shortlist failed, and the ranking from the scoring program did not predict which half. The authors state plainly that computational scores could filter candidates but did not reproduce the experimental order of affinities. That is the most useful result here for anyone running a similar pipeline. A filter that enriches hits is worth having even when it cannot rank them, but it means the wet laboratory step stays in the loop and the number of designs you can afford to test still sets the pace. Ten peptides is a small sample from which to judge any method, and the authors present the work as a workflow rather than a benchmark.

The reason for part of the error is written into the setup. Because DNA was absent from the design model, de novo peptide design here could not account for contacts between a peptide and the DNA wrapped around the particle, nor for the neighbouring histones. Peptides that look good against a bare dimer may clash or lose contacts on an assembled nucleosome. Other recent design work on cyclic binders and on venom neutralising proteins has had the advantage of targets that fold on their own, which chromatin does not. The gap between the model and the real particle is the honest explanation for most of the misses.

What the experiment cannot show

Everything here happened in solution with purified components. No cells were treated, no animal was dosed, and nothing in the paper shows that these nucleosome binding peptides reach a nucleus, survive inside one, or change how a gene is read. Specificity is inferred rather than demonstrated, because the assay only asks whether a design can displace a known patch binder. The authors also report that their measured value for the benchmark peptide differs from published figures, which they attribute to the charge of the fluorescent label. That is a reminder that affinities measured in different assays are not directly comparable across papers.

The practical use is more modest than a drug. Tight patch binders are tools for blocking one class of chromatin interaction in a dish, and the group suggests, without testing it, that such peptides might stabilise nucleosomes. Turning any of this into a therapeutic would mean solving delivery into the nucleus, selectivity against the many natural proteins that use the same site, and stability in a cell, none of which this study addressed. The authors themselves frame the paper as a proof of concept and a workflow. Readers should treat the result as evidence that a design program can produce usable chemical probes, not as evidence of a medicine in waiting.

Conclusion

The study is small, clearly reported and useful mostly as a method check. Design software produced nucleosome binding peptides that work in a tube, two of them as tightly as the natural benchmark, while half the batch did nothing and no score predicted which. That mixture is the normal shape of early design work. The next questions are whether the same peptides hold up against full nucleosomes, inside cells and against the proteins that already occupy the site.

Source Oleinikov PD and colleagues. Design and validation of nucleosome binding peptides using BindCraft. Frontiers in Bioengineering and Biotechnology, published 22 October 2026. DOI 10.3389/fbioe.2026.1946220

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