- What happened: Anthropic's biology research team experimentally confirmed that a bacteriophage-derived protein called "L05" assembles 96 copies of itself inside E. coli to form a giant hollow "vault particle."
- Why it matters: Vault particles have been studied mainly in eukaryotes. This work shows that proteins associated with bacteria and viruses can also form them. The team also used the AI model Claude across the project, from searching for proteins to analyzing their three-dimensional structure.
- What to watch: The findings are an unreviewed preprint, and it has not been confirmed that the particle forms naturally in its native host. Identifying the biological function of vault particles, which has remained unresolved for about 40 years since their discovery in 1986, is still an open challenge.
On November 4, 2026, Anthropic's biology research team announced that a protein derived from a bacteriophage (a virus that infects bacteria) self-assembles inside E. coli cells into a giant hollow structure known as a "vault particle." Cryo-electron microscopy showed that 96 protein molecules fit together in a regular arrangement to form a barrel-shaped particle with a molecular weight of about 9.5 megadaltons.
Vault particles are huge protein complexes discovered in 1986 and have been studied mainly in eukaryotes, including mammals. Even now, about 40 years later, their biological role is still not well understood. The significance of the new study is that it shows prokaryotic and phage-derived vaults, whose existence had mostly been predicted from amino acid sequence similarity, can actually form particles.
The team used Anthropic's AI model Claude for sequence searches, experimental design, and analysis of electron microscopy images. However, the results rest on a preprint posted to bioRxiv that has not been peer reviewed, and particle formation was confirmed only in E. coli engineered to express the protein. It is not known whether the same structure forms in the native host or during phage infection. The team also reported that suggestions from Claude delayed one experiment by two weeks and led to errors in structural analysis.
A giant protein complex whose function has been unknown for 40 years
Vault particles were discovered in 1986. Nancy Kedersha and Leonard Rome examined samples extracted from rat liver under an electron microscope and found a previously unknown, giant hollow structure. Early observations showed an elongated shape of about 35 × 65 nanometers with a distinctive seam in the middle, reminiscent of a Gothic cross-vaulted ceiling, hence the name "vault." The discovery was reported in the 1986 Journal of Cell Biology.
In mammalian vault particles, 78 copies of a protein called Major Vault Protein (MVP) form the outer shell. Two half-shells of 39 molecules each face one another at the midpoint, and the interior contains small amounts of vault-specific RNA (vRNA) and associated proteins. The particle's molecular weight is about three times that of a ribosome, making it one of the largest protein complexes in mammalian cells.
- 構成コピー数
- 半殻あたりのコピー数
データを表で見る
| 構成コピー数 (分子) | 半殻あたりのコピー数 (分子) | |
|---|---|---|
| 哺乳類ヴォールト | 78 | 39 |
| ファージ由来L05ヴォールト | 96 | 48 |
Vault particles are found in a wide range of eukaryotes, from slime molds to humans, yet no related genes have been found in budding yeast, the fruit fly Drosophila melanogaster, or the nematode C. elegans. A 2013 evolutionary analysis suggested that the common ancestor of eukaryotes acquired vaults and that related genes were later lost in several lineages. Knockout mice lacking vault genes, however, develop and survive normally. Proposed roles include multidrug resistance to anticancer drugs, transport of substances through nuclear pores, and innate immunity, but no decisive evidence explains a primary function.
Advances in genome analysis have recently suggested that MVP-like genes may also exist in prokaryotes and bacteriophages. A 2019 preprint and 2023 research report pointed to their presence based on amino acid sequence similarity and structure prediction. But having a similar sequence is a different matter from actually being able to form a giant vault particle. By producing candidate proteins in the lab and directly confirming particle formation by electron microscopy, this study goes a step beyond earlier sequence analyses.
Narrowing candidates from about 15 billion protein sequences
Anthropic's team searched about 15 billion predicted protein sequences in databases for proteins distantly related to MVP. The work used "Claude Science," the company's setup for applying its AI models to scientific research.
The sequence analysis found MVP-homologous sequences across 29 bacterial phyla and 4 archaeal phyla. Their detection frequency among genome-sequenced species, however, was under 1%. This figure is derived from sequences registered in databases and does not represent the actual prevalence of vault particles among microbes in nature.
Considering candidate diversity and experimental feasibility, the team selected 44 proteins. They first evaluated expression levels and full-length protein synthesis efficiency for each candidate using a cell-free expression system that makes proteins without cells. At this stage, working under human researchers' instructions, Claude assisted with back-translating amino acid sequences into DNA sequences, adjusting codons for the expression host, designing regulatory sequences such as promoters, and preparing order data for DNA synthesis services.
Next, 24 candidates that performed well in the cell-free system were expressed in E. coli, and 15 could be recovered as soluble proteins rather than insoluble aggregates. When these were purified and observed by transmission electron microscopy, 7 showed the hollow, barrel-shaped particles characteristic of vaults. Further detailed cryo-electron microscopy analysis ultimately identified four vault structures: two from phages and two from bacteria.
| Screening stage | Number tested | Result | Main evaluation |
|---|---|---|---|
| Sequence search | ~15 billion | Homologous sequences detected in 29 bacterial and 4 archaeal phyla | Sequence homology to MVP |
| Cell-free expression test | 44 | Candidates selected based on expression results | Protein synthesis efficiency and expression level |
| Expression in E. coli | 24 | 15 recovered as soluble proteins | Solubility and expression state |
| Electron microscopy screening | 15 | Vault-like particles confirmed in 7 | Formation of hollow, barrel-shaped structures |
| Cryo-electron microscopy analysis | 7 | Vault structures identified in 4 | Three-dimensional structure and symmetry |
From these four, the team chose the phage-derived protein "L05," which efficiently forms uniform, fully closed particles, for detailed structural analysis. L05 was selected for its strong particle-forming ability, not as an average representative of all candidates.
A giant hollow structure built from 96 proteins
The L05 vault structure revealed by cryo-electron microscopy shares its basic form with the mammalian vault but differs in several important ways.
The biggest difference is the number of proteins forming the outer shell. Whereas the mammalian vault consists of 78 molecules (39 per half-shell), the L05 vault consists of 96 (48 per half-shell). Its estimated total molecular weight is about 9.5 megadaltons.
The outer dimensions also differ. The mammalian vault has a long axis of about 67 nanometers and a maximum diameter of about 40 nanometers, while the L05 vault has a long axis of about 54 nanometers and a diameter of about 45 nanometers. It is roughly 20% shorter but wider, giving a shape closer to a sphere. As a result, its interior space is estimated to be about 25% larger than that of the mammalian vault.
- 全長（長軸）
- 最大直径（短軸）
データを表で見る
| 全長（長軸） (nm) | 最大直径（短軸） (nm) | |
|---|---|---|
| 哺乳類ヴォールト | 67 | 40 |
| ファージ由来L05ヴォールト | 54 | 45 |
The shell proteins themselves also differ. Mammalian MVP has nine repeat domains, whereas L05 has only seven repeat segments. Although L05 has fewer repeat regions in the side wall, more molecules are arranged around the circumference, producing a thicker barrel than the mammalian vault.
Particularly interesting is the structure of the caps that close both ends of the particle. In L05, 48 protein chains with identical amino acid sequences each adopt a different conformation and gather from the center toward the end, forming a dome-shaped structure with three tiers. Of three neighboring chains, the first ends at the first tier, the second ends at the second tier, and only the third reaches the center. Arranged as 16 sets of three chains, the 48 chains close off the end.
That a single protein uses different conformations to assemble such a complex structure is notable from a structural biology perspective. Whether this unusual closing mechanism is linked to any biological function is unknown. The published technical summary also does not state the final resolution of the electron microscopy map, the total number of particle images used, statistical validation metrics for the structural model, or measurement errors for the dimensions.
| Structural feature | Phage-derived L05 vault | Mammalian vault |
|---|---|---|
| Number of constituent proteins | 96 (48 per half-shell) | 78 (39 per half-shell) |
| Estimated molecular weight | ~9.5 megadaltons | ~13 megadaltons (including associated molecules) |
| Particle dimensions | ~54 × 45 nm | ~67 × 40 nm |
| Interior space | ~25% larger than the mammalian vault | Baseline for comparison |
| Repeat segments in side wall | 7 | 9 |
| Cap structure | Stepwise termination in sets of three chains | Smoothly curved structure |
| Other components | Density suggesting an unidentified molecule bound to each subunit | vRNA and associated proteins |
Engineering the protein to enlarge the interior by 20%
After determining the three-dimensional structure of the L05 vault, the team tested whether modifying the protein could control the particle's shape and size.
First, focusing on the regions where neighboring proteins contact each other, they designed 15 amino acid substitution variants with Claude's assistance. Replacing just one amino acid in a region involved in cap formation prevented the normal barrel structure from closing. Under the electron microscope, many abnormal, flat, open rosette-shaped structures were seen instead of sealed particles. This showed that specific intermolecular interactions near the cap are important for completely closing the particle.
Next, they duplicated two of L05's seven side-wall repeat segments to design an artificial protein with nine repeat segments, the same number as mammalian MVP. When expressed in E. coli, the modified protein retained its ability to self-assemble and formed vault particles longer than the original L05. The interior volume increased by about 20% compared with wild-type L05.
This result shows that the repeat sequences forming the vault's side wall are flexible enough to be modified. However, the fact that artificially adding repeats enlarged the particle does not prove that vaults evolved by the same mechanism in nature. Clarifying their evolution will require further study of sequences from related species and intermediate structures.
Is the unknown bound molecule coenzyme A?
The structural analysis also uncovered an important clue to L05 vault function. Each of the 96 proteins forming the outer shell has a pocket-like structure capable of binding a small molecule, and electron density suggesting the presence of a non-protein molecule was detected there.
Claude pointed out that the shape of this density may resemble coenzyme A (CoA), a cofactor involved in processes such as acyl group transfer in cells. A similar shape alone, however, cannot establish that the bound molecule is CoA. Chemical identification by mass spectrometry and verification of binding properties using methods such as isothermal titration calorimetry (ITC) would be needed.
At present, it is unknown what occupies this binding pocket. The vault particle might be involved in transporting or storing some metabolite, but these results alone cannot support such a conclusion.
Research using Claude: 90-minute purification and 34 hours to a structure
Beyond the biological discovery, the study is notable as an example of AI being built into real scientific research. Anthropic says Claude was involved from protein sequence searches through experimental design, consideration of sample purification methods, and analysis of cryo-electron microscopy data.
The efficiency gains were especially clear in vault purification. The conventional method of extracting vaults from mammalian tissue requires repeated sucrose density gradient centrifugation and can take about three days. With the procedure the team built, everything from lysing E. coli to loading the sample onto electron microscopy grids could be completed in about 90 minutes.
This method uses size-exclusion chromatography, which separates molecules by size. However, earlier work (PubMed: 30036602) had already purified vaults by size-exclusion chromatography, so the technique itself is not new. The three-days-versus-90-minutes comparison is also not a direct benchmark using the same sample under the same conditions. The achievement here lies in greatly simplifying the purification by optimizing the resin and elution conditions for vaults produced in E. coli.
Interestingly, Claude initially argued against this purification method. When the team proposed size-exclusion chromatography, Claude warned that large protein aggregates might contaminate the same fraction. After the human researchers decided to adopt the method, however, Claude advised on selecting a suitable column and tracking the particles. It ultimately proposed loading the centrifuged supernatant from the E. coli lysate directly onto the column without pre-purification.
This approach risked contaminating or degrading the column resin, but the experiment achieved rapid purification. In effect, the human researchers' judgment and the AI's subsequent optimization of the procedure worked better than Claude's initial concern suggested.
Claude was also used for structural analysis. A multi-agent environment incorporating Claude handled setting up RELION, the open-source cryo-electron microscopy image analysis software, and running the data processing pipeline. With human researchers overseeing the analysis strategy, a first three-dimensional structural model was obtained 34 hours after data input, and the final electron density map was completed in about a week.
That said, this speed cannot be attributed to the AI's ability alone. The analysis used a computing environment with hundreds of GPUs, far beyond the resources of a typical university lab. The 34-hour result should be viewed as the product of AI-assisted work combined with a large-scale parallel computing environment.
AI errors caused a two-week delay and image analysis problems
Using Claude did not always make the research more efficient. Anthropic's report also records cases in which erroneous AI suggestions caused experimental delays and rework in analysis.
In an early experiment, there was a problem with the culture medium preparation procedure Claude produced. It included an additive that inhibits transformation, the introduction of plasmid DNA into E. coli, so the experiment did not proceed normally. The team reportedly needed about two weeks to identify the cause.
Problems also arose in cryo-electron microscopy image analysis. Claude skipped part of the standard analysis procedure on its own judgment, producing an unsuitable electron density map in which the true three-dimensional structure and its mirror image were mixed. Human researchers had to correct the analysis by instructing Claude to prioritize the standard procedure and use exceptional handling only when the standard method failed.
According to the team, in areas where the humans had sufficient expertise, they could quickly spot and correct Claude's errors. In areas where researchers' knowledge was thinner, they sometimes missed problems in the AI's suggestions and lost time. While showing that AI can support a wide range of steps in scientific research, the study also makes clear that expert judgment and verification remain essential.
Unresolved biological function and potential as nanocapsules
What has been confirmed is that a phage-derived protein can self-assemble into vault particles in artificially engineered E. coli. It has not been confirmed that the same 96-molecule vault forms in the bacteria that phages carrying the L05 gene normally infect, or during natural infection.
Even if vault particles are produced during phage infection, their role is unknown. Whether they are containers that protect some molecule from host defenses, structures that carry enzymes, or something with another function must be tested in future experiments. They may perform the same function as mammalian vaults, or they may have acquired a different role over the course of evolution.
On the other hand, the ability to spontaneously form a giant hollow structure whose interior can be enlarged by genetic modification could lead to engineering applications. Anthropic also mentions possible use in drug delivery systems (DDS) that enclose enzymes or nucleic acid drugs and carry them to target cells. Research on nanocarriers using mammalian vaults has a 2005 precedent, but this study did not encapsulate a drug in L05 vaults or achieve delivery to target cells. Therapeutic applications would require verifying particle stability, immunogenicity, and safety.
Leonard Rome, who discovered vault particles and took part in this study as an advisor, also noted that elucidating vault function remains a very difficult problem. He nonetheless expressed hope that Anthropic's detailed structural data will prompt more researchers to take it on.
Key future tasks are to confirm that vault particles form in natural hosts and phage-infected cells and to identify their actual biological function. That will involve identifying the unknown molecule in L05's binding pocket and examining how deleting or overexpressing the related gene affects infection efficiency and host survival.
The study provides experimental support that proteins capable of forming vault particles exist in a broader range of organisms than previously thought. It also shows concretely that, while Claude can support many steps of structural biology research, erroneous suggestions can slow the work. How far the newly obtained structural information and AI-assisted research methods will help in approaching the function of vault particles, unresolved for about 40 years, will be worth watching.