AI-Assisted Structural Analysis of the Prokaryotic Vault: From 15 Billion Sequences to a 34-Hour Cryo-EM Reconstruction

1. A 40-Year Mystery: What Exactly Is a Vault?

In 1986, researchers isolated a previously undescribed particle from rat liver. Under electron microscopy, its longitudinal section showed a continuous arched profile, resembling a miniature cathedral dome. This appearance led to the name “vault.”

Since then, a long-standing question has remained: the vault can be observed, but its function, evolutionary origin, and assembly rules are difficult to define.

The structural features of the vault are striking:

• Large size: Its linear dimensions are about three times those of a ribosome. If mass is compared, the order of magnitude and measurement basis must be stated separately; linear size and mass should not be directly conflated.
• High symmetry: Mammalian vaults typically consist of 78 protein chains, with 39 chains in each half. The major vault protein contributes most of the mass and encloses a hollow cavity.
• High abundance: A mammalian cell can assemble tens of thousands of similar particles.

The difficulty is not only structural complexity. The functional evidence is contradictory. Vaults are usually absent in common model organisms such as yeast, fruit flies, and nematodes. Mice lacking vault-related genes can still develop and survive normally. This creates a “missing paradox”: the structure is elaborate and apparently costly to build, yet it does not appear to be essential for some organisms.

Several hypotheses have been proposed, including roles in nucleocytoplasmic transport, specific mRNA localization, ribosome regulation, and stress or immune responses. However, no unified validation has emerged. A deeper bottleneck lies in the research paradigm: systems based mainly on mammalian cells have low expression levels, long purification cycles, and slow structural iteration. Testing a single hypothesis can take years.

2. Strategic Shift: From Asking What It Does to Tracing Where It Came From

The central shift in this research was to move from “ask function first” to “trace origin first.” If functional evidence is inconsistent in higher organisms, then where did the vault come from? Does it exist in more ancient lineages? Which organisms retain a version closer to an ancestral state?

The researchers focused on prokaryotes because prokaryotic systems offer engineering advantages:

Dimension	Mammalian-cell system	Prokaryotic system
Sequence resources	Relatively limited	Large-scale comparison using genomic and metagenomic data
Expression and purification	Low yield, long procedures, cycles measured in weeks	Fast expression and easy scale-up in hosts such as E. coli
Structural analysis	High sample-preparation and iteration costs	Clear particle sources, suitable for batch screening
Evolutionary information	Distant from the common ancestor, so signals are diluted	Closer to deep homology, with more primitive functional clues

This is a strategy of trading breadth for depth: first use large-scale sequence retrieval to identify candidate families, then create more room for trial and error in expression, purification, and structural analysis.

3. Screening Vault Candidates From 15 Billion Sequences

The research team used the major vault protein as the search target and performed an ultra-large-scale homology search across approximately 15 billion protein sequences. This task is not simple keyword matching; it requires a balance among sensitivity, specificity, and structural validation.

3.1 Technical Workflow

1. Homology identification: Known vault sequences and distant candidate sequences may have very low similarity. Exact matching alone would miss candidates, so sensitive searches based on hidden Markov models or sequence alignment are needed.
2. Domain validation: The major vault protein has characteristic domain combinations, including an N-terminal cap-like domain, coiled-coil regions, and domains associated with the inner side of the barrel wall. Only candidates with complete domain combinations and reasonable arrangement proceed to the next round.
3. Phylogenetic validation: Phylogenetic analysis is used to determine whether candidate sequences are truly homologous or merely appear similar due to convergent evolution.
4. Gene-adjacency analysis: Researchers examine whether candidate genes are located near antiviral defense genes, providing clues for later functional hypotheses.

3.2 Results and Statistical Definitions

The research team reported that the candidate family set expanded from 41 to 1,507, an increase of more than 36-fold. The related proteins were distributed across 29 bacterial phyla and 4 archaeal phyla. In bacteria, they remain rare and were found in less than 1% of species.

To avoid misreading these numbers, the statistical scope should be clarified:

• 15 billion sequences: The total number of public protein sequences included in the search, including bacterial, archaeal, genomic, and metagenomic sources. The specific database versions and release dates should follow the versions listed in the paper.
• 41 to 1,507: The number of independent candidate families after dereplication, domain-integrity validation, and phylogenetic filtering, not the raw number of hit sequences.
• 36-fold: Calculated as 1,507 divided by 41.
• 29 bacterial phyla / 4 archaeal phyla: Counted according to the taxonomic system and lineage divisions used in the paper.
• Less than 1%: An estimated occurrence frequency of candidate families among the bacterial species searched. This depends on species representativeness and sampling completeness.

3.3 Example Reproducible Code

In real engineering, tools such as HMMER, MMseqs2, or equivalent software are commonly used. The following code is only a workflow illustration. In actual execution, it should be replaced with the versions and parameters specified by the paper.

from pathlib import Path
import pandas as pd
SEEDS = Path("seeds/mvp_ref.fasta")
SEQ_DBS = ["prokaryote_genomes", "phage_genomes", "metagenomes"]
# 1) High-sensitivity homology search
# Real tool examples: hmmsearch (HMMER v3.4) or easy-search (MMseqs2)
raw_hits = []
for db in SEQ_DBS:
    raw_hits += profile_search(
        query=SEEDS,
        database=f"/data/{db}",
        method="hmm",
        evalue=1e-3,
        max_hits=1000,
    )
# 2) Merge and deduplicate
family = pd.concat(raw_hits)
family = family.drop_duplicates(subset=["query_id"])
family = family.drop_duplicates(subset=["subject_id"])
# 3) Domain validation
family = family[family["hmm_cov"] >= 0.8]
# 4) Expression and validation plan
plan = build_expr_plan(family, host="E_coli", mode="batch")
plan.to_csv("output/expr_candidates.csv", index=False)

Key parameters include the input FASTA format, tool version, E-value threshold, coverage threshold, clustering identity, dependency library versions, and error-log storage path. Only when these parameters are fully recorded can the number of candidate families be rechecked.

This workflow is a typical “data funnel”: hundreds of billions of sequences are progressively narrowed down into large candidate pools, and then into a list of batches suitable for cloning and expression. Because the search task is read-heavy and write-light and can be partitioned, it is suitable for deployment on elastic computing platforms, scaling out during peak periods and releasing resources when idle.

4. From Gene to Particle: Expression, Purification, and 90-Minute SEC

After the candidate library was established, the research team selected 44 candidates from the 1,507 for expression testing. AI assisted with synthetic DNA design, codon optimization, promoter and ribosome-binding-site configuration, and the design of induction-condition combinations.

Key engineering points included:

• Optimizing codons according to host preferences and avoiding high-GC regions, strong secondary-structure regions, and sequences likely to cause abnormal transcription.
• Generating multiple expression-condition sets in parallel rather than testing them one by one.
• For giant complexes, subunit pairing, barrel-wall geometry matching, and assembly order can all cause failure. Therefore, soluble expression rate is a key threshold.

In this batch, 15 candidates expressed soluble protein, a success rate of about one-third. For large assembly systems, this indicates that some candidates met the conditions for structural analysis.

For purification, the researchers used size-exclusion chromatography instead of traditional ultracentrifugation. Conventional sucrose-solution ultracentrifugation usually processes 0.5 to 1.0 mL of crude sample and takes about 36 to 72 hours, or roughly three days. In this study, the size-exclusion chromatography condition compressed the process to 90 minutes. To support reproducibility, the relevant conditions are as follows:

• Sample volume: About 200 microliters loaded.
• Column specification: A pre-packed size-exclusion column with a bed volume of about 24 mL.
• Flow rate: About 0.5 mL/min.
• Buffer: A fixed formulation version, with pH, salt concentration, and additives clearly specified.
• Peak determination: The target peak is judged using both the elution-volume window of marker proteins and the ultraviolet absorbance curve.
• Repeat validation: Each batch should have at least three technical replicates. Only when peak position, yield, and purity meet the criteria should the sample proceed to cryo-electron microscopy.

The automated workflow includes column equilibration, sample loading, constant-flow elution, real-time ultraviolet recording, peak determination, and fraction collection. Control scripts and scheduled tasks allow the system to run unattended.

It should be emphasized that AI should not independently determine the process route. Traditional practice often uses ultracentrifugation for purification and size-exclusion chromatography for analysis. In this study, the roles were reversed: size-exclusion chromatography was first used to quickly determine whether target particles had been obtained, and then parameters were iterated. The process route should be decided by researchers familiar with the system, while AI is responsible for compressing the trial-and-error cycle at the execution layer.

5. 34-Hour Cryo-EM Reconstruction: What Did AI Do?

Cryo-electron microscopy three-dimensional reconstruction usually includes motion correction, contrast transfer function estimation, particle picking, two-dimensional classification, three-dimensional reconstruction, model building, and refinement. Vaults are barrel-shaped and often appear as ring-like structures with distinct inner and outer contrast in two-dimensional views, which helps with particle screening. Known symmetry can also be used as a constraint to reduce the three-dimensional solution space.

Completing reconstruction in 34 hours came from three types of compression:

1. Parallel computing compressed queue time: Particle alignment, two-dimensional classification, and three-dimensional classification can be split across multiple GPUs.
2. Real-time feedback compressed waiting time: Data acquisition and processing were performed synchronously, allowing researchers to adjust acquisition strategies in time.
3. Automated workflows compressed human intervention: AI handled parameter configuration, anomaly detection, and initial result screening.

Acceptance must rely on objective metrics, especially FSC 0.143:

Acceptance item	Suggested metric	Acceptable standard	Verification method
Global resolution	Spatial frequency corresponding to FSC = 0.143	Reaches the preset target	Half-map FSC curve
Local resolution	Distribution of local sharpness	Key regions satisfy interpretation requirements	Local FSC or ResMap
Particle proportion	Proportion of target particles in two- and three-dimensional classification	Stable and without obvious contamination	Classification statistics
Resolution unit	Angstrom, Å	Clearly labeled	Reports and figures
Model plausibility	Agreement between geometric configuration and density	Passes validation	Model refinement report

Without objective gates, even a very fast workflow may simply produce results whose correctness cannot be judged more quickly.

The research team obtained a three-dimensional structure of the prokaryotic vault. Bacterial versions are usually composed of a single major vault protein, so their structural logic is relatively simple while retaining the basic assembly architecture of the vault. However, a first structure is not a final conclusion. Whether the resolution is sufficient for detailed functional interpretation, and whether local regions are clear enough, still require further experimental confirmation.

6. Engineering Vision: From Structural Analysis to Programmable Containers

The engineering appeal of the vault comes from its container properties: its large internal space can carry molecules; its inner and outer surfaces can be modified separately; subunit self-assembly can be regulated through genetic modification; and its prokaryotic origin is favorable for scalable expression.

Prokaryotic vaults are usually built by one protein chain repeatedly folding to construct the barrel wall, while eukaryotic versions involve multiple different protein chains. A more accurate statement is that both eukaryotic and prokaryotic vaults use major vault protein-type proteins assembled repetitively. The differences lie mainly in subunit domains, accessory proteins, and assembly hierarchy, rather than a simple dichotomy of “one monomer does everything” versus “multiple subunits divide labor.”

Based on this plasticity, the research team attempted to expand loading capacity through genetic modification, providing a basis for programmable nanocontainers.

Before discussing applications, biological safety and regulatory boundaries must be clarified. Directions such as drug delivery, nanoreactors, vaccine display, and synthetic-biology modules are still exploratory. Any experiments should satisfy the following requirements:

• Genetic modification and expression operations must be carried out in closed laboratories that meet the appropriate biosafety level.
• When pathogen-related elements or human samples are involved, ethical review and biosafety approval should be completed.
• Nanoreactor and enzyme-immobilization experiments need to assess leakage, toxicity, and waste disposal.
• Vaccine and display platforms require evaluation of immunogenicity, safety, and compliance.
• Without approval, related vectors should not be used in humans, clinical settings, or environmental release.

Potential applications include:

• Drug delivery: Loading therapeutic molecules into the hollow cavity to protect them from degradation.
• Nanoreactors: Immobilizing enzymes in the inner cavity to construct compartmentalized reaction spaces.
• Vaccine and display platforms: Displaying antigen fragments on the particle surface.
• Synthetic-biology modules: Serving as programmable intracellular compartments to separate metabolic pathways.

These applications still require substantial validation, but the designable and scalable prokaryotic vault already provides a basis for discussion.

7. A Reusable AI-Assisted Research System

The value of this research is not limited to solving one structure. It also demonstrates a transferable engineering path: data retrieval, candidate screening, design and synthesis, expression and purification, structural analysis, and engineering modification.

Stage	AI advantage	What humans must control	Speed-up approach
Sequence retrieval	Parallel search across massive datasets	Threshold setting and boundary review	Elastic computing and high-throughput scheduling
Candidate design	Batch generation of DNA and expression plans	Candidate selection and experimental priority	Template-based generation and automatic validation
Protein purification	Parameter search and workflow automation	Process route and acceptance criteria	Equipment-control software and scheduled tasks
Structural analysis	Parameter configuration and anomaly detection	Resolution judgment and model plausibility	GPU parallelism, real-time feedback, and automated gates
Engineering modification	Variant design and effect prediction	Target definition and risk assessment	Design-experiment-learn loop

Five engineering principles can be summarized:

1. Define standards before discussing speed. Both speed and resolution must have objective gates.
2. Divide responsibilities between humans and machines. AI handles search and execution, while humans handle direction and judgment.
3. Turn trial and error into parallel work. Experiments that can be batched and parallelized should not be run serially.
4. Turn experimental processes into data assets. Parameters, results, and conclusions should be structured and retained.
5. Make computing elastic and data tiered. Retrieval and reconstruction workloads are suitable for elastic cloud scheduling.

8. FAQ: What Is a Vault? Why Is Knockout Not Lethal? What Can AI Not Do?

What is a vault?
A vault is a large hollow protein complex with a barrel-shaped or dome-like structure. The mammalian version consists of multiple protein chains, while the prokaryotic version can be assembled by a single major vault protein folding repeatedly.

Why are knockout mice still normal?
This indicates that the vault is not essential for development in some model organisms. It may perform conditional, auxiliary, or environment-responsive functions. Loss of a structure does not mean it has no function; it may also indicate functional redundancy or functions that are not triggered under current experimental conditions.

What can AI not do in structural analysis?
AI cannot replace objective validation. It cannot prove function by itself, cannot replace quality gates such as FSC, local resolution, and model plausibility, and cannot independently determine process routes, experimental safety, or final scientific conclusions.

9. Conclusion: From Year-Level Iteration to Week-Level Iteration

From the “cathedral” observed in 1986, to a 40-year functional mystery, to the structural analysis of the prokaryotic vault, what changed was not only the answer but also the way of finding answers. The search scale expanded from a few known sequences to hundreds of billions of sequences. Purification changed from starting at three days to a 90-minute cycle. Structural analysis changed from starting at weeks to being measured in hours.

The functional mystery of the vault is not yet fully solved, but the emergence of the prokaryotic version opens a new direction. When AI handles retrieval, design, and execution, while humans handle direction, standards, and judgment, structural biology is entering an era of week-level iteration.