First 3D Structures of Bacterial Vault Particles Resolved: AI-Assisted Search, Size Engineering, and Experimental Validation

1. Why vault particles matter

Vault particles are among the largest ribonucleoprotein complexes in cells. Human cells contain roughly 10,000 copies of each vault, and a typical vault is about 67 nm long and 40 nm wide. This hollow structure is naturally suited to carrying molecular cargo, which is why vaults have long been considered for applications such as drug delivery and vaccine development.

Despite this promise, their biological function has remained unclear for about 40 years since their discovery in 1986. One major obstacle has been the lack of a suitable model system. Vaults derived from mammalian cells are difficult to produce at scale. Previous knockout studies of the major vault protein in model animals produced only mild phenotypes, making it hard to support detailed functional research.

To overcome this bottleneck, the researchers took a homology-search approach: looking for vault particles from other branches of life that are easier to prepare and can be engineered more flexibly. They also used large AI models to shorten the cycle from sequence search to structural analysis.

2. Structural results: 96 copies and tunable size

2.1 From 15 billion sequences to four complete structures

The research team first searched approximately 15 billion predicted protein sequences for homologs of the major vault protein. Candidate sequences covered 29 bacterial phyla and 4 archaeal phyla. From these, 44 proteins were selected for experimental validation, with results falling into three layers:

• 7 proteins successfully assembled into vault-like particles in E. coli;
• cryo-EM confirmed complete closed structures for 4 of these candidates, with 2 derived from bacteria and 2 from phages;
• L05, a phage-derived protein, was selected for detailed analysis and resolved to 3.2 Å.

The L05 vault consists of 96 identical protein chains. Each chain contains 7 wall repeat sequences, whereas the human major vault protein contains 9. The L05 particle measures 54.5 nm in length and 47.5 nm in width, giving it a more spherical appearance than the human vault. Its enclosed volume is about 25% larger than that of the human vault, and its total molecular mass is approximately 9.5 MDa. This work produced a bacterial vault particle library containing four complete 3D structures from bacterial and phage sources.

Feature	Human vault	L05 (phage-derived)	L05+2 (engineered variant)
Subunit composition	Major vault protein	96 identical proteins	L05 plus two inserted wall repeats
Wall repeats per chain	9	7	Inferred to be 9
Length	~67 nm	54.5 nm	Larger than L05
Width	~40 nm	47.5 nm	Larger than L05
Internal volume	Baseline	~1.25× human vault	~20% larger than L05; total volume ~1.5× human vault

2.2 Size engineering: wall repeats as a tuning knob

Wall repeats are structural units repeated along the axial direction of the major vault protein and form the barrel-shaped main body. To test whether these repeats can control particle size, the researchers inserted additional wall repeat sequences into L05, producing the L05+2 and L05+4 variants.

• The cryo-EM structure of L05+2 showed that both length and width increased simultaneously. Its internal volume increased by another ~20% relative to L05, and its total volume reached about 1.5× that of the human vault.
• These results suggest that the number of wall repeats can serve as a “knob” for tuning vault size.

It is important to distinguish “internal volume” from “total volume.” Internal volume refers to the cavity space available for molecular cargo, while total volume refers to the overall volume occupied by the particle. Conclusions about size comparisons are reliable only when these metrics are standardized.

2.3 Ligand clues inside the cavity

Structural analysis also revealed an unassigned region of electron density on the inner wall of L05. Computational searches suggested that the most likely ligand is coenzyme A or a derivative. The inner walls of human vaults similarly bind nucleotide-type molecules such as NAD+. Because these molecules are structurally similar, the inner cavity of vaults may be a functional region for binding small molecules and partner proteins.

However, this density has not yet been fully assigned, and further experiments are still required.

3. How large language models contributed across four stages

Large language models played a role in four key stages, always in the form of “generate candidates, let experiments decide”:

1. Search: screen candidates from ultra-large sequence libraries;
2. Design: generate DNA sequences for laboratory synthesis;
3. Optimization: improve protein purification workflows;
4. Data processing: reconstruct structures from cryo-EM data.

Stage	Work performed by AI	Measurable effect	Validation method
Sequence search	Screen candidates from ~15 billion predicted sequences	Candidates covered 29 bacterial phyla and 4 archaeal phyla; 44 proteins selected for validation	Wet-lab expression and assembly
DNA sequence design	Generate DNA sequences for synthesis and design	—	Synthetic cloning and expression validation
Purification optimization	Improve purification and sample preparation protocols	Reduced from ~3 days to ~90 minutes	Particle integrity and purity checks
Cryo-EM data processing	Parameter configuration, workflow orchestration, and result interpretation	First complete structural model generated in ~34 hours; conventional analysis usually takes weeks	Resolution assessment and structural review

3.1 Search: turning needle-in-a-haystack into a filterable funnel

The main challenge in the search stage is scale. Determining whether a protein sequence is homologous to the major vault protein cannot be done by pairwise comparison across all sequences. A common workflow is:

1. Build a position-specific scoring model, such as a hidden Markov model, from known sequences;
2. Scan sequence databases in parallel;
3. Filter candidates by score tiers;
4. Remove redundancy based on taxonomic diversity.

Large language models are well suited to tasks with many parameters and dispersed judgments, such as:

• generating and comparing combinations of search parameters;
• writing batch comparison scripts;
• stratifying candidates by phylum, length, domain composition, and other criteria;
• ranking validation priorities using sequence features and literature clues.

The core value of AI here is not that it is faster than specialized software, but that it can convert expert experience into executable, reproducible filtering rules. Ultimately, wet-lab experiments determine which proteins can assemble into particles: only 7 of the 44 candidates succeeded.

3.2 Design: generating DNA sequences for synthesis and expression

With gene synthesis and heterologous expression systems, researchers only need to obtain the DNA sequence to produce a target protein in an experimental host. In this area, large language models can:

• reverse-translate target protein sequences into coding sequences;
• adjust codon usage to match the translational preferences of the expression host;
• avoid features that may interfere with synthesis or expression, such as repeated fragments, abnormal GC content, strong secondary structures, and cryptic splice sites;
• design cloning strategies, such as inserting additional wall repeats to create size variants;
• modify designs through iterative dialogue and rapidly generate multiple alternatives.

All designs still require confirmation through synthetic cloning, expression, and assembly experiments.

3.3 Optimization: reducing purification from 3 days to 90 minutes

Purification time is often consumed by repeated trial and error. AI can move some of this trial and error into computational design. Conventional vault purification takes about 3 days. The team’s optimization strategy can be summarized as “coarse first, then fine”:

• Coarse extraction uses hypotonic lysis and density-gradient centrifugation. Because vault particles are among the largest cellular complexes, their density can be clearly separated from most impurities, enabling rapid enrichment.
• Fine purification uses size-exclusion chromatography. This method separates molecules by hydrodynamic radius and is well suited to particles with large size and regular shape, reducing purification time to about 90 minutes.

In such optimization work, large language models can help design protocols, explain parameters, troubleshoot anomalies, and iterate solutions. However, time is only one metric. Particle integrity, aggregation state, and reproducibility must also meet standards for the optimization to be considered valid.

3.4 Data processing: first structural model in ~34 hours

Cryo-EM image processing involves multiple dependent steps, such as:

• motion correction;
• contrast transfer function parameter estimation;
• particle picking;
• 2D classification;
• 3D classification and refinement;
• resolution evaluation.

Each step has multiple algorithm and parameter choices and relies heavily on experience. Large language models can assist with parameter recommendation, workflow orchestration, log interpretation, and preliminary evaluation of results. In this study, the first complete structural model was generated in about 34 hours, whereas conventional analysis would typically take several weeks.

For highly symmetric particles such as vaults, two additional issues deserve attention:

• Symmetry can improve resolution. Vaults are assembled from two mirror-symmetric halves, and subunits within each half are also arranged repeatedly. These constraints can be used during reconstruction.
• Symmetry constraints can create artifacts. Local asymmetric features may be “averaged out.” Manual inspection of local resolution, density continuity, and reasonable model conformations is still necessary.

4. Engineering implications: making AI participation reusable

4.1 Correct collaboration: models generate options, experiments decide

This study reflects a typical division of labor in AI-assisted research:

• AI’s role is to generate candidate solutions quickly and at low cost, including sequences, conditions, parameters, and code;
• Human and experimental work defines filtering criteria, executes validation, and provides feedback;
• Optimization is an iterative loop that converges based on feedback.

Illustrative code:

def ai_assisted_cycle(problem, ai_generate, evaluate, max_rounds=5):
    history = []
    for r in range(max_rounds):
        # Provide historical results and constraints so that candidates converge across rounds
        candidate = ai_generate(problem, history)
        # Validation is handled by deterministic rules and experiments; AI does not act as judge
        result = evaluate(candidate)
        history.append((candidate, result))
        if result.passed:
            return candidate
    raise RuntimeError("Maximum rounds reached without validation; human intervention is required")

4.2 Ultra-large-scale retrieval: parallel sharding and layered filtering

For sequence databases at the trillion-byte scale, processing can be organized around storage, computation, and filtering:

def screen_homologs(seq_db, hmm_profile, score_cutoff, identity_cutoff):
    """Screen candidate homologs in an ultra-large sequence database (illustrative)."""
    # 1. Scan records in parallel to reduce single-node memory pressure
    raw_hits = seq_db.parallel_map(
        lambda rec: hmm_profile.scan(rec),
        batch_size=100_000,
    )
    # 2. Coarse filter by score threshold to reduce the scale of precise comparisons
    scored = [h for h in raw_hits if h.evalue <= score_cutoff]
    # 3. Perform precise comparison and redundancy removal; retain representative sequences by phylum diversity
    return diversity_dedup(scored, identity_cutoff, by="phylum")

Two engineering issues are important:

• Reproducibility: fix model versions, parameters, and random seeds, and record the retrieval version associated with each result.
• Traceability: each candidate should be traceable to its source database and version.

4.3 Protocol design: set hard constraints first, then optimize scoring

The design of L05+2 and L05+4 variants is a typical example of “specification-driven option generation.” If experience is written as executable hard constraints and followed by scoring-based ranking, optimization is less likely to drift:

def evaluate_purification_plan(plan):
    """Evaluate a purification plan: apply hard constraints first, then compute a composite score (illustrative)."""
    # Hard constraints: reject immediately if any constraint is not met
    if plan.buffer_ph not in range(6, 9):          # Buffer conditions must maintain particle stability
        return None
    if plan.additives_compatible_with_downstream is False:
        return None                                 # Additives that interfere with downstream experiments are not allowed
    if plan.duration_minutes > 180:
        return None                                 # Maximum processing time
    # Composite score: yield, purity, and duration
    score = (3 * plan.yield_ratio + 2 * plan.purity
             - plan.duration_minutes / 60.0)
    return score

4.4 Workflow execution: standard steps can be filled, not skipped

The reliability of experimental results depends on whether the workflow is verifiable. Standard operating procedures can be fixed as a checklist, and AI-generated steps can be merged into a controlled process:

REQUIRED_STEPS = [
    "motion_correction", "ctf_estimation", "particle_picking",
    "class2d", "class3d", "refinement", "resolution_eval",
]
def run_processing(steps, ai_param_tuner):
    """AI only recommends parameters; step order and completeness are constrained by the workflow (illustrative)."""
    params = ai_param_tuner(steps)                  # Parameter suggestions provided by the model
    executed = []
    for name in REQUIRED_STEPS:                     # Execute in fixed order; steps cannot be skipped
        executed.append((name, params.get(name)))
    if set(executed_names(executed)) != set(REQUIRED_STEPS):
        raise RuntimeError("Workflow incomplete; results are not admissible")
    return executed

4.5 Infrastructure: data, compute, and observability

This type of research is essentially data-intensive engineering. When moving to the cloud, capabilities are usually selected according to the workload:

• object storage for retaining raw images, sequences, and model snapshots;
• container platforms for elastic batch computing;
• message queues for task splitting and status synchronization;
• logging and monitoring services for end-to-end observability.

As an example, Baidu AI Cloud’s general product family can correspond to these scenarios. Exact specifications should be confirmed in the official documentation. In addition, if compute resources expose external interfaces, authentication and access control should be planned in advance to avoid unauthorized access.

5. Limits of AI: lessons from two failures

The study also exposed the boundaries of AI participation in scientific research.

Failure 1: An additive recommendation interfered with DNA delivery and delayed the research by two weeks.
The AI-generated plan looked reasonable on local metrics but broke assumptions required by downstream steps. Optimization objectives must cover the entire chain: purification not only supports structural analysis but must also be compatible with subsequent loading and delivery experiments.

Failure 2: Standard steps were skipped during cryo-EM data processing, resulting in an inaccurate structural map.
When a workflow lacks mandatory checks, a model may produce seemingly reasonable results while omitting required steps. The team later corrected the issue through closer human supervision.

Failure type	Symptom	Root cause	Improvement
Cross-stage side effect	An additive interfered with DNA delivery, causing a two-week delay	Optimization focused on single-stage metrics and ignored downstream constraints	Model the full-pipeline objective and set downstream compatibility as a hard constraint
Missing workflow integrity	Standard steps were skipped, producing an inaccurate structural map	Lack of mandatory validation and independent review	Fix the standard workflow, introduce resolution assessment, and add human review

These lessons can be summarized into three principles:

• Leave unverifiable parts to humans. Model errors often “look reasonable” and are difficult to detect through intuition alone, so independent and quantifiable checks are required.
• Set explainable boundaries for models. Use models freely in clearly constrained tasks, but retain human decision-making where conclusions depend on trustworthiness.
• Record the process, not only the results. Complete logs of prompts, parameters, and intermediate outputs make it possible to locate problems quickly and reproduce conclusions.

6. Significance and outlook

From a structural biology perspective, bacterial vault particles provide a more tractable experimental system for studying the assembly mechanisms of vault-type proteins. Because the number of wall repeats can be adjusted, they also serve as a size-controlled scaffold for cargo encapsulation. Applications such as displaying chimeric antigens, loading therapeutic molecules, or incorporating imaging probes all depend on structural characterization and size control.

At the same time, the biological function of vaults still needs to be determined. Assignment of inner-wall density and ligand-binding mechanisms require further experimental validation.

For R&D teams, the lessons from this study can be transferred directly:

• AI is well suited to high-frequency tasks such as retrieval, design, optimization, and data analysis;
• experimental validation, workflow integrity, and conclusion review must remain under human control;
• the true efficiency ceiling is determined not by the model alone, but by the data pipeline, validation loop, and governance framework built around it.

The role of AI in research is shifting from “answering questions” to “executing workflows.” The more standard the workflow and the clearer the validation criteria, the more stable the efficiency gains from AI become, and the more controllable the costs remain.