AI-Driven Structural Biology: Solving a 40-Year Vault Protein Structure Mystery

1. Introduction: A structural mystery that stayed open for 40 years

In 1986, researchers discovered an unusually large class of particles in crude fractions of mammalian cells. The curved shell resembled the arches of a cathedral vault, so the particle was named a “vault.” For decades afterward, it remained a long-standing puzzle in structural biology:

• Remarkable size: The vault particle has a mass of roughly three times that of a ribosome and is one of the largest protein complexes inside cells. In mammals, it is built from 78 polypeptides, dominated by 78 copies of the major vault protein (MVP). These polypeptides come together in upper and lower halves, enclosing a large hollow interior.
• High abundance: A single mammalian cell can contain tens of thousands of these hollow barrel-like particles, making it difficult to believe that they serve no purpose.
• Unknown function: Vault particles are absent from common model organisms such as yeast, fruit flies, and nematodes. Even after the major vault protein gene was knocked out in mice, the animals still developed normally.

Humans could observe this container, but could not explain what it holds or why it exists. Conventional approaches reached an impasse. A different route then began to take shape:

1. Use AI-driven research tools to mine clues from massive sequence datasets.
2. Use automated or semi-automated experiments for rapid trial and error.
3. Use cloud computing with elastic compute capacity to complete structural analysis.

Following this route, the team obtained the first three-dimensional structure of a prokaryotic vault particle in less than four months. This article breaks down the technical workflow step by step and extracts transferable engineering lessons.

2. Why this giant protein complex was difficult to solve

The problem of determining a structure with unknown function can be divided into four stages, each with its own difficulties:

• Difficult to discover: Vault proteins are highly divergent during evolution. Conventional sequence-similarity searches struggle to collect homologous sequences scattered across many species.
• Difficult to express: Prokaryotic vault particles may assemble differently from their eukaryotic counterparts. It is difficult to predict in advance whether heterologous expression will produce soluble, correctly assembled particles.
• Difficult to purify: The particles are huge and prone to dissociation or aggregation. Traditional separation methods are slow and low-throughput, making it hard to screen dozens of candidates in parallel.
• Difficult to resolve structurally: Cryo-EM sample preparation, data screening, three-dimensional reconstruction, and model building form a long chain. Every step involves parameters that often require human judgment.

The breakthroughs also appeared at these four stages.

3. Overall workflow: human-AI division of labor and closed-loop validation

At its core, this research is an end-to-end practice of AI-assisted scientific discovery. The division of labor can be summarized as follows:

• What AI handled: generating and screening candidates in high-dimensional spaces, designing and scheduling experiments, assisting with data analysis and model building, and rapidly retrieving relevant literature and knowledge.
• What humans handled: defining the scientific question, setting acceptance criteria, reviewing key checkpoints, and making final judgments on AI outputs.

The closed-loop validation process was:

AI generates a plan → automated or semi-automated experiments are executed → results are returned in structured form → AI revises the next plan based on those results → repeat and iterate

The performance at each stage changed significantly:

Stage	Traditional approach	Performance with AI collaboration
Sequence discovery	Manual comparison against limited sequence libraries; family size remained at 41	Searched about 15 billion sequences; family expanded to 1,507 members, about a 36-fold increase
Cloning and expression	Designing synthesis and expression plans one by one; long turnaround time	DNA design, expression configuration, and ordering plans for 44 candidates were coordinated by AI
Sample purification	Sucrose density gradient ultracentrifugation took about 3 days	Size-exclusion chromatography required about 90 minutes per round
Structure determination	Three-dimensional reconstruction and model building often took weeks or months	Three-dimensional reconstruction and model building completed in about 34 hours

4. Step one: sequence mining — finding a needle in 15 billion sequences

4.1 Shift the battlefield: from higher organisms to microbes

Since the function of vault particles in mammalian cells remained difficult to determine, it made sense to look for answers in microbes. Microbes reproduce quickly, have clearer genetic backgrounds, and are easier to use for heterologous expression. In addition, once a vault protein gene is identified, hosts such as Escherichia coli can be used to produce the protein on demand.

Over the past decade, environmental metagenomic sequencing data has grown exponentially. From samples such as soil, seawater, and the human gut, researchers have obtained large numbers of sequences from uncultured microorganisms. These datasets contain many protein families that have not yet been studied.

4.2 Mining strategy

In a systematic search for vault proteins, the researchers searched through approximately 15 billion protein sequences. The general strategy was as follows:

• Layered screening: Fast comparison tools were first used to identify distant candidates. More sensitive sequence comparison methods were then applied iteratively to expand the candidate set and reduce false negatives.
• Structural cross-validation: Candidates were subjected to three-dimensional structure prediction and compared with the known vault structure backbone. Sequence can vary greatly, but the folding backbone is relatively conserved. This helped filter out false positives that looked similar in sequence but did not fold like a vault protein.
• Gene neighborhood analysis: The researchers examined the genes surrounding each candidate to determine whether it appeared near gene clusters associated with defense functions.

4.3 Mining results

• Family size: The known vault protein family expanded from 41 members to 1,507 members, an increase of about 36-fold.
• Distribution: These proteins were found across 29 bacterial phyla and 4 archaeal phyla. Homologous genes were also present in the genomes of phages that infect bacteria.
• Occurrence frequency: They are rare in the bacterial world, appearing in less than 1% of species.
• Gene neighborhood: They are often located near antiviral defense genes, suggesting that they may be specialized equipment used by bacteria to resist viruses.

4.4 Method notes

• Algorithm level: Iterative sequence alignment and hidden Markov model-based lineage searches are mainstream methods for mining distant homologous families.
• Engineering level: When dealing with tens of billions of sequences, comparison, clustering, and structure prediction need to be divided into shardable batch tasks, scheduled in parallel across GPU resources, and tracked with full data-version and parameter records to ensure reproducibility.
• Role of large models: Large language models can help understand the search intent, plan the strategy, orchestrate tasks, summarize candidates, and explain the rationale. This turns parameter selection from guesswork into evidence-based engineering decisions.

It is important to emphasize that sequence-level clues are only a starting point. The presence of a gene does not prove function, and even less does it allow detailed structural inference.

5. Step two: from sequence to protein — expression design and clone construction

• Candidate selection: From the 1,507 candidates, 44 were selected for experimental validation based on structure-prediction confidence, gene-neighborhood clues, and sequence quality.
• Plan generation: For each candidate, AI helped prepare executable plans, including DNA synthesis sequences with host-optimized codons, adjusted GC content, and reduced secondary structure; expression vector and induction condition configurations; and gene synthesis orders.
• Heterologous expression: The synthesized genes were introduced into E. coli. In the end, 15 candidates, about one-third of the total, expressed soluble protein.

This hit rate shows two things. First, structure-prediction-based prioritization was effective because it concentrated experimental resources on more promising candidates. Second, moving from a gene to soluble protein remains the main bottleneck. If automated workstations and microplate-based screening are introduced, the number of candidates that can be validated in parallel could increase substantially.

6. Step three: faster purification — from 3 days to 90 minutes

6.1 Why the method had to change

Correct separation of vault particles commonly relies on sucrose density gradient ultracentrifugation. The principle is that particles with different sedimentation coefficients migrate in a centrifugal field to positions where their buoyant density is close to that of the surrounding medium, forming bands.

• Problem: This workflow takes about 3 days and processes only a limited amount of sample per run. It cannot support parallel screening of dozens of candidates.
• New approach: The team switched to size-exclusion chromatography (SEC), also known as gel filtration chromatography.

6.2 Why SEC is better suited to giant particles

The stationary phase of SEC consists of porous beads packed in a column. Molecules smaller than the pore size enter the pores and travel a longer path through the column, so they elute later. Particles larger than the pore size are excluded from the beads and move with the mobile phase, eluting first.

Schematic of chromatographic elution order
Mobile phase ──► [ porous bead-packed column ] ──► detector / fraction collector
             o o o o o
             pore size  ← small molecules enter pores → migrate slowly, elute later
             o o o o o
Vault particles (diameter much larger than pore size) are excluded from beads → migrate fastest, elute first
Typical order: empty cores / viral particles → giant protein complexes → proteins → peptides → salts / small molecules

Vault particles are among the largest complexes in cells, so they fall into the earliest elution region. This physical property makes them easy to distinguish from other proteins on an SEC chromatogram. A single SEC run takes about 90 minutes, allowing multiple rounds per day and enabling parallel evaluation of candidate proteins. In the end, 5 candidates formed structurally homogeneous and morphologically intact particles that met the requirements for cryo-EM observation.

6.3 Challenges and the value of “contrarian” feedback

Huge particles are not ideal subjects for SEC. Several problems may occur:

• Dissociation: Shear forces and dilution may partially disassemble the complex.
• Nonspecific adsorption: Particles may adsorb onto the column resin and be lost.
• Aggregation: Elution peaks may show tailing.

Common countermeasures include reducing flow rate, using more rigid resin, adjusting the ionic strength and additives of the buffer, or performing a quick negative-stain EM check immediately after chromatography.

During this process, the AI tool did not simply agree with the human plan. It first questioned the SEC approach and then offered targeted suggestions. This is the proper posture for human-AI collaboration: replacing “quick affirmation” with “first attempt to falsify, then optimize.” Ultimately, the final judgment still came from chromatographic profiles and electron microscopy images.

7. Step four: structure determination — recovering the 3D structure in 34 hours

Once homogeneous particles were obtained, the cryo-EM workflow began:

1. Sample preparation: The sample was rapidly frozen into vitreous ice to preserve the native conformation of the particles.
2. Imaging and preprocessing: Low-dose imaging recorded two-dimensional projections. Motion correction and contrast transfer function (CTF) estimation were then performed.
3. Particle picking: Individual particles were picked from micrographs.
4. 2D and 3D classification: Particles with abnormal orientations or incomplete structures were removed.
5. 3D reconstruction and refinement: A high-resolution density map was generated.
6. Model building: The amino acid sequence was placed back into the density map.

7.1 Symmetry constraints

The vault shell is usually formed by multiple copies of the same protein arranged around a barrel-like structure, giving it relatively high symmetry. Applying appropriate symmetry constraints during reconstruction effectively averages information from each repeating unit, significantly improving the effective signal-to-noise ratio. For large complexes where single-particle analysis struggles to reach high resolution, this is a common strategy.

7.2 Assistance from artificial intelligence and machine learning

• Particle picking: Deep-learning-based particle picking methods are widely used in cryo-EM data processing. They can reduce manual bias and workload.
• Model building: Machine-learning-based modeling tools can generate initial models, after which human experts correct them using the density map. Unreasonable side-chain conformations, mismatched secondary structures, and similar errors are exactly where human oversight is needed.

7.3 The role of compute

Two-dimensional classification and three-dimensional reconstruction in cryo-EM are computationally intensive tasks and are naturally suited to parallel execution on multi-GPU nodes.

• Data uploaded to the cloud: Raw images amounting to several terabytes were uploaded to cloud storage, and a workflow engine automatically triggered each processing step.
• Elastic resources: When multiple processing tasks were queued at the same time, the cloud could temporarily expand the GPU resource pool and release it after task completion.
• Speed-up result: With AI coordinating the workflow, the process from data to a three-dimensional structure took only about 34 hours.

8. Interpreting the result: more than just seeing the shape

The value of this achievement has three layers:

1. Scientific level: For 40 years, vault particles had only been observed in mammalian cells. For the first time, a homologous complex in prokaryotes was confirmed and its structure was resolved, filling in an important piece of the evolutionary puzzle.
2. Functional clues: Bacterial vault protein genes are often located near antiviral defense genes, suggesting that vault particles may be involved in bacterial immunity. However, the exact mechanism still requires further experimental confirmation.
3. Methodological level: The study shows that “AI-assisted sequence mining + automated experiments + cloud computing” can complete work in months that previously would have taken years.

A deeper implication is that the vault particle itself is a programmable hollow nanoscale container. Once its assembly rules are understood, it may be engineered into drug-delivery vehicles, vaccine antigen-display platforms, or nanoreactors. Structural determination is a prerequisite for these applications.

9. Engineering perspective: turning AI research capability into infrastructure

Viewed from a cloud platform, this research forms a clear data pipeline:

• Data layer: Massive sequence and cryo-EM data are stored in object storage. Data can be tiered by access frequency: hot data is kept in standard storage, while historical data can be moved to low-frequency or archive storage. Metadata and experimental records are managed centrally in databases.
• Compute layer: Sequence search, structure prediction, and cryo-EM data-processing tasks can be containerized and submitted to a container platform or batch-processing service. Tasks that tolerate interruption can use preemptible instances to reduce costs. Task progress, resource utilization, and failure rates can be connected to monitoring and alerting.
• Scheduling layer: Experiments with many candidates in parallel are essentially aggregations of large numbers of short tasks, which are suitable for distribution and retry through a message queue.

A typical pipeline organization (schematic):

[Sequence database] ──► [Sequence mining / structure prediction] ──► [Candidate ranking]
                                             │
                                             ▼
[Experiment design] ──► [Automated / manual experiments] ──► [Quality gate: soluble? homogeneous?]
                                             │ pass
                                             ▼
[Cryo-EM sample prep] ──► [Upload EM data to object storage] ──► [GPU cluster: 2D/3D classification and reconstruction]
                                             │
                                             ▼
                              [Model building and validation] ──► [Archive results and experimental records]

Key engineering principles:

• Define explicit quality gates at every stage. For example, is the SEC peak symmetric? Are particles homogeneous in negative-stain EM?
• Keep data and parameters traceable throughout, including sequence versions, buffer formulations, chromatography parameters, and cryo-EM processing parameters.
• Record failures as well. Expression conditions and results for every candidate should enter a knowledge base and inform the next round of decisions.

10. Five lessons for human-AI collaborative research

1. Let humans make judgments and let AI handle retrieval, calculation, and documentation: Researchers decide which scientific questions to pursue; AI handles pattern recognition, search, computation, and document organization.
2. Perform large-scale computational screening first, then run expensive experiments: Leaving wet-lab experiments to a small number of top-ranked candidates can significantly save time and reagents.
3. Question AI conclusions: Sequence homology does not mean identical function, and model prediction does not equal experimental fact. Experimental data must be the final authority.
4. Automation is not optional; it is an amplifier: The improvement from 3 days to 90 minutes in purification, and from weeks to 34 hours in structure determination, comes not from a single algorithm but from an end-to-end pipeline.
5. Data governance must come first: Traceable sequence-library versions, experimental records, and model versions are the foundation of reproducible AI-driven research.

11. Risks and limitations

• AI hallucination: Large language models may generate plausible but nonexistent gene IDs, references, or parameters. Any factual content must be checked against original databases.
• Screening bias: Sequence mining depends on existing data. Uncultured and unsequenced microorganisms remain blind spots.
• Experimental bottleneck: Only about one-third of candidates yielded soluble protein from gene to protein. Automation can alleviate this bottleneck but cannot eliminate it.
• Functional inference requires caution: “Adjacent to defense genes” is only an associative clue. Gene knockout, infection challenge, and other experiments are needed for validation.
• Compliance and safety: Work involving gene synthesis and microbial operations should be carried out in laboratories with appropriate biosafety conditions and must comply with local regulations and ethical requirements.

12. Conclusion: from solving one structure to redesigning the research workflow

The vault particle structure mystery that troubled the field for 40 years was not solved by a single “black-box technology.” It was solved by a human-AI collaborative pipeline in which computation and experimentation ran in parallel: AI handled retrieval, reasoning, and orchestration; automated experiments enabled efficient trial and error; cloud compute turned data into structure. For any R&D team facing “too many candidates, too-slow experiments, and too-messy data,” this methodology is worth borrowing: give repetitive work to machines and reserve key judgment for humans.