Solving a 40-Year Structural Mystery: How AI Assisted the Discovery and Cryo-EM Analysis of a Prokaryotic Vault Protein

A Long-Unsolved Structural Puzzle

Vault proteins are large, barrel-shaped ribonucleoprotein complexes found in most eukaryotic cells. First described in 1986, they have been implicated in various cellular processes, including intracellular transport, signal transduction, and multidrug resistance, yet their precise biological function remains undefined. A key challenge in studying vault biology is the lack of a simplified, controllable model system: eukaryotic vault particles are large, complex to assemble, and require lengthy purification protocols, making structural and functional studies both slow and difficult.

A promising avenue emerged from evolutionary analysis. Previous studies identified proteins with sequence homology to eukaryotic vaults in bacteria and bacteriophages; however, no direct experimental evidence confirmed their ability to assemble into complete vault-like particles. Identifying a prokaryotic candidate would enable researchers to leverage the speed, simplicity, and short production cycles of prokaryotic expression systems, establishing a far more efficient research platform than its eukaryotic counterpart.

This work spans multiple domains: database mining, experimental design, recombinant expression, protein purification, cryo-EM data collection, and image analysis. This long and complex pipeline makes it an ideal scenario for deploying large language models—not as decision-makers replacing human judgment, but as full-workflow collaborative assistants capable of accelerating retrieval, design, troubleshooting, and interpretation tasks at each stage.

Three Core Stages of AI Involvement

1. Large-Scale Sequence Retrieval and Candidate Screening

Vault-related sequences are scattered across massive public protein databases. In this study, the search space covered approximately 15 billion predicted protein sequences. In this phase, the language model contributed in the following ways:

• Search strategy construction: Assisted researchers in formulating keyword systems for homology searches, defining sequence features such as conserved domains, repeated motifs, and oligomerization interfaces—translating vague biological objectives into executable search conditions.
• Candidate tiering: Rank-tagged preliminary hits according to evolutionary source, domain integrity, and expressibility predictions, helping the team converge from vast results to 44 candidate proteins for experimental validation.
• Literature cross-checking: Rapidly aggregated existing annotations for relevant proteins, flagging which candidates had prior biochemical evidence and which were entirely uncharacterized, thus informing validation priorities.

Ultimately, experimental results confirmed that candidate L05, derived from a bacteriophage, successfully assembles into a vault-like particle. Notably, the AI’s role here was to accelerate convergence—it did not decide which protein would work; rather, it organized the evidence required for making that judgment more rapidly and comprehensively.

2. Experimental Design and Construct Assembly

At the early stages of the project, the model also supported multiple scheme design tasks:

• DNA construct design: Generated multiple expression construct proposals based on codon preference, domain boundaries, and purification tag placement, allowing researchers to evaluate distinct alternatives.
• Purification optimization: Traditional vault purification typically requires ~3 days and involves multi-step centrifugation and chromatography. With AI assistance, the team revisited the critical bottlenecks in the purification path and compressed the total time to 90 minutes. The significance of this improvement is practical: the faster the purification, the lower the risk of protein degradation or aggregation, directly increasing downstream structural analysis success rates.
• Validation pipeline planning: For all 44 candidates, the model helped design a rapid trial-and-error validation workflow, defining pass/fail criteria and alternative strategies for each step, thus preventing time wasted on low-yield paths.

3. Structure Determination and Result Interpretation

After the L05 protein was recombinantly expressed in E. coli, the team determined its three-dimensional structure via cryo-EM. Within the cryo-EM image processing chain—particle picking, 2D classification, 3D reconstruction, and model building—the language model assisted in writing and debugging processing scripts, explaining parameter selections, interpreting classification outputs, and providing initial assessments of structural models. Thanks to the efficiency gains achieved in earlier stages, the first structural model was completed within 34 hours.

Results: An Unprecedented Prokaryotic Vault Structure

The structural analysis yielded several key findings:

• Assembly scale: The vault-like particle consists of 96 identical copies of the same protein, with a total molecular weight of approximately 9.5 MDa, making it one of the largest protein shells reported in prokaryotes.
• Comparison with human vaults: Human vaults are composed of 78 major vault protein copies. Clear differences were observed in overall dimensions, internal cavity volume, and the arrangement of repeating protein subunits, providing direct evidence for understanding structural changes in vault evolution.
• A new starting point for functional studies: With a stable prokaryotic expression system and atomic-resolution structure in hand, researchers can now conduct follow-up experiments—including mutation studies, ligand loading assays, and assembly dynamics investigations—to gradually uncover the true biological function of vaults.

Methodological Takeaways: A General Model for AI-Assisted Research

Beyond this specific case, a reusable collaboration framework can be extracted:

1. Data layer: AI assists in pre-screening and structuring large-scale public data, resolving the issue of information overload.
2. Design layer: AI translates research objectives into concrete experimental schemes (constructs, pipelines, controls), while researchers review and make final decisions—forming a “model drafts—human approves” closed loop.
3. Execution layer: AI supports script development and workflow optimization, shortening the iteration cycle from hypothesis to validation.
4. Interpretation layer: AI helps bridge structural, biochemical, and imaging data, lowering cross-disciplinary barriers to knowledge integration.

The value of this model extends well beyond structural biology; it is equally applicable to target discovery, protein engineering, and drug screening—any domain requiring extensive experimental iteration. When experimental cycles shrink from days to hours, the very organization of scientific work shifts: researchers can dedicate more effort to formulating questions and interpreting mechanisms, while delegating procedural tasks such as retrieval, design, and troubleshooting to model-assisted collaboration.

From a broader perspective, this study demonstrates not a single technological breakthrough, but the leverage created by combining general reasoning capability, domain knowledge, and experimental verification. A structural conundrum pending for nearly four decades was systematically addressed through a deeply AI-integrated workflow. As this collaboration model matures, “AI-assisted fundamental science” is moving from proof of concept to daily practice.