AI-Assisted Structural Elucidation of Prokaryotic Vault Particles: A Closed-Loop Practice from Sequence Mining to Cryo-EM Modeling

Vault particles are large, vesicular protein complexes that have long attracted attention in structural and cell biology. Widely present in human cells since their discovery in the 1980s, their exact biological functions remain insufficiently explained. While researchers have proposed their involvement in intracellular transport and cellular protection, definitive conclusions are lacking. Previously, sequences evolutionarily related to vault proteins were found in prokaryotes and related viruses, but direct evidence of their assembly into complete vault structures was missing. A recent public study, integrating AI models into the research pipeline, successfully resolved the complete 3D structure of a prokaryote-related vault for the first time, offering new insights into the assembly mechanisms and structural diversity of such protein complexes.

Unlike merely “letting the model generate conclusions,” the value of this work lies in demonstrating how AI can be embedded into a real scientific research loop: connecting large-scale data retrieval, experimental protocol design, protein purification, and cryo-EM image analysis, while participating in experimental iteration under the guidance of researchers. This article dissects the approach, key phases, outcome characteristics, and engineering implications of this case from a technical perspective, and discusses the validation boundaries necessary when AI participates in wet lab experiments.

1. Background: Challenges in Vault Structure Elucidation

The difficulty in resolving vault structures first stems from a “large candidate space.” Finding proteins that might assemble into vaults from massive predicted protein sequences requires comprehensive judgment using evolutionary correlation, structural domains, repetitive units, and expression feasibility. Secondly, the “experimental pipeline is long”: candidate proteins must undergo gene construction, expression, purification, sample quality assessment, cryo-EM data acquisition, 3D reconstruction, and model refinement. Failure at any stage can prolong the cycle. Thirdly, “structural verification requirements are high”: computational predictions can narrow the search scope, but the final structure must be supported by experimental data.

The choice to start with bacteriophages was pragmatic. Bacteriophages are viruses that infect bacteria, and their proteins need bacterial hosts for expression. Utilizing the rapid growth rate, relatively short experimental cycles, and mature genetic manipulation of prokaryotic expression systems, the research team could establish a more iterable experimental framework to verify whether candidate proteins form the target structures in a shorter time.

2. Technical Pipeline: From Sequence Space to Structural Model

This case can be abstracted into a “computational screening—wet lab verification—structural elucidation—model iteration” pipeline:

1. Problem Definition and Sequence Retrieval: Define the goal of finding prokaryote-related candidate proteins capable of assembling into vaults, and retrieve and filter from a large-scale set of predicted protein sequences.
2. Candidate Ranking and Experimental Selection: Rank candidate proteins based on sequence features, domain composition, repetitive unit probability, and expression/purification feasibility to narrow the wet lab scope.
3. Construct Design and Expression/Purification: Design DNA constructs for candidate proteins, formulate expression and purification protocols, and assess sample homogeneity and stability experimentally.
4. Cryo-EM Data Acquisition and Modeling: Conduct cryo-EM data collection on qualified samples, followed by 3D reconstruction, model building, and refinement.
5. Result Review and Iteration: Cross-validate the structural models with biochemical data and sequence analysis results; return to construct or purification condition optimization if necessary.

In this pipeline, the AI model’s role is not to replace experiments but to reduce information integration costs, accelerate the generation of protocol drafts, and help researchers make verifiable choices among numerous possibilities more quickly.

3. Key Phases of AI Participation

3.1 Large-Scale Sequence Retrieval and Candidate Screening

The research team searched for targets among approximately 15 billion predicted protein sequences, selected 44 candidate proteins for experimental verification, and ultimately confirmed that the L05 protein from a bacteriophage could assemble into vault particles. This process demonstrates AI’s value in “large-space search”: facing a sequence volume far exceeding human reading capacity, models can assist in constructing search strategies, summarizing candidate features, and generating candidate lists, which researchers then review using biological priors.

It must be emphasized that candidate screening is not simply “scoring and ranking.” Factors like sequence similarity, domain boundaries, repetitive unit counts, potential expression difficulties, and purification tag positions can all affect subsequent experimental success rates. AI is suitable for multi-feature integration and protocol drafting, but the final selection still requires experimental data and expert judgment.

3.2 Construct and Experimental Protocol Design

In the early stages of the experiment, AI models can assist in searching protein data resources, designing DNA constructs, and helping researchers identify candidate protocols. The “design” here is more akin to protocol generation and constraint checking: for example, providing construct drafts based on target protein sequences, indicating potential tag positions, expression host compatibility, and purification steps for researchers to modify before wet lab work.

In this case, the AI-assisted protocol reduced the purification time of vault particles from approximately 3 days in traditional workflows to 90 minutes, with the first structural model completed in 34 hours. These metrics indicate that AI’s value lies not just in “computing fast,” but in organizing scattered knowledge, procedural steps, and parameter choices into executable protocols, thereby reducing trial and error.

3.3 Purification and Cryo-EM Image Analysis

Cryo-EM structure elucidation relies on high-quality samples and extensive image data processing. AI can provide assistance in image screening, particle picking, 2D classification, and 3D reconstruction parameter recommendations, but model outputs must be reviewed by experimental personnel. The reliability of structural models ultimately depends on sample quality, data resolution, and modeling standards, rather than the text or predictions generated by the model itself.

4. Result Interpretation: Structural Features of Prokaryotic Vaults

The researchers expressed the L05 protein in Escherichia coli and resolved its structure using cryo-EM. The results showed that the vault consists of 96 identical protein copies, with a total molecular weight of about 9.5 million Daltons, making it one of the larger protein shells reported in prokaryotes. In contrast, the human vault consists of 78 major vault protein copies, and the two differ in size, internal space, and protein repeat structures.

The significance of these results is twofold: on one hand, it provides direct evidence that prokaryote-related vaults can assemble into complete structures; on the other hand, it suggests that the vault family possesses structural diversity in copy number, shell size, and internal space. For subsequent research, such structural information can be used to propose assembly mechanism hypotheses, design mutation verification experiments, and provide referential backbone features for protein engineering.

5. Engineering Implications: How to Implement the AI + Wet Lab Closed Loop

From an engineering perspective, the most reusable aspect of this case is that “AI participates in process orchestration rather than point-to-point Q&A.” An abstract implementation framework is as follows:

Input: Large-scale candidate sequence set, research goals, experimental constraints
Steps:
1. Retrieval and Filtering: Narrow down the candidate set based on evolutionary correlation, structural domains, repetitive units, etc.
2. Candidate Ranking: Integrate expression feasibility, purification difficulty, and structural stability risks.
3. Protocol Generation: Output construct drafts, purification step drafts, and key quality control points.
4. Wet Lab Verification: Expression, purification, and quality assessment.
5. Structural Elucidation: Cryo-EM data acquisition, reconstruction, modeling, and refinement.
6. Review and Iteration: Feed back experimental results to refine screening rules and experimental parameters.
Output: Experimentally verified structural models and reusable workflow knowledge.

When implementing, focus on three points: First, establish a “human review + experimental verification” gate; any model recommendation must be reviewed by experts before entering wet lab work. Second, retain complete experimental records and version information to ensure traceability and reproducibility. Third, treat AI output as a source of hypotheses rather than factual conclusions, cross-validating with control experiments and independent data.

6. Boundaries and Risks: Verifiability First

AI participation in biological research still has clear boundaries. Models may be affected by training data biases, offering overly confident suggestions for rare protein families or specific assembly modes. Parameter choices in experimental protocols involve extensive implicit experience, and model-generated content must be verified item by item. Structural models must conform to cryo-EM data and physicochemical constraints. More importantly, AI cannot replace quality control and ethical compliance reviews in wet experiments.

Therefore, a reasonable positioning should be: AI is responsible for expanding search scopes, integrating knowledge, generating drafts, and accelerating analysis; researchers are responsible for problem definition, critical judgment, experimental execution, and result endorsement. Only by prioritizing “verifiability” can AI-assisted structural elucidation transition from a demonstrative case to a stable, reusable scientific research infrastructure.

Conclusion

This case demonstrates a realistic path for AI participation in fundamental biological research: narrowing down the candidate range from approximately 15 billion predicted sequences, selecting 44 candidate proteins for verification, confirming that L05 can assemble into vaults, and, with AI assistance, shortening the purification process to 90 minutes and completing the first structural model in 34 hours. Its core insight is not that “AI replaces scientists,” but that AI can connect data retrieval, experimental design, and result analysis, helping research teams form verifiable hypotheses faster. As workflows standardize and validation mechanisms improve, such capabilities are expected to extend further into experimental iteration phases in target research, protein engineering, and drug discovery.