AI Models Decode Prokaryotic Vault Proteins, Solving a 40-Year Structural Mystery

3 hr ago0 views

By searching roughly 15 billion protein sequences and accelerating protein purification and cryo-EM analysis, AI models helped identify prokaryotic vault proteins across bacteria, archaea, and bacteriophages. The work expanded the known vault protein family from 41 to 1,507 members and produced the first three-dimensional structure of a prokaryotic vault in just 34 hours.

1. The “Ghost Cathedral” That Challenged Biology for 40 Years

In 1986, a well-known research team isolated a mysterious intracellular particle from rat liver. Under an electron microscope, its interior displayed elegant curves resembling the vaulted ceiling of a cathedral. The structure was therefore named a vault.

Vaults are among the largest protein complexes found inside cells, with a mass approximately three times that of a ribosome. In mammals, a vault consists of 78 protein chains arranged as two symmetrical halves that form a large, hollow container.

Yet each human cell produces tens of thousands of these hollow particles, while their precise physiological function has remained unclear. Common model organisms, including yeast, fruit flies, and nematodes, do not contain vaults. Even when researchers knock out vault-related genes in mice, the animals can still develop normally.

This sophisticated “protein barrel” consequently remained a major biological mystery for four decades.

2. AI-Powered Discovery Across 15 Billion Protein Sequences

Because complex animals offered few clear clues, the research team turned to the simpler world of bacteria and archaea. Using advanced AI models, the researchers conducted a broad search across approximately 15 billion protein sequences.

This effort required substantial computational infrastructure. In practical engineering environments, processing biological sequence datasets at this scale typically depends on high-performance computing clusters and distributed storage systems, such as those provided by Baidu AI Cloud. The models did more than perform keyword-based matching. Through deep sequence analysis and structural prediction, they identified potential homologous genes that might otherwise remain hidden.

# Pseudocode: AI-assisted bulk search and filtering of vault protein homologs
def search_vault_proteins(sequence_db, ai_model):
    candidates = []
    for seq_chunk in sequence_db.stream(chunk_size=1000000):
        # Extract features and calculate similarity with an AI model
        embeddings = ai_model.extract_features(seq_chunk)
        # Perform preliminary filtering with structure prediction scores
        predictions = ai_model.predict_structure(embeddings)
        for seq, pred in zip(seq_chunk, predictions):
            if pred.score > THRESHOLD and pred.is_spherical:
                candidates.append(seq)
    return candidates

The search revealed that these proteins are distributed across 29 bacterial phyla and four archaeal phyla. They were also found in bacteriophages, viruses that infect bacteria.

The discovery expanded the known vault protein family from 41 members to 1,507—an increase of approximately 36 times. More importantly, these genes are extremely rare in bacteria: fewer than 1% of bacterial species possess them. They are also frequently located near antiviral defense genes, suggesting that vaults may serve as specialized bacterial defense structures.

3. Optimizing Protein Expression and Purification with SEC

After identifying the genetic candidates, the next challenge was to express and isolate the proteins. The team selected 44 candidate proteins, while AI models handled preparatory tasks such as synthetic DNA design and configuration of expression controls. After the sequences were introduced into Escherichia coli, 15 candidates successfully produced soluble proteins.

The researchers then had to separate particles that were hundreds of times larger than ordinary proteins. Traditional vault purification relies on ultracentrifugation through a sucrose solution. The complete process can take up to three days, making it unsuitable for high-throughput screening.

To overcome this bottleneck, the team used size-exclusion chromatography (SEC). In SEC, a sample passes through a column filled with porous beads. Small molecules enter the pores, taking a longer path and eluting later. Large vault particles cannot enter the pores, so they travel through the spaces between the beads and emerge first.

With AI-assisted modeling and optimization of hydrodynamic parameters and chromatography conditions, the purification process was reduced from three days to approximately 90 minutes. This dramatic improvement made large-scale protein characterization and functional validation possible.

4. Rapid Structural Reconstruction from Cryo-EM Data

Once highly purified protein samples were available, the final step was to determine their three-dimensional structures using cryogenic electron microscopy, or cryo-EM.

Cryo-EM generates vast numbers of two-dimensional projection images. Reconstructing a three-dimensional structure from these images requires complex computational processing, which traditionally can take weeks or even months.

In this study, AI models again played a central role. Deep learning algorithms were used to denoise electron microscopy images, identify particles, and classify different conformations, greatly reducing the need for manual intervention.

Supported by cloud-based GPU computing, the team reconstructed the first three-dimensional structure of a vault from a prokaryotic organism in just 34 hours.

5. Conclusion and Outlook

From searching for rare genes to optimizing purification and processing cryo-EM data, AI models participated in nearly every stage of this landmark biological investigation. The work not only addressed a 40-year structural mystery involving prokaryotic giant proteins, but also demonstrated the potential of AI for interdisciplinary life-science research.