AI-Accelerated Structural Analysis of Prokaryotic Vault Proteins

Background

Vault particles are among the largest protein complexes in the cell. Their mass is about three times that of a ribosome. Dozens of polypeptide chains assemble into a symmetric, hollow, barrel-shaped structure, and the interior is large enough to hold various biological macromolecules. Researchers first isolated vaults from mammalian tissue in the mid-1980s, but their exact function remained unknown for a long time. Knocking out the associated genes did not stop model organisms from developing normally, and common model organisms often lack homologs. As a result, research stalled.

Searching prokaryotes for homologous systems with simpler structures offered a way forward.

Large-Scale Sequence Mining

Finding prokaryotic homologs starts with large-scale sequence mining. With tens of billions of protein sequences available, comparing sequences one by one is highly inefficient. A family search strategy based on hidden Markov models solves this problem. Starting from a small set of known seed sequences, the method iteratively expands the homologous family.

This approach expanded the known families by dozens of times. The homologs span several bacterial and archaeal phyla, and related genes were also found in phage genomes. These genes often sit next to antiviral defense systems, which suggests that prokaryotic vault proteins may take part in immune defense. Sequence diversity analysis shows that prokaryotic vault proteins occur in less than 1% of species. They are rare but functionally conserved modules.

Candidate Protein Expression and Purification

Experimental validation of candidate proteins requires efficient gene synthesis and expression strategies. The team selected several dozen representative sequences from the expanded family, designed codon-optimized synthetic genes, and placed them under inducible promoters. After introduction into a prokaryotic expression host, about one-third of the candidates produced soluble, high-molecular-weight particles.

Traditional sucrose density gradient ultracentrifugation takes several days, which makes high-throughput screening impractical. Size-exclusion chromatography offers a faster alternative. Because giant particles cannot enter the pores of the resin, they elute first. This reduces purification time to a few hours. The method works especially well for samples with very different particle sizes. Combined with online UV detection, it allows real-time monitoring.

Cryo-EM Data Collection and Processing

Cryo-electron microscopy is central to obtaining three-dimensional structures. Purified particles are prepared by negative staining or cryo-preparation, and thousands of micrographs are collected on a high-resolution electron microscope. Steps such as motion correction, contrast transfer function estimation, and particle picking can run through automated workflows.

After two-dimensional classification selects high-quality particles, the team performs three-dimensional reconstruction and refinement. For highly symmetric barrel structures, applying the corresponding point group symmetry constraint significantly improves resolution. The entire process, from raw data to density map, can be shortened to tens of hours. Manual processing typically takes several weeks.

Structure-Guided Engineering

After a structure is solved, the protein can be engineered further. Truncating, inserting, or fusing functional modules can change the size of the barrel opening or the charge environment inside it. These changes could allow the container to carry small molecules or nucleic acids. Molecular dynamics simulations can predict the stability of engineered variants and guide follow-up experiments.

Together, these steps form a closed loop: computational screening, experimental validation, structural feedback, and functional engineering. This loop can reduce exploration that once took years to a matter of months.

Building a Reusable Bioinformatics Pipeline

The workflow can be packaged as a reusable bioinformatics pipeline with three modules:

• Sequence search module: Takes seed sequences as input and outputs the expanded family.
• Expression design module: Generates optimized nucleic acid sequences and primers.
• Data processing module: Takes raw cryo-EM images as input and outputs density maps.

The modules pass data to one another through standard file formats, so they can run in parallel on elastic cloud computing resources. Monitoring and logging services track task status in real time, which keeps large-scale screening traceable.

Conclusion

This work shows that combining machine learning-driven sequence analysis with traditional biochemistry and structural biology can address a long-standing structural question about giant protein complexes. As sequence databases keep growing and algorithms improve, similar strategies may extend to the study of other unknown protein complexes. This could speed up the translation of basic life science discoveries into applications.