AI Helps Crack Prokaryotic Vaults: A Breakthrough in a Forty-Year-Old Structural Mystery

A Protein Container That Baffled Biologists for Decades

Inside cells, proteins rarely work in isolation. Many essential life processes depend on enormous complexes assembled from multiple subunits—ribosomes, proteasomes, and various membrane machines. The vault is one of the more unusual examples: it is massive, hollow inside, and shaped like a container formed by two halves clasped together.

When early researchers isolated these particles from mammalian tissue, they observed prominent arch and vault-like features, hence the name. Unlike common globular proteins, mammalian vaults are built from large numbers of identical or near-identical protein subunits arranged symmetrically, with the upper and lower halves connected through specific interfaces, enclosing a large internal cavity.

The problem is that despite this structural complexity, the vault never exhibited a sufficiently clear physiological function. It is abundant in some animal cells, absent from many common model organisms, and even when researchers weakened the relevant genes genetically, experimental animals could still develop normally. This combination—highly complex structure, hard-to-observe function—made the vault a long-standing open question in structural biology.

Why the Search Shifted to Prokaryotes

A common strategy for studying complex protein assemblies is to start with organisms that are structurally simpler and genetically better characterized. Compared with mammals, bacteria and archaea have more compact genomes, easier-to-establish protein expression systems, and lower culture costs. If vault-like structures could be found in prokaryotes, researchers would have much more direct genetic and biochemical tools to study their assembly and biological function.

However, the relevant proteins in prokaryotes are not easy to find using conventional sequence searches, for several reasons:

1. Sequence divergence can be extreme. The same structural family may have undergone long evolutionary divergence across species, leaving very low amino acid sequence similarity. Simple sequence alignment easily misses distant homologs.

2. Giant complexes are often built from repeated subunits. A single subunit may not carry distinctive sequence features; what is truly recognizable may be the repeat architecture, interface residues, and overall fold.

3. Gene annotation is often incomplete. Some microbial genomes are of limited quality; relevant genes may be annotated as “proteins of unknown function,” and start-site predictions can even be wrong.

Researchers therefore need to combine multiple lines of evidence—sequence similarity, structure prediction, gene neighborhood, species distribution, and functional associations—to filter credible candidates from vast protein sequence databases.

How AI Joins the “Needle in a Haystack” Search

In projects like this, AI does not simply replace experimentalists. Instead, it handles much of the information organization, candidate ranking, and iterative planning.

First, the system can cluster and compare large-scale protein sequences to find potential distant homologs. Second, by predicting secondary structure, overall fold, and likely oligomerization interfaces, it can prioritize candidates with “assemblable” characteristics. Third, analyzing the gene neighborhoods of candidate proteins helps determine whether they co-occur with genes involved in antiviral defense, stress response, or other composite functions.

According to the research roadmap described, the team found that vault-like proteins are likely distributed across multiple bacterial and archaeal groups, and even appear in some phage-associated genomes. These proteins are not widespread across the microbial world, but they can appear in clusters in particular species and are often located near defense-related genes on the genome.

Such patterns do not prove function directly, but they offer an important clue: prokaryotic vaults may not be a universally present “basic organelle,” but rather a specialized structure that certain microbes deploy under specific ecological pressures. For example, they could participate in sequestering harmful molecules, organizing defensive responses, or serving as a temporary “molecular warehouse” during viral infection.

From Candidate Sequences to Experimentally Testable Proteins

Computational screening only produces a shortlist. The real challenge is whether these proteins can be expressed stably, purified, and assembled into complete particles.

A complete experimental workflow typically includes:

• Selecting candidate genes and checking sequence quality;
• Designing coding sequences suited to the expression host;
• Configuring expression control elements and necessary purification tags;
• Introducing the constructs into expression strains;
• Optimizing induction conditions for protein yield;
• Checking whether the protein enters the soluble fraction;
• Determining whether it forms a high-molecular-weight complex;
• Purifying and structurally characterizing the complex.

From multiple candidates, the researchers selected a set of sequences for experiments. Some produced soluble proteins in standard bacterial expression systems. This step sounds basic, but it already eliminated a large number of unsuitable candidates. Many giant proteins misfold, aggregate, or degrade during expression, and even when individual subunits are produced, they may not assemble into stable particles.

AI’s value at this stage lies mainly in rapid iteration of experimental design. For example, it can adjust expression temperature, induction duration, buffer composition, salt concentration, and tag position based on prior results, and organize each round’s outcomes into a traceable decision record. Ultimately, researchers still confirm the approach with experimental data, but the trial-and-error path can be shortened considerably.

Purification: Compressing a Three-Day Process into Hour-Scale Screening

The difficulty of purifying giant protein complexes is not just extracting the protein—it is preserving the native assembly state throughout the process.

The traditional method is density-gradient ultracentrifugation. Under high centrifugal force, the sample separates into layers by particle size and density, and researchers recover the target fraction from a specific position. This method separates well, but the workflow is long, typically involving sample preparation, gradient setup, extended centrifugation, fraction recovery, and downstream analysis—not suitable for processing many candidates at once.

Another common method is size-exclusion chromatography. The column is packed with porous media of controlled pore size: small molecules can enter the pores and therefore travel a longer path, while giant complexes are excluded and elute earlier. For vault-like particles, which are much larger than ordinary proteins, this method offers high screening efficiency.

But several key risks remain:

1. Particles may be sheared or dissociated. Excessive flow rates or pressure can damage the complex.

2. The buffer system may be unsuitable. Mismatched pH, salt concentration, or additives can destabilize the interfaces between subunits.

3. Empty shells and complete particles are hard to distinguish. Even with the correct elution peak position, further confirmation is needed that intact assemblies are present.

4. Aggregates can cause misjudgment. Aggregated proteins can also show large apparent sizes, so verification with electron microscopy, scattering, or gel analysis is essential.

Reportedly, after redesigning the workflow, purification steps that once took days were compressed to roughly an hour and a half. The significance is not just time saved—more importantly, researchers can now evaluate many more candidate proteins in parallel, shifting from “single-project validation” to “batch screening.”

Cryo-EM: Reconstructing 3D Structure from Particle Images

With stable samples in hand, the next step is structure determination. The core idea of cryo-electron microscopy is to rapidly freeze protein samples in vitreous ice close to their native state, collect large numbers of two-dimensional particle images from different orientations, and computationally reconstruct a 3D density map.

This step typically involves:

1. Sample preparation: loading the purified complex onto holey grids and vitrifying it by rapid freezing;
2. Data collection: acquiring large numbers of microscope images at cryogenic temperatures;
3. Particle picking: identifying target particles against complex backgrounds;
4. 2D classification: grouping particles with similar views and discarding damaged or contaminated images;
5. 3D reconstruction: computing a 3D density map from 2D projections at different orientations;
6. Model fitting: placing the protein sequence and structural models into the density map for interpretation;
7. Quality assessment: checking resolution, local clarity, symmetry, and model reliability.

The vault-like structure’s high symmetry is both an advantage and a challenge. Symmetry can improve reconstruction efficiency, but incorrectly imposed symmetry constraints can mask genuine asymmetric states. Researchers therefore need to run both symmetric and asymmetric reconstructions and compare results across processing approaches.

In the related work, 3D structure processing time was compressed to the tens-of-hours scale. The key was not a single algorithm suddenly “automating everything,” but AI handling repetitive tasks such as particle picking, parameter exploration, outlier identification, and model comparison—freeing researchers to focus on structural interpretation and experimental validation.

New Questions Raised by Artificial Expansion

Once researchers obtained the prokaryotic vault structure, they could go further and engineer it—for example, by modifying surface sites, connecting regions, or cavity entrances on subunits to explore whether the complex can carry different molecules or to alter its assembly stability.

But artificial expansion does not simply mean “making the protein bigger.” Structural engineering must simultaneously consider:

• Whether the overall subunit fold remains stable;
• Whether subunit interfaces still pair correctly;
• Whether changes in cavity size affect assembly dynamics;
• Whether the engineered particles are prone to aggregation;
• Whether the structure retains sufficient thermal and chemical stability;
• Whether the modification introduces new immunogenicity or biosafety risks.

Structural design should therefore follow the principle of “small iterative steps, verification at each layer”: modify a single site first, then check expression, assembly, purification, and structural changes—never declare success based on computational models alone.

What This Means for Technology Development

The most noteworthy aspect of this work is not any single protein structure, but the new kind of workflow that emerges when AI is combined with experimental science.

First, the research process shifts from linear to closed-loop. The traditional flow was “hypothesize—design experiment—wait for results—replan,” whereas AI can continuously organize results, generate candidate plans, and help compare alternative experimental paths.

Second, the boundary between computational screening and experimental validation is being redrawn. In the past, only a few candidates could enter the experimental stage; now, computational methods can expand the search space first, and automated experiments plus rapid purification can deliver high-throughput validation.

Third, the importance of data recording has risen sharply. Every failed expression, purification anomaly, or unstable reconstruction becomes valuable input for the next round of screening. Without standardized records, AI cannot learn or compare effectively.

Fourth, structural biology is moving from “solving one structure” toward “designing a class of structures.” When protein sequences, 3D structures, assembly mechanisms, and experimental results can be analyzed in a unified way, research goals extend beyond explaining nature’s existing molecular machines to their controlled engineering.

Staying Cautious: AI Cannot Replace Scientific Validation

Although AI can dramatically shorten search, design, and data processing times, it cannot by itself prove a structure’s biological function. Sequence correlation does not equal functional correlation; gene neighborhood association does not prove shared pathway membership; and structure predictions are no substitute for real experiments.

Follow-up studies must at least answer several questions: Do prokaryotic vaults participate in antiviral defense? How do their assembly timing and copy numbers change in the cell? Are specific binding molecules present inside the cavity? Do cells lacking the relevant genes show defects under particular environmental stresses? These questions require combined validation through genetics, cell biology, biochemical analysis, and structural studies.

Overall, the discovery of prokaryotic vaults demonstrates a broadly applicable technical path: first use AI to uncover hidden signals in an ultra-large sequence space, then close the experimental loop through expression, purification, and cryo-EM. It not only helps explain a structural puzzle that has persisted for decades, but also signals a future in which life science research increasingly relies on the coordinated model of “computational prediction—experimental validation—data feedback—structure design.”