How AI Solved a 40-Year Mystery Surrounding a Giant Protein Complex

1. A 40-Year Biological Mystery

In 1986, a research team at the University of California, Los Angeles, isolated an unusual particle-like structure from rat liver. Under an electron microscope, its interior showed curved shapes resembling the arches of a cathedral, leading researchers to name it the vault.

Several characteristics of this structure have puzzled biologists ever since:

• Remarkable size: A vault weighs roughly three times as much as a ribosome, making it one of the largest known protein complexes inside cells.
• Intricate architecture: A mammalian vault consists of 78 protein chains—39 on each side—that fold together to form a huge hollow container. A single cell can assemble tens of thousands of these “protein barrels.”
• Unclear function: Vaults have not been found in common model organisms such as yeast, fruit flies, or worms. Even mice whose vault-related genes were knocked out developed normally.

A complex that consumes substantial cellular resources yet has no clearly defined biological function became one of structural biology’s most persistent mysteries. Two major obstacles stood in the way of solving it: vault samples were difficult to obtain and purify, while structural analysis required extensive experimental and computational effort.

2. Looking for Answers in Simpler Organisms

Because progress in complex animals had been difficult, the research team turned to the simpler world of prokaryotes. At the same time, it brought a large language model into the research workflow—not merely for writing papers or code, but as an active collaborator throughout the process.

1. Sequence Mining: Finding a Needle in 15 Billion Proteins

The team conducted a systematic search across a database containing approximately 15 billion protein sequences. AI supported the search in several ways:

• Pattern recognition and hypothesis generation: Based on the structural characteristics of known vault proteins, the system inferred sequence patterns that might appear in distant homologs.
• Systematic screening: It expanded the search into remote evolutionary regions that conventional homology searches often miss.
• Result organization and prioritization: It ranked the large number of hits according to factors such as confidence and the likelihood of forming an assembled structure.

The results exceeded expectations. These proteins were found across 29 bacterial phyla and four archaeal phyla. Numerous homologous genes also appeared in bacteriophages, viruses that infect bacteria. The known vault protein family grew from 41 members to 1,507—an increase of 36 times.

The ecological distribution provided another important clue. These proteins are extremely rare in bacteria, appearing in fewer than 1% of species, and their genes are often located next to antiviral defense genes in the genome. This pattern suggests that vaults may serve as specialized equipment for bacterial antiviral defense, offering a promising direction for understanding their function.

2. Expression Validation: AI Handles the Experimental Engineering

Identifying candidate sequences was only the first step. The team selected 44 candidate proteins. AI then assisted with the practical engineering work, including synthetic DNA design, configuration of expression-control elements, and preparation of reagent purchasing lists.

This stage illustrates an important methodological point: the bottleneck in scientific research is often not thinking, but doing. In a conventional workflow, a researcher may spend several days reviewing literature, designing primers, comparing vectors, and calculating expression conditions. A language model can produce multiple candidate plans within minutes and explain the rationale behind each one, allowing human experts to focus on critical decisions.

After the candidate genes were introduced into Escherichia coli, 15 sequences successfully produced soluble proteins. A new challenge immediately emerged: how could these particles—hundreds of times larger than ordinary proteins—be cleanly separated from cell lysates?

3. Purification and Structural Analysis: Compressing the Timeline by Two Orders of Magnitude

1. AI-Accelerated Purification

Purifying giant protein complexes is notoriously difficult. They are large, prone to aggregation, and highly sensitive to buffer conditions. Traditionally, researchers need extensive trial and error to determine suitable chromatography conditions, and a complete optimization cycle can take several days.

With AI assistance, the process was reorganized into a more systematic workflow:

1. Condition prediction: Based on characteristics such as isoelectric point, molecular weight, and stability, AI recommended combinations of buffer pH, salt concentration, and chromatography media, substantially narrowing the experimental search space.
2. Experimental matrix design: Instead of testing conditions sequentially, the team ran parallel small-scale screens that covered a wider range of conditions with fewer experiments.
3. Real-time interpretation: AI rapidly analyzed chromatography peak profiles and electrophoresis results, allowing the next steps to be adjusted dynamically.

As a result, a purification experiment that previously required approximately three days was completed in 90 minutes—an efficiency improvement of several dozen times.

2. Intelligent Processing of Cryo-EM Data

Once a highly purified sample was available, the team moved to cryo-electron microscopy for structural analysis. The traditional challenge at this stage is data processing. Hundreds of thousands of particle images must pass through particle picking, two-dimensional classification, three-dimensional reconstruction, and model refinement. The process typically takes weeks or even months.

By using AI to coordinate and optimize the workflow, the team resolved the three-dimensional structure in approximately 34 hours. The research timeline was reduced from months to days. It also produced the first high-resolution three-dimensional structure of a vault from a prokaryotic organism, finally revealing the nature of a structure that had remained mysterious for four decades.

4. Reusable Lessons for AI-Driven Science

The value of this work extends beyond the analysis of a single structure. It provides a transferable engineering model for AI-driven scientific research:

• AI as a full-workflow collaborator: AI participated directly in hypothesis generation, sequence mining, experimental design, and data analysis rather than merely offering suggestions.
• A changing role for human experts: Scientists shifted from performing every task themselves to making decisions and validating hypotheses, while delegating repetitive work to the model.
• Prioritizing bottlenecks: The team first identified the most time-consuming stages—purification and data processing—and then applied AI specifically where it could have the greatest impact, instead of automating indiscriminately.
• A data-driven feedback loop: Each experimental result was quickly fed back into the model, creating an iterative cycle of prediction, validation, and correction.

For scientific and engineering fields more broadly, the lesson is clear: when large models are deeply integrated with domain expertise and automated experiments, the timescale of research can be redefined. Reducing a process from months to hours does more than save time—it shortens the distance between a hypothesis and a discovery.