AI Agents for Protein Structure: How a 40-Year Vault Mystery Was Solved

1. Unsolved for 40 Years: A Cellular Protein Barrel with No Known Purpose

A hollow barrel shaped like a cathedral

In 1986, a structural biology team isolated a rare, large particle from the cytoplasm of mammalian cells. Under the electron microscope, its cross-section showed an arc reminiscent of a cathedral dome, and it was named the “vault.” At roughly three times the mass of a ribosome, it is one of the largest known non-membranous protein complexes in the cell.

Structure known, function a mystery

In mammals, the vault is built mainly from 78 copies of a single major vault protein (MVP). The copies join end to end to form two barrel-shaped halves that cap a large hollow interior. The particle also incorporates small amounts of other proteins and a short piece of non-coding RNA. Decades later, however, what it actually does inside the cell remains unsettled:

• Knockouts show no obvious phenotype. Mice engineered to lack the main vault gene still develop normally, with no significant phenotypic change.
• Absent from common model organisms. Vaults cannot be found in workhorse experimental models such as yeast, Drosophila, and C. elegans.
• Findings are hard to reproduce. Evidence on whether vaults reconstituted in vitro can enter the nucleus, or whether they are linked to drug resistance, has been neither reproducible nor convincing.

At the heart of the mystery is a question: if the vault is not essential to the cell, why has evolution retained such a large, energy-consuming, and intricately assembled system? The research team’s answer is that the vault may be “silent” under ordinary culture conditions and may act only in specific stress scenarios. That suggests a new approach to finding its “true identity”: perhaps we should look for it in another corner of the biological world.

2. Doing AI-Assisted Science Right: From Answer-Giver to Collaborator

In this study, AI agents took on a large amount of concrete work: scanning sequence databases, predicting structures, designing expression vectors, optimizing purification protocols, and processing cryo-EM images. Their value, though, went beyond “doing the calculations for people.” Acting as collaborators, they changed how the research itself was organized.

What LLMs do well: search, combine, and rank across massive datasets

Database searches, initial structure screening, vector element combinations, and chromatography parameter design are all essentially search problems under combinatorial explosion. Which proteins are worth synthesizing, and which purification conditions are most likely to work? A model can propose large numbers of candidates and rank them quickly, something that is hard to achieve by checking options one by one by hand.

Where AI falls short: judging “what is real”

Models are good at producing something “that looks reasonable.” They do not naturally understand that something “must be proven experimentally.” Several key details from this study illustrate the point:

• Pushing back on a purification plan. The human researchers planned to switch to size-exclusion chromatography (SEC), and the AI objected. It pointed out that small molecules can get trapped in the pores of an SEC column, and the enormous, hollow vault is prone to shear, clogging, or dissociation inside the column. Instead, it suggested first looking for ways to speed up the centrifugation workflow substantially. In the end, the method that cut the process from 3 days to 90 minutes was density-based centrifugation, not chromatography. A final SEC step confirmed the purity and homogeneity of the product.
• Spotting gaps unprompted. It identified and corrected flaws in an experimental protocol drafted by humans.
• Reading negative results. After an expression attempt failed, the AI could read the grayscale distribution on SDS-PAGE gels, work out whether the problem lay in expression level or solubility, and adjust its strategy accordingly.

Division of labor: AI generates plans, humans set goals and take responsibility

Agents are not all-powerful. They can be “overly aggressive when modifying experimental protocols,” and the researchers had to “pull them back from dangerous pitfalls.” Efficiency gains are possible precisely because humans hold the boundaries around the goals and safety.

3. Inside the Pipeline: Where AI Stepped In at Every Stage

3.1 Sequence mining: finding a needle in 15 billion sequences

The problem: Mammalian vaults are assembled from 78 copies of the same MVP. The team had previously tried to “build” vault-like particles in vitro from recombinant MVP, but could not reproduce the native particles. What made the search harder was that this class of proteins is highly repetitive in sequence, so conventional homology searches tend to miss many of them.

The approach: The team used AI agents to search about 15 billion protein sequences exhaustively. They started from sequence similarity, combined it with features of repeated domains, and added structure prediction to identify genuine homologs.

The results:

• Homologous proteins were found across 29 bacterial phyla and 4 archaeal phyla. Large numbers of related genes were even hiding in the genomes of phages that infect bacteria.
• The known vault protein family grew from 41 to 1,507 members, an increase of about 36-fold.
• These proteins are quite rare in the bacterial domain, appearing in fewer than 1% of species.
• They often sit next to antiviral defense genes. They may well be “special equipment” bacteria use to fend off viruses.

Engineering considerations:

• Resolution control: Place a “coarse screen” between the large-scale search and fine-grained structure prediction. First scan the entire database quickly using profile models (such as profile HMMs) or sequence similarity, then run structure prediction only on the matched fragments.
• Rank the results: Score and rank hits by sequence completeness, number of repeat units, transmembrane hydrophobic regions, and predicted structure confidence. Repeatedly seeing low-scoring hits often means the profile model is too broad and needs to be tightened.
• Neighborhood analysis: Survey the gene annotations within a defined range on either side of each homolog. Significant enrichment of defense-related genes is useful supporting evidence for inferring function. Of course, genomic adjacency is only a clue. It does not amount to proof of function.

3.2 Molecular design and synthesis: from 1,507 candidates to 44 experiments

The approach: From the 1,507 homologous sequences, the team selected 44 candidate proteins. The AI agents handled everything from synthetic DNA design and expression control configuration to drafting the order.

Key points:

• Representative sampling: Candidates were chosen in stratified fashion by phylum and phylogenetic branch, so that the results would represent the family as a whole rather than the incidental traits of a single branch.
• Engineering design: Codons were optimized to suit the host. Promoters, ribosome binding sites, and terminators were chosen to balance expression level against soluble fraction. Affinity purification tags, solubility tags, or protease cleavage sites could also be added.
• Batch management: Like controlling software versions, each candidate was bound to a versioned build record (sequence source, vector map, tagging strategy), so that any later anomaly could be traced back to a specific design.

3.3 Expression and solubility screening: 15 of 44 candidates made the cut

After transformation into E. coli, 15 of the candidate sequences produced soluble protein. But the real test was only beginning.

The vault is not a single protein molecule. It is a hollow particle organized from dozens or even hundreds of copies assembled in an orderly fashion, and it often needs to bind RNA to remain stable. A protein dissolving in solution does not mean it can self-assemble into a complete particle. Soluble expression is only an “entry ticket.”

Engineering approach:

• Log the expression level and soluble fraction of each construct round by round, building a structured screening matrix.
• Distinguish two kinds of failure. If a construct shows no expression at all, the problem usually lies in codons, transcription, or translation. If it expresses but forms inclusion bodies, the problem usually lies in folding and assembly.
• Run a quick morphological check on soluble samples using negative-stain EM first. Only move on to cryo-EM once regular barrel-shaped particles are visible. This avoids a lot of wasted data collection.

3.4 Purification: from 3 days to 90 minutes

Why traditional methods are slow: Conventional vault isolation relies on sucrose density gradient ultracentrifugation. Target fractions are collected according to where the particles sediment within the gradient. The whole process takes about 3 days, with low throughput and large batch-to-batch variation. That could never support batch screening of dozens of candidates.

The AI’s contribution:

1. It vetoed the SEC route (for the reasons described in Part 2).
2. It shifted to optimizing the centrifugation protocol, shortening each individual run and increasing daily throughput.
3. Combined with the gel electrophoresis results, it adjusted the strategy round by round for constructs that failed to express.

Why the speed-up matters: In AI-driven science, the bottleneck is often not “generating candidates” but “validating candidates.” A 3-day purification run meant that at most two or three samples could be validated per week. Once it was compressed to 90 minutes, several rounds of the “design-build-test-learn” cycle could be completed within a single day. The value of an agent rapidly producing plans can only be realized if the validation end has enough throughput.

3.5 Cryo-EM: a 3D structure in 34 hours

Why the traditional workflow is slow: A cryo-EM structure determination typically goes through sample vitrification, data collection, particle picking, 2D classification, initial model generation, 3D classification, refinement, and model building and validation. Many steps rely heavily on experience and require repeated manual parameter tuning. Vault particles are large, have a hollow interior, and are sensitive to the electron beam, which adds to the difficulty. In the past, solving a 3D structure could take weeks or even months.

What AI does well: Tasks such as particle picking, 2D classification, symmetry selection, mask tuning, and parameter optimization have clear rules and dense feedback. They can also be evaluated offline against gold-standard datasets, which makes them well suited to repeated iteration by agents.

Final timing: Just 34 hours from data acquisition to a completed 3D structure.

A reusable, efficient workflow:

1. Dataset setup and preprocessing: Import the images, perform motion correction, and estimate the contrast transfer function (CTF).
2. Particle selection: Run a quick initial selection with a lightweight model, then gradually tighten the selection criteria.
3. 2D classification: Confirm that particle morphology is uniform, and remove fragments and aggregates.
4. Initial model and 3D classification: Identify different conformations and reveal differences between “empty-shell” and “cargo-containing” particles.
5. Symmetry-constrained refinement: Barrel-shaped vaults have a high degree of internal repeat symmetry. Applying this constraint and performing symmetry averaging can significantly improve resolution.
6. Local resolution assessment and model building: Use half-map independent validation (for example, Fourier shell correlation between independently refined half-maps) to avoid spuriously high resolution caused by “overfitting to noise.” Then fit the atomic model and check its geometric plausibility.
7. Pre-release review: Optionally bring in a third-party validation service for an independent check.

The other side of efficiency: Cryo-EM datasets can easily reach hundreds of gigabytes or even terabytes. Raw data volume, read/write speed for intermediate results, and GPU scheduling and queue times can all have a real effect on whether the “34 hours” is achievable. Planning compute resources deserves as much attention as optimizing the algorithms.

4. Five Principles for Deploying AI Research Agents

Principle 1: Break big science questions into small, automatically executable tasks

“Cracking a 40-year mystery” cannot be executed directly. “Find homologous proteins among 15 billion sequences” and “design synthesis plans for 44 candidates” can. The clearer the task boundaries, the higher the degree of automation. Use a structured description to constrain the agent:

{
  "task": "Heterologous expression and purification of candidate proteins",
  "inputs": {
    "constructs": "44 codon-optimized synthetic gene sequences (version v1)",
    "host": "E. coli expression strain",
    "equipment": ["Automated liquid handling workstation", "Centrifuge", "Analytical SEC system"]
  },
  "constraints": {
    "biosafety_level": "BSL-1",
    "max_ucl_min_per_batch": 90
  },
  "evaluation": {
    "metrics": ["Soluble protein yield", "Particle homogeneity", "Negative-stain EM morphology"],
    "pass_criteria": "Soluble yield > X mg/L and well-formed particles visible by EM"
  },
  "on_failure": "Output a failure attribution report and generate the next round of candidates; do not change safety boundaries on its own"
}

Writing constraints and acceptance criteria into a configuration is more reliable than writing them into a chat conversation. It also makes auditing and reproduction easier.

Principle 2: Validate in layers; never let the model grade its own work

• Offline evaluation: Measure the agent’s precision and recall against datasets with known ground-truth answers.
• Independent validation: Validate against data that were not used for training or screening.
• Experimental loop: Treat wet-lab results as the only judge.

Be especially wary of “leakage bias” caused by overlap between sequence databases and the training sets of structure prediction models. Such overlap can make evaluation metrics look artificially high.

Principle 3: Write the boundaries of human-AI collaboration into the workflow

• AI is responsible for: Generating candidates, executing standardized subtasks, interpreting structured data, and proposing hypotheses for improvement.
• Humans are responsible for: Defining problems and acceptance criteria, reviewing high-risk operations, and course-correcting when the AI goes off track.

Principle 4: Treat experimental data as first-class citizens

Sequences, build records, purification parameters, gel electrophoresis images, and cryo-EM data should be catalogued uniformly and stored in a data repository with version control. Only when every conclusion can be traced back to the raw data can each of the AI’s “reasoning” steps be reviewed.

Principle 5: Run the AI research pipeline on general-purpose cloud capabilities

These workflows can usually be mapped onto a few general capabilities:

• Containerized compute platform: Packages images for structure prediction, image classification, 3D reconstruction, and other tasks, and schedules them through a job queue.
• GPU compute: Used for structure prediction inference, neural-network-based particle picking, and 3D reconstruction.
• Object storage: Holds terabyte-scale raw cryo-EM data and intermediate results.
• Task orchestration service: Links the “design → build → test → learn” loop together.
• Logging, monitoring, and alerting: Tracks task status, failed-job retries, and resource usage.

On Baidu AI Cloud, you can use serverless services such as Function Compute, which are event-driven and billed per invocation. For example, an object storage event can trigger a 3D reconstruction job, so there is no need to keep GPU instances running around the clock for occasional tasks.

# Example: Function Compute handler that triggers 3D reconstruction from object storage events
import json
import os
import subprocess
def handler(event, context):
    for record in event:                       # A single event can contain multiple records
        if not record["name"].endswith(".mrc"):  # Only process image files
            continue
        # Bucket name and object path come only from the trusted event source; do not concatenate external input directly into a command
        obj_path = os.path.join("/data/raw", record["name"])
        out_path = obj_path.replace("/data/raw", "/data/results")
        # Call the reconstruction program with an argument array to avoid shell injection; the program image is pre-packaged in the container environment
        subprocess.run(["reconstruct", "--input", obj_path, "--output", out_path], check=True)
        # Afterwards, write the job status back to a task table, or emit a message to trigger the quality-control workflow
    return {"status": "ok"}

5. Risks and Boundaries: AI Science Is Not “One Click to Results”

1. Hallucination risk: Models can produce structures, pathways, or parameters that seem plausible but are wrong. Experiments and independent validation must be the final authority.
2. Reproducibility: Research results must be reproducible. That requires fixing the model version, parameters, random seeds, and data snapshots. It also requires recording which steps were performed by AI, so that conclusions remain traceable.
3. Cost and quotas: Large-scale sequence searches and all-atom structure prediction are computationally expensive. Set budget limits, task priorities, and failure-retry policies.
4. Biosafety and compliance: Mining genes from environmental samples and expressing them synthetically must go through synthetic DNA sequence screening and laboratory biosafety review. When human or environmental samples are involved, data and privacy regulations must also be followed. An AI’s boldness must not turn into crossing red lines.
5. Intellectual property: Whether AI-generated designs can be patented, and how rights should be assigned, is still debated. Conduct a dedicated assessment before commercialization.

6. Conclusion: Beyond Efficiency, See the Bottlenecks Clearly

The vault breakthrough offers two lessons. First, AI agents turned the serial mode, in which “a single structure determination could take months,” into a parallel mode of “batch generation, rapid validation, and rolling iteration,” shrinking the overall timeline to under 4 months. Second, they moved a functional puzzle that had remained open for 40 years to a stage where there are “candidate functional hypotheses that can be tested experimentally.”

The real bottlenecks sit at two ends. At the front end lies “proposing questions worth validating.” At the back end lies “the throughput of experimental validation.” AI can dramatically compress the time spent on information processing and plan generation in the middle. However, it cannot replace the experimental verdict, nor can it decide for humans “what questions are worth studying.”

Action recommendations for R&D teams:

• Start with a standardized workflow that gives fast feedback, such as purification optimization, image classification, or parameter tuning.
• Build a version-controlled data foundation and define clear acceptance metrics.
• Keep human review checkpoints in place.
• Bring safety, compliance, and IP assessments forward to the start of the project.