AI Agents in Giant Protein Structure Determination: The Full Pipeline, From Homolog Search to 3D Reconstruction

A Structural Biology Problem That Sat Unsolved for Forty Years

In 1986, a research team isolated a previously undescribed ribonucleoprotein particle from rat liver. Under the electron microscope its outline was a barrel-shaped dome assembled from multiple arcs, resembling cathedral vaulting, so it was named the “vault.”

Its physical parameters make it one of the most conspicuous objects in the cell:

• Its mass is roughly three times that of the ribosome, making it one of the largest known protein complexes.
• The mammalian homolog consists of 78 polypeptide chains (two sets of 39) that pair off to enclose a sizeable, sealed interior cavity.
• A single mammalian cell can assemble tens of thousands of copies.

Conspicuous, yet it long had no answer. Three obstacles stacked together to create the field’s classic “high-input, low-output” bind.

First, classical genetics could not pin down its function. The three most common model organisms—yeast, Drosophila, and C. elegans—have no homologous gene at all, and knocking out the corresponding gene in mice still yields animals that develop normally. When a knockout produces no clear phenotype, the “delete-and-observe” route, the most effective path to functional inference, breaks down entirely.

Second, sample size was hard to scale. Natural sources are low in abundance, the purification workflow is long, and batch-to-batch variation is large; a single experiment cycle routinely runs for days. But structural biology is a statistical science—it needs dozens or hundreds of batches iterated repeatedly, and the per-batch cost alone decides whether a project is even viable.

Third, sequence homology was sparse. Members of this family are reasonably similar to one another within eukaryotes, but distant cross-species searches have yielded little for a long time. Without evolutionary coordinates, researchers cannot tell how ancient the structure is or which regions are functionally critical.

The empirical rule in structural biology is this: when a complex is “structurally clear but functionally unknown,” the breakthrough rarely comes from a more precise instrument. It comes from expanding the sampling space. That is precisely where this work found its opening.

The Breakthrough Idea: Move the Search From Eukaryotes to Prokaryotes

Since the genetic backgrounds of complex animals held no functional clues, the team turned its attention to prokaryotes—organisms that are structurally simpler, easier to manipulate genetically, and cheaper to express.

2.1 A Carpet Search Across an Enormous Sequence Space

Running a distant-homology search over a collection on the order of 15 billion protein sequences is fundamentally a large-scale computational task, not an interactive query. The key lies in a layered retrieval strategy:

1. Primary screening: sequence-level homology detection. Short sequence fragments perform an initial pass that quickly eliminates obviously irrelevant entries. This step favors recall—better to keep too much than too little.
2. Secondary screening: profile HMM alignment. Candidate clusters are turned into multiple sequence alignments and scored with position-specific models, surfacing distant homologs whose similarity falls below conventional thresholds but whose topology is consistent.
3. Tertiary screening: structural and fold consistency validation. Candidate sequences are modeled with protein structure prediction and compared by fold topology and the arrangement of key domains, filtering out false positives caused by coincidental sequence similarity.
4. Quaternary screening: genomic context analysis. The search checks a candidate gene’s genomic neighbors, whether it is co-transcribed, and whether it sits near mobile genetic elements.

On the results side, the search’s value shows up along three dimensions:

• A leap in scale. The known prokaryotic vault protein family grew from 41 members to 1,507—roughly a 36-fold expansion. Once a family has enough members, sequence conservation analysis, coevolution signal analysis, and phylogenetic reconstruction finally carry real statistical weight.
• Breadth of distribution. The hits span 29 bacterial phyla and 4 archaeal phyla, indicating that the structure is an ancient, repeatedly rediscovered solution on the prokaryotic tree of life rather than the accidental product of a single lineage.
• Functional clues. These genes are not widespread in bacteria—they appear in fewer than 1% of species—and they are often physically adjacent to antiviral defense gene clusters. The positional information makes a compelling case: this is not a basic metabolic component but more likely conditional equipment that comes into play only when the cell is infected by a bacteriophage.

That last point explains the earlier phenotype puzzle. If a system’s function is “response-type,” knocking it out under non-stress conditions will naturally reveal nothing. The prerequisite for functional inference is to first create the correct stress condition—not to knock the gene out over and over.

From Candidates to Samples: The Experimental Design Stage Gets Automated

With 1,507 candidates in hand, the next bottleneck was deciding “which one is worth doing experiments on.” Researchers moved 44 candidate proteins into the wet-lab workflow, and the synthetic-DNA sequence design, expression-switch configuration, and even the ordering plan were all generated by agents.

The engineering value of this stage is routinely underestimated. Every step of molecular cloning is a formalizable decision:

# Illustrative pseudocode: automated decision chain from candidate to expression construct
def design_construct(candidate_seq, host="E. coli"):
    # 1. Codon optimization: adapt to host preferences, avoid secondary structure and rare codons
    optimized = codon_optimize(candidate_seq, host=host,
                               avoid=["hairpin", "rare_codon", "repeat_region"])
    # 2. Tag and purification strategy: pick an affinity-tag combination by predicted pI/hydrophobicity
    tag = choose_tag(predicted_pI=estimate_pI(optimized),
                     predicted_hydrophobicity=estimate_hydrophobicity(optimized))
    # 3. Expression switch: promoter strength and inducer concentration tuned by expression difficulty
    promoter = "T7" if predict_soluble(optimized) > 0.5 else "tunable_mid_strength"
    inducer_conc = grid_search_range(low=0.05, high=1.0, unit="mM")
    # 4. Ordering plan under cost and throughput constraints
    order = batch_and_split(constructs=optimized, max_cost_per_batch=BUDGET)
    return Construct(backbone="pET-derived", insert=optimized,
                     tag=tag, promoter=promoter, inducer=inducer_conc, order=order)

Note that this is not “AI doing the experiments.” It converts a batch of design decisions that are clearly ruled, combinatorially explosive, and reliant on hard-won experience from manual judgment into batch-generatable plans. The value is twofold: the design space can be opened up across dozens of candidates without adding headcount, and every decision’s rationale is recorded explicitly, so that a failure can be traced precisely back to the step that caused it.

The efficiency change in the expression and purification stage is the most striking. By report, a purification workflow that once took about three days was compressed to roughly 90 minutes. A change of this magnitude rarely comes from a single trick; it comes from three things happening at once:

• Better candidate selection, which cuts the sunk cost of “finishing the work only to find the target never expressed at all.”
• Buffer formulations, chromatography gradients, and centrifugation parameters derived by regression on historical experimental data, replacing a great deal of trial and error.
• A small-volume, high-throughput parallel layout that lets all 44 candidates be processed in a single batch rather than queued serially.

The Structure Determination Pipeline: What Makes Up the 34 Hours, and Its Limits

Cryo-EM 3D reconstruction is a standard pipeline, and the parameter choice at every stage directly affects final resolution:

Stage	Main Work	Parameter Sensitivity
Motion correction	Frame-by-frame drift compensation, dose weighting	High
CTF estimation	Determine the location of imaging phase inversion	Extremely high—one mistake ruins everything
Particle picking	Cut target particles from micrographs	High; determines stack quality
2D classification	Remove anomalous-orientation and degraded particles	Medium
Initial model	Build from scratch to avoid model bias	Medium
3D classification & refinement	Separate conformational heterogeneity, improve resolution round by round	High
Local refinement	Handle flexible subunits separately	Medium
Model building & validation	Build the atomic model, assess geometric plausibility	Determines result credibility

Reports state that a 3D reconstruction that once took weeks or even months was compressed to 34 hours. This number must be interpreted correctly, or it becomes easy to mislead:

1. The 34 hours is compute time, not the project cycle. It measures “how long reconstruction takes once the data is ready,” and excludes sample preparation, batch screening, cryo-sample prep, and image acquisition. The full project still takes about four months.
2. The speedup comes mainly from compute orchestration, not from a new algorithmic principle. Every stage is a parallelizable batch task: particle picking can be split across multiple machines, 2D classification can run in parallel by batch, and the GPU load of 3D refinement can shift from serial queuing to elastic scheduling. Turning what was a serial chain into a directed acyclic graph (DAG) and auto-scaling when compute is tight is the real driver of the time compression.
3. It cannot be extrapolated into a universal performance promise. Giant symmetric complexes have large particles, high contrast, and relatively easy particle picking—they are inherently among the objects most friendly to compute optimization. Swap in a small-molecule, highly flexible, or low-abundance sample, and the same orchestration delivers a markedly smaller speedup.
4. Resolution must be backed by validation metrics. A reported resolution should come with a Fourier Shell Correlation (FSC) curve, a local-resolution distribution, model-to-map fit, and stereochemical plausibility checks. Without them, no matter how fast the reconstruction, it is just “a picture.”

The System Architecture Behind Agent-Assisted Research

Abstracted into engineering terms, the pipeline above divides into five layers. This layered design is the most useful reference for teams that want to reuse the paradigm.

┌──────────────────────────────────────────────────────┐
│ Decision layer: hypothesis generation / plan review / iteration decisions │
├──────────────────────────────────────────────────────┤
│ Orchestration layer: task decomposition · DAG scheduling · failure retry │
├──────────────────────────────────────────────────────┤
│ Tool layer: sequence retrieval · structure prediction · chromatography parameter recommendation │
│            cryo-EM data processing pipeline · model validation │
├──────────────────────────────────────────────────────┤
│ Data layer: sequence libraries · experiment records · raw image data · metadata │
├──────────────────────────────────────────────────────┤
│ Infrastructure: elastic compute · object storage · message queue │
└──────────────────────────────────────────────────────┘

The corresponding general-purpose way to host this on the cloud:

• Object storage holds the raw cryo-EM frames and intermediate products, which can reach the multi-TB scale for a single project; a lifecycle policy is needed to automatically tier cold data down.
• A container platform hosts heterogeneous images (software-stack dependencies differ sharply between stages) and elastically scales GPU nodes by queue depth.
• A message queue decouples “submitting a task” from “executing a task,” so that long-running jobs and interactive queries never block each other.
• Log service and monitoring/alerting record each step’s parameter inputs and outputs—the prerequisite for failure attribution and result reproduction.
• A metadata and experiment-record system links “which parameter set a given image used, which construct it corresponds to, which purification batch it came from” into a single traceable chain.

That last point matters most: the biggest risk of agents in research is not miscalculation but an untraceable process. Once a conclusion cannot be traced back to its specific inputs and parameters, the credibility of the entire body of work is compromised. On platforms such as Baidu AI Cloud, these capabilities are typically delivered as combinations of standard cloud products, so teams do not have to build the whole infrastructure from scratch.

From “Understanding” to “Engineering”: The Point of Artificial Capacity Expansion

Solving the structure is only the starting point. For a naturally formed hollow container like the vault, the direction with more applied value is engineering its cavity and opening size. The vault’s barrel opening is relatively narrow, so its native cargo is size-limited. By mutating the arrangement at subunit interfaces and widening the opening diameter, it becomes possible to load larger functional molecules—for example, therapeutic proteins, enzyme complexes, or antigen fragments.

This step is called “artificial capacity expansion engineering.” It shifts the research from structural description to engineering customization, pointing directly at three classes of application:

• Nanodelivery carriers. A protein shell protects the payload and shields it from some proteolytic degradation.
• Vaccine scaffolds. Antigens displayed at high density in the hollow interior or on the outer surface raise immunogenicity.
• Artificial compartmentalization. A multi-enzyme cascade is sealed inside a single cavity, raising local substrate concentration and pathway efficiency.

It must be stressed that these directions are still early-stage exploration. Engineering a protein–protein interface can easily break assembly, so any size adjustment must be validated against assembly integrity, thermodynamic stability, and in vivo pharmacokinetic behavior. Feasibility cannot be inferred from the structural model alone.

A Reusable Methodology and a Risk Checklist

From this practice, a set of general lessons can be drawn that is independent of the specific biological subject.

Reusable practices

1. Prioritize expanding the sampling space over deepening a single model. When “knockout shows no phenotype + sparse homology + expensive samples” appear together, continuing to optimize the existing system yields limited returns; moving to a genetically simpler species system usually reaches a conclusion faster.
2. Family scale is the precondition for analytical capability. 41 members cannot support a reliable conservation analysis; 1,507 can. The value of a computational discovery often takes the form of “making a class of analysis go from infeasible to feasible” rather than handing over an answer directly.
3. Use genomic context to fill in the functional hypothesis. Sequence answers “what it is”; neighboring genes answer “under what conditions it is useful.” Locating a gene next to a defense cluster delivers both a functional hypothesis and a lead on how to design the experiment.
4. Turn the pipeline into a DAG with elastic scheduling. This is the highest-ROI optimization, and it is almost entirely discipline-independent.
5. Keep the whole chain traceable. Every design decision, parameter set, and intermediate product must be traceable.

Risks to watch

• Sequence-library bias. The distribution in search results reflects the database’s sampling bias, not the true distribution in nature; “rare” may simply mean “not yet sequenced.”
• Structure-prediction confidence must be assessed locally. A high overall score does not mean the functional-site region is reliable; low-confidence regions at flexible loops and interfaces should be flagged separately and not used as a design basis.
• The wet lab cannot be bypassed. Computation narrows the search space; any conclusion still needs closed-loop validation through biochemical and functional experiments.
• Efficiency numbers must come with conditions. Discussing a speedup without the sample properties, compute scale, and degree of parallelism easily leads to conclusions that cannot be reproduced.
• Define the boundary of closed-loop automation. Which stages may the agent decide autonomously, and which require human confirmation (for example, animal handling, reagent-order budgets, and final adoption of a structural model), should be spelled out during process design.

Conclusion

The significance of this work does not lie in how eye-catching a given automation number is. It lies in demonstrating a complete paradigm shift: when a scientific problem’s bottleneck is “the sampling space is too small and the iteration cycle is too long,” computational intelligence can step directly into the experimental-design decision stage instead of merely serving as a data-analysis tool.

The takeaway for engineering teams is just as direct. To run this kind of chain, three things genuinely need to be built: an orchestratable task pipeline, traceable data and experiment records, and an elastically scalable compute base. With all three in place, the cycle from hypothesis to structure can drop from “months” to “days.” Miss any one of them, and the speedup fails exactly when it is needed most.