Large-Scale Protein Homology Search: Discovering Distant Vault Protein Homologs in 15 Billion Predicted Sequences

Introduction: Turning “needle in a haystack” into an engineering pipeline

In molecular biology, can a large protein complex with a long-unknown structure and uncertain function have “distant relatives” at the sequence level? A biology research team at an AI company used AI-assisted tools to explore this question and identified clues within a massive collection of about 15 billion predicted protein sequences.

Their approach can be summarized in three steps:

1. Find clues: Use computational methods to search for remote homologs of the major vault protein in a database containing tens of billions of sequences.
2. Filter candidates: Rank candidate proteins according to multiple indicators and select the ones most worth experimental validation.
3. Validate conclusions: Express and purify candidate proteins, then use cryo-electron microscopy (cryo-EM) to reconstruct their three-dimensional structures.

This route ultimately led to the first three-dimensional structure of a vault-like particle encoded by a bacteriophage, a virus that infects bacteria. For bioinformatics engineers and structural biology researchers, the value of this work goes beyond the biological result. It also provides a reusable engineering method:

• How to perform high-sensitivity retrieval on an extremely large sequence library;
• How to control false positives;
• How to close the loop efficiently between computational experiments (“dry lab”) and biological experiments (“wet lab”).

This article follows this distant-homolog discovery chain, moving from biological background to computational methods, experimental validation, and engineering practice.

---

1. Vault particles: a large “barrel” that has puzzled structural biology for nearly four decades

1.1 Structure and naming: one of the largest protein complexes in cells

In 1986, two cell biologists discovered a distinctive barrel-like structure in multiple organelles of mammalian cells. They named it the vault particle after the architectural vault that connects walls and roofs.

• Shape: A hollow barrel assembled from two symmetric halves, resembling an arched architectural vault.
• Composition: The core component of mammalian vaults is a large structural protein, commonly abbreviated as major vault protein, or MVP. Each vault particle contains 78 copies, with 39 copies in each half. In addition to the core protein, vault particles contain smaller amounts of other proteins and RNA.
• Scale: MVP consists of more than 800 amino acids, making the vault one of the largest known protein complexes by volume in human cells.

Unlike many other large complexes, the vault has not yet been resolved into a complete high-resolution atomic structure. Without structural details, its function is difficult to explain mechanistically. This is the structural basis for the long-standing stagnation in vault research.

1.2 Functional hypotheses: many directions, insufficient evidence

Several major hypotheses have been proposed for vault function:

• Nucleocytoplasmic transport: Vaults are often located near nuclear pores and have been proposed to participate in transport of macromolecules between the nucleus and cytoplasm.
• Multidrug resistance: MVP expression is significantly elevated in many drug-resistant tumor cell lines, suggesting that vaults may help cells pump out drugs.
• Innate immunity: The size of pores on the vault wall is close to that of nucleotides, leading to the hypothesis that vaults may be involved in nucleic acid recognition or sequestration.

However, after decades of research, these hypotheses still lack conclusive genetic evidence. Two major obstacles remain:

• Vaults are present at low abundance in cells, making isolation and purification difficult.
• Researchers have long been unable to reconstitute complete vault particles in vitro.

Being “visible, hard to capture, and difficult to rebuild” has been a key obstacle to understanding vault function.

1.3 Broad distribution and the model-organism gap

Vaults are widely distributed across eukaryotes. They have been found in diverse groups, from slime molds—social amoebae often used as research models—to humans. This broad distribution suggests that vaults may have important functions, because structures that repeatedly appear over long evolutionary timescales are often retained by natural selection.

However, a contradiction arises: common model organisms such as yeast, nematodes, and fruit flies generally lack vaults. Without vaults in these systems, researchers cannot directly test function through gene knockout or use mature genetic systems to dissect assembly. This created a research deadlock.

This deadlock also points to a way out: researchers must move beyond a small number of model organisms and search for study subjects across a broader tree of life.

---

2. Why search for distant relatives in 15 billion sequences?

2.1 The basic logic of homology search

Homology search is one of the most fundamental inference tools in bioinformatics. Its logic is:

1. Sequence homology implies common ancestry.
2. During evolution, sequences accumulate mutations, while protein three-dimensional structures are relatively conserved.
3. Therefore, even if two proteins have very low sequence similarity, finding a distant relative can provide structural clues for inferring the target protein’s folding, domain composition, and potential function.

A useful example is the use of a “decoy protein.” Researchers used the sequence of a bacterial longevity protein, a widely distributed protease, as a clue and found a distant relative in an archaeon. This led to the structure of the archaeal protease. The finding also improved the engineered production of a bovine protease commonly used in cell culture. This shows that finding distant relatives at the sequence level is often the starting point for opening the door to structure and function.

2.2 Large-scale sequencing gives distant homology search new leverage

Over the past decade, sequencing costs have continued to fall, and high-throughput sequencing has become increasingly common. Researchers have performed large-scale metagenomic sequencing on environmental samples such as soil, ocean, and human gut, combined with viral sequencing and gene prediction. This has produced an astonishing number of predicted protein sequences. The target database in this search contained about 15 billion sequences.

• A sense of scale: If each phone stored 2,000 songs, comparing these sequences to songs would correspond to the storage capacity of about 7.5 billion phones.
• Search scale: The sequence database contained about 15 billion predicted proteins, far exceeding any previous homology search for vaults.

Conclusion 1: The “sea” faced by traditional homology search has expanded to an unprecedented scale. The inability to find distant relatives in the past may simply have been because the sea was not yet large enough.

2.3 What makes distant homology detection difficult?

As evolutionary distance increases, sequence differences become larger. When the identity between two sequences falls below about 25%, sequence similarity may no longer distinguish true homology from chance similarity. This region is commonly called the “twilight zone.” More sensitive methods are needed:

• Profile-based search: Multiple homologous sequences are aligned into a position-specific scoring model, such as PSI-BLAST. Unknown positions are determined by “amino acids commonly observed in this column” rather than by a single sequence, giving higher tolerance to mutations.
• Hidden Markov models: PSI-BLAST is limited in capturing insertions and deletions. Hidden Markov models, or HMMs, further model the probability of deletions and insertions at each position, offering higher sensitivity.
• Combining structure and profiles: Tools such as HHsearch first perform sequence alignment, then compare hidden Markov profiles with hidden Markov profiles, also known as HMM-HMM comparison. This is more sensitive than sequence-level search.

Conclusion 2: The more distant the sequence, the more important high-sensitivity algorithms become. At the same time, the larger the dataset, the more engineering effort is needed to run high-sensitivity methods at acceptable cost.

---

3. Computational pipeline: fishing for distant homologs in 15 billion sequences

This section does not rely on one specific result. Instead, it breaks down the technical points of ultra-large-scale distant homolog retrieval from a general methodological perspective. In one sentence: use layered filtering to control scale, iterative profile search to improve sensitivity, structure prediction as a backup, and strict false-positive control to maintain reliability.

3.1 Overall architecture: a five-stage funnel

15 billion sequences  ──① Deduplication and decontamination──▶ Independent sequence set
    │
    ▼  ② Fast vector/similarity prefilter (millisecond-level recall)
Million-scale candidates
    │
    ▼  ③ Iterative hidden Markov profile search (PSI-BLAST / Jackhmmer-like algorithms)
Thousand-scale homologous hits (E-value ≤ 1e-3)
    │
    ▼  ④ Structure prediction + three-dimensional structure comparison (backbone metrics such as TM-score)
Hundred-scale high-confidence candidates (domain composition + structural fold consistency)
    │
    ▼  ⑤ Diversity-based deduplication + candidate priority scoring
Dozens of experimental candidates
    │
    ▼
Expression and purification → Assembly identification → Cryo-EM reconstruction

3.2 Data governance: how to store and compute with tens of billions of sequences

• Deduplication clustering: Cluster sequences using a similarity threshold, such as 40%–50%, and keep only a representative sequence for each cluster. The major vault protein is a large protein with relatively few repeated sequences, so a threshold in the 30–50 range can provide good compression.
• Decontamination: Remove vector sequences, host sequences, adapter sequences, and low-complexity regions to avoid false hits.
• Storage strategy: Each sequence is only a few hundred bytes, so the raw data for tens of billions of sequences is on the order of several terabytes to tens of terabytes. Object storage is suitable for centralized management. Intermediate results needed for comparison, such as candidate lists, scores, and predicted structures, should also be archived with source and version information.

3.3 Search strategy: balancing speed, sensitivity, and coverage

Common starting points are summarized below:

Parameter	Common starting point	Role and trade-off
E-value threshold	1e-3	Smaller expected values give more reliable results; too strict a threshold may miss true distant relatives
Query coverage	≥30%–50%	When only part of the protein is compared, the whole-protein E-value may be underestimated, so coverage must also be considered
Number of iterations	2–5	More iterations improve sensitivity but increase the risk of error amplification
Maximum target sequences	Set according to library size	Prevents performance fluctuations and result truncation in very large target libraries

In practice, a common approach is “run multiple strategies in parallel and take the union of results”:

• Use a known single sequence as a seed for iterative retrieval.
• Use a domain-family profile database as a seed for hidden Markov retrieval.

The results from both strategies are merged and then rechecked at the structural level. A single sequence can lose signal because of mutations, while family profiles are closer to the “common features” that the major vault protein has retained during evolution.

3.4 Structure prediction: the second engine of distant homology retrieval

Protein structure prediction has advanced rapidly in recent years. Given an amino acid sequence, models can output three-dimensional coordinates and per-residue confidence metrics such as pLDDT. In distant homolog retrieval, predicted structures can play three roles:

1. Structure-level search: Use predicted structures to search experimentally solved structure databases, such as vault cryo-EM density maps and MVP structure models, for structurally similar proteins. A TM-score around 0.5 is often considered the boundary for “same fold,” while below 0.2 is essentially unrelated.
2. Domain composition comparison: The core region of the major vault protein contains about 10 tandem repeat domains, each folding into a “winged helix” module, along with a cap domain and a stalk domain. Whether the “domain fingerprint”—the number, type, and arrangement of domains—matches is important evidence for candidate selection.
3. Oligomeric assembly prediction: Vault function depends on ordered assembly of dozens to hundreds of copies. Predicting multimeric structures for candidates can help assess whether they are likely to assemble.

Full-scale structure prediction for a very large library is expensive. A more practical approach is to keep only high-confidence hits during sequence search, then batch-predict structures for those hits.

3.5 False-positive control: the step most likely to fail

As retrieval sensitivity increases, false positives also increase. Common control methods include:

• Random decoy testing: Generate random sequences with similar length and amino acid composition to the target protein, then estimate the probability of achieving the same score to assess the actual false-positive level.
• Shuffle testing: Randomly shuffle the residue order of the query sequence, rerun the search, and observe how many hits remain.
• Domain completeness verification: Confirm that a candidate contains an N-terminal cap region, tandem repeat region, and C-terminal stalk region, with a domain count close to that of known vaults.
• Phylogenetic deduplication: Keep only representative members from groups of closely related sequences to avoid counting the same discovery multiple times.

In a search across about 15 billion sequences, candidate vault proteins were found across 29 major bacterial groups and four archaeal groups. In the bacterial domain, where there are often dozens of phyla, vault proteins were detected in fewer than one-third of the groups. This indicates that candidate filtering was quite strict rather than accepting every barrel-shaped protein.

3.6 Candidate priority scoring (sample code)

import math
# Weights are illustrative; in real projects, they should be tuned based on data characteristics and experimental feedback.
def priority_score(hit: dict) -> float:
    """Score a single candidate protein; higher scores indicate higher priority for experimental validation."""
    evalue   = hit["evalue"]                        # E-value from homology search
    coverage = hit["query_coverage"]                # Fraction of query sequence covered by alignment
    plddt    = hit["mean_plddt"]                    # Mean confidence of predicted model
    tm_score = hit["tm_score"]                      # Structural alignment score against reference structure
    domains  = hit["core_domain_count"]             # Number of core repeat domains
    novel    = 1.0 if hit["phylum_representative"] else 0.0  # Whether it comes from a representative group
    evalue_term   = min(-math.log10(max(evalue, 1e-300)) / 20.0, 1.0)
    coverage_term = min(coverage / 0.8, 1.0)
    model_term    = min(plddt / 100.0, 1.0)
    fold_term     = min(tm_score / 0.8, 1.0)
    domain_term   = 1.0 if 5 <= domains <= 15 else 0.0
    return (0.25 * evalue_term + 0.15 * coverage_term + 0.10 * model_term
            + 0.25 * fold_term + 0.15 * domain_term + 0.10 * novel)
# Keep a small number of representatives per phylogenetic group to balance diversity and validation cost.
def select_candidates(hits: list[dict], per_phylum: int = 3) -> list[dict]:
    ranked = sorted(hits, key=priority_score, reverse=True)
    selected, phylum_count = [], {}
    for hit in ranked:
        phylum = hit["phylum"]
        if phylum_count.get(phylum, 0) >= per_phylum:
            continue
        selected.append(hit)
        phylum_count[phylum] = phylum_count.get(phylum, 0) + 1
    return selected

The core ideas of this scoring scheme are twofold:

• Balance sensitivity and reliability: E-value and coverage answer “does it look similar,” while structural comparison and domain fingerprints answer “is it really a true homolog.”
• Balance diversity and cost: Keeping only a small number of representatives per group covers different lineages while keeping the experimental workload manageable.

3.7 Engineering implementation: making large-scale retrieval more stable and cost-efficient

• Task orchestration: Split retrieval into stages such as prefiltering, iterative search, structure prediction, and scoring. Decouple stages with message queues, schedule asynchronously, and retry failed tasks automatically.
• Resource scheduling: Package computational environments in containers. Prioritize GPU resources for compute-intensive stages and use elastic scaling to control idle costs.
• Observability: Use logging, monitoring, and alerting services to track task progress, candidate counts at each stage, and failure rates. Trigger alerts promptly when anomalies occur.
• Reproducibility: Fix random seeds and fully record tool versions, parameters, and data versions so that any candidate can be traced back to the original sequence and run records.
• Result management: Store hit lists, scores, predicted structures, and other outputs in a unified result repository with version management.

3.8 Cloud practice: building a retrieval pipeline with general-purpose capabilities

Large-scale computational tasks of this type have two resource characteristics: high peak demand and clearly staged usage. This makes them suitable for elastic cloud resources. From an engineering perspective, a complete pipeline can be built from general-purpose cloud capabilities. For example, common capabilities available on Baidu AI Cloud include:

• Using object storage to centrally manage sequence libraries and intermediate products;
• Using elastic containers and GPU computing power for compute-intensive steps such as structure prediction;
• Using message queues to connect tasks across stages;
• Using logging and monitoring/alerting services to maintain runtime observability.

In research scenarios where cost is sensitive, “scaling up during peaks and releasing resources during troughs” is often more economical than maintaining a large self-built cluster over the long term.

3.9 What role does AI play: assistant, not a black box replacing experiments

The original report repeatedly mentions the use of AI tools during this process. Their role can be seen in four areas:

• Retrieval and comparison: Quickly screening massive sequence data and generating candidate lists for human review;
• Structure prediction: Providing three-dimensional models and confidence estimates for candidates without experimental structures;
• Candidate ranking: Assisting prioritization based on multi-dimensional information such as domain composition and cluster representativeness;
• Analysis and writing support: Helping researchers organize, summarize, and present results.

It must be emphasized that predicted structures are only “clues,” not facts. The study ultimately confirmed cryo-EM structures for four candidates: two from bacteria and two from bacteriophages. This ratio—4/15 for expression and 4/44 for the full pipeline—shows that computational screening cannot replace experimental validation. A strict closed-loop validation process is required.

---

4. Experimental validation: from gene to atomic model

4.1 Four gates: a typical funnel

44 candidates ──▶ 15 successfully expressed and purified (about 34%)
              ──▶  7 assembled into vault-like particles (about 47% of successful expression)
              ──▶  4 confirmed by cryo-EM structures (about 57% of successful assembly)
Overall success rate: 4 / 44 ≈ 9%

Each gate corresponds to a different biological threshold:

1. Expression gate: Whether genes from diverse sources—bacteria, archaea, and bacteriophages—can fold correctly and express solubly in a host.
2. Assembly gate: MVP is a highly ordered assembly that depends strongly on intermolecular interactions. Only proteins with complete assembly signals and correct folding can spontaneously form particles.
3. Imaging gate: Particle homogeneity, ice quality, and particle orientation distribution determine whether a resolvable structure can be obtained.

4.2 Cryo-EM: a mainstream method for visualizing nanoscale “barrels”

Cryo-electron microscopy is suitable for resolving large complexes that are difficult to crystallize. Its main workflow includes:

1. Rapidly freezing the sample in a thin layer of vitreous ice to preserve its native conformation.
2. Collecting micrographs under low-dose conditions.
3. Performing image correction and particle picking.
4. Removing impurities and degraded particles through two-dimensional classification, performing three-dimensional reconstruction, and using symmetry information to constrain and optimize the model.
5. Fitting amino acid sequences into the density map, building an atomic model, and refining it.

For highly symmetric barrel-like particles, symmetry constraints can significantly improve resolution. This is one reason why large complexes such as vault particles are well suited to cryo-EM.

4.3 L05: an “improved vault” from a bacteriophage

The researchers ultimately selected a phage-derived protein, designated L05 in the study, for detailed structural analysis. Bacteriophages are viruses that specifically infect bacteria. After invading a host, they hijack host machinery such as transcription and translation to replicate themselves. The L05 protein assembles into a complete particle composed of 96 copies of itself. Its geometric parameters are summarized below:

Parameter	L05 particle	Mammalian vault
Major protein copy number	96	78
Length	About 54 nm	Longer
Width	About 45 nm	—
Total molecular mass	About 9.5 MDa	—
Enclosed volume	About 25% more than mammalian vault	Baseline

The most intriguing feature is the enclosed volume. Although the L05 particle is shorter, it encloses about 25% more volume. A reasonable explanation is that it gains greater internal space by increasing its cross-section without increasing length. Using a barrel analogy: for a slender container, increasing wall height adds capacity; for a short, wide container, increasing width also adds capacity. This “short and wide” assembly mode suggests that vault-like particles have geometric plasticity and provides new clues for understanding vault assembly rules.

---

5. Scientific value: a new branch of the cellular “barrel” family

5.1 Evolutionary perspective: vault-like particles may be an ancient family

Vaults were previously thought to exist mainly in eukaryotes. Finding distant relatives in bacteria, archaea, and bacteriophages suggests that vault-like particles may belong to a broader and older protein family.

• Bacteria and archaea have been evolving for billions of years. A protein family that spans bacteria, archaea, eukaryotes, and even viruses usually implies either a long shared ancestry or widespread horizontal gene transfer.
• Comparing gene trees with species trees can help distinguish vertical inheritance from horizontal transfer, enabling more reliable reconstruction of evolutionary history.

5.2 Functional perspective: a new route out of the model-organism gap

• Simplified models: Bacteria and bacteriophages are genetically simple to manipulate, grow quickly, and are inexpensive to culture. They offer hope for systematically testing functional hypotheses such as “transport” and “immunity” through gene knockout, mutagenesis, and in vitro assembly experiments.
• Assembly mechanism: Searching for vault-like particles in bacteria and bacteriophages may help distinguish the core assembly signals required to maintain the entire particle from accessory components specific to eukaryotic cells, thereby locating key steps in assembly.
• Assembly factors: Eukaryotic vault assembly also depends on several helper proteins, such as a co-chaperone that binds MVP and assists its folding. It is also worth investigating whether bacterial homologs have corresponding “helpers.”

5.3 Application perspective: potential nanoscale containers

Vaults are hollow, self-assembling, and have pores in their walls, making them naturally suitable for loading and exchanging molecules. They are therefore considered candidate carriers for drug delivery, vaccine platforms, and nanoreactors. The bacterial and phage-derived particles resolved in this work have advantages: their sources are clear, their components are relatively simple, and they can be expressed in large quantities in conventional hosts. This makes them promising synthetic biology chassis. However, these remain research prospects, and their application value still requires further experimental validation.

---

6. Reusable engineering checklist for distant protein homolog discovery

Stage	Key actions	Common risks and countermeasures
Data preparation	Deduplication, decontamination, version management	Contaminating sequences cause false hits; lost intermediate results make reproduction impossible
Search strategy	Run multiple strategies in parallel, such as profile search plus structural comparison	A single strategy may miss candidates; overly strict thresholds may remove true distant relatives
Quality control	Decoy tests, domain verification, phylogenetic deduplication	Duplicate counting of homologous proteins overestimates the number of discoveries
Candidate ranking	Multi-indicator weighting plus group quotas	Candidates may be too concentrated, lowering experimental success rate
Compute scheduling	Staged elastic resources plus queue decoupling	Peak resources may remain idle; failed tasks may need reruns
Observability	Logs, metrics, and alerts	Large batch failures may go undetected
Experimental closure	Progressive filtering through expression, assembly, and electron microscopy	Drawing conclusions solely from predictions lengthens validation cycles
AI collaboration	Human review of AI-generated candidates and explanations	Blindly accepting model outputs ignores the evidence chain

From “searching for distant relatives” to an intelligent research platform: a general capability reference

AI platforms for scientific research can usually be abstracted into the following general capability layers. Specific products and interfaces vary by vendor; the following is for reference only:

• Data layer: Unified storage, version management, and access control for experimental data, sequence data, and model outputs;
• Compute layer: Elastic scheduling of heterogeneous CPU/GPU resources, job queues, and failure retries;
• Tool layer: Containerized packaging of common bioinformatics tools and structure prediction models, enabling standard workflows to be launched easily;
• Orchestration layer: Workflow engines that connect steps such as retrieval, prediction, ranking, and experiment registration, with support for resuming from checkpoints;
• Model service layer: Standard interfaces for serving trained models, making batch inference easier;
• Collaboration and governance layer: Electronic experiment records, reproducible results, traceable data, and audit and cost control capabilities.

---

7. Limitations and future directions

7.1 Current limitations

1. The “twilight zone” of retrieval: Highly divergent homologs, or homologs lacking key domains, may still be missed. Negative results do not prove nonexistence.
2. Candidate coverage: Different sample sources determine the coverage bias of the sequence library. Many microorganisms in environmental samples have still not been cultured or sequenced.
3. Structural limitations: Cryo-EM resolution is affected by particle homogeneity, conformational flexibility, and other factors. Some regions may remain difficult to model.
4. Undefined function: Finding distant relatives does not equal understanding function. The biological function of vaults still requires experimental validation.

7.2 Future directions

• Continue expanding the search space: Extend retrieval to more environmental samples and uncultured microorganisms.
• Introduce stronger structure-level search: Structure-based comparison using predicted structures may help discover distant relatives in the sequence “twilight zone.”
• Use protein language models to assist retrieval: Self-supervised pretrained vector representations of protein sequences can move similarity comparison from “literal matching” to a “semantic level,” potentially capturing remote relationships that traditional alignment misses.
• Build automated experimental closed loops: Combining automated workstations with active learning strategies may allow models to decide “who to validate next,” while experiments continuously feed results back into the model.
• Move structure prediction toward assemblies: Progress from single-chain prediction to multichain assembly prediction can help predict whether candidates can form vault-like particles before experiments, further improving candidate filtering efficiency.

---

Conclusion

The core difficulty in vault research has been the lack of a tractable research object. The value of this work lies in connecting computation and experiment into a real discovery chain:

• 15 billion sequences represent scale, bringing previously unreachable “distant relatives” into the search space.
• 29 bacterial groups and four archaeal groups represent breadth, showing that vault-like particles form an ancient family spanning the tree of life.
• 44 candidates, 15 successful expressions, 7 assemblies, and 4 structures represent rigor: every conclusion was experimentally validated.
• 96 copies, 9.5 MDa, and about 25% more enclosed volume represent new understanding: the geometry of vault-like particles is plastic.

For R&D teams, the significance of this case goes beyond vaults themselves. It shows that in the era of big data, life science discovery is shifting from “hypothesis-driven linear research” toward a cycle of “computation generating hypotheses and experiments validating hypotheses.” Those who can better connect large-scale computing, intelligent models, and experimental platforms—while maintaining reproducibility—are more likely to find the next “distant relative” in vast biological datasets.