AlphaFind v2: Similarity Search in AlphaFold DB and TED Domains across Structural Contexts

AlphaFind v2 is a scalable application that enables rapid, biologically relevant structural similarity searches across the AlphaFold Database and TED domains by combining fast protein embeddings with US-align refinement to support diverse search modes, including pLDDT-aware filtering and multidomain matching.

The Gist

Imagine you have a massive library containing 240 million books, but instead of words, every book is a 3D blueprint of a tiny machine called a protein. These machines run the show inside every living thing, from bacteria to humans.

The problem? You can't just flip through 240 million books to find the one you need. If you try to compare every single blueprint page-by-page (which is what computers usually do), it would take longer than the age of the universe.

Enter AlphaFind v2. Think of it as a super-smart, high-speed librarian that can instantly find the books in this massive library that look the most like the one you are holding, even if the text (the DNA sequence) inside them is completely different.

Here is how it works, broken down into simple concepts:

1. The "Fingerprint" Trick (The Fast Search)

Imagine you have a new mystery book and you want to find similar ones. Instead of reading every page, you take a quick "fingerprint" of the book's cover and style.

• AlphaFind v2 does this by creating a digital "fingerprint" (called an embedding) of your protein's shape.
• It instantly scans the library using these fingerprints to find a shortlist of 100 "look-alikes." This happens in a blink of an eye.

2. The "Fine-Tooth Comb" (The Precise Check)

The fingerprint is fast, but it's not perfect. It might miss a tiny detail.

• Once the librarian finds the shortlist, she pulls those specific books off the shelf and does a deep, detailed comparison (using a tool called US-align).
• She measures exactly how well the pages line up, calculating a "similarity score" (TM-score) to tell you how closely related they really are.

3. The Special Features (Why v2 is Better)

The previous version of this librarian (AlphaFind v1) was good, but v2 has some cool new superpowers:

• The "Blur Filter" (pLDDT Search):
  Sometimes, the blueprint of a protein has parts that are shaky or uncertain (like a sketch that wasn't finished). If you try to compare the whole thing, those shaky parts mess up the match.

  • The Analogy: Imagine trying to match two faces, but one person is wearing a foggy mask. AlphaFind v2 lets you say, "Ignore the foggy parts; only compare the clear, sharp features." This helps find matches that would otherwise be hidden by "noise."

• The "Lego Block" Search (TED Domains):
  Proteins are often made of smaller functional blocks (like Lego bricks) called domains. A protein might be a long chain, but only one specific block does the important work.

  • The Analogy: If you are looking for a specific type of Lego wheel, you don't want to search for whole cars. You want to search just for the wheels. AlphaFind v2 lets you search specifically for these "blocks" across the entire library.

• The "Mix-and-Match" Search (Multidomain):
  Some proteins are like a sandwich: Bread-Top, Meat, Cheese, Bread-Bottom. Sometimes you want to find other sandwiches that have the same order of ingredients, even if the bread looks different.

  • The Analogy: This mode lets you search for the entire sandwich structure (the specific order of domains) rather than just looking for the meat or the cheese separately. You can even adjust sliders to say, "I care more about the cheese alignment than the bread."

Why Does This Matter?

In the past, scientists had to guess how a protein worked based on its DNA code (the recipe). But two proteins can have very different recipes but build the exact same machine.

AlphaFind v2 allows scientists to:

1. Find hidden cousins: Discover proteins that look alike but have different names.
2. Understand function: If a new protein looks structurally like a known "door opener," it's probably a door opener too.
3. Move fast: It turns a search that used to take days into one that takes seconds.

The Bottom Line

AlphaFind v2 is a high-tech, shape-shifting search engine for the world's largest collection of protein blueprints. It uses smart shortcuts to find candidates quickly and then does a precise check to ensure they are a perfect match, helping scientists understand life's building blocks faster than ever before.

You can try it out yourself for free at their website, just like using a search engine, but for 3D protein shapes!

Technical Summary

1. Problem Statement

The rapid expansion of protein structural data, particularly with the AlphaFold Protein Structure Database (AlphaFold DB) containing over 240 million predicted structures, has created a bottleneck in structural biology. While large-scale structure comparison is essential for understanding protein function and evolution, traditional alignment-based methods (e.g., US-align) are computationally expensive and too slow for searching hundreds of millions of structures.

• Key Challenges:
  • Scalability: Exact 3D structural alignment cannot be performed in real-time across massive datasets.
  • Noise: AlphaFold predictions often contain low-confidence regions (indicated by low pLDDT scores) or intrinsically disordered regions that obscure true structural homology.
  • Granularity: Existing tools often focus on full chains, missing the functional relevance of specific domains or multi-domain arrangements.

2. Methodology

AlphaFind v2 addresses these challenges through a hybrid search architecture that combines fast approximate vector search with precise structural alignment refinement.

A. Data Preparation & Embeddings

• Source Data: Utilizes AlphaFold DB (v4) and the TED (The Encyclopedia of Domains) database.
• Embedding Generation:
  • Full Chains: Structures are compressed into 1536-dimensional vectors using the ESM3 generative model combined with a transformer neural network.
  • Confidence Filtering: Separate embeddings are generated for chains with unstable regions removed based on pLDDT thresholds (70, 80, and 90).
  • Domains: Pre-computed 128-dimensional Foldclass embeddings are used for TED domains.
• Vector Database: Embeddings and metadata are stored in OpenSearch using HNSW (Hierarchical Navigable Small World) indexes with 16x compression, enabling efficient approximate nearest neighbor (k-NN) search.

B. Three-Phase Search Workflow

1. Phase 1 (Query Processing): The user query (protein ID or structure) is converted into a fixed-length embedding.
2. Phase 2 (Approximate Retrieval): The vector database performs a k-NN search (k=100) using cosine distance to return a ranked list of candidate structures immediately. This provides near-instant results.
3. Phase 3 (Refinement): Running asynchronously in the background, this phase performs exact pairwise structural alignments using US-align on the top candidates. It computes precise metrics (TM-score, RMSD, aligned residues) and links results to experimental structures (PDB).

C. Search Modes

AlphaFind v2 offers six distinct search modes:

1. Full-Chain: Standard search over entire protein chains.
2. pLDDT-Filtered Chains: Restricts alignment to regions with confidence scores $\ge$ 70, 80, or 90, filtering out structural noise.
3. TED Domains: Searches specifically against domain definitions from the TED database.
4. TED Multidomain: A novel mode that searches for proteins containing a specific combination and order of domains, allowing users to weight the contribution of individual domain pairs to the overall alignment score.

3. Key Contributions

• Multi-Context Search: Unlike predecessors, AlphaFind v2 allows searching across full chains, confidence-filtered regions, and specific domain contexts (single or multi-domain).
• Interactive Refinement: The "Multidomain" mode includes interactive sliders allowing users to adjust the weighting of specific domain pairs in a 3D alignment, facilitating the analysis of complex structural arrangements.
• Hybrid Speed-Accuracy Trade-off: By decoupling the initial retrieval (vector search) from the final ranking (structural alignment), the system delivers immediate approximate results followed by high-precision metrics without blocking the user interface.
• Integration of Experimental Data: The system automatically links predicted AlphaFold structures with known experimental structures (PDB) where available.

4. Results and Performance

The authors evaluated AlphaFind v2 against FoldSeek Server (for chains) and Merizo-search (for domains) using a test set of 2,050 multidomain proteins and 4,420 TED domains.

• Speed:
  • Approximate Results: AlphaFind v2 is an order of magnitude faster than competitors.
    • Chains: ~2.4s (vs. 42s for FoldSeek).
    • Domains: ~0.49s (vs. 146s for Merizo-search).
  • Refinement: The time to compute refined TM-scores is comparable to FoldSeek but significantly faster than AlphaFind v1.
• Accuracy (TM-Score):
  • AlphaFind v2 achieved statistically superior TM-scores (p < 0.05) compared to all baselines.
  • Top 10 Chains: 0.733 (v2) vs. 0.596 (FoldSeek).
  • Top 10 Domains: 0.947 (v2) vs. 0.865 (Merizo-search).
• Case Studies:
  • PIN3 (pLDDT Filtering): Successfully identified homologs in soybean that were missed by full-chain searches due to disordered loops, achieving a TM-score of 0.947 by focusing only on high-confidence regions.
  • NCAM1 (Multidomain): Successfully identified proteins with the specific modular architecture of NCAM1 (Ig domains + Fibronectin domains), demonstrating the ability to capture complex domain arrangements rather than just individual domain matches.

5. Significance

AlphaFind v2 represents a significant advancement in the accessibility and utility of the AlphaFold DB.

• Biological Relevance: By enabling searches based on confidence thresholds and specific domain architectures, it allows researchers to bypass the "noise" of low-confidence predictions and focus on biologically meaningful structural features.
• Scalability: It proves that approximate vector search followed by targeted refinement is a viable strategy for navigating the "zoo" of predicted protein structures, making large-scale structural genomics feasible for interactive exploration.
• Accessibility: The tool is freely available as a web application, requiring no login, and provides rich metadata and interactive 3D visualizations (via Mol*), democratizing access to advanced structural bioinformatics.

Access: The tool is available at https://alphafind.ics.muni.cz/.