There is a vast, uncharted universe inside biological databases known as the Dark Proteome—millions of proteins that have been sequenced but never characterized. We don’t know what they look like, and we don’t know what they do.
Historically, scientists used tools like BLAST to find proteins that look similar based on their amino acid sequence (letters). But what happens when evolution diverges so far that two proteins share only 18% of their sequence, yet still fold into the exact same 3D shape and perform the same job? Sequence alignment fails us; this is known as the Twilight Zone of homology.
Today, we’ll explore how we used our custom PostgreSQL extension, pg_bio, coupled with Deep Learning vector embeddings (ESM-2), to search the Dark Proteome by structure rather than sequence.
A Primer for Newcomers: Why Structure Matters
If you’ve never worked in bioinformatics, you might wonder: What exactly is a protein embedding vector?
To understand this, let’s look at one of the most famous proteins in biology: Green Fluorescent Protein (GFP). Originally found in jellyfish, GFP is a tiny molecular lantern that glows green under UV light.
Below is the 3D structure of GFP. Notice how the entire protein forms a hollow cylinder made of flat, ribbon-like walls. This specific architectural shape is called a Beta-Barrel. Hidden perfectly inside the center of this protective barrel is the chemical “bulb” that emits the green light.
If the amino acid “letters” of GFP mutated through millions of years of evolution, traditional sequence-matching tools (which act like a text search) would fail to recognize the mutated protein. However, a deep learning AI doesn’t just read the letters—it “learns” the laws of physics and understands that the sequence will fold into this exact hollow barrel.
The AI then converts that 3D structural understanding into a Vector Embedding: a long list of numbers (e.g., [0.45, -0.12, 0.89...]). A vector is essentially a mathematical barcode for the protein’s 3D shape. When we store these barcodes in a database like PostgreSQL and use pg_bio to search them, we aren’t searching text. We are instantly finding proteins that look and function the same, even if their text is completely unrecognizable!
The Mission: Telomerase
We set out to find hidden homologues of Telomerase—the enzyme responsible for maintaining the ends of chromosomes. Using pg_bio, we loaded the 3D embedding vectors of known Telomerase proteins and performed a massive HNSW (Hierarchical Navigable Small World) index scan against uncharacterized proteins.
The results were instant, and startling.
We discovered an uncharacterized orphan protein in a species of Archaea (Methanofollis formosanus, UniProt ID A0A8G1A137) that sat remarkably close to the mouse Telomerase Component 1 (TEP1, UniProt ID P97499) in the high-dimensional latent space.
Let’s do a sequence alignment to see if we could have found this the old-fashioned way:
Bait Length: 2629 aa | Orphan Length: 2227 aa
Global Alignment Score: -71.5
Sequence Identity: 18.49%
Warning: Sequence identity is < 20%. This is in the 'twilight zone' of sequence homology.
At 18.49% sequence identity, traditional sequence alignment like BLAST would likely miss this connection entirely or flag it as noise. But pg_bio found it in seconds.
Visualizing the Architecture (3Dmol.js)
Why did the AI embeddings link them? We couldn’t find an AlphaFold model for the new Archaeal orphan because it’s so obscure it hasn’t been modeled publicly yet! However, analyzing the known bait (TEP1) reveals the secret.
TEP1 is composed of WD40 repeats—massive, all-beta strand structural motifs that fold into circular “beta-propellers”. InterPro database predictions confirm our newly discovered Archaea orphan is built of Parallel beta-helix repeats (Pectin lyase-like folds). Both are large, repetitive, all-beta solenoids used for macromolecular scaffolding! The neural network successfully clustered their 3D global topologies together.
Let’s look at the incredible 3D structure of the Mouse TEP1 Telomerase component.
Interactive Structural Breakdown
The beauty of 3D visualization is that we can dissect this massive 2600-amino-acid engine into its core functional parts. pg_bio matched this protein to the Archaea orphan specifically because they share repetitive scaffolding folds.
Use the interactive controls below to explore the architecture of TEP1 and understand exactly what the AI found:
What do these colors mean?
- Yellow (Beta-Strands): These are flat, sheet-like structures. In TEP1, they arrange themselves into massive circular repeating patterns (Beta-propellers). This is the exact architectural “scaffolding” feature that
pg_biodetected and linked to the Archaea orphan! - Cyan (Alpha-Helices): These coiled, spring-like structures often form the active sites or flexible hinge regions of the protein.
- Gray (Loops & Intrinsically Disordered Regions): Notice the long, sweeping, string-like curves flying out into empty space? These are flexible “tentacles” that do not have a fixed 3D shape. In the Telomerase complex, these act like molecular lassos, flopping around dynamically until they find and bind to RNA!
- Red (WD40 Propeller): When you click “Zoom to WD40”, you are isolating the C-terminus of TEP1. This massive ring of beta-strands acts as a docking station for other proteins in the Telomerase complex. This shape is universally used in biology for structural assembly.
The Power of PostgreSQL + Vectors
By converting protein sequences into structural vectors, pg_bio allows scientists to bypass the sequence “twilight zone”. We aren’t just searching text; we are querying the physical reality of the molecules directly inside the database using optimized <=> cosine distance operations.
This workflow guarantees we can:
- Find structural homologues instantly.
- Filter by biological constraints (e.g., sequence length bounds to prevent matching fragments against massive chains).
- Validate through domain architecture (InterPro) and 3D visualization.
The Dark Proteome is no longer dark. With AI embeddings and Postgres, we have a flashlight!
