The human genome is not a flat, 2D string of letters. Inside the nucleus, DNA twists, folds, and loops back on itself in complex 3D topological structures. This architecture is critical: a piece of DNA (an Enhancer) can physically loop backward in 3D space to touch a gene (a Promoter) located millions of base pairs away, turning that gene on or off.
To study this, geneticists use Hi-C Sequencing or deep learning prediction models like ORCA, which output massive “Contact Maps”.
The $N \times N$ Matrix Problem
A Hi-C Contact Map is a giant $N \times N$ matrix tracking the physical distance between every single base pair in a chromosome. For large genomes, these matrices are gigabytes in size. Most of the matrix is empty (zeros), because most base pairs never touch each other.
Loading .hic or .cool files into R or Python to find specific enhancer-promoter loops requires massive amounts of RAM.
Enter pg_bio: Relational Topologies
Instead of treating the genome as a flat file, we can treat it as a relational graph using pg_bio.
pg_bio introduces the SparseAttentionMap PostgreSQL type. Under the hood, this uses a Compressed Sparse Row (CSR) structure natively in Rust. It entirely discards the empty space (the zeros) and only stores the exact base pairs that are physically interacting.
🎬 Companion Voiced Video & CSR Interactive Studio See how
SparseAttentionMapcompresses 90%-zero single-cell RNA matrices, Hi-C chromosomal loops, and 500 ns MD residue networks in Episode 03: Transcriptomics & CSR Sparse Attention Graphs:
Mining for Genomic Loops in SQL
By storing the genomic architecture in PostgreSQL, finding complex 3D DNA loops becomes a trivial SQL query.
If we want to ask the database: “Which base pairs are physically wrapping around and touching the Promoter located at base pair 25?”, we can use the get_top_interacting_residues function:
SELECT
sequence_name,
get_top_interacting_residues(orca_hic_map, 25, 5) as closest_physical_dna_contacts
FROM genomic_predictions
WHERE sequence_name = 'Synthetic_Plasmid_Construct';
Output:
sequence_name | closest_physical_dna_contacts
---------------------------+-------------------------------
Synthetic_Plasmid_Construct | {1300, 1301, 1302, 26, 24}
In a fraction of a second, the database proves that base pair 1300 is physically touching base pair 25.
By pushing 3D genomic matrices into a highly optimized PostgreSQL extension, computational biologists can instantly join topological data against their existing gene annotation tables, mapping the regulatory network of the cell using nothing but standard SQL.
