Skip to content

KBank

Bacterial surface glycan database, biosynthetic gene-to-glycan alignment, and de novo polysaccharide structure prediction.

Built on Structure-of-Arrays (SoA) Compressed Sparse Row (CSR) chemical graphs and Numba-accelerated message-passing kernels, KBank bridges SeroGen biosynthetic gene cluster predictions directly to chemical polysaccharide structures (CSDB / IUPAC / KEGG KCF).

Python 3.14+ 1,530 Curated Records Relational MPNN GPLv3 License

Key Capabilities

  • Curated Geno-Glycomic Database (kbank.db)

    Provides 1,530 curated bacterial polysaccharide structures (CSDB, IUPAC Linear, and KEGG KCF) paired with complete biosynthetic gene clusters and MCL protein homology families across Klebsiella, Acinetobacter, Escherichia, Salmonella, Streptococcus, and Pseudomonas.

  • CSR Directed Glycan Graphs (kbank.glycan, kbank.sugars)

    Represents repeating-unit oligosaccharides and modifications inside contiguous SoA CSR arrays with root-invariant topological hashing, Weisfeiler-Lehman graph kernels, and direct CSDB / IUPAC / KEGG KCF parsing and serialization.

  • Bipartite Gene-to-Glycan Alignment (kbank.biosynthesis.align)

    Maps individual glycosyltransferases, acetyl/pyruvyltransferases, and Wzy polymerases onto specific sugar linkages via BiosyntheticAssigner and PanGraphConsensusPrior Expectation-Maximization homology priors.

  • De Novo Polysaccharide Structure Prediction (kbank.biosynthesis.predictor)

    Combines Operon Precursor Pool gating, factorized 1D-CNN enzyme classifiers (BiosyntheticClassifierSuite), a Relational Message-Passing GNN (RelationalMPNN), and a branch-and-bound Constrained Spanning Arborescence solver (GlycanPredictor).

End-to-End Geno-Glycomic Pipeline

flowchart LR
    LOC["<b>Segmented BGC Locus</b><br/><i>SeroGen LocusPredictor</i>"]
    CLS["<b>Factorized 1D-CNN Suite</b><br/><i>GT Donor, Mechanism &amp; Wzy</i>"]
    ALN["<b>Bipartite EM Aligner</b><br/><i>BiosyntheticAssigner + PanGraph Prior</i>"]
    MPNN["<b>Relational MPNN + Arborescence</b><br/><i>GlycanPredictor De Novo Solver</i>"]
    GLY["<b>Directed Glycan Graph</b><br/><i>CSDB / IUPAC / KEGG KCF</i>"]

    LOC --> CLS --> ALN --> MPNN --> GLY

Quickstart

from kbank.biosynthesis.predictor import GlycanPredictor
from kbank.db.records import GlycanRecords
from serogen.core.io.fasta import GenomeAssembly
from serogen.loci.predictor import LocusPredictor

# Load curated bacterial glycan-locus reference database
records = GlycanRecords.load()

# Segment capsule/O-antigen locus from a genome assembly
assembly = GenomeAssembly.from_fasta("KL106.fasta")
locus = LocusPredictor().predict(assembly)[0]

# Predict top candidate polysaccharide structures via Relational MPNN
predictor = GlycanPredictor.from_records(records)
top_hit = predictor.predict(locus, top_k=1)[0]
print(top_hit.glycan.to_csdb().decode(), f"(p={top_hit.probability:.2%})")

from kbank.io.csdb import CSDBParser
from kbank.io.iupac import IUPACParser

csdb_parser = CSDBParser()
glycan = csdb_parser.parse("-3)[bDGlcpA(1-2)aLRhap(1-3)]aDGalp(1-")

print(f"Residues: {glycan.num_nodes}, Linkages: {glycan.num_edges}")

from kbank.biosynthesis.align import BiosyntheticAssigner
from kbank.db.records import GlycanRecords

records = GlycanRecords.load()
assigner = BiosyntheticAssigner()

# Align a biosynthetic gene cluster to its paired glycan structure
record = records[0]
assignment = assigner.assign_record(record.glycan, record.locus)

Comments