Skip to content

Architecture & Algorithms

KBank is architected on the same zero-copy Structure-of-Arrays (SoA) and Compressed Sparse Row (CSR) principles as SeroGen, extending genomic interval and sequence containers to chemical polysaccharide graphs and biochemically constrained structure prediction.

1. Chemical Monosaccharide Registry & CSR Glycan Graphs

  • kbank.sugars.Sugars: Immutable SoA chemical registry encoding monosaccharide base rings, stereochemical configurations (Anomer, Chirality), ring forms, and non-stochiometric substituents (\(O\)-acetyl, pyruvyl, lactyl, formyl, phosphate, sulfate).
  • kbank.glycan.Glycan & Glycans: Packed CSR directed multigraphs representing repeating-unit oligosaccharides, side-chain branches, and polymer backbone (Wzy) linkages, supporting canonical root-invariant topological hashing and Weisfeiler-Lehman subtree kernels.

2. Pan-Graph Expectation-Maximization Enzyme-to-Linkage Alignment

  • kbank.biosynthesis.align.BiosyntheticAssigner: Formulates enzyme-to-linkage assignment as a minimum-cost bipartite matching problem (LOCUS_SUGAR_EXT) coupling:
  • Biochemical rule compatibility (BiosyntheticRuleRegistry),
  • Multi-record pan-graph EM homology priors (PanGraphConsensusPrior),
  • Syntenic operon transition penalties, and
  • Biological Und-PP reducing-end frame rotation.

3. De Novo Polysaccharide Structure Prediction (GlycanPredictor)

  1. Operon Precursor Pool Gating: Identifies nucleotide-sugar biosynthesis pathways present in the locus (e.g., rhamnose rmlABCD, fucose gmd/fcl, uronic acid ugd).
  2. Factorized 1D-CNN Sequence Classifiers (BiosyntheticClassifierSuite): Predicts GT donor specificity, stereochemical mechanism (retaining vs. inverting), Wzy polymerase linkage, and initiating transferase directly from protein sequence.
  3. Relational Message-Passing Neural Network (RelationalMPNN): Refines pairwise donor-acceptor linkage probabilities over the candidate sugar-enzyme graph.
  4. Constrained Spanning Arborescence Solver: Extracts the top-\(K\) biochemically valid repeating-unit trees via Numba-accelerated branch-and-bound search.

Comments