Skip to content

MPNN Glycan Structure Predictor (GlycanPredictor)

This presentation documents the Two-Phase Optuna TPE hyperparameter optimization, held-out multi-regime structural validation, and out-of-sample Klebsiella pneumoniae KL106 (LT174576.1) and KL107 (KF793263.1) capsule benchmarks of GlycanPredictor and RelationalMPNN (src/kbank/data/glycan_predictor_mpnn.npz).

Key Takeaways

  • Two-Phase Optuna TPE Hyperparameter Optimization:
  • Phase 1 (RelationalMPNN): Tunes message-passing hidden dimension, layer depth, factored bilinear edge-scoring rank, 2-hop syntenic edges, learning rate, weight decay, and single-free-GT clamping probability on pre-extracted biosynthetic arborescences.
  • Phase 2 (GlycanPredictorParams): Jointly tunes all 14 Stage ¾ node and pairwise arborescence potential blending weights across Cross-Validated (unseen record) and Pure Sequence De Novo (cluster_id=0) evaluation regimes.
  • Exact Out-of-Sample K. pneumoniae KL106 -> K106 & KL107 -> K107 Recovery: Fetches public NCBI GenBank accessions LT174576.1 (KL106) and KF793263.1 (KL107), segments their CPS loci de novo via SeroGen's LocusPredictor, and recovers both complete capsule structures with 100% topological and stereochemical isomorphism (WWL = 0.0000).
  • Automatic Production Model Export: Persists both the Optuna-selected RelationalMPNN weights and the 14 tuned GlycanPredictorParams potentials in src/kbank/data/glycan_predictor_mpnn.npz.

Run Locally

# Present as interactive slideshow
uv run --with marimo --with altair --with optuna marimo run scripts/train_glycan_predictor.py

# Edit notebook cells
uv run --with marimo --with altair --with optuna marimo edit scripts/train_glycan_predictor.py

Comments