Skip to content

10-Head 1D-CNN Cluster Classifiers (BiosyntheticClassifierSuite)

This presentation documents the 2-pass Expectation-Maximization (EM) self-training, 5-fold sequence-stratified cross-validation, homology-cluster log-probability pooling, and 1D-CNN sequence embedding geometry of ClusterClassifier and BiosyntheticClassifierSuite (src/kbank/data/cluster_classifiers.npz).

Key Takeaways

  • 10 Factorized Enzymatic Specificity Heads: Jointly predicts glycosyltransferase donor_linkage, donor, linkage, acceptor_sugar, acceptor_motif, is_backbone, catalytic mechanism (retaining vs. inverting), initiator_sugar (WbaP/WcaJ/WecA/Itr), wzy_linkage polymerase specificity, and modifier_linkage regioselectivity directly from amino acid sequences.
  • 2-Pass EM Self-Training & Warm-Started Multi-Scale 1D-CNN: Warm-starts multi-scale (5, 9, 13) convolutional filters from SeroGen's GeneRoleClassifier backbone and refines non-locked record alignments in an E-step before refitting all 10 heads (src/kbank/data/cluster_classifiers.npz).
  • Homology-Cluster Pooling (predict_cluster_proba): Combining sequence-level CNN log-probabilities across MCL homology clusters boosts reaction specificity accuracy over single-sequence CV across all 10 enzymatic heads.

Run Locally

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

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

Comments