Polysaccharide-Specific CNN Gene Role Classifier¶

This presentation documents the multi-source reference curation, Optuna TPE hyperparameter optimization, Numba-accelerated layer representation sweep, and held-out validation of GeneRoleClassifier across non-redundant (<= 80% CD-HIT) bacterial (taxonomy_id:2) and bacteriophage/prophage (taxonomy_id:10239) proteins spanning all 29 functional GeneRoles, a joint 1-bit bacterial/phage source head, and open-set UNKNOWN (Role 29) biological decoy rejection.

Why Polysaccharide-Specific CNN Classification Beats Remote Homology¶

Bacterial surface polysaccharide loci (capsular polysaccharides CPS, O-antigens O-antigen, wall teichoic acids WTA, and exopolysaccharides EPS) evolve under intense diversifying selection from bacteriophages and host immunity:

1. Sequence Divergence Beyond BLAST / MMseqs2 / Profile-HMM Sensitivity: Essential membrane machinery — such as Wzx/GtrA flippases (FLIPPASE), Wzy polymerases (POLYMERASE), WaaL O-antigen ligases (LIGASE), and serotype-defining glycosyltransferases (GLYCOSYLTRANSFERASE) — routinely diverge into the sequence "twilight and midnight zones" (< 20–25% amino-acid identity) where k-mer seeding and alignment bitscores fail.
2. Conserved Catalytic Motifs & Emergent Biophysical Topology: Despite extreme global sequence divergence, these enzymes strictly conserve short local catalytic signatures (e.g., DxD, ED, and EX7E glycosyltransferase catalytic triads, Rossmann-fold GxGxxG nucleotide-sugar binding loops, and LytR-CpsA-Psr motifs) alongside whole-protein transmembrane hydropathy and fold repeat profiles.
3. Pure Emergent Dual-Pooled 1D-CNN + Joint 1-Bit Source Head: GeneRoleClassifier combines learnable amino-acid embeddings (embed_dim=64), parallel multi-scale 1D convolutional branches (num_filters=192, kernel_size=(5, 9, 13)), and dual global max (conv_max, peak local catalytic and viral motifs) + length-normalized mean (conv_avg, global fold/transmembrane density) pooling (feature_dim = 2 * num_filters = 384), driving both the 29-class GeneRole readout (dense_w_) and a joint 1-bit genomic source readout (source_w_, predict_source / predict_source_proba: 0 = bacterial, 1 = bacteriophage/prophage) while keeping ArchetypeSpecs 32-bit uint32 role bitmasks untouched.
4. Active-Site Embeddings for Twilight-Zone Clustering Rescue: Intermediate representations extracted via GeneRoleClassifier.transform(seqs, layer="conv_max") (default in Loci.cluster_genes) capture catalytic motif activations and feed directly into seroverse.core.seq.clustering.Clusterer (embedding_weight, motif_rescue_threshold) to rescue divergent intra-role homologs during Markov Clustering (MCL).

Architectural Assessments & Conclusions¶

1. 11-Layer Representation Sweep (transform Benchmark on 1,565 Held-Out Proteins)¶

Across all 11 representations exposed by GeneRoleClassifier.transform, unweighted penultimate ([conv_max, conv_avg], 384-D) and unweighted conv_max (192-D) achieve the highest overall clustering and retrieval scores, outperforming class-separation-weighted task_weighted:

2. Dual-Head Weight-Norm Division of Labor (dense_w_ vs. source_w_)¶

Inspecting the Frobenius squared weight norms (||W||_F²) of the trained linear readouts reveals how the shared [conv_max, conv_avg] representation serves two distinct biological tasks simultaneously:

Run Locally¶

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

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

Comments