Skip to content

Benchmarks & Interactive Slide Decks

SeroRef maintains reproducible Marimo notebooks in scripts/ covering every stage of the pipeline: 4-tier de novo SeroRefRecords reference database curation, single-pass bacterial & prophage ORF prediction, polysaccharide-specific 1D-CNN gene role training and layer sweeping, and de novo capsular polysaccharide (CPS) locus discovery, hybrid MCL clustering, and pan-locus database construction.

Each notebook is configured with a native 16:9 Marimo slide layout (scripts/layouts/*.slides.json) that exports directly into interactive slide decks.

  • 4-Tier De Novo SeroRef Reference Curation

    Harvests, caches, and predicts surface polysaccharide loci de novo across ENA NCTC3000 (PRJEB6403), DSMZ StrainInfo v2 & BacDive v2, NCBI Landmark BGC sweeps, and NCBI accession, PubMed elink & strain assembly resolution, splitting multi-antigen serotype formulas (O1:K1:H7) into src/seroref/data/seroref.npz.

    Explore SeroRef Curation Deck

  • Single-Pass Bacterial & Prophage ORF Benchmark

    Evaluates OrfPredictor on Escherichia coli K-12 MG1655 (CP014225.1, 4.66 Mb) and its 9 embedded cryptic/lambdoid prophages, demonstrating how 29-state frameshift/overlap decoding and sliding 4-kb GC adaptation recover 95.6% of prophage 3'-stops (+17 rescued ORFs) and 99.1% of host CDSs in a single pass at >60 Mb/s.

    Explore ORF Benchmark

  • Polysaccharide-Specific CNN Gene Role Classifier

    Uncapped multi-source reference curation (7,832 non-redundant <= 80% CD-HIT sequences across UniProtKB, UniRef50, NCBI IPG, and KEGG), Optuna TPE hyperparameter optimization, Numba-parallel 11-layer representation sweep, and dual-head validation (87.89% 29-role Macro F1 + 90.4% 1-bit bacterial/phage source accuracy, src/seroref/data/model.npz), beating remote homology in the < 25% identity twilight zone.

    Explore Classifier Deck

  • S. pneumoniae De Novo CPS Locus Discovery & Database

    4-stage validation of de novo CPS locus discovery (LocusPredictor), hybrid CNN-rescued MCL orthogroup clustering (Clusterer), directed pan-locus graph construction (Loci.as_pan_graph), contig-split stitching (100% recovery across intergenic & mid-ORF breaks), and leave-one-out novelty detection (LocusMatcher).

    Explore De Novo CPS Database Deck

Running Notebooks Locally

All notebooks can be launched interactively as reactive web apps, slide presentations, or editable Marimo notebooks via uv:

# Launch any notebook directly in Marimo's fullscreen slide presentation mode
uv run --with marimo --with altair marimo run scripts/curate_seroref.py
uv run --with marimo --with altair marimo run scripts/benchmark_prophage_orfs.py
uv run --with marimo --with altair marimo run scripts/benchmark_spneumo_db.py
uv run --with marimo --with altair --with optuna marimo run scripts/train_gene_classifier.py

# Open in the reactive Marimo notebook editor
uv run --with marimo --with altair marimo edit scripts/curate_seroref.py

Comments