C.elegans

Minimal code and result bundle for the C. elegans division-prediction project.

Included

• code/geo_matched_gen/scripts/: key scripts for single-gene baselines, multi-gene alignment, classifier training, and GNN gene-feature tests.
• code/geo_matched_gen/experiment*/: core experiment summaries and the main train_exp3.py implementation.
• results/report0317/: current report, plots, manifests, and summary CSVs.
• results/report0309/: gene-coverage summary tables used for panel selection.

Not Included

• Raw EPIC CSV data.
• Large generated intermediate datasets.
• Temporary logs / caches / local-only artifacts.

Main Current Findings

• Single-gene expression is weak; size is a strong baseline.
• Multi-gene alignment adds signal, especially with donor ensembling.
• Classifier performance peaks around a focused 10--11 gene panel.
• In the GNN, gene features help when injected into event heads rather than concatenated into the trunk.