miyuiu/predictability-gradient
0
predictability-gradient
Research showcase for the paper:
When Does Attention Help? A Predictability Gradient for Genomic Trait Prediction, Validated by Virulence-Factor Attribution
This Space is a paper companion, not a production predictor. It presents the main scientific result as an interactive exhibit:
- attention-pooling helps machinery traits more than compositional traits;
- the advantage collapses at family-level holdout;
- pathogenicity attention concentrates on a few proteins;
- top-attended proteins are enriched for VFDB virulence factors;
- top-protein ablation flips pathogenic predictions.
Refreshing assets
From the repository root:
python3 spaces/research_showcase/scripts/build_assets.pyThe script reads runs/*.json and writes:
spaces/research_showcase/assets/predictability_gradient.jsonThat JSON is intentionally small so the Space can run on CPU-basic without access to the full 105 GB per-protein embedding corpus.
