README.md

---
license: cc-by-4.0
tags:
- rna
- rna-rna-interaction
- attention
- interpretability
---

# MARRI headline model — test-set interpretability export

Per-sample interpretability/diagnostics for the MARRI headline configuration
(`dyn ft-rnafm`, seed 42 — RNA-FM fine-tuned, RNet2D frozen, dynamic negative
resampling), evaluated on the fixed held-out test split (n=16,058 pairs,
seed 42).

Code and training/evidence logs: https://github.com/GainGod-Xu/MARRI

## Files

- `marri_ft-rnafm_dyn_bs4_seed42_test_results_interpretability.h5` — per-sample
  diagnostics: cross-attention maps (A→B, B→A) at every fusion layer (global
  and local-window scope), the hybridization-gain (H_ij) matrices, fused
  logits, opening-penalty terms, and RNet2D per-base structural features, all
  in fp16 with gzip compression, plus per-sample metadata (sequence IDs, raw
  sequences, local-window coordinates, gene type/genomic context).
- `marri_ft-rnafm_dyn_bs4_seed42_test_results.h5` — the corresponding
  predictions/probabilities/targets/ROC and PR curves for the same test run
  (test ROC-AUC 0.7530, PR-AUC 0.7346, Accuracy 0.6830, Precision 0.6663,
  Recall 0.7255, F1 0.6947, MCC 0.3677).

Generated by re-running the test/export phase against the already-trained
checkpoint (no retraining); metrics are identical to the ones reported for
this run in the paper.