Upload README.md

README.md

@@ -1,13 +1,65 @@
 ---
-title: Mutation XAI
-emoji: 👀
-colorFrom: purple
-colorTo: yellow
+title: Mutation Explainability Intelligence System
+emoji: 🧬
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.6.0
+sdk_version: 4.31.5
 app_file: app.py
 pinned: false
-license: apache-2.0
+license: mit
+tags:
+  - bioinformatics
+  - genomics
+  - explainability
+  - clinical-genomics
+  - pytorch
+  - xai
+short_description: Explanation-first tri-model genomic variant analysis
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🧬 Mutation Explainability Intelligence System
+
+**Explanation before prediction. Always.**
+
+## Models used
+
+| Model | Repo |
+|---|---|
+| Splice disruption (MutationPredictorCNN_v2) | `nileshhanotia/mutation-predictor-splice` |
+| Protein context v4 (MutationPredictorCNN_v2) | `nileshhanotia/mutation-predictor-v4` |
+| Classic pathogenicity (MutationPredictorCNN) | `nileshhanotia/mutation-pathogenicity-predictor` |
+
+## Internal signals extracted
+
+From each model — nothing redesigned, internal hooks only:
+
+- **conv3 activation norm profile** — per-nucleotide CNN signal (99 positions)
+- **Gradient attribution map** — input-gradient backward pass
+- **Mutation-centered peak** — activation value at the specific mutation position
+- **Splice aura distance** — distance to nearest GT donor / AG acceptor
+- **Counterfactual delta** — probability range across all alternative substitutions
+- **Feature ablation response** — causal effect of zeroing splice / region / mutation features
+- **Risk tier classification** — PATHOGENIC → BENIGN
+
+## Cross-model explainability metrics
+
+| Metric | Description |
+|---|---|
+| Mutation Peak Ratio | peak_signal / mean_signal across window |
+| Counterfactual Magnitude | max probability range across alternatives |
+| Cross-Model Locality Score | cosine similarity of conv3 profiles |
+| Signal Concentration Index | fraction of activation within ±5bp of mutation |
+| Explainability Strength Score | composite 0–1 score |
+| Activation Pattern Type | Sharp / Broad / Flat |
+
+## Dominant mechanism classification
+
+- **Splice-driven** — splice model dominant + proximity to splice site
+- **Protein-driven** — v4 + classic dominant
+- **Consensus** — all three models agree
+- **Ambiguous** — divergent signals
+
+## ⚠️ Disclaimer
+
+For research use only. Not a clinical diagnostic tool.