README.md

---
title: Mutation Explainability Intelligence System
emoji: 🧬
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
pinned: false
license: mit
tags:
  - genomics
  - bioinformatics
  - explainability
  - pathogenicity
  - splice
  - XAI
---

# Mutation Explainability Intelligence System

**Explainability-first** three-model ensemble for SNV pathogenicity prediction.

The system answers five core questions **before** presenting any prediction score:

1. Why was this variant predicted pathogenic / benign?
2. Which internal model signals drove that decision?
3. Is the signal localised at the mutation site?
4. Did removing the mutation change the prediction?
5. Do multiple models agree mechanistically?

## Models

| Model | Repo | Architecture |
|---|---|---|
| Splice | `nileshhanotia/mutation-predictor-splice` | MutationPredictorCNN_v2 |
| V4 | `nileshhanotia/mutation-predictor-v4` | MutationPredictorCNN_v4 |
| Classic | `nileshhanotia/mutation-pathogenicity-predictor` | MutationPredictorClassic |

## Input

- Chromosome, Position (hg38), Ref base, Alt base, Exon/Intron flag
- Sequence fetched automatically from **Ensembl REST API** (99-bp window)

## Explainability Signals

- **conv3 activation norm profiles** — per-nucleotide activation intensity
- **Mutation-centred activation peak** — peak at mutation site vs mean
- **Gradient attribution maps** — input-gradient backward pass
- **Splice aura distance** — proximity to GT/AG splice dinucleotides
- **Counterfactual delta** — all alternative bases tested
- **Feature ablation** — splice / region / mutation / sequence groups
- **Cross-model locality score** — Pearson correlation of activation profiles
- **Explainability Strength Score** — 0–1 composite quality metric

## Confidence Levels

`High` / `Moderate` / `Low` based on model agreement, ESS, and counterfactual magnitude.

## Disclaimer

For **research use only**. Not a clinical diagnostic tool.