🧬 Tier-10 Clinical Splice-Aware Mutation Predictor

Model: MutationPredictorCNN_v2 | Val Accuracy: 74.8% | Input dim: 1106 | Trained on: 100k ClinVar variants

Explainability stack: conv3 activation · gradient attribution · counterfactual analysis · feature ablation · splice distance · clinical tier classification.

Reference Sequence (99 bp, centered on variant)

Mutated Sequence (same window with SNV applied)

Chromosome

Genomic Position (GRCh38)

Exon flag (1 = in exon)

0 1

Intron flag (1 = in intron)

0 1

Clinical Explainability Report

CNN Activation Heatmap

Splice Distance Risk Heatmap

Mutation Rank Percentile

🔬 Causal Mutation Analysis (Counterfactual)

Counterfactual Analysis Summary

Counterfactual Probability Chart

🎯 Gradient Attribution Map

Gradient Attribution Heatmap

⚗️ Causal Feature Impact (Ablation Analysis)

Feature Ablation Causal Chart

Feature Ablation Summary

Clinical Examples (ClinVar confirmed)

Risk tier classification

Tier	Probability
PATHOGENIC	≥ 0.90
LIKELY PATHOGENIC	≥ 0.70
POSSIBLY PATHOGENIC	≥ 0.50
LIKELY BENIGN	≥ 0.20
BENIGN	< 0.20

Tier-10 explainability stack

CNN Activation Heatmap — conv3 L2 norm per nucleotide (blue=low → red=high)
Gradient Attribution Map — input-gradient backward pass, showing which bases the model was most sensitive to
Counterfactual Analysis — all 3 alternative substitutions at the mutation position tested, probability range = causal importance score
Feature Ablation — splice / region / mutation features zeroed one at a time; probability delta = causal contribution

⚠ Disclaimer: Research use only. Not a clinical diagnostic tool.