🧬 HelixMind — Mutation Impact Predictor

Prototype Learning Core v0.1 — A 1D CNN that learns to detect biologically significant point mutations from DNA sequence context.

This is not a rule-based simulator. It's a learned biological signal detector — trained to recognize sequence patterns that correlate with functional impact, the same way a real mutation effect predictor would work at scale.

Built as the first learning module in the HelixMind pipeline: a system for predicting microbial mutation behaviour and antimicrobial resistance.

---

What It Does

Takes a DNA sequence + a point mutation and outputs a probability that the mutation disrupts protein function.

Input:  DNA sequence (50 bases) + mutation position + new base
Model:  1D CNN → GlobalMaxPooling → Dense → Sigmoid
Output: Impact probability  e.g. 0.87 → HIGH FUNCTIONAL IMPACT 🔴

---

Model Architecture

DNA Sequence Input  (50, 4)  ← one-hot encoded
    └─ Conv1D(32, kernel=3) + BatchNorm + ReLU
    └─ Conv1D(64, kernel=3) + BatchNorm + ReLU
    └─ GlobalMaxPooling1D
    └─ Dense(32) + Dropout(0.3)
         ↕  Concatenate
Mutation Position  (1) ──┐
New Base Identity  (1) ──┘
    └─ Dense(8) → Dense(16) → Dense(1, sigmoid)

Two-branch design: the CNN reads full sequence context, while a second branch encodes where the mutation is and what the new base is. They merge before the final classification head.

---

Biological Logic

Training data is synthetic but encodes real biological reasoning:

Condition	Label
Mutation falls inside a functional motif (ATGCG, GCATG, TTAGG)	High Impact (1)
Mutation falls outside all motifs	Low Impact (0)
~10% random label flip	Biological ambiguity noise

The model learns to associate sequence context around the mutation site with functional consequence — it doesn't just memorise positions.

---

Results

Trained on 5,000 synthetic samples, 80/10/10 split.

Metric	Score
Test ROC-AUC	~0.92
Accuracy	~88%
Training time	~3 min on Colab CPU

---

Quickstart

Run in Google Colab — no setup needed.

1. Upload HelixMind_Mutation_Predictor.ipynb to Colab
2. Runtime → Run All
3. Edit the interactive cell at the bottom to test your own sequences

---

Project Structure

helixmind-mutation-predictor/
├── HelixMind_Mutation_Predictor.ipynb   # Main notebook (data → train → eval → demo)
├── helixmind_demo.html                  # Standalone browser demo (no Python needed)
├── helixmind_mutation_core_v0.keras     # Saved model (generated after training)
└── README.md

---

Live Demo

Open helixmind_demo.html in any browser. Paste a 50-base DNA sequence, pick a mutation position and new base, and get an instant prediction with a sequence map highlighting the mutated position.

The demo runs the same biological signal logic client-side — no server required.

---

Using the Predictor

# After running all notebook cells:

result = predict_mutation_impact(
    sequence='AAAAAAAAAAAAAAAAAAAATGCGAAAAAAAAAAAAAAAAAAAAAAAAAAA'[:50],
    mutation_position=20,
    new_base='C'
)

# Output:
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
#   🧬 HelixMind Mutation Impact Assessment
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
#   Mutation:     A→C at position 20
#   Impact Prob:  0.8731
#   Prediction:   HIGH IMPACT 🔴
#   Confidence:   87.3%
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Important — always build test sequences programmatically to avoid off-by-one errors:

motif = 'ATGCG'
pos = 20
sequence = 'A' * pos + motif + 'A' * (SEQ_LEN - pos - len(motif))  # guaranteed 50 bases

---

Roadmap

This is v0.1 of the HelixMind learning core. Planned expansions:

• Swap synthetic data for real ClinVar / UniProt variant datasets
• Sliding window input for full-length gene sequences
• Resistance phenotype prediction head (AMR focus)
• FastAPI wrapper → HelixMind backend module
• Streamlit UI for wet-lab researcher use

---

Stack

• Python 3.10+
• TensorFlow / Keras
• NumPy, scikit-learn, matplotlib, seaborn
• Google Colab (free tier compatible)

---

---

Built with the goal of merging ML engineering and computational biology into something that actually matters.