🧬 AMR Multi-Antibiotic Resistance Predictor

AI-powered genomic resistance prediction for E. coli — Ciprofloxacin, Ceftriaxone & Amoxicillin

---

🔬 What This Does

This project is a multi-task machine learning system that predicts antibiotic resistance in E. coli genomes. Given 40 genomic features extracted from whole-genome sequencing data (gene presence/absence markers, mutation counts, and proportion scores), the model simultaneously predicts resistance outcomes for three clinically important antibiotics:

Antibiotic	Drug Family	Treats
Ciprofloxacin	Fluoroquinolone	UTIs, respiratory infections
Ceftriaxone	3rd-gen Cephalosporin	Pneumonia, meningitis, hospital infections
Amoxicillin	Penicillin	Ear infections, strep throat, chest infections

Each prediction returns one of three standardised AST labels:

• 🟢 S — Susceptible (antibiotic kills the bacteria)
• 🟡 I — Intermediate (uncertain, dose/context dependent)
• 🔴 R — Resistant (antibiotic fails, bacteria survives)

---

📊 Model Performance

Antibiotic	Accuracy	MCC	F1 (weighted)	AUC
Ciprofloxacin	78.3%	0.553	0.717	0.934
Ceftriaxone	83.5%	0.632	0.783	0.935
Amoxicillin	79.8%	0.572	0.732	0.934

Algorithm: MultiOutputClassifier wrapping RandomForestClassifier (200 trees, max depth 12)

---

🗄️ Dataset & Data Pipeline

Source

Genomic data and Antimicrobial Susceptibility Testing (AST) results were sourced from the PATRIC / BV-BRC public database — the largest freely available bacterial genomics repository maintained by the U.S. Department of Energy.

• Organism: Escherichia coli (Gram-negative, clinically relevant)
• Genomes collected: 3,000 whole-genome sequencing (WGS) records
• AST labels: R / I / S for Ciprofloxacin, Ceftriaxone, and Amoxicillin

Data Cleaning Steps

Raw data from PATRIC required significant cleaning before it was usable:

1. Duplicate removal — genomes with identical PATRIC IDs or redundant AST entries were dropped
2. Missing label handling — genomes missing AST results for any of the 3 antibiotics were excluded
3. Outlier filtering — genomes with abnormal feature distributions (>3 SD from mean) were flagged and reviewed
4. Label standardisation — raw MIC values and non-standard labels were mapped to S / I / R using EUCAST 2023 breakpoints
5. Class balance check — final label distribution confirmed: S ≈ 65%, I ≈ 15%, R ≈ 20%

Feature Engineering

Features were extracted from cleaned genome assemblies using two bioinformatics tools:

a) AMRFinderPlus — identifies resistance genes:

amrfinderplus -i genome.fasta -o amr_genes.tsv

Feature Group	Columns	What Was Extracted
F1–F15	Binary (0/1)	Presence/absence of resistance genes & key mutations (e.g., gyrA S83L, bla_TEM, bla_CTX-M-15)
F16–F30	Integer (0–10)	Counts of resistance elements, mobile genetic elements, efflux pump genes
F31–F40	Real (0.00–1.00)	Proportional scores — fraction of genome with resistance markers, intact target sites

b) bcftools — identifies point mutations:

bcftools mpileup -f reference.fasta genome.bam | bcftools call -mv -o variants.vcf

Mutations in gyrA, parC, ompF, and beta-lactamase regions were recorded as binary features.

---

🚀 Quick Start

1. Clone the repo

git clone https://github.com/Rishu-raj-02/AMR-Multi-Antibiotic-Resistance-Predictor.git
cd AMR-Multi-Antibiotic-Resistance-Predictor

2. Install dependencies

pip install -r requirements.txt

3. Train the model

python train_model.py

Reads data/amr_dataset.csv, trains the model, saves artifacts to models/.

4. Start the web app

python app.py

Open http://localhost:5000 in your browser.

---

🗂️ Project Structure

AMR-Multi-Antibiotic-Resistance-Predictor/
│
├── data/
│   └── amr_dataset.csv          # Cleaned E. coli dataset (3,000 genomes × 44 cols)
│
├── models/                      # Auto-generated by train_model.py
│   ├── amr_model.pkl            # Trained MultiOutputClassifier
│   ├── encoders.pkl             # LabelEncoders for S/I/R
│   ├── metrics.json             # Per-antibiotic performance stats
│   ├── feature_importance.json  # Feature ranking per antibiotic
│   └── feature_cols.json        # Ordered feature column list
│
├── templates/
│   └── index.html               # Dark biotech dashboard UI
│
├── app.py                       # Flask web server + REST API
├── train_model.py               # Model training & evaluation script
├── requirements.txt
├── render.yaml                  # One-click Render.com deployment
├── Procfile                     # Gunicorn process config
└── README.md

---

🌐 API Endpoints

Method	Endpoint	Description
GET	/	Interactive web dashboard
POST	/predict	Run resistance prediction
GET	/metrics	Model performance stats
GET	/random_sample	Load a sample genome for demo

Example /predict request:

POST /predict
{
  "features": {
    "F1": 1, "F2": 0, "F3": 1,
    "F4": 1, "F11": 1,
    "F16": 7, "F17": 3,
    "F31": 0.85, "F32": 0.20, "F33": 0.70
  }
}

Example response:

{
  "predictions": {
    "CIPRO": {
      "label": "R",
      "full": "Resistant",
      "confidence": 87.3,
      "probabilities": { "S": 0.07, "I": 0.06, "R": 0.87 }
    },
    "CEFTRIAXONE": { "label": "S", "confidence": 72.1 },
    "AMOXICILLIN": { "label": "R", "confidence": 81.5 }
  }
}

---

🧬 Feature Reference

F1–F15 · Binary (0 = absent, 1 = present)

Gene presence / mutation markers extracted via AMRFinderPlus:

Feature	Biological Meaning	Linked Antibiotic
F1	gyrA S83L mutation	Ciprofloxacin ↑R
F2	gyrA D87G mutation	Ciprofloxacin ↑R
F3	bla_TEM gene presence	Amoxicillin ↑R
F4	Multi-drug resistance plasmid marker	All three ↑R
F7	bla_SHV gene	Ceftriaxone ↑R
F11	bla_CTX-M-15 (ESBL) gene	Ceftriaxone ↑R

F16–F30 · Integer (0–10)

Count features — number of resistance mutations, mobile elements, efflux pump genes detected per genome assembly.

F31–F40 · Real (0.00–1.00)

Proportion scores — fraction of genome containing resistance markers, proportion of intact drug-binding sites, etc.

---

🏥 Top Resistance Drivers (Feature Importance)

Rank	Feature	Importance	Biological Role
1	F3	0.1018	bla_TEM — Amoxicillin resistance
2	F2	0.0719	gyrA mutation — Cipro resistance
3	F4	0.0570	Multi-drug plasmid marker
4	F7	0.0553	bla_SHV — Ceftriaxone resistance
5	F1	0.0536	gyrA S83L — Cipro resistance

---

🏥 Clinical Decision Support

The web app generates real-time clinical guidance:

• Flags Resistant predictions with evidence-based alternative drug suggestions
• Recommends confirmatory MIC lab testing for borderline (I) results
• Displays probability distributions across all three resistance classes

⚠️ Disclaimer: This tool is intended for research and educational purposes. Clinical treatment decisions must always be confirmed with certified laboratory antimicrobial susceptibility testing.

---

📖 How the Model Was Built

1. Data Collection — 3,000 E. coli WGS records with AST results from PATRIC/BV-BRC
2. Cleaning — duplicate removal, missing label exclusion, EUCAST breakpoint standardisation
3. Feature Extraction — AMRFinderPlus (resistance genes) + bcftools (point mutations)
4. Feature Engineering — 40 columns: 15 binary + 15 integer + 10 real-valued proportions
5. Modelling — MultiOutputClassifier (RandomForest, 200 estimators, depth 12)
6. Evaluation — 80/20 stratified split; Accuracy, MCC, F1-weighted, AUC-ROC (OvR)

---

🤝 Contributing

Pull requests welcome. For major changes, please open an issue first.

---

📄 License

MIT — see LICENSE

---

Built for IIT Mandi Hackathon — E. coli · Ciprofloxacin · Ceftriaxone · Amoxicillin