Predicting antimicrobial resistance from bacterial genomes using Graph Attention Networks

GEN_RESIST predicts whether bacteria are resistant to antibiotics by analyzing their genome sequences. Instead of waiting 2-3 days for lab culture results, you upload a FASTA file and get predictions in seconds.

The model uses Graph Attention Networks (GAT) to analyze k-mer patterns in bacterial genomes and predict resistance across 30 different antibiotics.

Traditional methods either:

• Search for known resistance genes (misses novel variants)
• Count k-mer frequencies (loses sequence context)

We build a graph where k-mers are nodes and edges connect k-mers that appear near each other. This preserves local sequence structure and lets the attention mechanism learn which patterns matter for resistance.

1. Parse the genome - Extract DNA sequence from FASTA
2. Generate k-mers - Sliding window of length 11, canonical form (min of forward/reverse complement)
3. Build graph - Nodes = k-mers, edges = co-occurrence within 10-position window
4. Add features - GC content, genome length, presence of known resistance genes
5. Run model - 2-layer GAT processes the graph
6. Get predictions - Probabilities for 30 antibiotics, threshold at 0.5

Predictions:

• Resistant or Susceptible for 30 antibiotics
• Confidence score (probability) for each

Extra info:

• Detected resistance genes (from CARD database)
• Genome stats (GC content, length)

1.Clone the repo and setup

• cd GEN_RESIST
• pip install -r requirements.txt

2. Prepare your genome file:

   • Must be in FASTA format (.fasta, .fna, or .fa)
   • Whole bacterial genome sequence
   • Should be assembled (not raw reads)
   • Works best with complete genomes

3. Upload and predict:

   • Click "Choose File" or drag-and-drop your FASTA
   • Click "Predict Resistance"
   • Wait a few seconds for results

4. Interpret results:

   • You'll see predictions for 30 antibiotics
   • "Resistant" = bacteria likely resistant (probability > 0.5)
   • "Susceptible" = antibiotic likely effective (probability < 0.5)
   • Check probability scores for confidence:
     • 0.9+ = very confident
     • 0.6-0.9 = confident
     • 0.4-0.6 = uncertain (borderline)
     • <0.4 = confident susceptible
   • Review detected resistance genes for supporting evidence

β-lactams: Ampicillin, Amoxicillin, Piperacillin, Ceftriaxone, Cefotaxime, Ceftazidime, Cefepime

Carbapenems: Meropenem, Imipenem, Ertapenem

Fluoroquinolones: Ciprofloxacin, Levofloxacin, Nalidixic acid

Tetracyclines: Tetracycline, Doxycycline

Aminoglycosides: Gentamicin, Streptomycin, Kanamycin, Tobramycin

Others: Azithromycin, Erythromycin, Vancomycin, Colistin, Chloramphenicol, Rifampicin, Trimethoprim, Sulfamethoxazole

We check for 15 high-impact genes:

• β-lactamases: blaCTX-M-15, blaTEM-1, blaNDM-1, blaKPC
• Quinolone: qnrS1, qnrB, gyrA
• Tetracycline: tetA, tetM
• Colistin: mcr-1, mcr-2
• Others: vanA, ermB, aac(6')-Ib

• ML: PyTorch, PyTorch Geometric
• Backend: FastAPI
• Data: Biopython, NumPy
• Frontend: Vanilla HTML/CSS/JS
• Deploy: Docker, Render

• Clinical: Fast guidance for empirical antibiotic therapy
• Surveillance: Track resistance trends in populations
• Research: Study genotype-phenotype relationships
• Education: Demonstrate ML in genomics