Multi-Level Taxonomic Novelty Detection with Increased Genomic Depth

This project investigates the performance of a Naive Bayes Classifier (NBC) for taxonomic novelty detection using an expanded genomic database with increased species representation depth.

Overview

Building upon our previous work which analyzed NBC's novelty detection capabilities using k-mer counting on a single-genome-per-species database (4,634 species), this project significantly expands the scope by:

• Utilizing a comprehensive database of 58,979 unique species
• Including multiple genomes per taxonomic class (319,554 unique genomes in total)
• Analyzing how increased genomic depth affects classification performance

Methodology

Dataset creation

To ensure statistical robustness and balanced representation, we implemented a two-stage filtering process:

1. Initial Filtering

   • Excluded species with fewer than 400 genome representatives
   • This threshold ensures sufficient data for meaningful model training

2. Random Sampling

   • Generated balanced training datasets through random sampling
   • Each trial maintained consistent genome counts across different species configurations
   • Number of species varied between trials due to natural variation in genome availability per species

k-mer counting

Models are trained on k-mer frequencies. All k-mer count files were generated using Jellyfish. The k-mers used in this project were of length 3, 6, 9, 12 and 15.

Testing data

[... ...]

Post-data analysis and ROC/AUC generation

[... ...]

Results