Setup

• Python 3
• SciKit Learn

Process

Pipeline

1. Get the data from NCBI
2. Read it in, and output as annotated JSON
3. Read in data, Kmer-ise it, and perform Bag Of Words
4. Train the classifier with random set
5. Test the classifier on unseen set

Using Genomic Data for ML

• How to select?
• How to pass to a classifier?
• How much data to use?

Classifier Choice

• Support Vector Machine Classifier (SVC)
• Supervised, fast for small numbers of examples
• Tries to separate training data, maximising the margins correctly
• Can tune this if needed

https://monkeylearn.com

Tuning

• Kmer Size: 6
• Step Size: 3
• Max Features: 5000
• Genome Count per Organism: 100

K = 3, Step = 6, genomes per organism = 100, Max Features=1500

K = 3, Step = 6, genomes per organism = 100, Max Features=5000

DOI: 10.1186/1471-2148-12-171

Testing other classifiers

The Good

K Neighbours

Random Forest

Gaussian Naive Bayes

The OK

Decision Tree

Random Forest

The Ugly

Quadratic Discriminant Analysis

Gaussian Process

Conclusions & Future Work

1. Idea: Encouraging results with diverse inputs
2. Tuning is important to get best results, but is difficult
3. Some classifiers are more accurate than others

Questions?