VirLab

Contributers: Gabriel Steinberg (gsteinb1@binghamton.edu), Kenneth Chiu (kchiu@binghamton.edu), Anna Levenberg (alevenb1@binghamton.edu), Janis Louie (jlouie2@binghamton.edu), Len Kaupert (lkauper1@binghamton.edu)

Past Contributers: Hayden Brown (hbrown10@binghamton.edu), Yan Ma (yma73@binghamton.edu)

Vector identification in metagenomic data using k-mers

• The goal is to find the vector of a disease with an unknown vector by training an SVM on genomes with known vectors.
• Training on complete genomes of diseases vectored by Aedes and Culex mosquitoes
• Testing on simulated reads of complete genomes of diseases vectored by Aedes and Culex mosquitoes

Main Files

1. fasta_parser.py
   • Parses input data and returns an array "genomes" with genome objects populated with vector, disease, and sequence for each genome
2. k_mer_creator.py
   • Populates each genome object with all k-mers for that sequence
   • For every k-mer in all genome objects, it adds a placeholder k-mer (k-mer: 0) for that k-mer all other genomes. So every genome has every k-mer that every other genome has.
   • Split into training & testing sets
   • Make fasta files with testing data for BBMap
   • Simulate reads from test genomes
   • Find significant kmers for testing and training sets and create CSV's with them
3. kruskal_wallis.py
   • Tests if a kmer can help distinguish two classes by analyzing the difference in the medians
4. SVM.py
   • CLassification using a Support Vector Machine

Instructions

1. Clone this repository into your home directory
2. Install BioPython and skikit learn
   • https://biopython.org/
   • https://scikit-learn.org/stable/
   • To install the dependencies, use pip3 install -r requirements.txt
3. For Windows: Install Java in your Linux subsystem for the BBMap script to work
4. Run python3 current_code/k_mer_creator.py to generate files
5. Run SVM.py to classify

TODO

• Need a way for SVM to return an "i dont know" - distance from the vector
• Turning reads into contigs and test on those
• Convert some nested for loops to list comprehensions for speed and readability
• Adding script for one hot encoding

Janis

• Reproduce stats for full genomes and record them
• Reproduce stats for reads (should have high error rate
• Produce results for classification with reads

Anna

• Fix the cnn input shape sizes. Get CNN working for 300 samples
• Get it to work for a bunch of genomes of different sizes