This repository contains my work in tetramer nucleotide frequency dimension reduction. The purpose of these scripts is to take any given fasta input, chop it up, encode each fragment as a 256 dimension vector, then graph/cluster them.
I came across the concept of shared nearest neighbor which should work better than KNN in high dimension space. Might be interesting to implement it and see what happens. This technique hinges on the Jarvis-Patrick neighborhood metric. This may be a repository that implemented the approach and made it available with scipy KNN https://github.com/albert-espin/snn-clustering.
Currently it is 8/28/2022, and the last time I worked on this was in 10/2021, and I didn't make a readme or conda env then.
I believe the workflow is to use genome_to_tetramer.ipynb to turn a given .fa file into a csv table where rows are fragements and columns are tetramers. I have some clustering tutorials that I followed, but I want to start fresh now that I've used Seurat.
Input: csv of tetranucleotide frequencies Step1: Normalize the columns into Z-scores Step2: Perform PCA Step3: Peform t-SNE on PCs rather than raw data Step4: Find clusters using any method I want! K-means, heiarchical, or whatever seurat does.
I will put this into a file called HoG_workflow.ipynb
I also added the stats.yml which should set up the proper conda env with the command conda env create -f "stats.yml".