Also see my website page
The aim of this project is to combine small cell free RNA sequencing techniques with machine learning to identify promising candidate biomarker for cancers. We construct the whole pipeline from mapping to expression matrix construction( full length and domain feature) to feature selection and propose a robust feature selection method to identify potential biomarkers.
This repo stops updating since we are writing a paper based on the work.
We use sequential mapping strategy for small RNA data (HCC exRNA and GSE71008)
We use fast mapping order test method to map in thousands of different orders in a very short time and decide the order.
Considering the characteristic of small RNA and exRNA. We think it is better to use domain features instead of full length RNA. So we also do peak calling to construct both full length and peak expression matrix.
All the datasets need to be normalized and remove batch effect.
I test several normalization method:
- CPM
- remove piRNA and miRNA
- Use several reference miRNA
- SCNorm
- TMM
- RLE
I also use batch removal method including RUVs and Combat.
Our aim is to establish a robust feature selection machine learning model to select very few feature as candidate cancer biomarkers.
- Normalize domain coverage
- Scale each feature (log) independently (using z-scores, min-max, robust normalization)
- Run a classifier (Random Forest, Logistic Regression, Linear SVM) to select features based on feature importance. Optimize hyper-parameters by 3-fold cross-validation.
- Optionally, use a recursive feature elimination (RFE) to eliminate features.
- Repeat feature selection for 100 times. Randomly remove 10% samples in each run (Jackknife resampling) or do shuflle split.
- Select features that are recurrently selected across resampling runs (>50%)
- Refit the classifier on selected features
For detailed illustration of our pipeline please see Work Summary
As teaching assistant of Bioinformatics Basics. I use part of our work to assign the final project of this semester's course.