Hello, everyone. I obtained DEGs from RNAseq analysis for normal and infected samples. Then I decreased the number of them by some downstream analysis. Now I have 120 DEGs, and I want to select between them the best combination of biomarkers that can recognize normal from infected samples (biomarker panel). So I want to use machine learning methods (At first, I want to perform feature selection and then draw ROC curve, count MCC, Spe, Sen, ....for the combined set of selected biomarkers by different algorithms such as the neural network and random forest). Because I don't have experience in machine learning, I have some questions. And please let me know if you think I am doing any steps that explain here wrong!
1- What kind of RNAseq files should I enter into machine learning software? count file, FPKM, tpm, or any other files?
2- Should that be normalized?
3- Should the entry be log2 transformed?
4- Can the training and discovery dataset be the same?
5- Is what I write below a correct study design?: The use of a dataset for obtaining DEGs then, partitioning it into k subsets of equal size. Of the k subsets, a single subset is retained as the test data set. The remaining k - 1 subset is used as training data sets. The cross-validation process is then repeated k times, with each of the k subsets used exactly once as the test data. The k results from the k iterations are averaged (or otherwise combined) to produce a single estimation. And then performing a test for the model with an external dataset to validate the model.
6- Can the validation dataset be from a different technology like microarray? Is any pre-processing needed for the datasets to be tuned before performing machine learning methods in this case?
Thank you to answer my questions
Have a look to find the answers of your questions at:
https://www.nature.com/articles/s41467-020-17678-4
Article Biological classification with RNA-Seq data: Can alternative...
https://core.ac.uk/download/pdf/228470434.pdf
Good luck.
Thank you Imtiaz Husain for introducing these articles.
In the second article, It is stated that the normalization should not be performed before machine learning but it has written a sentence that had made me confused:
TCGA expression values were calculated based on (i) raw counts and (ii) log2 normalized counts with respect to the gene length and sequencing depth. The results showed that there was a strong preference, in terms of accuracy, in raw counts for the gene-based classifiers, but to a lesser extent for the transcript-based models. However, the opposite was observed where transcript-based models were more accurate when using log2 normalized counts
I think the sentence before "however" is in conflict with the sentence after it!
In the sentence before "however" it says for both gene-based classifier and transcript-based models raw count is preferred. But after the word "however" it says in transcript-based models using log2 normalized counts were more accurate!!!!
Do you know what is the meaning of this part of the article?
I would recommend normalized counts data as input to machine learning. The goal is to find trends/patterns in the data, but to do so, you need the sample-level data to work from (the more samples the better). I would recommend keeping the data limited to one type (RNA-seq OR microarray) since the numerical values and biases differ substantially.
If you use a random forest machine learning approach you use, you may not need to perform cross-validation, but otherwise it's a good idea. I would recommend subsetting the original dataset into 70/30 or 80/20 for training and testing.
Brett E Pickett
Thank you to answer. Shouldn't it be log2 transformed? What type of normalization method do you recommend?
fpkm values of your DEGs
Have a look on data from TCGA database. The same way you can prepare your data.
FPKM is not a perfect normalization method. I'd suggest you extract normalized counts from DESeq2 then prepare the dataset for ML.
Best!!
- Badges
- Science topic
- Similar topics
- Filing