Hello David Kim ,

If there is no defined "control group", it is possible that you are dealing with a dual-channel experiment (provided values are already a result of the "control vs experiment"). In those cases, the analysis protocol is different from the single-channel ones. Nevertheless, they still provide a DEG list. If you are using R to extract such results, use the limma package, it has a defined protocol for dual-channel experiments. If the dataset is not dual-channel, and the authors only provided the experimental samples, there is no valid way to get a DEG list. Your results would be severely compromised in numerous ways.

You mentioned a "machine learning dataset" you want to make. I can help you with some advice on this topic. Care to elucidate?

I have some articles on this topic:

[1] Grisci, BI, Feltes BC, Dorn M. Neuroevolution as a Tool for Microarray Gene Expression Pattern Identification in Cancer Research. Link: Article Neuroevolution as a Tool for Microarray Gene Expression Patt...

- In this article, we created a classification tool for microarray data, based on FS-NEAT.

[2] Article CuMiDa: An Extensively Curated Microarray Database for Bench...

- This is a microarray database we created, exclusively made to be used in machine learning protocols. All datasets were previously curated, treated with state-of-the-art expression analyzes, and treated to be used as inputs for both classification and clusterization tools. It might be useful to you.

Cheers.

The problem without a reference of some sort is all you have is normalized expression values. But you have no means of knowing what the treatment actually altered or affected in that expression data set, so no way to have any context to any inferred changes in pathway elements. The point of a control sample in ontology analyses is to provide the context of what changed in the treated samples, so you can then analyze those altered transcription profiles for what they indicate functionally in terms of ontology enrichment. So anything you do with the expression data you have and any ontology is immediately suspect as you have no idea if you're looking at a treatment effect or a normal transcriptional response. Imagine trying to infer processes indicative of cellular proliferation or induction of cholesterol biosynthesis due to treatment by something, without knowing anything about the untreated expression levels for those same genes in that same animal, tissue or cell line? Cannot be done.

You must have some associated "feature" for the samples. In layman terms say you got expression values for a gene X for 100 samples and out of these 100 say 30 have some y1 property (phenotypical probably) and other 70 have y2 property. You could, in theory use DEG for these two property groups.

Also, what kind of machine learning study are you intending to perform: a classification problem or a regression problem?

Chakit Arora It is a classification problem(up-regulated, down-regulated)

Do you have defined classes for every sample?

i.e for any sample you must have gene expression values as well as a class (up-regulated or down-regulated) that has been attributed a priori based on your study design and/or literature survey (this step is the most important one). If yes, only then you could proceed to develop ML based classification models. DEG based analysis can only help you to filter out important genes (i.e genes that are differentially expressed between the two classes with a significant p-value) which could then be used as features for ML training and validation purposes.

Chakit Arora We don't have dataset like you mentioned. So we should find dataset for training a ML model. This is why i posted this question. I try to search databases for this purpose.

There is no shortage of suitable microarray and RNA-seq gene expression studies suitable for your proposed sort of work. Just search NCBI's GEO database to start with. But the issue is you need two groups - one to serve as the baseline for changes in the other. Then you can use those changes to infer what cellular pathway(s) have been affected by the perturbation.

Dear David Kim, in the scope of microarray data for machine learning, may I suggest CuMiDa (Curated Microarray Database), a collection of cancer gene expression datasets created for the purpose of ML and bioinformatics research: http://sbcb.inf.ufrgs.br/cumida

Article CuMiDa: An Extensively Curated Microarray Database for Bench...

Cheers.