Since CNN has been widely applied in DNA sequence data, I'm wondering why CNN is not often used for predicting phenotype from SNP data, given that SNPs are essentially parts of DNA sequences and retain the nature of ordering.

For instance, this paper https://arxiv.org/abs/1611.09340 states that "... most of these techniques(CNNs) are based on sequence data where convolutional or recurrent networks are appropriate. When the full DNA sequence is unavailable, such as when data is acquired through genotyping, other methods need to be used..."

Please explain regarding the algorithms rather than the biology side(e.g. SNPs only explains a small portion of genetic variation, etc)

Hey guys, I just wanna remind that I'm addressing the use of SNPs(or you can include other types of variants as well)...not full DNA sequence data. It's very common to use CNN on full DNA sequences to predict variation loci. And it's easy to understand as it's natural to use CNN on sequential data.

But what I don't understand is why not using CNN (1)on SNPs (since SNPs retain the major spatial structure of DNA sequences) (2) to predict phenotype?


1 Answer 1


CNN has been used many times in genomics, what you described is not new. A 5s quick search on Google gave me this paper:


...predict phenotypes from genotypes by using a deep convolutional neural network (CNN)...

"Genotype" is a better terminology than "SNP", as short indels are also important. Your idea is certainly not new, and in fact insufficient as a research topic anymore.

Predicting phenotype from just genotype is old and quite limitied. No good journal will take a topic like that anymore. No matter how good your model is; no matter how many layers you have (e.g. 1000000 trillion layers), how many GPUs you have (e.g. every single GPU in the world), how many engineers you have (e.g. every single ML engineer in the world). CNN will never solve problem perfectly as it neglects the vital information from proteins. Whatever you see from a paper would be limited in applications (e.g. strong assumptions). Genotype data don't tell you how exactly your genes are transcribed.

Instead, the current research trend is on applying deep learning networks integrating mutliple data sources (proteins, transcripts, genes, genotypes etc), and form a single unifed model. This is why we have AI for precision medicine.

  • $\begingroup$ I've read that paper and it is the only paper I've seen using CNN to predict phenotype from genotype. I'll appreciate if you can find a few more papers doing this. $\endgroup$
    – Momoko
    Jul 19, 2018 at 5:36
  • $\begingroup$ @Momoko I need more time to search the literature, but there is a reason you're not finding any high impact paper because genotype itself is generally not enough for predicting phenotype. $\endgroup$
    – SmallChess
    Jul 19, 2018 at 5:42
  • $\begingroup$ And, it's better to focus on what I'm asking about -- regardless of the biology context, why is CNN inappropriate for predicting phenotype from part of DNA sequences (SNPs, indels, inversions, other polymorphisms, you name it)? Since so far people use CNN to predict genetic variants from full DNA sequences rather than what i'm saying. $\endgroup$
    – Momoko
    Jul 19, 2018 at 5:44
  • $\begingroup$ @Momoko I don't want to start a fight, but as a data scientist we need to understand the problem and data. There's no point wasting hours on training a model regardlessly it's CNN, Random Forest, RNN, whatever if the data can't possibility explain the outcomes. Biology is complicated, you need more data to predict phenotypes, at least gene expression!!!! $\endgroup$
    – SmallChess
    Jul 19, 2018 at 5:46
  • $\begingroup$ I think there's a reason regarding the data structure and algorithms $\endgroup$
    – Momoko
    Jul 19, 2018 at 5:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.