Deep learning for genetics model of polygenic complex disease.

Question

[michealsmith]

We know in recent several years very quickly deep learning (We could see as advanced multi-layer neural network) has achieved success in picture recognition, which could revolutionize medical diagnosis for example neurological disease diagnosis based on fMRI data.

I'm genetics person for complex disease, so I'm wondering if we could ever use deep learning to generate extremely complicated model for those polygenic complex disease? (Or you can say this is genetic biomarker)

For example, given hundreds of thousands of GWAS genotypes, could we generate complicated deep learning model to accurately predict disease? I know many biostatisticians have tried tons using different machine learning technique like: http://journals.plos.org/plosgenetics/article?id=10.1371/journal.pgen.1003264

But deep learning enjoys higher degree of complexity in modeling, so given their success in fMRI, maybe we could try in complex genetics?

Answer 1

[Philipp Bayer]

That should certainly be possible as neural networks can approximate any function (universality theorem, see this explainer). In fact, people have done what you're proposing: DeepWAS: Directly integrating regulatory information into GWAS using deep learning supports master regulator MEF2C as risk factor for major depressive disorder

Have a look at this list of modern deep learning papers in biology: https://github.com/gokceneraslan/awesome-deepbio

[michealsmith]

Many thanks. I just briefly read DeepWAS, it seems it first implement deep-learning-based DeepSEA to annotate functional SNP, then simply apply logistic regression to classify disease/control. This is good enough, but my question is looking for work that directly apply deep learning for classification between disease and control.