1
$\begingroup$

I am looking for publicly available data for a genomics deep learning project. My goal is to compare different architectures to predict biological insights from DNA sequences.

I have heard about Janggu before, but I am looking for a few datasets which are like the MNIST equivalents in genomics. Kaggle doesn't seem to have many such datasets, and I would be happy if you share your insights about potential datasets that one could use.

For example, it could be datasets that map 100-Mers to their ChIP-seq values or datasets that annotate a DNA sequence with the probability that a transcription factor would bind to it.

Choong and Lee lists a few datasets on their IEEE paper from 2017:

enter image description here

$\endgroup$
5
  • $\begingroup$ related: researchgate.net/post/… $\endgroup$
    – 0x90
    Nov 22 '20 at 21:34
  • 1
    $\begingroup$ If you add more details about the type of deep learning you're evaluating we will be better equipped to direct you to resources. If you're doing something like compositional data analysis, check out this paper and the list of 45 datasets with links on the last page. $\endgroup$ Nov 23 '20 at 5:38
  • $\begingroup$ @jared_mamrot thank you. I would like to train networks that take chip-seq data and DNA sequences as inputs. The prediction layer may be any biological signal one could think of? $\endgroup$
    – 0x90
    Nov 23 '20 at 5:46
  • 1
    $\begingroup$ Genomics is a huge field; you need to be more specific. If you're looking for everything with regards to publicly-available datasets, then there's GenBank $\endgroup$
    – gringer
    Nov 23 '20 at 21:07
  • $\begingroup$ David Reich Lab reich.hms.harvard.edu/datasets $\endgroup$
    – 0x90
    Nov 24 '20 at 17:48
2
$\begingroup$

Good question and the Janggu link is cool. Thanks!

Kaggle is/was heavily image recognition. There isn't an MNIST type data set in genomics, in case anyone is wondering this is a classic natural language processing (NLP) type analysis to identify handwritten single digit numbers.

There has been several FDA precision challenges, one being described here:

FDA precision challenges

This could provide something towards a 'standard'.

A key issue remains the accurate vectorisation of sequence and whether you think that is/was satisfactory in my personal opinion.

There is an old review article here

$\endgroup$
1
  • $\begingroup$ Since I have a new method to address this problem. I would like to find a dataset which is like MNIST to evaluate my idea. $\endgroup$
    – 0x90
    Nov 22 '20 at 22:03

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