I'm looking for biological datasets with specific properties that are available for me to analyze. Informally put, I have a hammer and I am looking for a nail. That is to say, I have an algorithm and I am looking for suitable data to test if it can tell us anything interesting about biology. Let me start by characterizing the properties that I'd like this data to have.

Math Properties

The data must be representable as signed directed graphs, which in less formal terms means that I am looking for a network where the connections between the nodes are arrows with a +/- sign associated with them.

The signed property of the edges must represent a property of the relation between discrete objects that only takes on one of two values. One way to get signed values from quantitative data would be to use the sign function, keeping only the relations whose sign of the quantitative variable were not equal to zero. Another way is to encode Boolean values, where False is - and True is +. I'd like to further clarify that the data should not only conform to this criteria, but should have instances of both positive and negative paths within the data.

The directed property of the edges mean that we care to distinguish 'A relates to B' from 'B relates to A'; order matters. Both, one, or neither of these relations can hold in these networks between any two nodes A and B.

Data Properties

The data must be analyzable in the sense that I can obtain the data in a form where I can either (1) immediately analyze the data or (2) transform/manipulate the data into a form that I can analyze. As a computer programmer, I'm likely to be able to find packages for (or code for myself) moderately arbitrary transformations as long as they are computable and interpretable. That is to say, in your answer I would like you to not worry too much about whether you have a ready-to-use software solution for analyzing or transforming the data. What I would like considered is whether there is (1) a direct download for a/the datafile(s) or (2) an application programming interface (API) for downloading the data in either pieces or as a whole.

Preference will be given to answers that do not cost money, and do not require creating accounts (although either are acceptable).

All else being equal, I'd prefer larger datasets (n > 1000) over smaller ones.

Preference will also be given to answers where the dataset has additional variables are available on the vertices. For example, if the vertices represent genes, then additional variables on the vertices would mean more data is available that describe those genes.

Academic Properties

The dataset(s) should be citable.


While I'm open to other biological subjects, there are two biological topics that I feel should be representable in the mathematical way I've described above.

The first is gene regulation data. The vertices would be genes, and for any two genes A and B we could have any of the following relations

  • Gene A downregulates Gene B
  • Gene A upregulates Gene B
  • Gene B downregulates Gene A
  • Gene B upregulates Gene A

The second example I've thought of is in biological signaling as the signals are sometimes directional (sender to receiver) and might have some sort of sign (turn on/off, increase/decrease).

Bonus points if the data is topical (COVID-19/coronavirus-related for example).

Data Sources

I've downloaded a few datasets from BioGrid, but I didn't see the kinds of mathematical properties I was looking for in those datasets. If there are datasets here that you think meet my criterion, please mention them.


3 Answers 3


While this does look like a very nice approach, I would actually argue that there are (almost) no biological datasets, that can be directly used for your algorithm - you can only do a sort of meta analysis using results/predictions based on other datasets.

The reason for this is that the nature of a signed graph requires you to look at effects that can have opposite directions, which in biology pretty much restricts you to regulatory functions*. The problem with that is, that is nearly impossible to directly measure regulatory effects. In all cases I am aware of, one has to measure the magnitude of an effect under different conditions and then calculate regulatory interactions based on that data.

This means any set of regulatory interactions you can put out into your network is already based on a given analysis with inherent assumptions and (potentially) biases.

This doesn't make the approach impossible, but I do think, it is something you need to be aware of, especially if you want to use multiple datasets.

As for actual datasets that would fit well:

  • transcription factors have already been mentioned in another answer
  • other regulatory effects on transcriptions (promoters / enhancers) may also work, though in this case the 'class' of the objects effected (genes/mRNA) is different than the effector
  • Another type of biological network that might fit are metabolic / enzymatic pathways. For these you could either also look at regulatory interactions (which again may be tricky because both proteins and metabolites can have regulatory effects) or maybe also directly at metabolic flow: in this case nodes in the graph would represent metabolites and edges reactions that produce or consume them. The KEGG database has large numbers of metabolic pathways for many different species.

*I was trying to think of other things, but presence/absence of a signal probably doesn't fit and neither does classification into strong/weak signals (i.e. interactions)

  • $\begingroup$ Thank you for the suggestions. I appreciate your laying out of the difficulty of getting such datasets. I didn't mention in the original post, but it was a similar line of reasoning that led me to put this question to the SE community. $\endgroup$
    – Galen
    Commented Apr 3, 2020 at 14:47
  • $\begingroup$ @Galen I think you need a collaborator in the field. It's super dangerous (or at least embarrassing) when people with an algorithm enter a domain they don't understand without help. SE is not a replacement for collaboration. $\endgroup$
    – Bryan Krause
    Commented Apr 3, 2020 at 15:12

This is a great biological question! It asks a lot about how empirical science is done in the field of modern biology. I'm glad we encourage such questions from curious people who want to learn more.

Here's some directed network graph data for regulatory networks of transcription factors (TFs), via http://www.regulatorynetworks.org/:

• Mouse TF: http://www.regulatorynetworks.org/results/networks/mm9/networks.v12032013.tgz

If you click on the site's About button, there is more information about this file. From the README, for instance:

Regulatory interactions were identified in each cell type by scanning the proximal DHSs (+/-5kb from the canonical transcriptional start site) of each transcription factor gene DNaseI footprints corresponding to the recognition sequences of known TFs.

The file genes-regulate-genes.txt contains two columns of gene names, at column 4 and column 5 of the text file.

The gene in the first column is bound (i.e., contains a DNaseI footprint in proximal regulatory DNA) corresponding to the second column's TF gene product.

The final column (column 6) represents the number of times a motif for the regulating gene is found in the regulatory region (promoter) for the regulated gene.

The gene names are HGNC symbols and are easy to look up in resources available through UCSC, GeneCards, Ensembl, etc.

DS-numbers are internal identifiers for cell lines. However, the prefix uses that cell line's more general name and is usually easily searchable or parsable.

Disclaimer: I created the regulatorynetworks.org site and I was a contributing author on a couple papers which used the datasets rendered in it (1, 2).

  • $\begingroup$ Thanks for the suggestion. It looks like this data has directed edges. Can you elaborate on how the edges are also signed? $\endgroup$
    – Galen
    Commented Apr 3, 2020 at 14:45
  • $\begingroup$ These are all positive values. $\endgroup$ Commented Apr 3, 2020 at 16:16
  • $\begingroup$ Ah, I see what you mean, you're quite right. I'll clarify in the question above that I'm looking for data that contain instances of both positive and negative paths. Sorry for the confusion. $\endgroup$
    – Galen
    Commented Apr 3, 2020 at 16:20
  • $\begingroup$ To augment this with data that contain negative values, ChIP-seq peak overlaps may work. For TFs for which there are ChIP-seq peak datasets (on ENCODE, for instance: encodeproject.org), you could use set operations (e.g., BEDOPS: bedops.readthedocs.io/en/latest) to call overlaps of these proximal promoters with TF-specific ChIP-seq peaks as +1/-1, to specify overlap/non-overlap, true/false categories. ChIP-seq peak data would be useful, because it lends biological evidence to support that there is TF binding in the promoter (and associated regulation of the target gene). $\endgroup$ Commented Apr 3, 2020 at 18:04
  • $\begingroup$ That's just spitballing, though. The networks in my answer contain genomic intervals, and you could do set operations against any number of interval-based datasets, in order to get "signed" data. For instance: overlap/non-overlap with SNPs, overlaps with conservation values averaged over the TF binding site, etc. Conservation might be useful if you want to associate nodes with +/- values; I think either the phyloP or the phastcons conservation scores are ranged positive to negative, to specify the selection pressure over genomic regions: from highly conserved, to neutral, to highly evolving. $\endgroup$ Commented Apr 3, 2020 at 18:11

Enrichr has a huge (164 libraries) collection of term-to-gene relations in machine-readable format.

X2K can be more interesting for you, as it partially implements networks approach. It has a large collection of libraries as well. And two more - KEA3 (kinases to genes), and ChEA3 (transcription factors to genes).


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .