Data Bias in protein-protein interactions data

Question

Databases like STRING have so many predicted as well as experimental interactions data for many organisms.However there is a bias in the data which they have?.

a)The first is gene co-occurrence which is more like a circular argument.

b)And the second one is that all interactions that are predicted computationally are tested for experimental interactions.This leads to a bias in the interaction data available for organisms in STRING.

How do we overcome this data bias while using STRING interaction data for analysis purpose?

a)Avoiding evidence channels of gene co-occurrence is one way.Is there any other way other than this?

b)How to statistically say that the bias is not governing the data or any way to test statistical significance of the data bias?

Answer 1

I'm one of the STRING maintainers, so I hope I can clear things up a bit.

Gene co-occurrence: genomic context methods are very useful for bacterial genomes, but not so much for eukaryotes. However, they are actually the most unbiased data we have, since there's no human influence (beyond selecting species for sequencing).

Benchmarking: Interactions predicted from text-mining etc. are benchmarked against the KEGG database. However, this doesn't necessarily lead to biases, as this benchmarking is basically a calibration step.

Actual biases: For eukaryotes, most interaction data comes from the experimental, database and text-mining channel. All of these are biased in various ways: e.g. there are study biases (disease genes will get more attention) and technical biases (some proteins are hard to work with)

As we don't know the actual set of protein-protein interactions, we can probably not test for biases in the data.