The sources of data underlying biological networks

It is important to emphasise that significant challenges arise not only from the sheer size of the dataset used, but also due to the fact that biological datasets are inherently noisy and incomplete. Often, different types of evidence will not overlap or will be contradictory. The way the data was obtained is an important aspect to considere here, with the information typically coming from the following sources:

Manual curation of scientific literature: Scientific curators or domain experts evaluate existing published evidence and store it in a database. This provides high-quality, well represented information, but curation is an expensive and time consuming task and the size of the datasets is limited by these factors.

High-throughput datasets: Some experimental approaches can generate large amounts of data, such as large-scale PPI datasets generated via yeast two-hybrid or affinity purification plus mass spectrometry identification. They provide large, systematically produced datasets but the information suffers the inherent biases of the chosen technique and they vary in quality.

Computational predictions: Many methods use existing experimental evidence as their basis and aim to predict unexplored relationships between biological entities. For example, protein interactions in humans can be used to predict a similar interactions in mice if there are close enough orthologues in this organism. They provide a tool to broaden and even refine the space of experimentally derived interactions, but the datasets produced are understandably noisier than with the previous sources.

Literature text-mining: Multiple algorithms are used to computationally extract systematically represented relationships from the published literature. As with the previous case, although they can greatly increase the coverage of the data, natural language processing is a tricky business and the results tend to be rather noisy.