Network inference approaches that are based on layering sample-specific information on a fixed set of known network interactions have helped identify sample-specific network properties. However, these methods, by themselves, cannot capture new interactions, which in turn need to be inferred from downstream analyses. In contrast, methods that explicitly infer a single-sample network do not suffer from this limitation. They also can incorporate information from additional related samples to boost sample size, thus enhancing statistical power. This can be helpful when a limited number of samples are available for a biological condition of interest.

Single-sample networks provide a distinct advantage compared to differential network approaches, since they enable comparisons when explicit groups are not available. This can, for example, be the identification of network changes associated with continuous clinical variables or the detection of new subtypes. In addition, analysis of single-sample networks can be corrected for potential confounders in a dataset, which may arise due to technical variation (batch effects) or known clinical features that contribute to network heterogeneity.

One caveat for methods that explicitly infer single-sample networks is that they work by borrowing information from a set of background samples. Therefore, single-sample networks inferred using different backgrounds may differ. Thus, it is important to carefully consider what samples to include in the background. This may be particularly challenging when dealing with heterogeneous datasets. Including samples from different groups (e.g., samples from several disease subtypes) in the background is commonly done when performing comparative network analyses between multiple groups. When the aim is to characterize networks across a homogeneous subgroup of samples, it may be best to only include samples from that subgroup, as the differences between samples may be easier to interpret.

Interpreting single-sample network edge weights can also be challenging, as the inferred weights typically do not explicitly follow the distribution of weights obtained for the aggregate network model. For example, it is unclear what “sample-specific correlation” means as single-sample edge weights inferred from correlations may not necessarily be bounded by [-1, 1]. These different distributions may impact downstream network analysis.

Finally, single-sample networks can be more sensitive to the preprocessing of the input data compared to aggregate networks, as a perturbation in a single sample is used to estimate network edges. This makes it challenging to distinguish between perturbations (and therefore edges) that are true signals compared to random noise, a problem exacerbated when data are very sparse.