Protein interaction networks are a standard resource in target discovery and computational biology. The idea is straightforward: a protein that interacts with many disease-relevant proteins, or that sits at a central node in a disease-relevant pathway, has a higher prior probability of being relevant to that disease than a protein with no known interactions in the relevant pathway. In practice, the network evidence is more complicated and more biased than this framing suggests.
The Detection Bias Problem
Protein interaction databases are not surveys of the actual interaction landscape of the proteome. They are records of what researchers have chosen to study. The proteins with the most annotated interactions in databases like STRING, BioGRID, or IntAct are predominantly the proteins that have been most extensively studied: well-known signaling proteins, transcription factors, proteins involved in major disease areas like cancer and neurodegeneration, and proteins with commercial reagent support that makes them easy to study.
This creates a systematic bias: highly studied proteins appear to be more highly connected not because they are more biologically connected but because more researchers have looked at their interactions. A protein at a similar biological hub position in a less-studied pathway will appear to have fewer interactions simply because fewer researchers have run the experiments needed to detect those interactions.
The practical consequence for target discovery is that using interaction network connectivity as a proxy for target importance will favor targets in well-characterized biological domains and systematically undervalue targets in understudied areas. This is the opposite of what a competitive drug discovery strategy often wants: the most heavily networked targets in the databases are typically the most pursued targets in pharmaceutical pipelines.
Different Interaction Types Have Different Evidentiary Value
Protein interaction data is heterogeneous in a way that aggregate network analyses often obscure. A physical binding interaction detected by co-immunoprecipitation in a relevant cell type is a different quality of evidence than a computational prediction based on sequence homology to a yeast protein pair. A genetic interaction detected in a two-hybrid screen is different from a functional dependency measured in a CRISPR screen. Proximity ligation data from BioID experiments captures transient and spatial proximity that may or may not reflect functional interaction.
Most protein interaction databases aggregate these evidence types into a single network, often with a confidence score that combines evidence from multiple sources. That confidence score is useful as a first pass, but it can obscure important distinctions. A high combined confidence score might reflect ten low-quality experimental observations rather than one high-quality observation. Understanding what the underlying evidence actually is, rather than trusting the aggregate score, requires going back to the primary data.
Pathway Context Matters More Than Degree
Network topology metrics like node degree (number of interactions), betweenness centrality (frequency of appearance on shortest paths between other nodes), and eigenvector centrality are often used as proxies for functional importance. The assumption is that more central nodes in the network are more essential to network function and therefore more likely to be disease-relevant targets.
This assumption has partial empirical support: essential genes do tend to be more connected in protein interaction networks than non-essential genes. But the relationship is noisy, and the detection bias problem described above means that the most connected genes in databases are not necessarily the most biologically central genes -- they are the most studied genes.
A more useful approach is to ask not about network topology globally but about the relationship between a specific target and the specific disease process of interest. Is the target in a pathway whose activity is dysregulated in the disease? Does the target interact with proteins already established as disease drivers? Are there interaction partners of the target that have genetic associations with the disease? These disease-contextualized questions use the network data more selectively and are more likely to yield actionable insights than global topology metrics.
Using Interaction Networks Appropriately
The right use of protein interaction network data in target discovery is as one evidence layer among several, not as a primary ranking criterion. Network connectivity evidence should be weighted proportionally to the quality of the underlying interaction data, filtered to interactions detected in biologically relevant experimental contexts (not just any cell line), and interpreted in the context of what is known about the specific pathway's role in the disease.
When interaction data and omics data agree -- a target that is upregulated in disease tissue and sits at a central position in a relevant pathway -- the agreement is meaningful. When they disagree -- a highly connected protein with no expression change in disease -- the disagreement is also meaningful and should prompt specific questions about why the apparent network importance does not manifest as a transcriptomic or proteomic signal in the disease context.
The Understudied Target Opportunity
The detection bias in protein interaction networks points to an opportunity that is easy to overlook. Targets in understudied biological domains will systematically appear less connected in network databases than their actual biological relevance warrants. A systematic analysis that specifically seeks targets with strong genetic evidence for disease involvement but low database connectivity -- indicating that the target is understudied rather than unconnected -- can identify candidates with high potential and lower competitive density.
This requires looking beyond the network data to the underlying experimental evidence and being willing to act on evidence that has not been extensively corroborated in the literature, with appropriate acknowledgment of the higher uncertainty. That tradeoff is often worth making when the competitive landscape and the specifics of the disease biology suggest that understudied targets are genuinely undervalued.
Correcting for Study Bias Computationally
Several approaches have been proposed to correct for the study-bias problem in protein interaction networks. One is to weight interaction evidence by the number of independent experimental studies that have detected the interaction, preferring interactions with multiple independent confirmations over those detected in a single study. Another is to apply a prior correction that downweights interactions involving highly-studied proteins and upweights interactions in less-studied regions of the network, on the assumption that the study bias is correlated with degree.
These corrections are imperfect but useful. They reduce the extent to which analysis conclusions are driven by which proteins happen to have been studied most, making the results more likely to reflect actual biological connectivity rather than historical research priorities. When comparing analysis results with and without study-bias correction, the differences are often substantial, which is itself informative: if the conclusions change significantly depending on whether you correct for study bias, those conclusions deserve extra scrutiny.