Metascape provides a rather unique protein-protein interaction (PPI) network analysis capability. In many gene list analysis resources, PPI analysis results in a rather massy hairball network. Besides stating such networks are statistically significant, there is not much biologists can say about such networks. To infer more biologically interpretable results, Metascape applies a mature complex identification algorithm called MCODE to automatically extract protein complexes embedded in such large network. Then taking advantage of Metascape’s functional enrichment analysis capability, it automatically assigns putative biological roles of each MCODE complex. Such analyses are very computational intensive and cannot be easily computed even by bioinformaticians. Regardless of its advanced PPI analysis algorithm, the results still heavily determined by the quality of its underlying PPI database.
Analyzing the publications citing Metascape, we found many users use STRING database for PPI analysis. Indeed STRING is probably the most comprehensive PPI data source, therefore, tend to provide a denser and oftentimes better looking network. The main reason Metascape has not included STRING is because we have not found a good way to cross compare STRING with other PPI data sources not yet included in STRING, especially we believe data sources such as OmniPath and InWeb_DB (the latter is no longer accessible to the public, therefore Metascape only uses an old snapshot) are presumably of higher quality than most STRING data. All interactions in STRING has a quality score, therefore, one can prioritize and use only the high-quality subset, however, we are not able to assign similar scores to interactions not yet captured by STRING. In the latest Metascape release, we now propose a way to compile an integrated PPI database including STRING, BioGrid, OmniPath and InWeb_DB. We believe this is an important step forward to significantly bring greater value of Metascape PPI analysis to our users.
Physical Interactions and Genetic Interactions
There are two types of protein-protein interactions: physical interactions and genetic interactions. “Genetic interactions capture functional relationships between genes using phenotypic readouts, while protein-protein interactions identify physical connections between gene products” [ref]. Physical interaction means two proteins are biochemically bond, either directly or through a complex. Genetic interaction more refers to functional interaction, such as regulation, so we will call them functional interactions as well. Oftentimes, genetic interactions include observations derived through computational means, therefore, they tend to be less accurate and potentially are more agreeable with Gene Ontology (therefore, less of a truly orthogonal data source). In BioGrid, these two types are counted independently [see link] and we often only use physical interactions to get results that are more conservative. Many STRING users tend to ignore the differences and apply both sets to their data, therefore, their STRING networks do appear denser. We do not believe there is a straightforward answer on either using physical only or combining both interaction types . If the physical-only network is already sufficiently dense, we should use it as it is more reliable and provides evidence more independent from the GO enrichment analysis. However, if the physical-only network is too sparse, a combined network is needed in order to gain useful biological insights.
Evidence Score for Non-STRING Data
STRING provides a probabilistic framework to assign a confidence score for each PPI pair, by assuming all evidences are independent. We therefore can assign both a physical score and a combined score for its data record. But how to assign a score to data not captured in STRING, so they can be combined?
First, for those PPI pairs that are already included in STRING, we check their STRING physical scores. The figure below shows the physical score distribution of BioGrid physical subset, BioGrid functional subset, OmniPath, InWebDB and STRING physical subset itself using human data. Notice these are accumulative curves for their score distributions. We can see about 50% of the PPI data in OmniPath and InWeb_DB have a physical score > 0.9, i.e., these two data source indeed are of high quality even by their STRING physical scores. Then BioGrid physical subset has better quality than its functional subset and STRING subset has the lowest quality. I.e., in terms of data source quality, we can say OmniPath > InWeb_DB > BioGrid (Physical) > BioGrid (Functional) > STRING (Physical), in line with we expected.
Now since we cannot assign individual STRING scores to those pairs that are not already in STRING, we can only assume all data in non-STRING data sources share the same STRING physical score. We subjectively choose the score corresponds to ~33% percentile (1/3 of the height in the accumulative curve) of the above distribution. That is we set OmniPath, InWeb_IM and BioGrid (Physical) a STRING physcial score of 0.537, 0.356, 0.260, respectively. Then we take the 33% percentile of the STRING physical distribution itself, 0.132, as the cutoff. Therefore, all physical interactions with STRING score > 0.132 are consider a reliable subset, which we call “Physical (Core)”. “Physical (Core)” include all of OmniPath, InWeb_DB, BioGrid Physical and 2/3 of STRING Physical. Then all physical interactions, regardless of their STRING scores are included in the “Physical (All)” dataset.
Similarly, if we use combined dataset, we can assigned STRING combined score of 0.537, 0.356, 0.260, 0.221 to OmniPath, InWeb_IM, BioGrid (Physical), BioGrid (Functional), respectively. We use a cutoff of 0.187, corresponding to 1/3 of the STRING Physical, to divide the combined dataset into “Combined (Core)” and “Combined (All)”, where 2/3 of STRING interactions are retained in the Core subset.
Note: Be aware that we derive these cutoffs based on human data and assume they are applicable to all organisms. Other organism contains fewer data records, therefore, we avoid making an organism-specific threshold.
Scope of the New Database
It is exciting to report that by including STRING in Metascape, the size of our PPI database has increased significantly. Below is the Venn diagram for human, where STRING contributes >2 million new human physical PPI pairs not covered by all previous data sources.
The same goes to the size of PPI dataset, when functional data are included. The figure below shows STRING totally contributes >5 million new PPI pairs for human.
Underlying Support
For all the networks generated by Metascape, we now include an edge property called “support”. This allows users to examine the origin of each interaction pair. An example support reads like:
This is a very confidence interaction that is supported by OmniPath, InWeb_IM, BioGrid and STRING. The STRING physical score is 0.896.
Summary
We combined all data from STRING, OmniPath, InWeb_IM and BioGrid to produce four datasets: Physical (Core), Physical (All), Combined (Core), and Combined (All). OminiPath, InWeb_IM and BioGrid Physical data are consider high quality and included in all datasets. Only physical interactions are in the Physical datasets. The Core dataset contains the 2/3 of higher-scoring corresponding STRING data. Metascape “Express Analysis” defaults to “Physical (Core)” to be conservative at this point (subject to change in the future), but savvy users can choose any of the four flavors through “Custom Analysis”.
Protein-Protein Interaction Data Sources
Why Include STRING Database
Metascape provides a rather unique protein-protein interaction (PPI) network analysis capability. In many gene list analysis resources, PPI analysis results in a rather massy hairball network. Besides stating such networks are statistically significant, there is not much biologists can say about such networks. To infer more biologically interpretable results, Metascape applies a mature complex identification algorithm called MCODE to automatically extract protein complexes embedded in such large network. Then taking advantage of Metascape’s functional enrichment analysis capability, it automatically assigns putative biological roles of each MCODE complex. Such analyses are very computational intensive and cannot be easily computed even by bioinformaticians. Regardless of its advanced PPI analysis algorithm, the results still heavily determined by the quality of its underlying PPI database.
Analyzing the publications citing Metascape, we found many users use STRING database for PPI analysis. Indeed STRING is probably the most comprehensive PPI data source, therefore, tend to provide a denser and oftentimes better looking network. The main reason Metascape has not included STRING is because we have not found a good way to cross compare STRING with other PPI data sources not yet included in STRING, especially we believe data sources such as OmniPath and InWeb_DB (the latter is no longer accessible to the public, therefore Metascape only uses an old snapshot) are presumably of higher quality than most STRING data. All interactions in STRING has a quality score, therefore, one can prioritize and use only the high-quality subset, however, we are not able to assign similar scores to interactions not yet captured by STRING. In the latest Metascape release, we now propose a way to compile an integrated PPI database including STRING, BioGrid, OmniPath and InWeb_DB. We believe this is an important step forward to significantly bring greater value of Metascape PPI analysis to our users.
Physical Interactions and Genetic Interactions
There are two types of protein-protein interactions: physical interactions and genetic interactions. “Genetic interactions capture functional relationships between genes using phenotypic readouts, while protein-protein interactions identify physical connections between gene products” [ref]. Physical interaction means two proteins are biochemically bond, either directly or through a complex. Genetic interaction more refers to functional interaction, such as regulation, so we will call them functional interactions as well. Oftentimes, genetic interactions include observations derived through computational means, therefore, they tend to be less accurate and potentially are more agreeable with Gene Ontology (therefore, less of a truly orthogonal data source). In BioGrid, these two types are counted independently [see link] and we often only use physical interactions to get results that are more conservative. Many STRING users tend to ignore the differences and apply both sets to their data, therefore, their STRING networks do appear denser. We do not believe there is a straightforward answer on either using physical only or combining both interaction types . If the physical-only network is already sufficiently dense, we should use it as it is more reliable and provides evidence more independent from the GO enrichment analysis. However, if the physical-only network is too sparse, a combined network is needed in order to gain useful biological insights.
Evidence Score for Non-STRING Data
STRING provides a probabilistic framework to assign a confidence score for each PPI pair, by assuming all evidences are independent. We therefore can assign both a physical score and a combined score for its data record. But how to assign a score to data not captured in STRING, so they can be combined?
First, for those PPI pairs that are already included in STRING, we check their STRING physical scores. The figure below shows the physical score distribution of BioGrid physical subset, BioGrid functional subset, OmniPath, InWebDB and STRING physical subset itself using human data. Notice these are accumulative curves for their score distributions. We can see about 50% of the PPI data in OmniPath and InWeb_DB have a physical score > 0.9, i.e., these two data source indeed are of high quality even by their STRING physical scores. Then BioGrid physical subset has better quality than its functional subset and STRING subset has the lowest quality. I.e., in terms of data source quality, we can say OmniPath > InWeb_DB > BioGrid (Physical) > BioGrid (Functional) > STRING (Physical), in line with we expected.
Now since we cannot assign individual STRING scores to those pairs that are not already in STRING, we can only assume all data in non-STRING data sources share the same STRING physical score. We subjectively choose the score corresponds to ~33% percentile (1/3 of the height in the accumulative curve) of the above distribution. That is we set OmniPath, InWeb_IM and BioGrid (Physical) a STRING physcial score of 0.537, 0.356, 0.260, respectively. Then we take the 33% percentile of the STRING physical distribution itself, 0.132, as the cutoff. Therefore, all physical interactions with STRING score > 0.132 are consider a reliable subset, which we call “Physical (Core)”. “Physical (Core)” include all of OmniPath, InWeb_DB, BioGrid Physical and 2/3 of STRING Physical. Then all physical interactions, regardless of their STRING scores are included in the “Physical (All)” dataset.
Similarly, if we use combined dataset, we can assigned STRING combined score of 0.537, 0.356, 0.260, 0.221 to OmniPath, InWeb_IM, BioGrid (Physical), BioGrid (Functional), respectively. We use a cutoff of 0.187, corresponding to 1/3 of the STRING Physical, to divide the combined dataset into “Combined (Core)” and “Combined (All)”, where 2/3 of STRING interactions are retained in the Core subset.
Note: Be aware that we derive these cutoffs based on human data and assume they are applicable to all organisms. Other organism contains fewer data records, therefore, we avoid making an organism-specific threshold.
Scope of the New Database
It is exciting to report that by including STRING in Metascape, the size of our PPI database has increased significantly. Below is the Venn diagram for human, where STRING contributes >2 million new human physical PPI pairs not covered by all previous data sources.
The same goes to the size of PPI dataset, when functional data are included. The figure below shows STRING totally contributes >5 million new PPI pairs for human.
Underlying Support
For all the networks generated by Metascape, we now include an edge property called “support”. This allows users to examine the origin of each interaction pair. An example support reads like:
{“StringDB”: “physical”, “StringEvidence”: “database:0.896,textmining:0.446,experiments:0.393,coexpression:0.023”, “OmniPath”: “omnipath”, “OmniPath_Reference”: “HPRD;SIGNOR”, “InWeb_IM”: “experimental”, “BioGrid”: “physical”, “BioGrid_PubMed”: “12634428”, “BioGrid_Type”: “Affinity Capture-Western”, “String_Score_Physical”: 939, “String_Score_Combined”: 967}
This is a very confidence interaction that is supported by OmniPath, InWeb_IM, BioGrid and STRING. The STRING physical score is 0.896.
Summary
We combined all data from STRING, OmniPath, InWeb_IM and BioGrid to produce four datasets: Physical (Core), Physical (All), Combined (Core), and Combined (All). OminiPath, InWeb_IM and BioGrid Physical data are consider high quality and included in all datasets. Only physical interactions are in the Physical datasets. The Core dataset contains the 2/3 of higher-scoring corresponding STRING data. Metascape “Express Analysis” defaults to “Physical (Core)” to be conservative at this point (subject to change in the future), but savvy users can choose any of the four flavors through “Custom Analysis”.