What is a STRING network?

STRING represents relationships between proteins as a network (graph). In the network visualization, proteins are shown as nodes (bubbles) and associations between them are shown as edges (lines). Each node represents a protein encoded by a single gene locus.

Edges represent predicted biological associations between proteins. These associations may correspond to direct physical interactions, such as proteins binding or forming a molecular complex, or to functional relationships, such as proteins participating in the same biological pathway or cellular process. STRING integrates many types of biological evidence to identify these relationships. Each edge is assigned a STRING score that reflects the confidence that the association exists given the available evidence.

The resulting network represents STRING’s current best estimate of protein associations within the organism. The connections captured in the network often reflect the underlying organization of biological systems, such as functional modules, pathways, signalling cascades and molecular complexes. As a result, proteins that work together in the cell frequently appear as connected groups in the network. Exploring these patterns can help identify novel functional relationships between proteins, interpret experimental results, generate new hypotheses about biological mechanisms, prioritize potential drug targets, or interactively explore the known biological context of your proteins.

1. What is a STRING score? Top ↑

The STRING score represents the confidence that a protein–protein association exists (either functional or physical), given all available evidence. In the STRING network visualization, these associations are represented as links (edges) connecting protein nodes.

Scores range from 0.150 to 0.999 in the STRING interface (expressed as 150–999 in downloadable files). A score closer to 1.0 indicates higher confidence that the association is real. The score is calibrated so that it approximately reflects the probability that the association exists, based on benchmarking against known biological pathways and complexes [1]. The final score is computed by probabilistically combining multiple independent evidence sources. The information on how we combine the individual scores into the combined STRING score see section (5).

Importantly, the STRING score does not measure interaction strength, binding affinity, or the magnitude of the biological effect. Instead, it reflects how strongly the available data support the presence of a biological association. A high STRING score therefore does not imply that the interaction is strong or stable at the biochemical level.

While higher scores are often observed for stable complexes and lower scores may correspond to transient or indirect associations, this distinction is not explicitly encoded in the score. An indirect functional association may therefore receive a higher score than a physical co-complex interaction if it is supported by more evidence.

Scores from all evidence sources are calibrated against the same external reference dataset, allowing them to be combined into a single probabilistic confidence score. Because of this calibration, scores from different evidence sources have the same interpretation and can be directly compared and combined.

For most analyses it is recommended to rely on the combined STRING score, which is also the default in the interface, rather than filtering interactions by individual evidence channels.

How scores are visualized in the network

In the STRING network, proteins are shown as nodes (bubbles), and associations between them are represented as links (edges). These links correspond to predicted protein–protein associations, and their appearance reflects the underlying STRING score and supporting evidence.

In the Settings tab you can change how the scores are visualized in the network.

  • Evidence view shows each evidence channel as a separate link between proteins, regardless of its individual score.
  • Confidence view shows a single link between proteins. The thickness and opacity reflect the confidence of the combined STRING score.

For single-protein searches, STRING by default tries to show the 10 highest-confidence interaction partners, if that many interactions are available above the selected score cutoff. This list represents STRING’s current best guess of the protein’s immediate functional neighborhood rather than a complete representation of its interactome. Additional interaction partners can be added and the score threshold adjusted in the Settings tab.

2. Functional vs physical networks Top ↑

The functional network is the default mode in STRING. It predicts whether two proteins are functionally associated, meaning that they contribute to the same biological process.

In practice, this usually means that the proteins participate in the same pathway or cellular process.

Because this definition is broad, the functional network can integrate many different sources of evidence that suggest a shared biological role. For most biological analyses, the functional network therefore provides the most informative view of protein relationships.

This type of network is particularly useful for tasks such as:

  • predicting gene function
  • identifying pathway members
  • prioritizing candidate disease genes
  • prioritizing potential drug targets
  • interpreting high-throughput experiments

Physical network (co-complex / physical interactions)

The physical network predicts whether two proteins are in physical proximity, for example through direct binding or membership in the same molecular complex.

For an interaction to appear in the physical network, STRING requires evidence indicating physical proximity between the proteins. This includes experimental interaction data, curated complex membership, or literature statements describing binding.

A physical link does not necessarily mean that the two proteins directly bind each other, since other proteins in the complex may mediate the interaction. Many literature statements and experimental methods do not distinguish between direct binding and co-complex membership. A physical interaction in STRING may therefore represent direct binding or indirect association within the same complex, and it may be stable or transient.

Regulatory events such as phosphorylation are not automatically interpreted as co-complex interactions unless they are detected by experimental proximity methods.

The physical network is effectively a subset of the functional network: interactions supported by physical evidence are also included in the broader functional network, which integrates both physical and functional associations. Because the physical network requires explicit physical evidence, many biologically real physical interactions may still appear only in the functional network if direct physical evidence has not yet been observed or curated.

The physical network is mainly useful when an analysis requires direct molecular proximity, such as when studying protein complexes or structural interactions. It should not be used as a proxy for correctness or as a general filtering strategy.

3. Evidence channels Top ↑

The STRING score is derived by combining multiple evidence channels. An evidence channel groups together multiple evidence sources of a similar type that suggest an association between proteins. Each channel can aggregate many individual sources, for example different experimental databases, different text-mining methods, or annotated pathway resources. Evidence channels are treated as largely independent when their scores are combined, whereas evidence sources within a channel are often not independent. STRING’s scoring pipeline therefore attempts to account for this dependence when combining evidence within each channel.

Each channel captures a different type of biological signal and contributes a partial score to the final combined STRING score.

All seven channels contribute to the functional network. Only a subset of channels — text mining, experiments, and curated databases — can provide evidence that supports physical interactions.

Each evidence channel is benchmarked against the same reference datasets:

  • KEGG pathways for functional associations
  • Complex Portal for physical interactions

Because the channels are benchmarked against the same reference datasets, a given score value is intended to represent a similar confidence level regardless of the evidence source. Each evidence channel has its own biases and error classes, and some signals (such as co-occurrence text mining) often capture more indirect functional relationships. However, combining evidence from multiple largely independent channels improves the overall predictive power of the network. For most analyses, disabling evidence channels is therefore not recommended unless there is a clear methodological reason.

The seven evidence channels in STRING are:

  • text mining
  • experiments
  • curated databases
  • co-expression
  • phylogenetic profiles (co-occurrence across genomes)
  • gene neighborhood
  • gene fusion

In the network view, each channel is represented by a distinct edge color.

The contribution of each channel can be inspected in the edge pop-up by clicking on the edge.

Text mining

The text-mining channel derives protein associations from the scientific literature. STRING analyzes both PubMed abstracts and full-text articles to detect relationships between proteins described in scientific publications.

Two main methods contribute to the text-mining score.

Co-occurrence analysis

The first method detects statistical co-occurrence of proteins in the literature. It measures how often two proteins are mentioned together within the same sentence, paragraph, or document compared to how often they would be expected to appear together by chance [2].

The strength of this signal is benchmarked according to how well it recovers known associations in the reference datasets used for STRING calibration.

A single co-mention, even within the same sentence, does not necessarily generate an association in STRING. The score depends on how frequently two proteins co-occur relative to their overall frequency in the literature. Highly studied proteins may appear in many contexts and therefore require stronger enrichment of co-occurrence to produce a link in STRING network.

This method predicts functional associations statistically and does not attempt to determine the exact type of interaction.

NLP / LLM extraction

The second method uses natural language processing to identify explicit statements about protein interactions in the literature.

The LLM model is fine-tuned on curator-annotated sentences describing physical associations. The system recognizes phrases describing physical interactions, such as binding or complex formation (for example “binds” or “forms a complex”). These statements can provide evidence for physical interactions [3].

Unlike the co-occurrence approach, which detects statistical associations in the literature, this method searches specifically for sentences describing physical interactions rather than general functional associations.

Experiments

The experimental channel integrates protein–protein interactions reported in major interaction databases such as BioGRID, IntAct, and MINT.

For each interaction, STRING parses the associated experimental evidence and determines whether the experimental method supports a physical interaction in addition to being a functional association [4].

High-throughput (HT) and low-throughput (LT) experiments are treated differently.

For high-throughput experiments, each dataset is benchmarked separately. This allows STRING to distinguish more reliable and less reliable subset of each dataset, resulting in a range of confidence scores. Interactions that fall below a minimal confidence threshold may not be included in STRING.

For low-throughput experiments, interactions are benchmarked based on the experimental method used. For example, interactions derived from X-ray crystallography typically receive very high confidence scores, while evidence from a single yeast two-hybrid experiment may correspond to a more moderate confidence score.

Evidence from multiple independent experiments is combined into a single experimental channel score. The contributing experiments and their individual contributions can be inspected in the experimental viewer in the edge pop-up.

Importantly, the STRING score does not correspond to the number of experiments supporting an interaction. Each experiment is weighted according to its reliability, estimated through benchmarking, and incorporated into the probabilistic confidence score.

Co-expression

The co-expression channel captures associations between proteins whose genes show similar expression patterns across many experiments and biological conditions.

The underlying assumption is that genes involved in the same biological process are often co-regulated and therefore tend to be expressed together. Co-expression does not necessarily imply direct regulatory interaction but reflects shared regulation or participation in related biological processes.

STRING derives co-expression signals from multiple large-scale datasets, including microarray experiments, bulk RNA-seq, single-cell RNA-seq, and proteomics datasets such as ProteomeHD.

To estimate co-expression, STRING compares expression patterns across many conditions and datasets. Rather than evaluating individual experiments separately, the signal is derived from the combined structure of many datasets, where proteins whose expression levels consistently correlate across diverse conditions receive higher co-expression scores.

Large expression collections often contain many similar experiments that could artificially inflate correlations. To reduce this redundancy, STRING applies dimensionality-reduction methods based on deep-learning variational autoencoders, which compress the data while preserving meaningful co-expression signals [4]

Databases

The database channel captures protein associations derived from curated pathway and complex databases, including KEGG, Reactome, and BioCyc.

When parsing these resources, STRING attempts to extract explicit relationships between individual proteins rather than linking all proteins belonging to the same pathway. Two proteins are connected only if the reference database indicates a relationship between them.

However, many pathway databases represent biological processes as interconnected functional modules or molecular complexes rather than simple pairwise relationships. Depending on how these structures are encoded, STRING translates them into pairwise associations between individual proteins by expanding multi-protein modules into individual links. This process takes into account the type of module and its size.

Phylogenetic profiles (co-occurrence across genomes)

The phylogenetic profile channel predicts functional associations based on the co-occurrence of genes across genomes. The underlying idea is that genes involved in the same biological process are often gained or lost together during evolution.

STRING constructs phylogenetic profiles by examining the presence or absence of orthologous genes across many genomes. Proteins whose genes show similar presence–absence patterns across species receive higher association scores [5]].

Because genomes in current databases are unevenly distributed across the taxonomic tree, the analysis accounts for the relatedness of species so that closely related organisms do not artificially inflate the signal.

This channel supports functional associations and does not imply physical interaction.

Gene neighborhood

The gene neighborhood channel predicts functional associations based on the proximity of genes in the genome.

In many genomes, especially in Bacteria and Archaea, genes involved in the same biological process are often located close to each other on the chromosome, sometimes in the same operon or gene cluster.

STRING derives this signal from the distance and relative orientation of genes in the genome. Genes that are consistently found in similar genomic arrangements receive higher association scores [6].

This signal supports functional associations and does not imply physical interaction.

Gene fusion

The gene fusion channel predicts functional associations based on gene fusion events.

This signal is derived from the analysis of COG/KOG orthologous groups. If two proteins belong to separate orthologous groups in most species but are found fused into a single protein in another species, this suggests that the proteins are functionally linked.

Fusion events indicate that the proteins can exist as a single polypeptide in some organisms, which strongly supports a shared biological role.

In principle, gene fusion could indicate physical interaction. However, in STRING this signal is treated only as evidence for functional association and does not contribute to the physical interaction network.

4. Transferred evidence (interologs) Top ↑

Many interactions in STRING are supported by evidence derived from other species. Because most organisms in STRING are not extensively studied, interaction evidence is often transferred across species using orthology relationships.

This approach is known as the interolog principle. If two proteins interact in one organism, their orthologs are likely to interact in another organism.

In STRING, orthology relationships are defined using COG and eggNOG orthologous groups [7]. Evidence for interactions observed between proteins in one species can, as well, contribute to interactions between the corresponding orthologs in another species. These relationships can be downloaded in the download section of STRING.

Transferred scores are scaled according to evolutionary distance and gene family divergence. Interactions transferred between closely related species receive higher confidence than those transferred across larger evolutionary distances [2].

Some evidence channels, such as phylogenetic profiles and gene neighborhood, naturally generate signals across many species. Other channels, including experiments and curated databases, are concentrated in a smaller number of well-studied organisms and therefore rely more heavily on orthology-based transfer.

When exploring an interaction in the edge pop-up, STRING indicates whether the evidence is derived directly from the species itself or transferred from another species.

In the web interface, transferred evidence cannot be disabled separately from direct evidence. However, downloadable datasets provide files that distinguish between direct and transferred interactions.

Transferred evidence reflects evolutionary conservation of interactions rather than direct experimental observation in the target species. For most non-model organisms, networks based only on direct evidence would be fairly sparse.

5. How the combined STRING score is calculated Top ↑

The combined STRING score is calculated by integrating the contributions from the different evidence channels. Each channel produces a probabilistic score reflecting the likelihood that an association exists based on that particular type of evidence.

To combine these scores, STRING uses a probabilistic framework that assumes the different evidence channels are largely independent. The combined score therefore increases when multiple independent channels support the same association.

Because even random pairs of proteins may appear associated by chance, the calculation incorporates a prior probability representing the baseline likelihood that two randomly selected proteins are associated. In the current STRING implementation this prior probability is p = 0.041.

The combined score is calculated in three steps.

  1. Remove the prior from each individual channel score

    channel_score_adjusted = (channel_score - prior_probability) / (1 - prior_probability)

  2. Combine the adjusted channel scores

    combined_score_adjusted = 1 - combined_score_adjusted * (1 - channel_score_adjusted)

    This step increases the combined confidence when multiple channels independently support the same association.

  3. Add the prior probability back once to obtain the final STRING score

    combined_score = combined_score_adjusted + prior_probability * (1 - combined_score_adjusted)

For a more detailed description of the scoring framework see [1].

A python implementation for combining all evidence scores is available here

6. Exploring protein networks in STRING Top ↑

STRING networks can be generated from either a single protein or from a set of proteins.

When a single protein is queried, STRING shows the highest-confidence interaction partners by default. These proteins represent STRING’s current best estimate of the immediate interaction neighborhood of the queried protein. By default, the network shows the 10 strongest associations. This provides a compact overview of the protein’s local network rather than a complete representation of its interactome.

When a set of proteins is queried, STRING displays the associations between proteins within that set. This allows users to see whether the proteins are connected and whether they form recognizable groups such as pathways, molecular complexes, or functional modules.

The density of the network depends on the score threshold used to display interactions. Higher thresholds show only the strongest associations and produce smaller networks, while lower thresholds reveal additional connections that may represent weaker but still informative signals.

These patterns often reveal the organization of biological pathways and complexes, helping users interpret experimental results and identify functional relationships between proteins.

7. Network expansion and interaction shells Top ↑

STRING allows networks to be expanded by adding proteins that are strongly associated with the proteins currently shown in the network. This helps users explore the broader interaction neighborhood of a protein or protein set.

Candidate proteins for expansion are evaluated using two criteria:

  • Strength of association, measured by the combined STRING score.
  • Specificity, meaning how strongly the candidate protein is connected to the current network compared with its connections to proteins outside the network.

By balancing these two measures, STRING selects proteins that are both strongly connected and specifically related to the current network.

Two layers of expansion are available.

First shell proteins are selected based on their associations with the original query proteins. These nodes are shown in color and represent proteins that are most specifically connected to the input set.

Second shell proteins extend the network further by selecting proteins associated with both the query proteins and the first-shell proteins. These nodes are shown in white and represent the next layer of the interaction neighborhood.

Network expansion is particularly useful when studying proteins with limited annotation, as it can reveal additional functional partners and biological context.

However, expanded proteins are selected using the STRING network itself. Because of this, statistical analyses such as functional enrichment tests should generally be performed on the original input set rather than the expanded network, since the expansion already incorporates information from STRING.

8. Selecting score thresholds and filtering evidence Top ↑

When using STRING, either in the web interface or when downloading data, you can choose a minimum score threshold. STRING scores range from 0.150 to 0.999 (or 150–999 in downloadable files).

In the web interface the threshold can be adjusted in the Settings tab. When downloading data it can be specified in the +more options panel.

Choosing thresholds in the STRING interface

When exploring networks in the STRING web interface, the score threshold mainly affects how clearly the network structure can be visualized.

If the threshold is set too low, the network may become very dense and appear as a “hairball”, especially for proteins belonging to well-studied pathways or complexes. In such cases, increasing the score threshold can help reveal the internal structure of the network and highlight the strongest and most informative associations.

Conversely, for proteins that belong to less-studied parts of the interactome, high score thresholds may produce very sparse networks with few visible connections. Lowering the threshold can reveal additional associations that help identify functional modules or connect otherwise isolated proteins.

For very small protein sets, it is often useful to reduce the threshold further in order to display more of the available evidence. Showing weaker associations can help reveal additional connections that may provide biological context for exploratory analysis.

The score threshold is also taken into account in the network statistics displayed by STRING. Changing the threshold therefore does not invalidate these statistics, but it may change their values because the set of displayed interactions also changes.

Choosing thresholds for downstream analysis

When downloading STRING data for computational analyses, all score values contain useful information. Even relatively low scores represent signals much above random expectation, although with lower confidence.

Lower thresholds provide more coverage but also introduce more noise. Higher thresholds provide higher confidence links but produce smaller networks that are more biased toward well-studied proteins and pathways.

Very high thresholds reduce the chance of identifying novel associations, particularly for less well-studied regions of the proteome.

As a general guideline:

  • 0.7 is often used as a conservative threshold for downstream computational analyses
  • 0.4 may be preferable when broader coverage is desired and the method can tolerate more noise

The optimal threshold ultimately depends on the sensitivity of the downstream method to false positives and on the size of the network that can be handled computationally.

Filtering by evidence channel

STRING allows users to enable or disable individual evidence channels in the interface.

In most cases, filtering interactions by channel is not recommended.

Because all channels are calibrated against the same reference datasets, a given score value has the same probabilistic interpretation regardless of the evidence source. A text-mining score of 0.5 is therefore expected to have roughly the same reliability as an experimental score of 0.5.

Different channels do have different biases and error classes, but removing channels generally reduces the predictive power of the network for downstream analyses. Channels should only be disabled when there is a clear methodological reason, for example if a particular evidence type introduces a known bias in the dataset being analyzed.

If the network appears too dense, it is usually better to increase the score threshold rather than remove individual evidence channels.

While individual evidence channels can be enabled or disabled in the interface, STRING does not support filtering interactions based on the requirement that evidence from a particular channel must be present. For example, it is not possible to require that an interaction has experimental evidence in order to be displayed. Such filtering is discouraged and does not necessarily produce more reliable networks. Because all channels are benchmarked against the same reference datasets, no individual channel should be assumed to be inherently more reliable than another at the same score level.

Why scores can change between STRING versions Top ↑

STRING scores are version-specific. For each STRING release, the entire network is recomputed from scratch using the data and inference pipeline available at that time.

Because of this, the score for a given protein pair may change between releases. The main reasons for these changes are updates to the underlying data sources and improvements to the prediction and data-extraction methods used in STRING.

As a result, an interaction may receive a higher or lower score in a newer STRING version. Some interactions may also appear or disappear entirely if their score moves above or below the inclusion threshold. If a link is not visible in the network, it may still exist in STRING but fall below the currently selected score cutoff.

For single-protein queries, STRING by default tries to display max 10 highest-confidence interaction partners at the given score cut-off. Because interaction scores may change between STRING versions, the composition or order of this top-10 list may also change between releases. This list represents STRING’s current best guess of the protein’s immediate interaction neighborhood rather than a complete representation of its interactome. To explore a broader network, additional interaction partners can be added by expanding the network in the Settings tab.

These changes do not mean that earlier scores were incorrect. Each STRING release represents the best prediction possible given the data and methods available at that time.

Importantly, STRING networks do not change within a release. Each version is fixed and fully reproducible. For reproducible analyses, users should therefore record the STRING version, species, network type, and score threshold used.

Previous STRING versions remain available, and accessable here. To reproduce an earlier analysis, simply use the same STRING version and protocol that was used in the original analysis.

References Top ↑

[1] von Mering C, Jensen LJ, Snel B, Hooper SD, Krupp M, Foglierini M, Jouffre N, Huynen MA, Bork P.
STRING: known and predicted protein–protein associations, integrated and transferred across organisms.
Nucleic Acids Research. 2005;33(Database issue):D433–D437.
https://doi.org/10.1093/nar/gki005

[2] Franceschini A, Szklarczyk D, Frankild S, Kuhn M, Simonovic M, Roth A, Lin J, Minguez P, Bork P, von Mering C, Jensen LJ.
STRING v9.1: protein–protein interaction networks, with increased coverage and integration.
Nucleic Acids Research. 2013;41(Database issue):D808–D815.
https://doi.org/10.1093/nar/gks1094

[3] Szklarczyk D, Gable AL, Nastou KC, Lyon D, Kirsch R, Pyysalo S, Doncheva NT, Legeay M, Fang T, Bork P, Jensen LJ, von Mering C.
The STRING database in 2021: customizable protein–protein networks, and functional characterization of user-uploaded gene/measurement sets.
Nucleic Acids Research. 2021;49(D1):D605–D612.
https://doi.org/10.1093/nar/gkaa1074

[4] Szklarczyk D, Kirsch R, Koutrouli M, Nastou K, Mehryary F, Hachilif R, Gable AL, Fang T, Doncheva NT, Pyysalo S, Bork P, Jensen LJ, von Mering C.
The STRING database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest.
Nucleic Acids Research. 2023;51(D1):D638–D646.
https://doi.org/10.1093/nar/gkac1000

[5] Franceschini A, Lin J, von Mering C, Jensen LJ.
SVD-phy: improved prediction of protein functional associations through singular value decomposition of phylogenetic profiles.
Bioinformatics. 2016;32(7):1085–1087.
https://doi.org/10.1093/bioinformatics/btv696

[6] Jensen LJ, Kuhn M, Stark M, Chaffron S, Creevey C, Muller J, Doerks T, Julien P, Roth A, Simonovic M, Bork P, von Mering C.
STRING 8—a global view on proteins and their functional interactions in 630 organisms.
Nucleic Acids Research. 2009;37(Database issue):D412–D416.
https://doi.org/10.1093/nar/gkn760

[7] Hernández-Plaza A, Szklarczyk D, Botas J, Cantalapiedra CP, Giner-Lamia J, Mende DR, Kirsch R, Rattei T, Letunic I, Jensen LJ.
eggNOG 6.0: enabling comparative genomics across 12,535 organisms.
Nucleic Acids Research. 2023;51(D1):D389–D394.
https://doi.org/10.1093/nar/gkac1022