STRING: Protein Interaction Networks

By the end of this lesson you will be able to pull a protein's interaction network out of STRING, read the confidence score on every edge, and say exactly what "these two proteins are associated" does — and does not — mean. That last part is where most people go wrong, so we will spend real time on it.

In the previous lesson on KEGG and Reactome you met the pathway map: an ordered picture of who acts on whom, with arrows for direction and reactions. STRING answers a different question. It does not draw the reaction. It draws the association — a weighted line saying "these two proteins tend to work together" — and it does so for known and predicted partners alike. Knowing the difference between a map and a network is the whole point of this lesson.

What STRING actually stores

STRING — the name stands for Search Tool for the Retrieval of Interacting Genes/proteins, currently version 12.5 at string-db.org — is a derived database. It invents no experiments of its own. Instead it aggregates evidence from many primary sources and turns each aggregate into one number: an edge between two proteins, weighted by how confident STRING is that a functional association exists.

"Functional association" is deliberately broad. It means the two proteins participate in the same process — they might bind directly, sit in the same complex, or simply be switched on together — without STRING claiming which. A pathway map commits to a mechanism; a STRING edge commits only to relatedness. That looseness is a feature: it lets STRING include predicted partners that no one has ever pipetted together.

The confidence score, 0 to 1000

Every edge carries a combined score — an integer from 0 to 1000 that estimates the probability the association is real, where 900 means STRING is ~90% sure. Four thresholds are conventional:

The score is combined, not additive. STRING treats each evidence type as an independent vote and merges them probabilistically, so an edge with several weak signals can still clear a high threshold if those signals are truly independent. The practical upshot: always set a threshold before you read a network. The default ≥ 400 already hides a swarm of low-confidence lines you would otherwise mistake for biology.

The evidence channels

The combined score is built from separate channels, each a distinct kind of evidence. You can see, and filter on, every channel individually:

• Experiments — biochemical or biophysical assays (two-hybrid, affinity capture) imported from primary interaction databases.
• Databases — curated pathway and complex knowledge from sources like Reactome and KEGG.
• Text-mining — co-mention of the two proteins across the abstracts of PubMed.
• Co-expression — the two genes rise and fall together across many transcriptomic datasets.
• Neighborhood, gene fusion, co-occurrence — genomic-context signals, most informative in bacteria.

Only three of these — experiments, databases, and text-mining — can support a claim of physical interaction. The rest tell you proteins are functionally linked, not that they touch.

The BRCA1–BARD1 edge

Make it concrete with our thread. BRCA1 (UniProt P38398, Ensembl gene ENSG00000012048) is a tumour-suppressor E3 ubiquitin ligase, and its closest STRING partner is BARD1. This edge is about as good as STRING edges get: its combined score sits in the highest tier because three independent channels all fire. Experiments show the two proteins pulled down together; curated databases record their complex; the literature co-mentions them constantly.

Here the network view and the structure view finally agree. The BRCA1 and BARD1 RING domains form a real heterodimer, solved by NMR as PDB 1JM7, and BRCA1's own BRCT repeats sit in PDB 1JNX — you met both in the PDB lesson. So the STRING edge is not merely co-mention: there is a 3D structure of these two proteins holding hands. That is the ideal case — every channel and the atoms agree.

Pulling the network with the API

STRING exposes a plain REST API at string-db.org/api, no key required. The interaction_partners method returns a protein's neighbours with per-channel and combined scores. Human is NCBI taxon 9606:

import requests

url = "https://string-db.org/api/tsv/interaction_partners"
params = {
    "identifiers": "BRCA1",   # STRING resolves the symbol
    "species": 9606,          # Homo sapiens
    "required_score": 700,    # high-confidence tier only
    "limit": 10,
}
resp = requests.get(url, params=params, timeout=30)

for line in resp.text.strip().splitlines()[1:]:
    cols = line.split("\t")
    partner, score = cols[3], cols[5]   # partner symbol, combined score
    print(f"{partner}\t{score}")
# BARD1, BRCA2, TP53, RAD51 ... each with its combined score

Two habits will save you. First, resolve your names with the get_string_ids method before a big query, so BRCA1 maps cleanly to STRING's internal protein ID rather than matching the wrong species. Second, always pass required_score — without it you get every flimsy edge STRING knows. To take the returned network further, load it into Cytoscape; our PPI network analysis walkthrough picks up exactly there.

You can now read a STRING network for what it is: a weighted map of association, scored 0–1000, assembled from channels you can inspect one at a time. It is not a pathway and not a proof of binding — but for asking "what does BRCA1 work with?", nothing gets you a defensible first answer faster.