bio-interaction-databases

What this skill does

## Version Compatibility

Reference examples tested with: requests 2.31+, pandas 2.2+, networkx 3.2+; STRING v12.0, BioGRID 4.4+, IntAct (live), SIGNOR 3.0+, OmniPath (live)

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show requests pandas networkx`
- API surface: confirm endpoint URLs match each resource's current docs

STRING URL is version-pinned (`version-12-0` as of 2024); older URLs (`version-11-5`) were deprecated 2023. The Cytoscape/Cytoscape.js ecosystem uses different version semantics; check the docs for the targeted version.

# Interaction Databases

**"Get protein-protein interactions for these genes"** -> The choice of database matters more than the choice of API. Different resources index different evidence (physical binding, functional association, genetic interaction, signed signaling), with different curation pipelines (manually curated vs high-throughput vs text-mined), different species coverage, and different licenses.

The decision matrix below is the postdoc-grade view: **what question is being asked, and which resource answers it best?**

- Python: `requests.get()` against REST endpoints; `pandas` for parsing; `networkx` for graphs
- R: `STRINGdb`, `OmnipathR` (mature Bioconductor clients)
- Web: STRING, BioGRID, IntAct, SIGNOR, OmniPath, ConsensusPathDB browsers

## Required Setup

```python
import requests
import pandas as pd
import networkx as nx
from io import StringIO

CALLER = 'bioskills-2026'  # STRING + OmniPath accept caller_identity for usage attribution
```

API key requirements:
- **BioGRID**: free key required (`https://webservice.thebiogrid.org/`)
- **STRING, IntAct, SIGNOR, OmniPath, Reactome**: no key

## Decision matrix: which resource for which question?

| Question | Best resource | Why |
|---|---|---|
| "Build a network around 10 genes" | STRING (medium confidence ~400) | Comprehensive; channels combinable; good viz integration |
| "Only physically interacting proteins" | IntAct or BioGRID physical | Curated physical interactions; PSI-MI standard |
| "Signed/directed signaling (phospho, ubiq, etc.)" | **SIGNOR** | Only major DB with mechanism types and direction |
| "Functional enrichment based on co-mentioned genes" | STRING functional (default) | Includes textmining channel |
| "Genetic interactions (synthetic lethality)" | BioGRID genetic | Largest curated genetic interaction set |
| "High-throughput Y2H interactome" | **HuRI** | Reference yeast-2-hybrid map of human |
| "Mass-spec-derived protein complexes" | **HuMAP v2** or BioPlex | AP-MS complex maps |
| "Curated pathways with interactions" | Reactome | Pathway-organized; gold standard for signaling |
| "Meta-database aggregating 100+ sources" | **OmniPath** | The modern "one-stop"; pre-aggregated |
| "Cross-species or non-human" | STRING | Species coverage broadest |
| "Bacterial interactome" | STRING bacterial | Limited curated alternatives |
| "Phosphorylation site-specific" | PhosphoSitePlus (commercial license) or SIGNOR | PSP has best PTM coverage but requires license |

**OmniPath** (Türei et al. 2021 *Mol Sys Biol* 17:e9923) deserves special mention: it aggregates >100 sources into a unified API with provenance tracking. For "give me all available interactions for X" workflows, OmniPath is the modern default.

## STRING (v12, channels, confidence)

STRING (Szklarczyk et al. 2023 *Nucleic Acids Res* 51:D638) aggregates evidence into a combined confidence score. The seven evidence **channels**:

| Channel | What it captures |
|---|---|
| `experiments` | Direct experimental evidence (BioGRID, IntAct, etc.) |
| `database` | Curated database (Reactome, KEGG) |
| `textmining` | Co-mention in PubMed |
| `coexpression` | Co-expression across conditions |
| `neighborhood` | Genomic neighborhood (prokaryotes mainly) |
| `fusion` | Gene fusion across species |
| `cooccurrence` | Phylogenetic profile co-occurrence |

The default `score` is the combined-evidence score (0-1000). Confidence tiers:

| Threshold | Tier | Use when |
|---|---|---|
| 150 | Low | Exploration; includes weak textmining |
| 400 | Medium | Default; balanced sensitivity/specificity |
| 700 | High | Publication-quality networks |
| 900 | Highest | Experimentally validated core only |

**Version pinning**: STRING URL is `https://version-12-0.string-db.org/api/...` as of 2024. Older `version-11-5` URLs were deprecated 2023 — code using them silently fails. Use the unversioned `https://string-db.org/api/...` to follow the current release, or pin to a specific version for reproducibility.

**`caller_identity`**: STRING requests all programmatic users pass `caller_identity=<app-name>` for usage attribution. The parameter helps STRING identify and contact heavy automated callers.

## BioGRID (physical + genetic, curated + HT)

BioGRID (Oughtred et al. 2021 *Protein Sci* 30:187) covers physical and genetic interactions across the broadest organism set of any major resource. Requires a **free API key** (`https://webservice.thebiogrid.org/`).

Key concepts:
- `EXPERIMENTAL_SYSTEM`: e.g. "Two-hybrid", "Affinity Capture-MS", "Synthetic Growth Defect"
- `THROUGHPUT`: "Low Throughput" vs "High Throughput" — the most important quality flag
- `EVIDENCE_TYPE`: physical vs genetic

For high-confidence physical interactions, filter to physical systems + Low Throughput:

```python
PHYSICAL_LT_SYSTEMS = {
    'Affinity Capture-MS', 'Affinity Capture-Western', 'Affinity Capture-RNA',
    'Co-fractionation', 'Co-purification', 'Reconstituted Complex',
    'Co-crystal Structure', 'Two-hybrid', 'Far Western', 'FRET', 'PCA',
}
```

## IntAct (PSI-MI curated physical)

IntAct (Del Toro et al. 2022 *Nucleic Acids Res* 50:D648) is the IMEx consortium reference for curated physical interactions in PSI-MI standard format. MINT was folded in ~2014; queries to MINT URLs now redirect to IntAct.

Direct REST API has changed multiple times; the most stable access is via OmniPath (which wraps IntAct) or via the PSICQUIC web services.

## SIGNOR (signed signaling)

SIGNOR (Lo Surdo et al. 2023 *Nucleic Acids Res* 51:D631) is **the only major curated database with signed, directed, mechanism-typed interactions**. Each edge has:

- `direction`: A->B
- `effect`: `up-regulates`, `down-regulates`, `unknown`
- `mechanism`: `phosphorylation`, `dephosphorylation`, `ubiquitination`, `binding`, etc.

Essential for any signaling pathway analysis or dynamic modeling. Coverage smaller than STRING/BioGRID but quality is high.

## Reactome (curated pathways)

Reactome (Milacic et al. 2024 *Nucleic Acids Res* 52:D672) is gold-standard for human pathway curation with full interaction reactions. Species-specific (human is by far the most complete).

Reactome ContentService REST: `https://reactome.org/ContentService/`.

## HuRI / HuMAP (human-specific interactomes)

- **HuRI** (Luck et al. 2020 *Nature* 580:402): yeast-2-hybrid interactome of human; ~53K binary interactions; biased toward binary high-confidence.
- **HuMAP v2** (Drew et al. 2021 *Mol Syst Biol* 17:e10016): AP-MS-derived complex map; integrates >15,000 proteomic experiments.

Both available for download; HuRI also has a web portal (`http://www.interactome-atlas.org/`).

## OmniPath (meta-database)

OmniPath (Türei et al. 2021 *Mol Syst Biol* 17:e9923) aggregates 100+ sources with provenance. Key endpoints:

| Endpoint | Content |
|---|---|
| `/interactions` | Signaling interactions (directed); the most useful for cross-DB consensus |
| `/enzsub` | Enzyme-substrate (kinase-substrate, etc.) |
| `/complexes` | Protein complexes |
| `/annotations` | Functional annotations |
| `/intercell` | Intercellular communication |

Each interaction includes `sources` (list of contributing databases) and `references` (PMIDs). For "give me everything anyone has said about A-B", OmniPath is the answer.

R users: `OmnipathR` (Bioconductor) is more ergonomic than raw HTTP.

## License gotchas (critical for commercial use)

| Resource | Licen