How an interactome is built.

ProPaths builds a protein interactome by reading the primary literature with language models and validating the result against the source papers. This page describes the pipeline, the sources and identifiers it uses, and which parts of the process are proprietary.

Sources and identifiers

Literature

Papers are retrieved from PubMed. Full text is taken from PubMed Central where the publisher has released it, through the public APIs.

Full text

Full text is never generated, reconstructed, or paraphrased to fill a gap. Where a publisher has not released full text, the paper is processed from its abstract and the record notes that this was the source.

Nomenclature

All protein and gene symbols are normalized to HGNC nomenclature before they are written to the database. Aliases, mutation prefixes, greek letters, and non-standard spellings resolve to the canonical symbol, so one protein occupies one node.

Citations

Every claim, measurement, and mechanism carries the PubMed identifiers of the papers it came from, with links back to the source records.

The pipeline

Five stages, run per protein.

{{ s.num }}

{{ s.name }}

{{ s.body }}

{{ u.n }} {{ u.label }} {{ u.text }}
{{ s.asideLabel }}
{{ s.aside }}

Extraction and validation are separate passes

The two stages run as separate calls with different instructions. Extraction proposes candidates from a paper and is permissive. Validation tests those candidates against the same paper and is restrictive. Because they are separate passes, a candidate is never accepted on the strength of the reading that produced it.

Rejection is the common outcome. Counts below are from the ATXN3 run.

791 papers retrieved
~2,000 candidate claims extracted
229 validated pairs
109 enriched interactions

Roughly 89% of extracted candidates are rejected before reaching the graph.

Evidence strength

Every claim in the graph carries a rating for the strength of the evidence behind it. The rating is derived from the type of experiment that produced the result rather than from the language used to describe it, and it is stored with the claim so it can be read alongside the finding.

The rating scale and the criteria behind it are proprietary.

The pathway ontology

Interactions are placed in a hierarchical ontology running from broad biological domain down to specific pathway. Its naming conventions follow the same principles as Gene Ontology, Reactome, and KEGG, and it extends their approach to hierarchy where interaction-level placement requires more specificity than those resources carry.

The ontology itself is proprietary. Placement is decided from the enriched content of an interaction, so a partner appears in every pathway where the literature reports it acting rather than being assigned to one.

One interaction, after the pipeline has run

ATXN3 and VCP, reconciled from 48 papers. The record below is the output of the enrichment stage: a named mechanism, measurements that retain their assay and system, and claims that remain separate where the papers disagree.

Reconciled claim
VCP allosterically activates ATXN3's deubiquitination activates

VCP binding raises ATXN3's catalytic efficiency on K63 chains about 3x, favoring branched-chain editing; polyQ impairs it about 50%.

Sources
Reinders 2025, EMBO J
Ub-AMC shows 3x kcat/Km
Potapenko 2024, Biochem J
ERAD luciferase reporter, knockdown cuts activity 40%
Mechanism, as ordered steps
The VBM docks onto the VCP N-domain, positioning the Josephin domain
VCP unfolds a substrate; ATXN3 trims its branched K48/K63 chains
Cofactor engagement competes with UFD1-NPL4, biasing VCP toward chain editing
PolyQ expansion over-stabilizes the interface and aborts the cycle
Measurements with assay and system
2.8x
deubiquitination boost on branched K48/K63 chains
HEK293T, Ub-AMC kinetics
45%
loss of TCRalpha turnover without ATXN3
HeLa, cycloheximide chase

Persistence and reuse

Each run writes into one cumulative database rather than producing a separate result set. A paper retrieved for one protein is available to every protein queried after it. A validated pair is not revalidated. An enriched interaction is rebuilt only when new papers change the evidence behind it.

Because symbols are normalized to HGNC before writing, the same protein cannot enter the graph twice under two names. Coverage accumulates across runs, and the marginal cost of each additional protein falls as the graph grows.

What we do not do

Interactions are not inferred from co-mention. A language model is not permitted to assert an interaction without a source paper that reports it. Candidates that fail validation are discarded rather than stored at reduced confidence.

Where papers disagree

Where the literature is inconsistent, the inconsistency is shown rather than hidden. Claims are consolidated and reconciled where there is a clear argument from strong physical evidence; where there is not, the conflicting findings are held as separate claims, each with its own evidence and sources.

When new papers arrive

An enriched record is rebuilt only when new papers change the evidence behind it. The rebuilt version is compared against the stored one and replaces it only if it is better supported.

See the pipeline's output on a real protein.

The ATXN3 interactome is open to explore. For specific questions on our process, or to have a protein mapped, get in touch.

Explore the Demo Get in touch