Skip to content

GIDEON / Between-Pathway Models: relevance assessment for dismech

Date: 2026-09-02 Source paper: Garcia JJ, Yu KM, Freudenreich CH, Cowen LJ. A novel ILP framework to identify compensatory pathways in genetic interaction networks with GIDEON. Bioinformatics 2026;42(Suppl 2):btag385. DOI: 10.1093/bioinformatics/btag385 (open access, CC-BY). Code and full BPM set: github.com/jocelynjgarcia/GIDEON, archived at doi:10.5281/zenodo.20130057. Status: analysis note — no KB content change proposed in this report.

What the paper does

GIDEON (Genetic Interaction-Driven Extraction of Optimal Networks) mines the genome-scale S. cerevisiae genetic interaction network (Costanzo et al. 2016, pairwise double-knockout fitness for nearly all non-essential gene pairs) for Between-Pathway Models (BPMs): paired gene sets where negative (synthetic-sick) interactions run between the two sets and positive interactions run within each set. That motif is the graph signature of compensatory pathways — knocking out one gene from each pathway disables both routes and makes the cell sick, while two knockouts within one pathway leave the other route intact.

Two methodological contributions:

  1. A gene-centered ILP. Rather than one global optimization that converges on the same best BPM repeatedly (the failure mode of the prior Liany et al. 2022 ILP, which had to delete discovered edges from the network to make progress), GIDEON solves a separate ILP per gene, constrained so that gene is the primary contributor to its BPM's objective. Interactions can therefore appear in multiple BPMs — biologically right, since real pathways share components.
  2. Distribution-informed (DI) edge weights. Epistasis for a pair is scored as the residual from a linear regression over each component gene's whole marginal distribution of double-knockout fitnesses, rather than from a two-gene multiplicative null alone. This needs only single- and double-mutant fitness values (no array-position/batch metadata), and improves competitor methods too.

Results: 3,215 diverse BPMs versus 1,027 (LocalCut) and 750 (Liany-ILP), with 1,220 BPMs whose two modules are enriched for the same function versus 301 and 33. A highlighted novel BPM ties aromatic amino acid biosynthesis to ergosterol biosynthesis (TKL1-centered; strong interactions among ERG3, ERG6, TRP3, ARV1, ARO7), which the authors read as a lead for antifungal combination targeting. The discussion flags the human direction: the new genome-scale human genetic interaction map (Billmann et al., Cell 2026) and cancer synthetic-lethality prediction (Fong et al. 2025; Liany et al. 2024 ASTER) as the translation path.

Why this is relevant to dismech

Pathway-level compensation is the systems-biology face of several things dismech already models explicitly:

  • Synthetic lethality as therapy. A clinically exploited BPM is exactly what kb/modules/dna_repair_synthetic_lethality.yaml curates: HRR/FA-BRCA deficiency in one "pathway" makes tumors dependent on the backup route that PARP inhibition then removes. The BRCA/PARP axis is the flagship human instance of the motif GIDEON finds at scale in yeast. Conforming entries (ovarian HGSC, TNBC, BRCA-mutant prostate cancer, etc.) already carry the treatment target_mechanisms pattern for it.
  • Digenic and oligogenic inheritance. A Mendelian digenic disorder (HP:0010984, worked exemplar PRPH2-Related_Retinopathy) is a germline BPM hit: one damaging variant in each of two compensating components, neither sufficient alone. The paper's Fig. 1 paralog pair (two genes that buffer each other so neither shows single-gene interactions) is likewise the abstract form of dismech's MODIFIER/SUSCEPTIBILITY/COOPERATING gene relationship types and the COOPERATING_HIT allelic role — and one mechanistic reading of incomplete penetrance.
  • Antifungal target modules. The ergosterol side of the highlighted BPM lands on modules dismech already has (fungal_ergosterol_synthesis_inhibition, fungal_membrane_ergosterol_binding, antifungal_intrinsic_resistance_gating). A validated aromatic-amino-acid-biosynthesis co-target would be a candidate future antimicrobial target module in the projects/ANTIMICROBIAL.md family — but today it is a yeast-model lead, not a curated therapy.
  • Mechanism modules generally. A BPM module (a coherent gene set acting as one functional unit) is a data-driven analog of a dismech mechanism module. The compensation relation between two modules, however, has no first-class slot in the schema — see the gap note below.

What dismech should and should not take from it

Not KB content, directly. GIDEON's 3,215 BPMs are computational predictions in baker's yeast, most without human disease correlates. dismech's evidence policy (PMID + exact snippet, evidence_source classifying the study) would grade any BPM-derived claim as COMPUTATIONAL at best, and the yeast→human step is precisely the HUMAN_MODEL_MISMATCH situation the discussion machinery exists for. Nothing in this paper warrants new disorder entries or new evidence items.

Plausible uses, in rough priority order:

  1. Hypothesis leads for curated entries. Where a dismech entry already carries a digenic/oligogenic claim or an unexplained modifier gene, a conserved yeast BPM containing the orthologs is legitimate supporting context for a discussions item (kind: KNOWLEDGE_GAP or HUMAN_MODEL_MISMATCH) with proposed_experiments — the published BPM set is browsable per gene, and the paper is citable (PMID pending indexing; DOI available now).
  2. A schema question worth recording, not solving here. dismech can say a treatment targets a mechanism and an exposure protects against one, but cannot say mechanism/pathway A compensates for B — the relation underlying synthetic lethality, digenicity, and paralog buffering alike. Today that lives in prose (description, module notes). If compensatory structure recurs as a curation need, it belongs in the design-decision register as an open decision (candidate shape: an inter-node or inter-module link class with a COMPENSATES_FOR predicate, evidence-bearing like ModelMechanismLink), rather than being improvised per entry.
  3. Human genetic interaction data as a future structured source. The Billmann et al. 2026 human GI map (and ASTER-style clinically oriented SL predictions) is the dataset to watch: if cancer entries start citing human GI screens for synthetic-lethality claims, a structured source or datasets: convention for GI screens would follow the existing verify-datasets discipline. Premature until a curation PR actually needs it.

Recommendation: file this as background reading for the synthetic-lethality and digenic-inheritance curation areas; take no schema or KB action now. If a curator hits the "compensates-for" expressivity wall in a real entry, that is the trigger to open the design-register question in item 2.

Follow-up (2026-09-16): item 3's "premature until needed" call was tested by actually running a BPM-style search on the Billmann 2026 HAP1 map — see experiments/human_gi_bpm/ for the pilot, which found 105 module pairs, 39 of them touching dismech-curated disease genes on both sides (OXPHOS × GATOR1/TSC mTORopathy machinery being the standout lead).