Gene-to-process jumps in the pathograph

Date: 2026-08-26 (figures re-measured 2026-09-25 against current main) Metrics: just compliance-connectivity (src/dismech/qc_plugins.py) Gate: just check-gene-activity-grounding (scripts/check_gene_activity_grounding.py) Corpus: kb/disorders, 3,149 entries

The question

GO names three levels between a gene and what a cell can no longer do: gene → molecular function → biological process. A pathograph edge running from a genetic: node to a pathophysiology node annotated only with biological_processes: skips the middle one — the graph says what the cell can no longer do without ever saying what the protein can no longer do.

That turns into two compliance questions, one after the other:

1. Is the gene in the pathograph at all? (genetic[].mechanism_outlink)
2. Does the node it lands on name a molecular function? (genetic[].mechanism_activity_grounding)

Both are graded coverage metrics in dismech.qc_plugins, alongside the existing phenotypes[].causal_inlink. They compose with weighted compliance and the conf/qc_config.yaml weights like any other field. Both gene metrics carry min_compliance: null: coverage is far too low for an absolute floor, so the second question is enforced as a ratchet instead — see Gating it below.

Where the KB stands

Phenotype connectivity:  21892/38839 nodes causally connected      (56.4%)
Gene-to-mechanism wiring: 3924/6444  causal genes wired            (60.9%)
Gene activity grounding:  1108/3924  wired genes land on an MF     (28.2%)

The grounding figure is what this branch leaves behind; on main it is 963/3924 (24.5%), so the tranches below move it 24.5% → 28.2% and cut the files carrying a gap from 1,330 to 1,201. The corpus moves several times a day, so read these as a dated snapshot — the recipes below print the current figures.

More than half of causal genes never reach the pathograph. A genetic[] entry connects only when some pathophysiology node carries the same gene in its gene:/genes: descriptor — that shared CURIE is the whole edge — so a gene with no such node sits in the genetic block and is invisible in the graph. 969 files have at least one. This is the larger of the two gaps, and it is the prior question: an unwired gene has no landing node to ground.

Of the genes that do reach the graph, more than seven in ten land on a node with no molecular function, across 1,201 files. The grounding denominator is the wired genes deliberately, so an unwired gene is charged once, against wiring, rather than twice.

Two related counts, for scale: 1,335 of 3,149 disorder files use molecular_functions: anywhere, and 1,068 pathophysiology nodes have an activity-shaped name (… molecular function deficiency, … Loss of Function, … Channel Dysfunction) — 567 of those, 53%, carry no MF term. The MF enum cache holds 870 terms against the BP cache's 2,777.

What the failures look like

Most often, the activity is already claimed in prose and only the term is missing. Carnitine-Acylcarnitine_Translocase_Deficiency has a node named SLC25A20 transporter molecular function deficiency, whose description says the variants "reduce mitochondrial inner membrane carnitine-acylcarnitine translocase activity" — annotated GO:0015879 carnitine transport and nothing else. Same shape in Lysosomal_Acid_Phosphatase_Deficiency (→ GO:0016311 dephosphorylation), NAGA_Deficiency_Type_3, MGAT2-congenital_disorder_of_glycosylation, Spinocerebellar_Ataxia_Type_2. Of the 1,492 single-gene landing nodes that still fail the check, 388 assert the activity in their own prose. Nothing new has to be established for those.

Sometimes the whole cascade is one node. Growth_Hormone_Insensitivity_Syndrome / "GH-IGF1 Axis Disruption" spells the chain out in its description — GH → GHR → JAK2 → STAT5B phosphorylation → dimerization → nuclear translocation → IGF1 transcription — while the graph holds a single node with two BP terms. 70 of the 181 pathway-landing nodes carry descriptions over 400 characters; the chain is often already written, one field away from being nodes.

Where a term is not the fix

• Many genes on one node. Primary_Ciliary_Dyskinesia / "Ciliary Dysfunction" carries 21 genes: DNAH5 and DNAH11 are dynein motors, RSPH1/4A/9 are radial-spoke structural constituents, CCDC39/40 are axonemal rulers, FOXJ1 and MCIDAS are transcription factors. No single MF term is true of that set. Also Dilated_Cardiomyopathy / "Sarcomeric and Cytoskeletal Dysfunction" (31 genes) and Autosomal_Recessive_Primary_Microcephaly / "Heterogeneous Biallelic MCPH Gene Dysfunction" (28). The check flags these, correctly, but the repair is to split the node.
• Classes with no shared molecular function. Autosomal_Recessive_Non-Syndromic_Intellectual_Disability has a node called "Loss of a Gene-Specific Molecular Function Required by Developing Neurons" whose description states outright that there are no prevalent genes, pathways or complexes and that the affected protein functions are very diverse. The absence of the term there is the finding.
• Genuinely genomic nodes. Dosage, imprinting and silencing claims have no activity step to name.

The fix is usually cheap

The molecular function is a property of the gene product, so it comes from GO directly rather than from disease literature. Ten fills drawn from the failing set, each verified against OLS as a live molecular_function:

entry	landing node (process-only today)	molecular_functions:
Carnitine-Acylcarnitine_Translocase_Deficiency	SLC25A20 transporter molecular function deficiency	GO:0015227 O-acyl-L-carnitine transmembrane transporter activity
Lysosomal_Acid_Phosphatase_Deficiency	Lysosomal Acid Phosphatase 2 Deficiency	GO:0003993 acid phosphatase activity
NAGA_Deficiency_Type_3	Alpha-N-Acetylgalactosaminidase Deficiency	GO:0008456 alpha-N-acetylgalactosaminidase activity
Aspartylglucosaminuria	AGA lysosomal enzyme deficiency	GO:0003948 N4-(beta-N-acetylglucosaminyl)-L-asparaginase activity
MGAT2-congenital_disorder_of_glycosylation	MGAT2 deficiency	GO:0008455 alpha-1,6-mannosylglycoprotein 2-beta-N-acetylglucosaminyltransferase activity
Familial_Hypercholesterolemia	LDLR Functional Defect	GO:0005041 low-density lipoprotein particle receptor activity
Brachyolmia	Deficient PAPS biosynthesis	GO:0004781 sulfate adenylyltransferase (ATP) activity
Ulnar-Mammary_Syndrome	TBX3 Haploinsufficiency	GO:0001227 DNA-binding transcription repressor activity, RNA polymerase II-specific
Spinocerebellar_Ataxia_Type_2	Loss of Ataxin-2 RNA-Binding and Translational Regulation	GO:0003729 mRNA binding
Growth_Hormone_Insensitivity_Syndrome	GH-IGF1 Axis Disruption	GO:0004903 growth hormone receptor activity (first link of a chain)

Two (GO:0005041, GO:0001227) are already in the MF enum cache and validate offline. The pattern to copy exists in 852 disorder entries — the inborn-errors files (ornithine_aminotransferase_deficiency, Primary_Carnitine_Deficiency, Trimethylaminuria) already carry MF and BP on the same node.

One caveat about inserting nodes

The node-classification work measured gene-grounded nodes at mean topological depth 0.94 and MF-grounded nodes at 0.81 — statistically the same place. GENOMIC and ACTIVITY read as alternative entry points in the current corpus, not sequential steps. So the metric should be read as "does the gene's landing name an activity", which is usually satisfied by adding a term to the node that already exists, rather than as a mandate to insert a node between every gene and its mechanism. See the node-classification brainstorm.

Running it

just compliance-connectivity                              # all three metrics, KB-wide
just compliance-connectivity --list-unconnected           # name the failing genes
uv run python -m dismech.qc_plugins kb/disorders/Asthma.yaml   # one entry
just compliance-connectivity --activity-fail-under 20     # gate on a threshold

Gating it

An absolute threshold cannot enforce this: at 28.2% coverage, a floor high enough to matter fails every PR and one low enough to pass gates nothing. So the second question is enforced as a ratchet, following check_environmental_evidence.py:

just check-gene-activity-grounding                 # runs in `just qc`
just check-gene-activity-grounding --count
just update-gene-activity-baseline                 # only ever to SHRINK

CI derives the grandfather set live from the base branch, so a PR fails only on genes it adds whose landing node names no molecular function. There is no snapshot to keep in sync and nothing for parallel curation PRs to race on.

Where it runs matters as much as what it checks. The gate is an ungated entry in the "Run whole-repo gates" step, passing --against-ref "origin/$BASE_REF", beside the pathograph connectivity floor. It cannot live in the path-filtered Python test lane: a curation PR touches only kb/, matches neither the python nor the schema filter, and so would never run the one check written for it — it would pass its own CI and then be ejected from the merge queue, which counts toward the ejection-strike hold. That is the same hole CLAUDE.md records for the entity-ref and duplicate-key sweeps (#9473). The pytest test_no_newly_ungrounded_genes is the exact twin of that step and is marked ci_step_twin, so it is deselected from the lane and still runs locally and in the nightly sweep; tests/test_ci_step_twins.py fails if the step is ever removed or path-gated, which brings the twin back into the lane. tests/gene_activity_grounding_baseline.txt (2,816 findings) is the local and shallow-checkout fallback, and is allowed to drift stale-high: a line for a gene since grounded grandfathers nothing.

Grandfathering a new line stays legitimate for the two shapes above — a many-gene bundle, and a class whose members share no molecular function. Neither is a reason to bind a term that overstates the node.

The phenotypes[].causal_inlink metric is separately gated by a corpus-level min_compliance floor (50.0, against a measured 56.4%); the two gene metrics stay advisory in the weighted score, with --genes-fail-under and --activity-fail-under for ad-hoc checks.