Regulator wiring: set-pool modes, inert-drug and conserved-participant switches (all default-neutral), specs/032-038 by adamjohnwright · Pull Request #76 · reactome/deltasignal

specs/033: pre-register one pool node per set-valued catalyst/regulator (1,881 catalyst groups x AND in 63 pathways; RAF self-requiring loop); mode chosen on tuning, judged on held-out

adamjohnwright

and

claude

committed

5934684

DS_SET_POOL_MODE=product|extreme|geomean|mean (specs/033): a node whose inputs are all set_member edges is a set pool; its member folds combine by the mode (product = the AND it replaces; extreme =…

adamjohnwright

and

claude

committed

ec2087c

specs/033 amendment before any arm (review): bench loaders skip set_pool nodes so readouts/pins are unchanged; pooled depletion specified and self-depletion fixed (LNG 81499ef); cycle gate restated…

adamjohnwright

and

claude

committed

039ce74

specs/033 result: pooled structure + product passes all gates (held-out +6, tuning +10, exp +3); mode chosen on tuning; RAF only partly fixed (14->17/49), second cause = inputs wired to recycled do…

adamjohnwright

and

claude

committed

5660f34

specs/034: regulator-wiring inventory -- 2,207 annotations by role x shape; generator folds Requirement and gene-expression regulation into positive; 32% of catalyst edges fed by a downstream (recy…

adamjohnwright

and

claude

committed

4ae9a88

specs/034: M6 traced (self-contained + exogenous inhibitors: SARS-CoV-2 N:M:PDPK1, ERBB2:trastuzumab, Inhibin:ACVR2A), exogenous = 202 drug + 66 pathogen entities; M4 counted: 7,989 input edges (9%…

adamjohnwright

and

claude

committed

4bc4341

specs/035: pre-register conserved loop participants held at baseline (D1): 1,722 loop-only inputs released unchanged by their loop (RAF: RAS GAPs, scaffold partners, F-actin, CNKSR)

adamjohnwright

and

claude

committed

69074f4

DS_CONSERVED_MODE=off|inert (specs/035): loop inputs released unchanged by their own loop held at baseline like cofactors; label-independent SCC detection via containment; solve reports conserved_r…

adamjohnwright

and

claude

committed

1fcc1bc

specs/035: D1 NOT ADOPTED -- held-out +42 (p=0.002) but experimental -122, TGF-b -30, MHC -18; PIP3 -308 traced: PIP3 itself met 'released unchanged' (second messenger bound/released) -- the criter…

adamjohnwright

and

claude

committed

46c038f

specs/035: RAF did not move under D1 (14/49 both); whole experimental loss is PIP3; next is an iteration trace of the RAF solve, not another rule

adamjohnwright

and

claude

committed

f09e9f8

specs/034 §8: RAF traced by iteration -- collapse starts at the scaffold binding reaction: PEBP1 self-contained inhibitor pushes, a 27-input AND of ~20 recycled set alternatives amplifies (gain ~20…

adamjohnwright

and

claude

committed

a9f4bb9

specs/036: pre-register over-cap set bundles as one input per curated component (96 cap hits / 21 pathways; RAF binding reaction 28 inputs from a 12000-combination cap)

adamjohnwright

and

claude

committed

7ab6d8c

specs/036: RAF sanity trace -- cap pools alone don't stop the collapse; compounding amplifiers: capped bundles, drug complexes, catalyst=substrate squaring; with all three addressed (mean) KO reads…

adamjohnwright

and

claude

committed

5c56e0c

Merge feat/inert-drugs + feat/set-pool into feat/combined-037 (both suites registered; CLAUDE.md keeps both config entries)

adamjohnwright

and

claude

committed

f1f9c20

specs/037: pre-register the combined structural fixes from the RAF trace (cap pools + set pools + inert drugs + dedup, mean pools) against base; attribution arms reported

adamjohnwright

and

claude

committed

88c3d8d

specs/037 revision after review, before any arm: fresh control build; experimental gate excludes RAF (the traced case); concentration gate; no pathway (incl. tuning) below -10; 033 +6/+3 was produc…

adamjohnwright

and

claude

committed

c135255

specs/037: cite LNG bddfe78

adamjohnwright

and

claude

committed

6f261e2

specs/037: combined NOT ADOPTED -- held-out -162, exp excl RAF -12; dedup is the damage (IFN a/b -112, TP53 -80); pools+drugs neutral; RAF 14->23/49 but nothing generalises

adamjohnwright

and

claude

committed

8c4962f

specs/034 §9: RAF vs MP-BioPath's own network -- theirs keeps one node per set with OR over curated alternatives, no member fan-out, no catalyst duplication, no scaffold release on dissociation; lo…

adamjohnwright

and

claude

committed

fbcb6a8

specs/034 §10: MP-BioPath hand edits analysed -- 1,800 edits / 71 pathways; cyclic 1,007->413; loop cutting = reaction->output deletions (84 of 148 in loops) and reaction->set-node (38 of 63); 607 …

adamjohnwright

and

claude

committed

617dbb4

DS_SET_POOL_MODE=max (the literal OR: strongest member fulfils the role) + specs/038 pre-registration: set pools for regulators/catalysts only, max vs product vs fresh control

adamjohnwright

and

claude

committed

5a2a927

specs/034 §11: UUIDs already resolve 86% of MPB's hand-cut loop artefacts; our cyclic nodes are mostly curated input/output reaction cycles (derived edges: depletion 18%, regulator 13%, catalyst 6%…

adamjohnwright

and

claude

committed

985cea2

specs/034 §12: MPB's 219 added inhibitions -- mostly reversal enzymes (phosphatase, demethylase, GAP, DUB) against the modified form; 187 absent from ours, 7 present as depletion; the automatic equ…

adamjohnwright

and

claude

committed

a0c4e62

specs/038 result: max NOT ADOPTED (held-out -120, exp -57, PIP3 -140); product passes all gates (held-out -2, exp +8 concentrated in 4 perturbations) -- set pools adoptable on faithfulness under pr…

adamjohnwright

and

claude

committed

09451ed

Merge remote-tracking branch 'origin/design/regulator-wiring' into feat/pool-max

adamjohnwright

committed

e750e43

Consolidate: drugs (032), set pools + max (033/038), conserved participants (035), regulator-wiring record (034), cap pools / combined (036/037); CLAUDE.md indexes 032-038; 16 asserting suites

adamjohnwright

and

claude

committed

9e8a9ad

Review of #76: specs/038 status, control name, predictions scored, self-inhibitor flags measured (1,215->1,192, 4 pathways); specs/033 pool->complex assembly note; CLAUDE.md wording (concentrated; …

adamjohnwright

and

claude

committed

9f9d26c

Navigation Menu

Repository navigation

Regulator wiring: set-pool modes, inert-drug and conserved-participant switches (all default-neutral), specs/032-038 - #76

Regulator wiring: set-pool modes, inert-drug and conserved-participant switches (all default-neutral), specs/032-038#76
adamjohnwright merged 31 commits into
mainfrom
feat/pool-max

Commits on Sep 26, 2026

specs/032: pre-register inert drug entities (RAF experimental 28.6% vs MP-BioPath 93.9% traced to MAP2K/MAPK-inhibitor drug complexes collapsing a loop)

DS_DRUG_MODE=propagate|inert (specs/032): drug-derived nodes from the bundle's drugs.csv held at baseline like cofactors; default propagate is byte-identical. Network carries drug_stids (nothing = …

Review of feat/inert-drugs: POSTed networks carry drug_stids (an edge-drop arm under inert had no drug table); ARM.json records the Drugs line; a confidence-0 observation no longer unpins a drug; P…

specs/032: inert drugs NOT ADOPTED -- experimental -5 (gate +10), RAF worse; held-out +30. RAF collapse traced to a positive loop closed by catalyst + dissociation recycling (p-MAP2K catalyses Phos…

Commits on Sep 27, 2026

specs/033: pre-register one pool node per set-valued catalyst/regulator (1,881 catalyst groups x AND in 63 pathways; RAF self-requiring loop); mode chosen on tuning, judged on held-out

DS_SET_POOL_MODE=product|extreme|geomean|mean (specs/033): a node whose inputs are all set_member edges is a set pool; its member folds combine by the mode (product = the AND it replaces; extreme =…

specs/033 amendment before any arm (review): bench loaders skip set_pool nodes so readouts/pins are unchanged; pooled depletion specified and self-depletion fixed (LNG 81499ef); cycle gate restated…

specs/033 result: pooled structure + product passes all gates (held-out +6, tuning +10, exp +3); mode chosen on tuning; RAF only partly fixed (14->17/49), second cause = inputs wired to recycled do…

specs/034: regulator-wiring inventory -- 2,207 annotations by role x shape; generator folds Requirement and gene-expression regulation into positive; 32% of catalyst edges fed by a downstream (recy…

specs/034: M6 traced (self-contained + exogenous inhibitors: SARS-CoV-2 N:M:PDPK1, ERBB2:trastuzumab, Inhibin:ACVR2A), exogenous = 202 drug + 66 pathogen entities; M4 counted: 7,989 input edges (9%…

specs/035: pre-register conserved loop participants held at baseline (D1): 1,722 loop-only inputs released unchanged by their loop (RAF: RAS GAPs, scaffold partners, F-actin, CNKSR)

DS_CONSERVED_MODE=off|inert (specs/035): loop inputs released unchanged by their own loop held at baseline like cofactors; label-independent SCC detection via containment; solve reports conserved_r…

specs/035: D1 NOT ADOPTED -- held-out +42 (p=0.002) but experimental -122, TGF-b -30, MHC -18; PIP3 -308 traced: PIP3 itself met 'released unchanged' (second messenger bound/released) -- the criter…

specs/035: RAF did not move under D1 (14/49 both); whole experimental loss is PIP3; next is an iteration trace of the RAF solve, not another rule

specs/034 §8: RAF traced by iteration -- collapse starts at the scaffold binding reaction: PEBP1 self-contained inhibitor pushes, a 27-input AND of ~20 recycled set alternatives amplifies (gain ~20…

specs/036: pre-register over-cap set bundles as one input per curated component (96 cap hits / 21 pathways; RAF binding reaction 28 inputs from a 12000-combination cap)

specs/036: RAF sanity trace -- cap pools alone don't stop the collapse; compounding amplifiers: capped bundles, drug complexes, catalyst=substrate squaring; with all three addressed (mean) KO reads…

Merge feat/inert-drugs + feat/set-pool into feat/combined-037 (both suites registered; CLAUDE.md keeps both config entries)

specs/037: pre-register the combined structural fixes from the RAF trace (cap pools + set pools + inert drugs + dedup, mean pools) against base; attribution arms reported

specs/037 revision after review, before any arm: fresh control build; experimental gate excludes RAF (the traced case); concentration gate; no pathway (incl. tuning) below -10; 033 +6/+3 was produc…

specs/037: cite LNG bddfe78

specs/037: combined NOT ADOPTED -- held-out -162, exp excl RAF -12; dedup is the damage (IFN a/b -112, TP53 -80); pools+drugs neutral; RAF 14->23/49 but nothing generalises

specs/034 §9: RAF vs MP-BioPath's own network -- theirs keeps one node per set with OR over curated alternatives, no member fan-out, no catalyst duplication, no scaffold release on dissociation; lo…

specs/034 §10: MP-BioPath hand edits analysed -- 1,800 edits / 71 pathways; cyclic 1,007->413; loop cutting = reaction->output deletions (84 of 148 in loops) and reaction->set-node (38 of 63); 607 …

DS_SET_POOL_MODE=max (the literal OR: strongest member fulfils the role) + specs/038 pre-registration: set pools for regulators/catalysts only, max vs product vs fresh control

specs/034 §11: UUIDs already resolve 86% of MPB's hand-cut loop artefacts; our cyclic nodes are mostly curated input/output reaction cycles (derived edges: depletion 18%, regulator 13%, catalyst 6%…

specs/034 §12: MPB's 219 added inhibitions -- mostly reversal enzymes (phosphatase, demethylase, GAP, DUB) against the modified form; 187 absent from ours, 7 present as depletion; the automatic equ…

specs/038 result: max NOT ADOPTED (held-out -120, exp -57, PIP3 -140); product passes all gates (held-out -2, exp +8 concentrated in 4 perturbations) -- set pools adoptable on faithfulness under pr…

Merge remote-tracking branch 'origin/design/regulator-wiring' into feat/pool-max

Consolidate: drugs (032), set pools + max (033/038), conserved participants (035), regulator-wiring record (034), cap pools / combined (036/037); CLAUDE.md indexes 032-038; 16 asserting suites

Review of #76: specs/038 status, control name, predictions scored, self-inhibitor flags measured (1,215->1,192, 4 pathways); specs/033 pool->complex assembly note; CLAUDE.md wording (concentrated; …