Reactions with zero inputs or zero outputs are dropped, taking their catalysts and regulators

[adamjohnwright]

src/reaction_generator.py:663-667 (decompose_by_reactions skip) and src/logic_network_generator.py:1817 (if not input_ids or not output_ids: continue).

Dropped reactions never enter reaction_id_map, so get_catalysts_for_reaction / get_*_regulators_for_reaction never see them — the catalyst and regulator edges vanish silently too.

Catalog-wide measurement (fresh v97, 92 pathways)

Now detected automatically by scripts/validate_logic_network.py after #57:

81 of 92 pathways: 11/11 clean
10 pathways missing reactions, 16 reactions total

pathway	missing
R-HSA-170834 Signaling by TGF-beta Receptor Complex	8
R-HSA-201451	2
R-HSA-74752, R-HSA-73894, R-HSA-6785807, R-HSA-5693606, R-HSA-5693532, R-HSA-446728, R-HSA-2559583, R-HSA-195721 Signaling by WNT	1 each

Every one is a BlackBoxEvent with in=1,out=0 or in=0,out=1:

R-HSA-9974400  in 1 out 0  26S Proteasome degrades polyubiquitinated BRCA1, BARD1
R-HSA-2176503  in 1 out 0  Degradation of SMAD2/3:SMAD4 complex
R-HSA-2169046  in 1 out 0  Degradation of TGF-beta receptor complex
R-HSA-4608855  in 1 out 0  PRICKLE1 is degraded by the proteasome
R-HSA-6790038  in 0 out 1  Expression of HSP90B1

Reactome-wide scope: 8 reactions with no input and 490 with no output, out of 16,344 human RLEs; those 498 carry 254 catalysts and 12 regulators that are lost with them.

Why it matters

Proteolytic degradation is how Reactome curates the dominant negative-feedback arm of many pathways. TGF-β loses its entire SMAD2/SMAD3/SMAD7/SKI/receptor degradation set, so the E3-ligase arm of that pathway does not exist in the network at all. On the input side, "Expression of …" BBEs disappear with their PositiveGeneExpressionRegulation regulators.

R-HSA-195721 (Signaling by WNT) is one of the ten GSoC evaluation pathways, and it is missing R-HSA-4608855 ("PRICKLE1 is degraded by the proteasome").

Design question

A reaction with an input and no output still has real modelling content: it consumes its input, and its catalyst controls that consumption. Skipping it is defensible only if the Hungarian input/output pairing cannot represent a one-sided reaction — which is an implementation constraint, not curator intent.

[adamjohnwright]

Measured before implementing, and the modelling loss is far smaller than I framed it. Downgrading this.

I set out to fix it by emitting catalyst -> input depletion edges (the existing edge type, whose divide-form inhibition is exactly "a knocked-out catalyst boosts the substrate"). Then I sized it.

The catalog-wide effect is one edge

Reactions where a depletion edge is actually representable — input-only, has a catalyst, and both the input and the catalyst exist as nodes in the generated network:

scope	representable reactions	new depletion edges
the 10 GSoC evaluation pathways	1	1
the whole 92-pathway catalog	1	1

The one is R-HSA-4608855 (PRICKLE1 degradation) in WNT.

Why: these reactions have no curated driver

My "TGF-β loses its entire E3-ligase arm" claim was wrong. Seven of TGF-β's eight dropped reactions have no catalyst and no regulator at all. Verified directly — their only outgoing relationships are:

R-HSA-2176503  Degradation of SMAD2/3:SMAD4 complex
   compartment, inferredTo, input, literatureReference,
   precedingEvent, reviewStatus, species, summation
R-HSA-2169046  Degradation of TGF-beta receptor complex      (same)
R-HSA-2186780  Degradation of Ub-SMAD7                       (same)

No catalystActivity, no regulatedBy, no output. They are terminal sinks with no curated driver. The eighth (R-HSA-173545) does have a catalyst, but its input is not a node in the network.

A depletion edge needs something to deplete with. With no driver, the only representable content is "this entity is constitutively consumed" — a constant sink, which is inert in a steady-state perturbation model because it cannot respond to any perturbation. So the network is not losing signal here; it is losing a node that would do nothing.

Reactome-wide the picture is different — 490 input-only reactions, 300 with catalysts — so this becomes worth building for a substantially larger catalog. It is not worth building for the current one, and it could not be validated: the benchmark would see a single edge.

What stands

• The coverage gap is real and the validator now reports it (81/92 pathways clean, 10 with missing reactions), which is the right place for it to live.
• The 8 output-only reactions catalog-wide (2 with a regulator) are similarly not worth special handling.
• Worth changing cheaply: the skip currently logs at warning for all cases, which conflates "dropped a driverless sink, harmless" with "dropped a reaction that had modelling content". Only the latter deserves attention.

Re-open at full priority when the catalog grows beyond the current 92 pathways, or if a pathway with driver-bearing degradation reactions enters the evaluation set.