ToxCast assay coverage of the dismech pathograph

How many ToxCast and Tox21 assay endpoints declare a gene target that a dismech pathophysiology node also names, counted from the side of the assays, the nodes, the genes and the diseases. Background: issue #12858; scope and interpretation: projects/TOXCAST.md.

A shared gene is a candidate, not a mapping. It says an endpoint and a node concern the same gene. It does not say they agree on direction, on the quantity measured, or on mechanism, so every figure below is an upper bound on what could be mapped.

Assay annotations retrieved from the EPA CTX API on 2026-09-30; kb/disorders/ as of commit 0aedd3ce8e. Both move, so treat the numbers as a dated snapshot and regenerate with just toxcast-coverage.

Pathograph node coverage

From the assay side

Assay endpoints	Count	Share of panel
In the panel	1570
Declare no usable gene target	462	29.4%
Declare a gene target	1108	70.6%
Candidates: a target is named by at least one node	552	35.2%
of which through the human gene	481	30.6%
of which through another species' ortholog only	71	4.5%

Of the endpoints that declare a gene target, 49.8% are candidates. An endpoint with no gene target cannot be found by this join at all, which is not the same as having no node it could speak to: a cell-viability counter-screen declares no gene and may still bear on a cell-death node.

By intended target family, largest first:

Intended target family	Endpoints	With a gene target	Candidates	Nodes reached	Entries reached
nuclear receptor	255	255	164	56	45
cell cycle	180	18	5	5	4
gpcr	158	158	47	39	33
kinase	112	112	55	92	79
cyp	78	78	11	6	6
cytokine	72	72	43	20	20
dna binding	71	71	42	104	84
neurodevelopment	64	0	0	0	0
background measurement	55	10	3	6	3
zebrafish development	46	0	0	0	0
channel 1	45	0	0	0	0
channel 2	45	0	0	0	0
transporter	45	45	26	18	15
protease	43	39	22	15	13
cell adhesion molecules	32	32	22	14	11
ion channel	32	32	22	28	21
phosphatase	27	27	8	23	17
cell morphology	26	3	3	1	1
esterase	26	26	5	7	4
steroid hormone	24	22	16	16	13
oxidoreductase	17	16	7	12	11
neuroactivity	15	0	0	0	0
transcription factor	15	15	10	7	6
transferase	14	14	8	6	5
growth factor	10	10	6	9	9
lyase	8	8	6	4	4
hydrolase	7	7	2	4	4
deiodinase	4	4	0	0	0
misc protein	4	4	3	7	6
cardiomyocyte function	3	0	0	0	0
cytokine receptor	3	3	3	7	7
metabolite	3	0	0	0	0
mitochondria	3	0	0	0	0
protease inhibitor	3	3	2	1	1
stress response	3	3	0	0	0
apoptosis	2	2	2	1	1
dehalogenase	2	2	1	1	1
epigenetic enzyme	2	2	1	1	1
growth factor receptor	2	2	0	0	0
ligase	2	2	2	2	1
membrane protein	2	2	0	0	0
mutagenicity response	2	2	0	0	0
apolipoprotein	1	1	1	2	2
catalase	1	1	1	1	1
enzyme	1	0	0	0	0
filaments	1	1	0	0	0
immunoglobulin	1	1	1	1	1
methyltransferase	1	1	1	2	2
microrna	1	1	0	0	0
oxidase	1	1	1	1	1

Nodes and entries are counted once per family, so those two columns do not sum to the totals below: one node can be reached from several families.

From the node side

Pathophysiology nodes in `kb/disorders/`	Count	Share of all nodes
All nodes	22403
Name at least one HGNC-bound gene	4021	17.9%
Candidates: name a gene some endpoint targets	452	2.0%

That is 11.2% of the nodes that name a gene. Pairing every candidate endpoint with every node it reaches gives 2080 endpoint-node pairs, which is the size of the job if each candidate were reviewed.

Candidate nodes by biological_scale, against all nodes:

`biological_scale`	Candidate nodes	All nodes	Share of that scale
MOLECULAR	186	4545	4.1%
CELLULAR	37	3981	0.9%
TISSUE	5	3271	0.2%
ORGANISM	2	2306	0.1%
(untagged)	222	8300	2.7%

64 of the 452 candidate nodes (14.2%) carry a genetic_context, meaning the node records a variant. Those are the nodes most likely to name a lesion where an assay reads a state, which is the first mismatch #12858 describes. The remainder are not thereby cleared: a node can name a lesion in prose without the slot.

Gene coverage

The panel names 496 distinct gene targets. 433 are named through the human gene by at least one endpoint; 63 are named only through another species' gene, and join to dismech only where the upper-cased symbol happens to be a human gene's symbol.

Where the target sits in `kb/disorders/`	Human targets	Ortholog-only targets	All	Endpoints naming one
`ON_NODE`: named by a pathophysiology node in at least one entry	180	14	194	552
`GENETIC_ONLY`: bound in a `genetic:` record, on no node	31	4	35	93
`ELSEWHERE_ONLY`: bound elsewhere in an entry (a subtype, a treatment, a biomarker), in no `genetic:` record and on no node	0	0	0	0
`ABSENT`: bound nowhere in `kb/disorders/`	222	45	267	486

Each target takes the first tier it satisfies. The endpoint column can count one endpoint in two rows, because 41 endpoints name more than one gene.

The 20 most-assayed targets, and how far each reaches:

Target	Endpoints	Tier	Nodes	Entries
ESR1	40	`ON_NODE`	8	7
AR	34	`ON_NODE`	6	4
PPARG	22	`ON_NODE`	4	4
NR3C1	16	`ON_NODE`	1	1
PGR	14	`ABSENT`	0	0
ESR2	13	`ON_NODE`	2	2
THRA	13	`ON_NODE`	1	1
THRB	13	`ON_NODE`	1	1
CXCL8	10	`GENETIC_ONLY`	0	0
NR1H4	10	`ON_NODE`	4	3
NR3C2	10	`ABSENT`	0	0
TP53	10	`ON_NODE`	38	36
CCL2	9	`ON_NODE`	1	1
CYP2C19	8	`GENETIC_ONLY`	0	0
NR1I2	8	`ABSENT`	0	0
PPARA	8	`ABSENT`	0	0
SRC	8	`ABSENT`	0	0
VCAM1	8	`ON_NODE`	1	1
CYP1A2	7	`ABSENT`	0	0
CYP2C9	7	`ABSENT`	0	0

The panel is very uneven: 247 targets appear in exactly one endpoint and 12 appear in ten or more.

Symbols the join rewrote, because ToxCast writes them differently from HGNC's approved symbol:

FOS|JUN is read as FOS and JUN
H2AFX is read as H2AX
H3F3A is read as H3-3A
PPP2R4 is read as PTPA

Disease coverage

Entries in `kb/disorders/`	Count	Share
All entries	3256
Carry at least one candidate node	317	9.7%
Bind a target somewhere, on no node	189	5.8%
Bind no target	2750	84.5%

In the middle row the gene is already bound in the entry, typically in genetic:, and no pathophysiology node names it. Whether it belongs on a node is a curation question for each entry, not a backfill: a gene on a node asserts that the gene drives that mechanism in that disease.

Candidate nodes in the entry	Entries
1	230
2-3	77
4 or more	10

The 20 targets reaching the most entries:

Target	Entries	Nodes	Nodes with a `genetic_context`	Endpoints
TP53	36	38	9	10
STAT3	20	21	0	2
PIK3CA	18	18	3	1
PTEN	10	15	0	3
FGFR3	9	10	1	2
FGFR1	8	10	0	2
GNAS	8	10	2	2
MYC	8	8	2	2
ESR1	7	8	1	40
JAK2	7	7	2	2
TNF	7	7	0	2
CACNA1A	5	8	1	4
FAS	5	5	0	1
HNF4A	5	5	0	1
IL6	5	5	0	3
NR5A1	5	6	0	1
PAX6	5	6	0	1
TGFB1	5	5	0	4
ABL1	4	4	2	2
AR	4	6	0	34

Mechanism modules

Counted apart, because a module is not a disease. 12 of the 898 pathophysiology nodes in kb/modules/ name a gene some endpoint targets, across 9 of 180 modules.

Regenerating

just toxcast-refresh                                 # cache the annotations; needs CTX_API_KEY
just toxcast-coverage                                # this report, to stdout
just toxcast-coverage --format tsv --table targets   # one row per gene target
just toxcast-coverage --format tsv --table nodes     # one row per candidate node
just toxcast-coverage --format tsv --table endpoints # one row per assay endpoint
just toxcast-coverage --format tsv --table diseases  # one row per entry with a candidate node