Cytochrome P450 Project

Bottom line: the 60 reviewed human cytochrome P450 entries carry 556
catalytic-activity lines, 528 of them experimental, covering 297 distinct
reactions — and GO represents that with 84 substrate-naming molecular-function
terms, of which 201 of the 297 reactions (68%) reach no GO term at all. The
annotation that does exist is sound: 48 of the 56 CYPs with a substrate-naming
term have experimental evidence for it, only 4 rest on IBA, and the
closure-filtered propagation gap is zero. So this family's problem is neither
bad annotation nor failed propagation — it is resolution. The sharpest
evidence is an asymmetry inside one protein: CYP4F3 has all three steps of
leukotriene-B4 ω-oxidation as GO terms, each mapped to its own RHEA reaction,
while the chemically identical C22/C26 fatty-acid ω-oxidation cascade on the same
protein has a term for step 1 and none for steps 2 and 3. GO is willing to
represent these reactions; nobody has requested the terms.

We did this because the P450s are where every general claim about enzyme
annotation quality gets tested: they are the most substrate-promiscuous family in
the human proteome, 21 of the 60 are disease genes, and the repository's own
coverage of them is badly skewed — all 13 CYP reviews here are
endogenous-substrate enzymes, and none of the drug-metabolising CYP1/2/3/4
enzymes has been reviewed at all.

Overview

This project audits the GO molecular-function representation of the human
cytochrome P450 superfamily as a whole, rather than one enzyme at a time. It
grew out of the family scan in the RHEA project
(RHEA-PROMISCUOUS-FAMILIES.md), which found
CYP450s to be the extreme case of reaction-to-term drop-out; this project asks
what that means for the annotations themselves and turns it into a curation
worklist.

Everything is computed live by cyp_go_audit.py
against UniProtKB, QuickGO, rhea2go and ec2go. Per-gene, per-reaction and
per-chemistry tables are in data/, starting with
cyp-summary.json. No result is
hardcoded.

How "informative" is decided

Calling a term "generic" by eye is where an audit like this usually goes wrong,
so the cut is derived from the data. For every catalytic MF term annotated
anywhere in the family, count how many other annotated terms are its is_a
descendants. The counts break sharply:

The gap between 24 and 6 separates five family-wide hub terms from the tail
of real activities that merely have a child or two. Terms at or above 20
descendants are treated as hubs (--hub-threshold, and the whole distribution is
emitted so the choice can be checked). Each CYP then falls in one of three tiers:

• substrate-resolved — carries a non-hub activity term that is also maximally
  specific within the family;
• partially-resolved — carries a non-hub term, but the family resolves it
  further on other members;
• unresolved — every catalytic term it carries is a hub, so GO does not say
  which reaction it catalyses.

Results

The annotation is curated, not inherited

The result that constrains everything else: of the 56 CYPs carrying a
substrate-naming activity term, 48 have experimental evidence for it and only
4 depend on IBA. Nor is anything failing to propagate — applying is_a closure,
no entry lacks a term that one of its mapped reactions would supply. This
family is not a propagation-failure or over-propagation story, and the review
actions one reaches for in those cases (REMOVE, MARK_AS_OVER_ANNOTATED) are
mostly not what its annotations need.

What it lacks is resolution. 297 experimentally curated reactions are represented
by 84 substrate-naming terms, a median of 3 per enzyme, and 184 of the
unmapped reactions carry experimental evidence. The information exists in
UniProt and stops at GO's door.

The cascade asymmetry: CYP4F3 against itself

The single clearest gap in the family, and the one that rules out "GO does not
model these reactions" as an explanation.

Fatty-acid ω-oxidation is three sequential oxidations by the same P450:
ω-methyl → ω-hydroxy → ω-oxo → dicarboxylate. For leukotriene B4, GO has all
three steps, and rhea2go maps all three:

For the C22 and C26 fatty-acid route — on the same enzymes, all experimental —
only step 1 has a term:

CYP4F3 carries both cascades. The LTB4 one is fully represented in GO; the
dicarboxylic-acid one, which is the route to the substrates of peroxisomal
β-oxidation and is implicated in ω-oxidation induction in fatty-acid oxidation
disorders, stops after the first step. GO:0010430 fatty acid omega-oxidation
exists as a biological process, so the pathway is known to GO — but a BP term
cannot receive a rhea2go mapping, which needs a molecular function.

The GO:0097258 / GO:0097259 pair is therefore an in-family template: two
new terms, built on a precedent from the same proteins, would close steps 2 and 3
for this route. Added as batch 8 of the RHEA mapping set (below).

Neither proposed name carries a chain-length qualifier, deliberately: each
pair's anchors are docosanoate (C22) and hexacosanoate (C26), and C22 is
long-chain by GO's own boundaries — GO:0102033 is defined for 13–22 carbons
and GO:0140692 for more than 22 — so a "very-long-chain" term would exclude
one anchor of each pair. GO already has an unqualified tier for exactly this
reason: GO:0120250 fatty acid omega-hydroxylase activity is the is_a parent
of both chain-length children. The step-3 name follows its template's suffix
(GO:0097259 is a monooxygenase) rather than reusing step 2's "ω-oxidase".

Zoomed out, the same shape recurs across the cascade steps, stated here from the
reactions themselves rather than from the coarse chemistry classification below.
Of the 22 reactions that form an oxo group, 18 are unmapped, and all 17 of
them outside the eicosanoid class are unmapped (17/17) — CYP24A1's vitamin-D
C24 oxidation chain, CYP19A1's 19-oxo aromatase intermediate, CYP51A1's
32-oxolanosterol step, CYP26C1's 4-oxoretinoate steps and CYP11B2's 18-oxocortisol
step. The four that are mapped are all eicosanoid: three
hydroperoxy-icosatetraenoate dehydrations onto GO:0106256, plus RHEA:48668,
which is step 2 of the LTB4 cascade that this page uses as its precedent. The two
reactions that form the second carboxylate (RHEA:39043, RHEA:39047, batch 8's
step 3) are both unmapped, while the LTB4 equivalent RHEA:48672 is mapped to
GO:0097259. So for both cascade steps the only GO coverage that exists is on the
one substrate the template comes from.

Where GO represents a cascade with a single term it uses a net-reaction one —
GO:0070330 aromatase activity and GO:0008398 sterol 14-demethylase activity
are both defined with 3 O₂ and 3 reductase equivalents, releasing formate — so
RHEA's partial reactions have nothing to land on there. The LTB4 trio proves the
per-step alternative is also acceptable to GO.

On the chemistry classification: a triage aid, not a measurement

cyp_go_audit.py also labels each reaction with a chemistry and a substrate class
by pattern-matching its equation, emitted as
data/cyp-chemistry.tsv and per-reaction
in data/cyp-reactions.tsv. No claim on
this page rests on those aggregates, deliberately: a regex over equation strings
cannot reliably decide which group a reaction forms, and this one has two known
failure modes in opposite directions.

It originally attributed a group the substrate already carried to the enzyme, so
ω-hydroxylation of an epoxy fatty acid (9,10-epoxyoctadecanoate → 18-hydroxy-9,10-epoxy-octadecanoate) counted as an epoxidation. Requiring the
pattern to match the product side and not the substrate side fixes that, but it
then misses poly-hydroxylation, where the substrate already has a hydroxyl
(20S-hydroxycholecalciferol → 20S,25-dihydroxycholecalciferol): 46 reactions
now fall to other/unclassified despite forming a hydroxyl, including
RHEA:48664, which this page cites above by id.

Three further corrections came out of the same pass. Saturated
docosanoate/hexacosanoate were being classed as PUFAs by a docosa pattern
meant for docosahexaenoate. The third step of ω-oxidation had no class at all
until a carboxylate formation (aldehyde to acid) one was added, and that
pattern initially matched only dioate/dicarboxylate, so RHEA:48672 — whose
product is 20-carboxy-leukotriene B4 — fell to other/unclassified for a
third reason rather than the poly-hydroxylation one. Matching carboxy as well
puts it where it belongs: the step-3 class now holds 3 reactions, and the only
mapped one is RHEA:48672 itself, the LTB4 template, with batch 8's
RHEA:39043/RHEA:39047 the two unmapped ones. The asymmetry this page is about
is now visible in the classification rather than contradicted by it.

Use the file to find candidate reactions, not to quantify a class. Per-reaction
labels are emitted precisely so they can be checked.

CYP4F22: the one enzyme GO does not describe

Four CYPs are unresolved, and three are explained: CYP4Z2P and CYP2G1P
are pseudogenes, and CYP20A1 is a genuine orphan with no curated reaction, so
GO's silence is correct. That leaves one:

CYP4F22 (Q6NT55) carries two experimental ultra-long-chain fatty acid
ω-hydroxylation reactions, EC 1.14.14.177 (which has no ec2go line), and its
entire GO molecular function is monooxygenase activity, the
oxidoreductase activity, acting on paired donors… parents, iron ion binding,
heme binding and protein binding. Nothing in GO states what it does. It is
the gene mutated in autosomal recessive congenital ichthyosis 5, where the
ω-hydroxylation of ultra-long-chain fatty acids for acylceramide synthesis is
the disease mechanism. GO:0140692 very long-chain fatty acid omega-hydroxylase activity is defined for chains longer than 22 carbons and covers both reactions
without a new term being needed — so this is a mapping gap, closed by batch 7 of
the RHEA set.

The four partially-resolved CYPs are a milder version: CYP26C1 and CYP2W1
have retinoic acid 4-hydroxylase activity by IDA, CYP2A7 and CYP2F1 have
arachidonate epoxygenase activity, in each case a term that other family
members resolve further.

Repository coverage is skewed to endogenous metabolism

The 13 reviewed CYPs — CYP11A1, CYP11B1, CYP11B2, CYP17A1, CYP19A1, CYP21A2,
CYP27A1, CYP39A1, CYP46A1, CYP51A1, CYP7A1, CYP7B1, CYP8B1 — are, without
exception, endogenous-substrate enzymes: steroidogenesis, bile-acid
synthesis, sterol and vitamin-D metabolism. Not one of the
xenobiotic-metabolising CYP1, CYP2, CYP3 or CYP4 enzymes has a review, and those
are precisely the promiscuous ones carrying the largest unmapped reaction sets
(CYP4F3 35, CYP3A4 29, CYP4F2 21, CYP1A1 18).

This is a defensible historical accident — the endogenous CYPs arrive through
pathway projects — but it means the family's hardest annotation questions are all
in the unreviewed half.

New GO term proposals

Consolidated from this audit and the RHEA family scan. All rows live in
rhea2go.sssom.yaml and validate with
just validate-rhea-mappings.

And two mapping-only fixes, needing no new term: CYP4F22's two ULCFA
reactions onto GO:0140692, and the CYP2E1/CYP4F ω and ω-1 hydroxylation
instance reactions onto GO:0102033 / GO:0120319 (34 reviewed annotations).

Curation Recommendations

1. Do not read low reaction-to-term ratios as over-annotation. Every review
   action that removes or downgrades an annotation is the wrong tool here: the
   existing terms are experimentally grounded and nothing is over-propagated. The
   actions this family needs are NEW (rarely, per the term proposals above) and
   mapping additions.
2. Check for an in-family template before proposing a term. The LTB4 trio
   settles the ω-oxidation proposals in a way that no argument from first
   principles could. Ask whether GO already represents this exact chemistry for a
   different substrate on the same protein.
3. Distinguish the cascade-step gap from the per-substrate gap. A missing
   term for an intermediate step of a multi-step transformation is a real gap
   with a precedent. A missing term for one more substrate of an activity GO
   already names is not — per
   RHEA-PROMISCUOUS-FAMILIES.md, GO
   obsoleted GO:0018777 for exactly that reason (provenance for the
   obsoletion is recorded there).
4. Review the xenobiotic CYPs. The unreviewed half holds the promiscuity, the
   pharmacogenomics and the unmapped reactions.
5. unresolved is a worklist, not a verdict. Three of the four unresolved
   CYPs should stay that way (two pseudogenes and an orphan); the tier is useful
   precisely because it is small enough to adjudicate by hand.

Follow-Up Targets

Methods

cyp_go_audit.py fetches reviewed human CYP entries (UniProt family query),
their CC CATALYTIC ACTIVITY RHEA cross-references and evidence codes, their
GOA molecular-function rows from QuickGO with evidence and reference, and is_a
ancestor closures for every term involved. Hub/tier classification, provenance,
reaction coverage and the chemistry classification are computed from those
inputs. The chemistry labels count a group as formed only when the pattern
matches the product side and not the substrate side; see the caveats above
before using them for anything but triage. Review coverage is read from genes/human/<GENE>/<GENE>-ai-review.yaml.

cd projects/CYTOCHROME_P450
uv run python cyp_go_audit.py --out data
uv run python cyp_go_audit.py --out data --hub-threshold 25   # move the hub cut

Each run stamps the GO release it read rhea2go/ec2go from, plus a
retrieved_at timestamp, into cyp-summary.json (rhea2go_go_release,
ec2go_go_release), so a committed table can be tied to a known snapshot; the
figures on this page are from releases/2026-07-26. The script refuses to
report rather than silently truncate if either source API returns a full page.

Outputs: data/cyp-summary.json, data/cyp-entries.tsv (per gene, with tier,
provenance and review status), data/cyp-reactions.tsv (per reaction, with the
rhea2go/ec2go comparison and chemistry class), data/cyp-chemistry.tsv,
data/cyp-propagation-gaps.tsv (empty).

Project Status

• Started: 2026-09-30
• Maturity: IN_PROGRESS — family-wide audit complete and reproducible; term
  proposals recorded in the RHEA SSSOM set; the exemplar gene reviews and the GO
  new-term requests are the open work.
• Computed live: 60 reviewed entries, 556 reaction annotations (528
  experimental), 297 distinct reactions, 96 mapped / 201 unmapped (191 without an
  EC), 107 MF terms (89 catalytic, 84 substrate-naming), tiers 52/4/4, provenance
  48 experimental / 4 IBA / 4 other, 0 closure-filtered propagation gaps, 13 of
  60 reviewed in this repository.
• Related: RHEA ·
  RHEA-PROMISCUOUS-FAMILIES.md ·
  ENZYME_SPECIFICITY