eggNOG-mapper on Non-model Genomes

eggNOG-mapper on non-model genomes

Back to the eggNOG-mapper evaluation

eggNOG-mapper is mostly used on genomes nobody has curated, so this page tests
it where that is true. It covers every completed review for three non-model
organisms in the corpus:
- Pseudomonas putida KT2440 (PSEPK, 847 proteins), a bacterial chassis
with almost no experimental GO;
- horse (HORSE, 30 proteins), a mammal absent from eggNOG 5;
- poplar (POPTR, 113 proteins), a tree whose proteins are in eggNOG but
carry almost no GO.

Field pennycress (THLAR) has only four reviewed genes, so it appears as
case studies rather than statistics. In total that is 994 proteins
(deep_dive/cohort.yaml).

The same pipeline as the main evaluation was used: emapper 2.1.13 with
database 5.0.2 and default settings, the same scoring, plus a manual check of
65 sampled terms (deep_dive/adjudication/).

Where the annotations come from

Genome Seed ortholog Median seed identity Level used for transfer Proteins given GO
P. putida KT2440 itself (798 of 847) 100% Gammaproteobacteria 483 of 847 (57%)
Horse other mammals (horse is not in eggNOG 5) 92% Metazoa (all 30) 29 of 30
Poplar poplar for 44, otherwise other plants 96% Streptophyta (109 of 113) 84 of 113 (74%)
Pennycress other Brassicales and Arabidopsis (not in eggNOG) 84% Streptophyta 4 of 4

Hiding each genome's own species changed almost nothing (P. putida 483 → 465
proteins with GO; horse unchanged). eggNOG holds essentially no GO for these
organisms, so every transferred term comes from other species. Unlike the
model organisms in the main cohort, the default run here is already an
honest test.

The transfer level is set by emapper's default auto scope list, and it is
broad: an entire kingdom-scale clade for horse and poplar. A horse protein can
therefore inherit terms from fly and worm orthologs, and a poplar protein from
moss and algae.

Automatic comparison

Default run, as % of predicted leaf terms:

Genome Leaf terms Median per protein EXACT BROADER NARROWER DOWNGRADED REJECTED UNDECIDED UNRELATED Core recall
P. putida 2,232 4 54.7 3.8 7.4 4.9 1.6 0.4 27.1 0.41
Horse 631 18 12.5 0.5 10.9 1.4 0.3 4.3 70.0 0.69
Poplar 388 4 26.8 7.0 16.2 5.4 0.0 0.0 44.6 0.35
Pennycress 12 2 25.0 0 8.3 16.7 0 0 50.0 0.27

UNDECIDED marks leaves related only to terms the review left UNDECIDED, which
records an absence of judgment rather than a negative one.

Two very different profiles appear.

Added value over GOA

Each review's core_functions terms were compared with what each source
covers:

Genome Core terms Already covered by GOA Recovered by emapper emapper only (not in GOA)
P. putida 2,455 1,878 (76%) 1,004 (41%) 85
Horse 45 38 31 1
Poplar 442 371 156 6
Pennycress 11 4 3 1

For a genome that already has UniProt's electronic annotation, eggNOG-mapper
recovers fewer of the reviewed core functions than GOA already holds. It adds
a core term GOA lacked for about one protein in ten (P. putida) or fewer.
Its value is as an independent, offline annotation for sequences that have no
UniProt entry, such as new assemblies and metagenomes, rather than as an
addition to GOA.

Manual check

Terms were sampled from the default run, as in the main evaluation: proteins
at random, then one term per protein. The rates are therefore per protein.
- P. putida: 24 terms (14 UNRELATED, 5 NARROWER, 5 REJECTED).
- Horse: 16 terms (12 UNRELATED, 4 NARROWER).
- Poplar: 16 terms (10 UNRELATED, 6 NARROWER).
- Pennycress: all 9 non-exact terms, a census of its 4 genes rather than a
sample.

The mix of strata differs by genome, and NARROWER terms are the safest, so
the rows are not directly comparable. The UNRELATED terms alone compare like
for like: wrong in 8 of 14 for P. putida, 5 of 12 for horse, 8 of 10 for
poplar and 3 of 6 for pennycress. These reviewers saw each term's automatic
category, and all 65 verdicts in the table below come from that pass. Ten
P. putida claims were later re-judged blind as well, in the main cohort's
blinded re-check: the five REJECTED claims, plus two NARROWER and three
UNRELATED claims used there as distractors. None of the ten crossed the
line between wrong and correct (two moved between PLI and NPI), so the table
is the same under either pass.

Genome Claims Correct (COR) True but vague (LSP) Undecided (UNC) Wrong (NPI/PLI/REP)
P. putida 24 8 2 0 14
Horse 16 7 0 4 5
Poplar 16 4 0 1 11
Pennycress 9 1 0 4 4

P. putida: phenotypes and paralogs from E. coli

The commonest error, 6 of 14, is a process term inferred from an indirect
effect in an enterobacterial ortholog:
- HtpG and RfbC given DNA damage response;
- IscU given intracellular iron ion homeostasis;
- SdhD given citrate metabolic process;
- SsuD given response to heat;
- PstC given regulation of phosphatase activity, an E. coli Pho-regulon
effect, carried onto a PstC from KT2440's second Pst-like operon.

The next is paralog pooling in large proteobacterial orthogroups:
- DavD, a glutarate-semialdehyde dehydrogenase, given the GABA shunt, which is
GabD's job;
- BetB given fermentation from other aldehyde dehydrogenases;
- the AcnD-family 2-methylcitrate dehydratase given aconitate hydratase.

All five terms the curators had removed were wrong again:
- IlvH, the regulatory subunit, given acetolactate synthase activity, which
belongs to the catalytic subunit;
- UbiX given 4-hydroxy-3-polyprenylbenzoate decarboxylase activity (it makes
the prenylated FMN that the real decarboxylase, UbiD, uses);
- Exonuclease I (SbcB) given RNA exonuclease activity;
- PrpD given 2Fe-2S binding, although the purified enzyme has no cluster.

Correct additions were concrete and structural: two 4Fe-4S motifs in PydA,
zinc binding in AroB, signal sequence binding for SecY, and homodimerization
for RlmH.

Horse: mostly right, wrong where mouse phenotypes leak in

Of the three genomes, horse is where orthology transfer adds most. The cost
is a long list per protein that still needs filtering.

Poplar: proteomics and paralogs from Arabidopsis

Poplar had the highest error rate (11 of 16).

Correct additions included SAM binding for the radical-SAM lipoyl synthase and
heme transport for ccsA.

Pennycress, gene by gene

Gene Leaves Verdict
FAE1 2 The core activity (fatty acid elongase) is an exact match, but ER membrane and very-long-chain fatty acid biosynthesis are missed. Captured, incompletely.
TFP 6 Glucosinolate catabolism is exact. It frames the protein as an enzyme regulator (outdated; TFP is an Fe²⁺-dependent lyase), misses the lyase activity, and imports Arabidopsis ESP's single-paper senescence, defence and nuclear terms. Partial and noisy.
CYP71B1 3 Generic P450 monooxygenase activity, which is all anyone knows. Its location and process terms are guesses.
matK 1 Only "chloroplast", which is correct. Group II intron splicing, the actual function, is missed.

Narrowing the taxonomic scope does not fix it

The broad transfer level suggests an obvious fix: emapper's pre-configured
--tax_scope all_narrow, which includes clades such as Mammalia and
Brassicales.

Genome Proteins with GO, default → narrow Core recall, default → narrow EXACT % of leaves, default → narrow
P. putida 483 → 57 0.41 → 0.04 55 → 60
Horse 29 → 12 0.69 → 0.22 13 → 29
Poplar 84 → 61 0.35 → 0.25 27 → 36

Precision rises, but coverage collapses: the narrowest clade containing the
seed usually has no GO-annotated members. In the main cohort with the own
species hidden, narrow scope gave GO to only 92 of 515 proteins (core recall
0.12), nearly all of them rat and mouse borrowing from each other. Neither
setting gets a broad scope's coverage without its errors.

Take-aways for annotating a new genome