MitoTOL: Mitochondrial Proteomes Across the Tree of Life

Bottom line: Chen et al. (Cell 2026, PMID:42822426)
take eight experimentally defined mitoproteomes (human, yeast, Arabidopsis, Acanthamoeba,
Giardia, Trypanosoma, Leishmania, Babesia) and place them on a phylogeny of 673
eukaryotes and 34 prokaryotic outgroups. From this they reconstruct a mitochondrion in the last
eukaryotic common ancestor (LECA) of about 1,438 proteins, with 70% of human mitochondrial
proteins tracing back to it, and nominate 33 mitochondrial protein families that are shared by
protozoan pathogens and absent from humans. For this repository the paper is a comparative and
computational resource. Its predicted localizations are hypotheses, not annotation evidence.
What it is good for is checking where an IBA "mitochondrion" (or a mitochondrial pathway term)
can and cannot be inherited: which mitochondrial systems are ancestral, which were gained in
animals, and which were lost in human or yeast. The experimental proteomes behind it are in five
companion MitoTOL papers, which are where any HDA-grade evidence for protist mitochondria lives.
Eight genes chosen from the paper have now been reviewed: human FH, CRLS1 and CLPP
(the human halves of three of the paper's non-homologous enzyme pairs), yeast NDI1 (NDH2), and
the four yeast ERMES subunits. All eight passed the lineage-loss and same-activity checks below.
The paper is cited in each review as evolutionary context only.

The paper

Chen MZ, Stefely JA, Bardon EG, MitoCarta Tree of Life Consortium, Calvo SE, Mootha VK.
Comparative analysis of mitochondrial proteomes across the tree of life. Cell 189(20):6285–6306
(2026). doi:10.1016/j.cell.2026.08.029. Open access
(CC BY).

Cached as publications/PMID_42822426.md. PubMed only had the abstract, so the full text
(summary, main text, limitations and methods) was added from the publisher PDF, tagged
full_text_provider: user_upload. supporting_text quotes from it can therefore be
machine-verified.

Code is at michaelzhuchen/mito-evolution and
data at Zenodo 17823713. The data archive is ~3.8 GB
(plus tens of GB of trees), so it is not mirrored here. Mitoproteome inventories are served at
mitocarta.org.

Companion MitoTOL papers (the experimental proteomes)

The human and yeast inventories are MitoCarta3.0 (PMID:33174596)
and the yeast high-confidence proteome of Morgenstern et al. These companion papers are not cached
here yet. They, not Chen et al., would back a mitochondrion annotation for a protist protein.

What the paper did

Curation implications

1. Predictions are not annotation evidence

DeepLoc2.0-mito predictions, and ancestral localizations inferred on reconciled trees, are
computational. As with MitoMatch:

• Never write a DeepLoc2.0-mito call or an ancestral-reconstruction inference into
  existing_annotations. If worth recording, it goes in GENE-predictions-review.yaml (source:
  DeepLoc2.0-mito) or a bioinformatics RESULTS.md.
• The training set is deliberately restricted to experimental proteomics "to avoid circularity",
  and the authors filter predictions to OGs predicted mitochondrial in at least 5 of 203 species
  because rarer calls were enriched in false positives. Any single-protein call should be read
  with that threshold in mind.
• The experimental proteomes in the companion papers are a different matter. A mitochondrial
  protein from those inventories is HDA-grade evidence for the protist protein itself. It is
  not evidence for its human or yeast homolog.

2. Where IBA "mitochondrion" can be inherited, and where it cannot

An IBA asserts that a PAINT curator placed the function at an ancestral node (see the IBA section
of CLAUDE.md). This paper gives an independent, proteome-scale estimate of which mitochondrial
systems are ancestral, which makes it a useful check on node placement.

• Deep placements are expected for core systems. The human pathways with the highest LECA
  share are translation factors, ISC, TCA cycle, solute transport, protein import and sorting,
  OXPHOS and the mitoribosome. An IBA mitochondrion or pathway term from a LECA-level node for
  these is consistent with the paper.
• Animal-gained systems should not carry pan-eukaryotic mitochondrial nodes. Apoptosis,
  fission, mitophagy, immune response and organelle contact sites have only a small LECA share,
  and 192 human mitochondrial OGs are Metazoa-specific. An IBA that puts a mitochondrial term for
  one of these on a node above Metazoa (or Opisthokonta) is worth checking.
• Known lineage losses (Figure 5B) give testable predictions for propagation errors:

Checked against the cached GOA files in this repo (2026-10-03): no human gene carries
GO:0032865 ERMES complex; GO:1990246 uniporter complex appears only on human MICU1; GO:0045271
respiratory chain complex I and GO:0003954 NADH dehydrogenase activity appear only on human,
mouse, bacterial and plant (chloroplast ndh) genes, never on a yeast gene. So none of the
predicted propagation errors is present among the genes reviewed so far. This should be
re-checked as yeast coverage grows.

3. Same activity, unrelated families

LECA had several pairs of non-homologous enzymes with the same activity, which were later lost in
complementary patterns:

Two consequences for curation:

• An MF term such as fumarate hydratase activity is correct for both members of a pair, but any
  family-based transfer (InterPro2GO, PANTHER) is only valid within one family. A protist protein
  annotated "fumarate hydratase" should be checked for which class it is before its other
  properties (cofactor, O₂ sensitivity) are assumed.
• "Absent in humans" in this paper means the family is absent, not the activity. The
  Discussion lists "CCA-adding enzyme" among LECA mitochondrial proteins absent from humans, but
  human TRNT1 provides CCA addition in mitochondria. This is presumably a PhROG-level distinction
  (a separate lineage of the enzyme), not a claim that human mitochondria lack the activity. A
  review should not cite this paper for the latter.

4. "Lineage-specific" and "unknown" need power estimates

Most single-clade mitochondrial OGs (56%) were not powered to detect outgroup homologs, so they may
be remote homologs rather than innovations. When a review says a mitochondrial protein is
"lineage-specific" or "has no homologs outside X", it should say whether detection power was
assessed. The 384 powered, Pfam-less OGs (Table S2) are the paper's best list of real unknown
mitochondrial families.

5. GO annotations were an input to the paper

To time cellular components (Figure 7D), the authors assigned GO cellular component terms to
ancestral PhROGs from "the literature/curation-based GO annotations of its extant descendants",
weighted per species. The mito-late conclusion (mitochondria acquired after the nucleus,
proteasome, cytoskeleton, Golgi and ER, at a similar time to peroxisomes) therefore rests partly on
GO CC curation quality. Over-annotated CC terms in well-studied species feed into this kind of
analysis directly.

Genes in this repository touched by the paper

Already reviewed here and discussed in the paper or its LECA model (Figure 6C):

Reviewed from this paper

Most removals are bare protein binding rows from high-throughput screens. The substantive calls:

• FH. Two NEW terms from PMID:26237645: positive regulation of NHEJ (GO:2001034) and site of
  double-strand break (GO:0035861). Nuclear FH binds H2A.Z at breaks and its local fumarate inhibits
  KDM2B. Urea cycle kept as non-core: cytosolic FH clears fumarate released by argininosuccinate
  lyase but does no step of the cycle.
• CRLS1. Two Reactome TAS rows removed (GO:0003841, GO:0047144). They map an unreplicated
  lysophosphatidylglycerol acyltransferase claim (Nie 2010, PMID:20025994) onto lysophosphatidic
  acid acyltransferase terms. PG acyl-chain remodeling, which rests on the same study, is undecided.
• CLPP. Endopeptidase Clp complex IDA/IPI made specific to the mitochondrial complex
  (GO:0009841); the IBA row keeps the family-level term because that node spans bacteria and
  plastids. No mitochondrial UPR term: mouse data show mammalian CLPP is dispensable for it.
• NDI1. Cytosol RCA (from a YeastPathways GO-CAM conversion) removed, since NDI1 faces the
  matrix and the external NDE1/NDE2 handle cytosolic NADH. Positive regulation of apoptosis
  marked over-annotated as an overexpression/ROS effect.
• ERMES. Tether activity (GO:0140474), ERMES complex and intermembrane lipid transfer accepted
  for all four. Inheritance, morphology, mtDNA and peroxisome phenotypes kept as non-core, because
  artificial tethers and VPS13 bypass alleles rescue them. Lipid transfer is core only for Mmm1 and
  Mdm12. Both MMM1 mitochondrial outer membrane rows (IDA PMID:11266455 and HDA PMID:16407407) are removed: Kornmann et al.
  2009 (PMID:19556461, full text cached) show Mmm1 is an ER protein "misannotated as a mitochondrial
  protein".

Not yet reviewed here, and prominent in the paper: human MCU, SMDT1 (EMRE), RHOT1/2
(MIRO), SFXN1–5, HCCS, SQOR, CIAO3 (the Nar1/NARF paralog of HydA), MPC1/2;
yeast NDE1/NDE2 (external NDH2) and GEM1 (MIRO).

Possible next steps

1. Cache the five companion papers and check whether GO consortium groups (TAIR, TriTrypDB,
   PlasmoDB-style resources) are ingesting them as HDA evidence.
2. Review the remaining genes listed above, starting with MCU/SMDT1 (the uniporter lost in
   yeast) and yeast GEM1, which links ERMES to the MIRO family the paper calls ancestral.
3. If Table S7 (per-OG gains and losses) can be obtained without the full 3.8 GB archive, join it
   to genes/human/* to list reviewed genes whose IBA mitochondrion node conflicts with the
   paper's inferred origin. Without that table, nothing in this page should be used to change an
   IBA action.