Proteins similar to mPGES-2 (PTGES2 / Q9H7Z7)

mPGES-2 · PTGES2 · Q9H7Z7 Similarity search across five complementary methods — every number traces to a committed tool output.

Proteins similar to mPGES-2 — methods and the convergent-vs-divergent conclusion

Target. Microsomal prostaglandin E synthase-2 (mPGES-2, gene PTGES2, human UniProt Q9H7Z7, 377 aa, EC 5.3.99.3). It is a GST-superfamily enzyme with an N-terminal thioredoxin/glutaredoxin fold carrying the 110-Cys-x-x-Cys-113 redox motif, and it binds a GSH–heme complex. Every number and every protein name below is drawn from this run's own full-database tool outputs; each is traceable to a row of results/master_similar_proteins.tsv (with a per-score pointer in results/master_provenance.tsv) or to the named per-method raw file. Nothing is imported from prior knowledge or the prior hermes run.

Why several complementary methods — and what each contributes

The proteins that are structurally similar to mPGES-2 and the proteins that are functionally similar to it are different sets (shown below). A single method therefore cannot answer the question, so five complementary axes were run, each recovering something the others miss.

A. Sequence homology — BLASTp, PSI-BLAST, jackhmmer

• BLASTp vs full Swiss-Prot (2026_02, 575,503 seqs) recovers only the PTGES2 orthologs: Macaca Q9N0A4 (97.3 % id), bovine Q66LN0 (88.1 %), mouse Q8BWM0 (83.3 %), zebrafish Q7ZUC7 (60.8 %, E = 1.14e-151); everything below is weak (GSTU6 E = 0.13, E. coli Grx2 P0AC59 E = 2.4). Single-pass sequence identity does not reach the wider GST superfamily — mPGES-2 is a divergent member (results/seq/blastp_swissprot.tsv; master rows 1–5, 9).
• PSI-BLAST (inclusion 0.5, 3 iterations, 38→97→305 hits) uses the position-specific profile to cross the twilight zone and newly reach GSTO1 (P78417, iter 2, E = 4.78e-4), GSTZ1/MAAI (O43708, iter 3, E = 1.43e-17), CLIC1 (iter 3), GDAP1 (Q8TB36, iter 3), CLIC6 (iter 3) and even H-PGDS (O60760, iter 3, E = 0.66) — remote homologs invisible to BLASTp (results/seq/psiblast.tsv; the psiblast_iter/psiblast_evalue columns of the master table).
• HMMER hmmscan vs Pfam-A (38.2) fixes mPGES-2's own domain architecture: GST_N PF13417 (i-E = 7.6e-12, env 104–174) + GST_C PF00043 (i-E = 0.0016, env 277–369) + Glutaredoxin PF00462 — a two-domain GST fold spanning Pfam clans CL0172 + CL0497 (results/seq/hmmscan_pfam.tsv).

A′. Domain / family-membership enumeration — InterPro + reverse-Pfam + Oakley benchmark

Running the domain call in reverse — "which human proteins carry mPGES-2's GST-fold Pfam families?" — enumerates 40 human GST-fold members and recovers 33/39 of the literature-curated Oakley 2023 "hidden GST" list (84.6 % recall, 100 % precision). The 6 misses (CLIC4, CLIC6, MTX1, MTX2, MTX3, GSTCD) carry GST-clan sibling Pfam families the query lacks and are explicitly delegated to the structural search (results/domain/membership_summary.json, oakley_benchmark_compare.md; the pfam_membership column of the master table).

B. Structural similarity — Foldseek + US-align/TM-align

• Foldseek vs full PDB (324,204 structures, TM-align mode, -s 9.5 -e 10) returns the holo target 2PBJ as the #1 hit (aligned-TM 0.996, self-control), then 296 glutathione-transferase structures, H-PGDS (3VI5 = human hematopoietic PGD synthase), glutaredoxins (human Grx2 2FLS, E. coli Grx2 4KX4), thioredoxin (6GN9) and CLIC1/CLIC4 (results/struct/README.md, foldseek_pdb_afmodel.tsv).
• Foldseek vs AFDB-SwissProt (590,183 models) returns 562 hits, of which 318 (57 %) are absent from the entire sequence-search hit set (BLASTp + PSI-BLAST union = 320 accessions) — the GST-fold neighbourhood is a structural signal, invisible to primary sequence (results/struct/afdb_vs_sequence_comparison.tsv).
• US-align/TM-align confirms every headline hit pairwise (26 structures) under a fixed classifier — same-fold = TM ≥ 0.5 by BOTH normalizations; domain-level-partial = ≥ 0.5 by the smaller-chain or the mPGES-2 N-domain (88–179) norm only; not-similar = < 0.5 by all. Because mPGES-2 is a two-domain, 377-aa protein with a ~46-aa insertion, a single-unit GST/glutaredoxin (~150–250 aa) matches ~one structural unit: the whole-mPGES-2 norm sits below 0.5 while the partner and N-domain norms sit well above it. The closest fold, H-PGDS, scores 0.431 / 0.751 (N-dom 0.669), RMSD 3.37 Å — domain-level-partial. All human GSTs (GSTO1 0.440/0.656, GSTO2 0.445/0.660, GSTP1 0.427/0.713, GSTA1 0.424/0.670, GSTM2 0.431/0.700, GSTZ1 0.423/0.685, GSTK1 N-dom 0.695), the CLICs (CLIC1 0.408/0.598, CLIC4 0.405/0.569, CLIC6 N-dom 0.732), GDAP1 (0.479/0.500), the three metaxins (MTX1 0.508/0.422, MTX2 0.491/0.666, MTX3 0.491/0.578), FAXC (0.512/0.476), and the glutaredoxin/thioredoxin partners (E. coli Grx2 0.445/0.714, human GLRX2 N-dom 0.661, TXN N-dom 0.562) classify domain-level-partial (results/struct/tmalign_summary.tsv + tmalign_extra_summary.tsv; master rows 6–26). Dali was best-effort supplementary; US-align/TM-align is the required oracle and is what the classification rests on.

C. Functional similarity — EC / GO / cofactor + PGES family

• RCSB EC-5.3.99.3 enumeration: the exact reaction "PGH2 = PGE2" is carried by 27 PDB polymer entities that fall into 3 unrelated folds — GST (2 entities: 1Z9H + 2PBJ, mPGES-2), MAPEG (18 entities, mPGES-1) and Hsp90-p23 (7 entities, cPGES). One enzyme activity, three independent structural solutions (results/func/ec_family_breakdown.tsv).
• Dual-GO intersection (GO:0043295 glutathione-binding) AND (GO:0020037 heme-binding) returns 7 hits, of which the 5 reviewed entries are all PTGES2 orthologs — the GSH+heme signature is PTGES2-unique (results/func/go_heme_gsh_dual.json).
• Cofactor site (2PBJ): the sole Fe axial ligand is the GSH cysteinyl-thiolate SG2 at 2.60–2.73 Å; a whole-PDB survey finds only 2PBJ co-binds GSH+heme with a thiolate–Fe bridge inside a GST fold (results/ligand/fe_coordination.tsv, bridge_survey.tsv).

The conclusion: the structural and functional neighbourhoods DIVERGE

Laying the two neighbourhoods side by side (all from the master table):

Structural neighbourhood — same fold, different function → divergent evolution. H-PGDS, the cytosolic GSTs (Omega/Pi/Alpha/Mu/Zeta/Kappa/Theta), the CLICs, GDAP1, the metaxins, FAXC, and the glutaredoxins/thioredoxins all share mPGES-2's GST/thioredoxin fold at the domain level (TM ≥ 0.5 by the partner or N-domain normalization) — yet they do different chemistry: GSH-conjugation detox (GSTs, EC 2.5.1.18), maleylacetoacetate isomerization (GSTZ1, EC 5.2.1.2), prostaglandin D synthesis (H-PGDS, EC 5.3.99.2), chloride-channel / glutaredoxin-like redox (CLICs, glutaredoxins), mitochondrial fission (GDAP1), protein import (metaxins). These proteins descend from a common ancestral GST/thioredoxin fold that diverged in function — classic divergent evolution.

Functional neighbourhood — same function, different fold → convergent evolution. The other enzymes that perform mPGES-2's reaction — mPGES-1 (PTGES, MAPEG fold) and cPGES (PTGES3, Hsp90-co-chaperone/p23 fold) — carry EC 5.3.99.3 (PGH2→PGE2) but are structurally unrelated to mPGES-2: US-align gives mPGES-1 0.184 / 0.353 / 0.377 and cPGES 0.152 / 0.283 / 0.240, < 0.5 on every normalization (master rows 55, 57; results/func/pges_family.tsv). (L-PGDS P41222, a lipocalin PGD synthase, is an additional different-fold negative control, 0.204/0.331/0.308.) Three unrelated folds independently arriving at the same PGH2-isomerase activity is convergent evolution — the same conclusion the RCSB "one EC → three folds" enumeration reaches.

Therefore the two sets are essentially disjoint. The proteins most similar to mPGES-2 in structure (GST-fold relatives) share almost none of its function; the proteins most similar to it in function (mPGES-1, cPGES) share none of its fold. The classifier is not merely permissive — it places the same-fold-different-function relatives at domain-level-partial and the same-function-different- fold analogs at not-similar (< 0.5 all norms). This divergence is exactly why several methods are required: sequence + structure recover the divergent GST-fold neighbourhood, while the functional axis (EC/GO/PGES-family) is the only way to surface mPGES-1 and cPGES, which every structural and sequence method correctly rejects.

Headline answer. The structurally similar proteins — the user's primary ask — are the GST-superfamily / thioredoxin-fold relatives foregrounded in the master table: closest is H-PGDS, then the human GSTs (Omega highest, then Pi/Alpha/Mu/Zeta/Kappa/Theta), the CLICs, GDAP1, the metaxins, FAXC, and the glutaredoxins/thioredoxin (N-domain match). mPGES-1 and cPGES are functional look-alikes only, on completely different folds.

Sources: results/master_similar_proteins.tsv (+ .md), results/master_provenance.tsv, and the per-method raw outputs under results/{seq,domain,struct,func,ligand}/. Methods, versions, DBs and exact commands: RUNBOOK.md and results/ENVIRONMENT.md. Reproducibility: results/REPRODUCIBILITY.md.

Structural superposition — mPGES-2 vs its closest structural hit (H-PGDS)

Master table — structural / domain neighbourhood (headline answer)

The structurally similar proteins, foregrounded first. Full machine-readable source: results/master_similar_proteins.tsv; per-score provenance: results/master_provenance.tsv. This table is generated at build time directly from that TSV, so it cannot diverge from the committed data.

Master table — functional analogs (same reaction, different fold — the divergence)

# sequence homology vs full Swiss-Prot
blastp -query data/Q9H7Z7.fasta -db swissprot -evalue 10 -max_target_seqs 5000 \
       -outfmt '6 saccver pident qcovs bitscore evalue'
psiblast -query data/Q9H7Z7.fasta -db swissprot -num_iterations 3 -inclusion_ethresh 0.5
hmmscan --domtblout hmmscan_pfam.domtbl Pfam-A.hmm data/Q9H7Z7.fasta

# structural search (TM-align mode, pinned sensitivity/e-value)
foldseek easy-search Q.pdb pdb aln.m8 tmp --alignment-type 1 -s 9.5 -e 10 \
    --format-output query,target,qtmscore,ttmscore,alntmscore,lddt,evalue,prob,alnlen,qlen,tlen,rmsd
foldseek easy-search AF-Q9H7Z7.pdb afdb_swissprot aln.m8 tmp --alignment-type 1 -s 9.5 -e 10 ...

# same-fold classifier oracle (whole-chain, both normalizations)
USalign data/AF-Q9H7Z7-F1-model_v6.pdb .pdb            # + vs N-domain 88-179

# regenerate this page's figure from committed structures
python3 pipeline/45_figure_superpose.py

# deterministic reproducibility spot-check (US-align pair + BLASTp)
bash pipeline/42_repro_check.sh