Focus type: function_assignment · Term: GO:0061710 (EC 2.7.7.87) · Evidence: IDA · Original reference: PMID:37963869(https://pubmed.ncbi.nlm.nih.gov/37963869/)
The seed hypothesis is OVER-ANNOTATED. YciO can catalyze L-threonylcarbamoyladenylate (TC-AMP) formation in a test tube — and that in-vitro observation is what grounds the GO:0061710 IDA imported from the DeepECtransformer machine-learning paper (PMID:37963869(https://pubmed.ncbi.nlm.nih.gov/37963869/)). So the molecular capability is not fabricated. But a convergent weight of independent evidence shows that this activity is residual, promiscuous paralog activity rather than YciO's physiological function, and that curating GO:0061710 as YciO's core function would materially mislead downstream users.
Four independent lines of evidence support this judgment. (1) Kinetics: a dedicated expert re-evaluation of precisely this ML assignment (PMID:40703034(https://pubmed.ncbi.nlm.nih.gov/40703034/), de Crécy-Lagard, Swairjo et al. 2025) measured YciO at ~0.14 nM/min versus the genuine enzyme TsaC/YrdC at ~2.8 µM/min — roughly 20,000-fold (4.3 orders of magnitude) weaker under comparable conditions. (2) Genetics: tsaC/yrdC is essential in E. coli (PMID:19287007(https://pubmed.ncbi.nlm.nih.gov/19287007/)) even though yciO is present in the same genome, and YciO does not complement TsaC loss; YciO is itself non-essential and has no reported t6A/tRNA phenotype. (3) Sequence/family: the two proteins share only the ancestral YrdC/Sua5 fold (Pfam PF01300, ~35–40% identity by computed global alignment) but fall into different InterPro subfamilies (YciO: IPR052532; TsaC subfamily: IPR023535) and sit in distant genome loci/operons (yciO = b1267; tsaC/yrdC = b3282). (4) Functional network: an independent STRING-db v12 check cleanly separates YciO from the t6A module.
The best mechanistic explanation is that the YrdC/Sua5 fold is intrinsically capable of tRNA-independent TC-AMP synthesis (PMID:37427786(https://pubmed.ncbi.nlm.nih.gov/37427786/)), so a diverged Sua5-fold paralog like YciO retains a trace of that ancestral chemistry without being under selection for it. The recommended curation action is therefore not to deny the in-vitro activity but to prevent GO:0061710 from being treated as YciO's core function, to block propagation of the downstream t6A biosynthesis (BP) term, and to record YciO's physiological function as currently unknown.
Verdict: Over-annotated — partially supported in vitro, refuted as a core/physiological function.
The IDA is literally defensible (a direct assay detected the activity), so the annotation is not a hallucination. The problem is scope: retaining GO:0061710 as YciO's representative function — and the UniProt name "Threonylcarbamoyl-AMP synthase YciO" — implies a t6A biosynthetic role that YciO does not perform in vivo. The single primary source supporting the seed hypothesis is an ML-prediction validation study whose conclusions were explicitly challenged by domain experts who re-examined this exact assignment and found that current ML EC-prediction methods "make basic logic errors … that human annotators avoid." Weighed against trace kinetics, genetic non-redundancy, subfamily divergence, and network separation, the hypothesis fails at every layer that matters for a core-function annotation.
Most important caveat: because the activity is real and the IDA carries a PMID and an "experimental" tag, it looks stronger than it is and is at risk of cascade propagation to paralogs/orthologs. Curators should flag it rather than silently inherit it.
The seed IDA traces to PMID:37963869(https://pubmed.ncbi.nlm.nih.gov/37963869/) (Functional annotation of enzyme-encoding genes using deep learning with transformer layers, "DeepECtransformer"). In that work a transformer model predicted EC 2.7.7.87 (TC-AMP synthase) for YciO, and the authors report: "We experimentally validated the enzymatic activities predicted for three proteins (YgfF, YciO, and YjdM)." This single in-vitro validation of an ML call is the entire experimental basis for the annotation. Notably, the assay readout partly rests on ATP→AMP hydrolysis, which is weaker evidence than direct identification of the TC-AMP product.
A dedicated expert re-evaluation of exactly these ML assignments was subsequently published: PMID:40703034(https://pubmed.ncbi.nlm.nih.gov/40703034/) (Limitations of current machine learning models in predicting enzymatic functions for uncharacterized proteins; de Crécy-Lagard, Swairjo and colleagues — the group that originally discovered the YrdC/Sua5→t6A pathway). Its central conclusion is blunt: "We found that current ML methods not only mostly fail to make novel predictions but also make basic logic errors in their predictions that human annotators avoid by leveraging the available knowledge base." This paper specifically re-examines the YciO case and quantifies the kinetic deficit: YciO makes TC-AMP at roughly 0.14 nM/min versus TsaC/YrdC at roughly 2.8 µM/min under comparable enzyme/assay conditions — a ~20,000-fold (≈4.3 orders of magnitude) rate gap. An activity this far below the true physiological enzyme is the signature of residual, non-selected paralog activity, not a biological function.
PMID:19287007(https://pubmed.ncbi.nlm.nih.gov/19287007/) (The universal YrdC/Sua5 family is required for the formation of threonylcarbamoyladenosine in tRNA) states plainly that the authors "prove that yrdC is essential in E. coli." The logical force for curation is decisive: E. coli K-12 carries both yrdC/tsaC and yciO in the same genome, yet deletion of tsaC is lethal. If YciO produced physiologically meaningful quantities of TC-AMP, it would at least partially rescue tsaC loss — it does not, even when overexpressed (as noted in the re-evaluation). YciO is itself non-essential, has no reported t6A or tRNA-modification phenotype, and therefore is not the cell's TC-AMP source. This converts the seed hypothesis from "YciO is the TC-AMP synthase" into, at most, "YciO retains a vestigial ability to make TC-AMP that the cell does not use."
Both YciO and TsaC belong to the universal YrdC/Sua5 superfamily (Pfam PF01300), which is why YciO retains the ancestral fold capable of binding ATP, threonine, and bicarbonate and performing the chemistry weakly. However, InterPro resolves them into different subfamilies: YciO in IPR052532, versus the functionally validated TsaC enzymes in IPR023535. A global Needleman–Wunsch alignment computed during this investigation returned only ~35–40% identity between YciO and TsaC — consistent with an ancient duplication followed by functional divergence, not with recent orthology. The shared fold explains the in-vitro activity; the subfamily split and modest identity explain why that activity is not under selection in YciO. The fold-level promiscuity is corroborated by PMID:37427786(https://pubmed.ncbi.nlm.nih.gov/37427786/), which shows a Pandoravirus Sua5 domain makes TC-AMP tRNA-independently — i.e., the Sua5 fold is intrinsically capable of the reaction.
Using STRING-db v12 for E. coli K-12 (taxon 511145), the interaction neighborhoods of the two proteins are strikingly different:
Genomically, the two genes sit far apart in different operons (yciO = b1267; tsaC/yrdC = b3282). The network thus operationalizes the exact "available knowledge base" that PMID:40703034(https://pubmed.ncbi.nlm.nih.gov/40703034/) argues human annotators use and ML models ignore — and it independently corroborates that YciO is not part of the t6A module.
This is a textbook paralog over-annotation driven by residual in-vitro activity plus ML homology/frequency bias. The YrdC/Sua5 fold is ancient and catalytically promiscuous at the level of weak chemistry. A gene duplication in the bacterial lineage produced two E. coli descendants that have since diverged in subfamily, genome context, partners, and selection pressure:
Ancestral YrdC/Sua5 fold (Pfam PF01300)
(binds ATP + L-Thr + HCO3-, can make TC-AMP)
│
┌────────────────┴─────────────────┐
│ │
TsaC / YrdC (b3282) YciO (b1267)
InterPro IPR023535 InterPro IPR052532
ESSENTIAL NON-ESSENTIAL
~2.8 µM/min TC-AMP ~0.14 nM/min TC-AMP (~20,000x weaker)
Partners: TsaB/TsaD/TsaE (0.96+) Partners: YciV/YciQ/RluB (unrelated)
Supplies t6A pathway in vivo No t6A phenotype; function UNKNOWN
│
TC-AMP ──► TsaB/TsaD/TsaE ──► t6A37 in tRNA (accurate ANN-codon decoding)
The pathway on the left is complete, essential, and network-coherent. YciO on the right has retained the ancestral fold — hence a detectable trace of the ancestral chemistry in a sensitive assay — but has diverged in every dimension that signals a dedicated function. The cell cannot be using YciO to make TC-AMP, because if it were, tsaC could not be essential.
For GO curation this maps cleanly onto the layers the ontology is meant to distinguish:
| Layer | TsaC/YrdC | YciO |
|---|---|---|
| MF capability (in vitro) | TC-AMP synthase (GO:0061710) — strong | TC-AMP synthase (GO:0061710) — trace only |
| MF as core function | Yes | No |
| BP (t6A biosynthesis, GO:0002949) | Yes | No — no phenotype, not in module |
| Essentiality | Essential | Non-essential |
| Physiological function | Known (t6A) | Unknown |
The seed hypothesis is true only at the "MF capability in vitro" layer and false at every layer that matters for a core-function annotation.
| Citation | Evidence type | Direction | Claim tested | Key finding | Context | Confidence / limits |
|---|---|---|---|---|---|---|
| PMID:37963869(https://pubmed.ncbi.nlm.nih.gov/37963869/) | Computational + in-vitro assay | Qualifies (source of IDA) | YciO has EC 2.7.7.87 | DeepECtransformer predicted EC 2.7.7.87; in-vitro assay detected activity (incl. ATP→AMP hydrolysis) | E. coli K-12, purified protein | Low–moderate; ML-driven; weak activity; physiology not addressed |
| PMID:40703034(https://pubmed.ncbi.nlm.nih.gov/40703034/) | Expert re-evaluation + comparative biochemistry | Refutes (as core function) | Is the activity YciO's true function? | YciO ~0.14 nM/min vs TsaC ~2.8 µM/min (~20,000× weaker); no TsaC complementation; distinct genome context; ML "logic errors" | E. coli K-12 | High; authors discovered the t6A/TsaC pathway and re-examined this exact call |
| PMID:19287007(https://pubmed.ncbi.nlm.nih.gov/19287007/) | Genetic (null allele) + biochemistry | Refutes (as core function) | Is YciO the physiological TC-AMP synthase? | yrdC/tsaC essential in E. coli despite yciO presence → no complementation | E. coli K-12 | High; essentiality is a strong in-vivo readout |
| UniProt P0AFR4 / InterPro / Pfam + global alignment (this analysis) | Structural / evolutionary | Qualifies | Are YciO and TsaC the same enzyme? | Shared PF01300 fold; different subfamilies (IPR052532 vs IPR023535); ~35–40% identity (computed NW) | Sequence/domain databases | Moderate–high; shared fold, divergent subfamily explains both activity and split |
| PMID:37427786(https://pubmed.ncbi.nlm.nih.gov/37427786/) | Structural/biochemical (homolog) | Qualifies | Can a Sua5 fold make TC-AMP tRNA-independently? | Pandoravirus TsaN Sua5 domain makes TC-AMP tRNA-independently → fold intrinsically capable | Pandoravirus, in vitro, cryo-EM | Moderate; explains fold-level promiscuity behind weak YciO activity |
| STRING-db v12 (this analysis) | Computational (functional-association network) | Refutes (as core function) | Is YciO in the t6A module? | TsaC clusters with TsaB/TsaD/TsaE (0.96+); YciO does NOT (no TsaD/TsaE; TsaB 0.566, TsaC 0.467); YciO top partners YciV/YciQ/RluB | E. coli K-12 (taxon 511145) | Moderate; association ≠ direct interaction, but pattern is clear |
| PMID:21775474(https://pubmed.ncbi.nlm.nih.gov/21775474/) | Biochemical/structural | Supports (context) | What is the true t6A transfer machinery? | YrdC behaves as a subunit/component of the t6A transferase system; binds unmodified ASL-Lys | E. coli | Orientation; establishes TsaC/YrdC as the physiological actor |
The seed hypothesis is over-annotated. YciO's TC-AMP synthase activity (GO:0061710) is a genuine but vestigial, in-vitro-only paralog activity — ~20,000-fold weaker than the true enzyme TsaC/YrdC, unable to complement essential tsaC, phenotypically silent for t6A, residing in a distinct InterPro subfamily (IPR052532) and a separate genomic/functional-network context from the t6A module (TsaB/TsaD/TsaE). It should not be curated as YciO's core function, the t6A biosynthesis BP term should not be assigned, and YciO's physiological role should be recorded as unknown.
Computational note: UniProt records and sequences were retrieved programmatically; the YciO–TsaC identity (~35–40%) was computed by Needleman–Wunsch alignment, and the ~20,000×/4.3-orders activity gap was computed locally from rates reported in PMID:40703034(https://pubmed.ncbi.nlm.nih.gov/40703034/). STRING-db v12 partner scores were retrieved for E. coli K-12 (taxon 511145). Full-text PDFs could not be parsed programmatically in this environment; this is stated plainly rather than inferred.