Validating E. coli ML predictions

Seven gene reviews against an expert audit of DeepECTransformer

AI Gene Review · projects/VALIDATING_ECOLI_PREDICTIONS · 2026

Bottom line

  • An expert audit (de Crécy-Lagard et al. 2025, PMID:40703034) found only 3 of 453 DeepECTransformer predictions for E. coli unknowns were correct and novel.
  • We reviewed 7 genes spanning five error categories; all 7 reviews are complete and our verdicts match the paper's.
  • The same errors sit in GOA: yciO's TsaC activity rows and a fepE tyrosine-kinase IBA row were marked REMOVE.

Why these seven

  • The paper sorted every prediction into a taxonomy: COR, CNN, LSP, UNC, PLI, NPI, REP.
  • We sampled across it to ask two questions:
    1. Does an agentic gene review reach the same verdict as the experts?
    2. Do the model's logic errors also appear in existing GO annotations?
  • Each gene got a full *-ai-review.yaml plus a structured *-det-predictions-review.yaml.

Prediction vs reality

A paralog error, in GOA too

yciO review, shown: the IEA GO:0061710 threonylcarbamoyladenylate synthase row (GO_REF:0000003, propagated from the EC number) is marked REMOVE. Not in frame: the IDA row for the same term from the DeepECTF validation paper (PMID:37963869) is also REMOVE.

Review actions across the seven genes

Lessons

  • Paralogs (yciO, yegV): a weak in vitro activity from a shared fold is not the biological function.
  • Pathway context (yjhQ, yrhB): a predicted enzyme is wrong if the host lacks the pathway or already has the enzyme (QueD).
  • In vitro vs in vivo (yjdM): phosphonoacetate hydrolase rows marked over-annotated.
  • Frequency bias (fepE): "histidine kinase" for a Wzz O-antigen chain-length regulator.

Status

  • ✅ 7/7 gene reviews and prediction reviews complete (March 2026).
  • The DeepECTF prediction table is browsable with the BioReason comparison material (BIOREASON_COMPARISON/deepectf-eval.html).

Read more: projects/VALIDATING_ECOLI_PREDICTIONS.md · genes/ECOLI/<gene>/