Methods

Draft for the curators to edit. This page was drafted from the project notes; check every figure before the site is published.

Where the papers come from

Papers were found in PubMed with the broad query ("Kluyveromyces marxianus"[All Fields] OR "K. marxianus"[All Fields]), followed by screening for papers about genetic or metabolic engineering. Only the title and abstract of each paper are read; full texts are not.

How results were extracted

A large language model (Claude Sonnet) read each abstract with a fixed written prompt (version 5) and returned structured results: the gene, the type of change (for example knockout, overexpression, point mutation), the effect, the phenotype, any stated numbers, the host organism and strain, and the sentence in the abstract that supports it. Each result was then reviewed by a person before it was accepted. Results marked manual were entered by a curator.

Gene identifiers

Gene names were matched to the K. marxianus DMKU3-1042 genome (RefSeq GCF_001417885.1) and given KLMA_ locus tags. Where a name matches several copies and the paper does not say which was used, the gene is marked undetermined paralog. S. cerevisiae orthologs were assigned by reciprocal BLAST against SGD. Names that are enzyme or protein labels rather than gene symbols are kept but marked, and have no locus tag.

Strains, products and conditions

Strain names were merged only when sources (culture collections and papers) show they are the same strain. Derived strains are shown under their parent strain. Products are grouped into classes. Culture conditions are recorded only when the abstract states them.

How accurate is it?

• Of the first 465 rows the model proposed, 95.5% were accepted by the reviewer. This says little about missing results.
• In a re-run on papers whose results people had already found, the current prompt recovered about 72% of results exactly and about 83% at the level of gene and effect. This figure is optimistic because the prompt was tuned on the same papers.
• On the first set of papers held back from tuning, recall was 0.82 and precision 0.82 (0.88 and 0.98 after adjudication). On a second held-back set, which contained very few results, the extraction found none of 3 results and invented none from 12 papers with no results; this set is too small to say more.
• Gene identifier enrichment was reviewed with an accept-all default; no edits were made, so the review is weak evidence of correctness.

What is not captured

• Anything stated only in the full text, figures or supplementary material.
• Heterologous protein or enzyme expression and promoter or signal-sequence work are the weakest areas.
• A strain is recorded for only about a third of results.
• Negative results are rarely reported in abstracts.

This release

Release 1.0 (2026-10-20): 651 results from 174 papers. Each release is a fixed, checked snapshot; see Downloads for the data files and a data dictionary.