Contributions
Annotation-Gap Discovery via Phenotype-Fitness-Pangenome-Gapfilling Integration combined metabolic-model gapfilling, Fitness Browser phenotypes, pangenome annotations, GapMind pathway evidence and BLAST homology to assign candidate genes to gapfilled reactions across 14 organisms and 18 carbon sources 1. The project built draft models from ModelSEED/RAST annotations and COBRApy and evaluated 574 organism-carbon source combinations with flux-balance analysis (FBA), a constraint-based method for predicting metabolic flux and growth 1. The report gives baseline accuracy as 42.5%, recall as 86.5% (244 of 282) and precision as 42.5% (244 of 574), with 330 false positives; because accuracy and precision share the same value, the report’s figures do not independently verify these as two metrics 1. The project ran conditional gapfilling on 38 false-negative cases and reported 219 added reactions, averaging 5.8 per case, but its listed categories (201 enzymatic, 14 transport, 12 exchange) do not sum to 219, and the report leaves this unresolved 1. The report notes that gapfilling is non-unique, since multiple valid solutions may exist for each false-negative case, and that the default ModelSEED gapfilling it used minimizes the number of added reactions without guaranteeing biological optimality 1.
The project resolved 96 of 201 gapfilled enzymatic reaction-organism pairs (47.8%), exceeding its pre-specified 30% H1 threshold 1. These comprised 44 high-confidence (21.9%), 19 medium-confidence (9.5%) and 33 low-confidence (16.4%) pairs, and 105 pairs (52.2%) remained unresolved 1. In leave-one-out cross-validation, BLAST alone resolved 70 pairs (34.8%), NB03 matching of EC (Enzyme Commission enzyme-classification) numbers alone resolved 51 (25.4%) and NB04 Bakta annotations alone resolved 22 (10.9%) 1. No single stream exceeded 35%, and the full pipeline added 13 percentage points over BLAST alone 1. Of the 50 reactions lacking an EC number (“dark reactions”), the project resolved only 8 (16%), compared with 88 of 151 (58.3%) reactions with known EC numbers 1. The project resolved rxn02185 (EC 2.2.1.6) and rxn03436 (EC 1.1.1.86), which are sequential branched-chain amino acid biosynthesis steps, with high confidence in 9 of 14 organisms each; the report notes this shows computational consistency, not experimental validation 1. The project identified 154 DIAMOND hits using 328 Swiss-Prot exemplar sequences covering 75/84 unique ECs, although the report describes the search direction inconsistently and does not reconcile the two descriptions 1.
Resolution rates in the project ranged from 20.0% in Bacteroides thetaiotaomicron (3 of 15) to 71.4% in Klebsiella michiganensis (5 of 7) 1. The report links this variation to reference annotation quality, Fitness Browser coverage and phylogenetic distance, but these are interpretations rather than controlled tests 1. Because 12 of 14 organisms were Proteobacteria, the lower effectiveness for divergent clades rests on a single Bacteroidetes organism and remains a hypothesis 1. The project found that GapMind flagged incomplete pathways for many of the same carbon sources that required gapfilling across 104 pairings, but it did not quantify the overlap 1. Exact concordance was limited because the KBase Data Lakehouse data give step counts rather than step identities 1. The project inserted 23 gene-protein-reaction (GPR) rules, formal links between genes and reactions, into SBML (Systems Biology Markup Language) models for knockout validation and found zero wildtype growth on minimal carbon-source media 1. This made the validation inconclusive, because the gapfilled reactions under test were themselves required for growth 1.