Accurate de novo identification of biosynthetic gene clusters with GECCO

Key results

1. ShownTrained on MIBiG v1.3 and tested on BGCs added only in MIBiG v2.0 (376-genome test set), GECCO outperformed original and retrained DeepBGC and ClusterFinder in domain-level receiver operating characteristic (ROC) and precision-recall (PR) analysis.

   Area under the ROC curve (AUROC) higher than all DeepBGC/ClusterFinder variants on the 376- and six-genome sets, tied with original DeepBGC on the nine-genome set (0.94 each); area under PR curve (AUPR) higher on all three sets. Versus retrained DeepBGC in 10-fold cross-validation (CV): GECCO AUROC 0.97, AUPR 0.73. AUROC/AUPR values for the 376- and six-genome sets are not given in the text. · GECCO provides superior precision and speed…; Validation of CRF performance on external test…; Figure legends

2. ShownAt the level of whole predicted clusters (segment overlap), GECCO's AUPR was substantially higher than every DeepBGC/ClusterFinder implementation. The authors attribute its accuracy mainly to more precise boundary placement, which also greatly reduces fragmented predictions.

   PR curves from segment overlap between predicted and known BGCs (minimum overlap 25, 50 and 75%) gave GECCO a substantially higher AUPR than all DeepBGC/ClusterFinder implementations; a histogram of predicted BGC lengths on the 376-genome set is used to show less fragmentation. No segment-level AUPR values are stated in the text. · GECCO provides superior precision and speed…; Validation of CRF performance on external test…; Figure legends

3. ShownWhen an entire biosynthetic class is withheld from training, GECCO still largely recovers it. Recovery is weaker for RiPP (ribosomally synthesized and post-translationally modified peptide) and saccharide clusters. A separate random forest assigns predicted clusters to MIBiG classes.

   Leave-one-type-out (LOTO) CV, withholding each MIBiG class from training: AUROC > 0.98 for four of six classes, 0.91 for RiPPs and 0.88 for Saccharides. The random forest class predictor reached AUROC > 0.90 for every class in five-fold CV. · GECCO provides superior precision and speed…; CRF training and cross-validation; Prediction of biosynthetic class

4. ShownOn 12,221 proGenomes2 representative genomes, GECCO's BGC count fell between the two other tools: close to double antiSMASH's total, yet well below DeepBGC's. On a single CPU, GECCO ran over three times faster than both.

   BGC counts: 115,131 GECCO, 59,041 antiSMASH (default parameters), 470,137 DeepBGC; most GECCO or DeepBGC calls were not detected by antiSMASH. Single-CPU runtime was compared on the three test sets (395 contigs) and on proGenomes2 (627,182 contigs from 12,221 genomes); no absolute times are given in the text. · GECCO provides superior precision and speed…; GECCO’s CRF-based approach provides insight into…; BGC identification in prokaryotic genomes

5. SuggestedInspecting the CRF shows that it leans on domains linked to secondary metabolism. Many high-weight domains, however, lie outside the BiG-SLiCE core biosynthetic set, including three domains of unknown function (PF19155, PF11379, PF17537) that the authors propose for functional study.

   937 of 2,027 BiG-SLiCE core domains (46.2%) are in GECCO's 5,255-domain model; core domains carry higher weights than non-core TIGRFAM domains (Mann-Whitney raw P = 3.12e-07) but not non-core Pfam ones (P = 0.10). Gene Ontology enrichment: 33 biological processes and 22 molecular functions (topGO P < 0.05); 3 Pfam clans (false discovery rate P < 0.10). · GECCO’s CRF-based approach provides insight into…; Comparison of GECCO and BiG-SLiCE domain sets; GO term enrichment; Pfam clan enrichment