Labs RSS GitHub โ€ข GPL-3.0
Tier:
Domain:
Topic:
Showing 10 of 50
โ† Return to CRV Science Labs / Natural Products & Microbial Genomics
#60 Tier 1: Exploratory ๐Ÿงช๐Ÿงช 2/5 โ€ข Technical Validation Natural Products & Microbial Genomics

Biosynthetic Gene Cluster Detection: Benchmarking a Domain-Architecture Rule Engine Against GECCO on Five Genomes with Curated MIBiG Loci

Comparing 34-family Pfam rule-based cluster calling with a conditional-random-field detector across 27.5 Mb of bacterial and fungal sequence and 33 experimentally curated clusters

By Bryan P. White โ€ข 13 min read โ€ข Medium (Undergraduate Biology & Biochemistry)
โœจ Executive Summary

We benchmarked a lightweight domain-architecture rule engine for biosynthetic gene cluster (BGC) detection against GECCO, a machine-learning detector, on five genomes (27.5 Mb) containing 33 experimentally curated MIBiG loci. The rule engine overlapped 15 of 33 loci with 29 predicted clusters, versus 27 of 33 with 108 for GECCO, trading sensitivity for fewer unconfirmed calls.

Update (1 November 2026): a follow-up calibration, Lab Discovery #60, scored this detector on five genomes withheld from tuning. On those genomes the detector described below overlapped 9 of 27 curated loci and called 66 clusters, of which 13.6% overlapped a curated locus, against 12.9% for GECCO (24 of 27 loci, 178 clusters). The larger overlap share reported in this article (14 of 29 clusters versus 24 of 108) therefore did not hold on new genomes and should be read as specific to this panel. The caveat that MIBiG is incomplete, so overlap is only a lower bound on precision, applies to both tools.

The follow-up also found that the sensitivity gap narrows after calibration (19 of 27 loci for the revised detector) and that the revised detector called 1 cluster in four BGC-free genomes against 18 for GECCO. This article is otherwise unchanged.

1. Why Cluster Detection Needs Domain Architecture, Not Just Domain Presence

Bacteria and fungi make antibiotics, siderophores and other specialised metabolites with assembly-line enzymes whose genes sit next to one another in biosynthetic gene clusters (BGCs). Two enzyme families dominate: polyketide synthases (PKS), built from ketosynthase (KS), acyltransferase (AT) and acyl-carrier-protein modules, and nonribosomal peptide synthetases (NRPS), built from adenylation (A), peptidyl-carrier (PCP) and condensation (C) modules (Fischbach & Walsh, 2006). A detector that simply counts these domains makes a predictable mistake: every bacterium also carries fatty-acid synthase (FAS) genes with KS, AT and carrier domains, and acyl-CoA ligases that contain an adenylation domain, yet neither produces a natural product.

Two design philosophies address this. GECCO scores every gene by its full Pfam domain profile and feeds the sequence of scores to a conditional random field (CRF), which learns cluster boundaries from known BGCs (Carroll et al., 2021). Our rule engine instead uses only 34 hand-chosen biosynthetic Pfam families and explicit architecture rules: a modular PKS gene must fuse a KS to an AT, reductase or dehydratase; an NRPS gene must carry a C domain or fuse A to a carrier protein. Neighbouring genes then contribute a windowed support score Si = Σj wj over the nine genes centred on gene i, and a cluster is seeded only when a core gene reaches Si ≥ 1.8 (Figure 1).

Figure 1. Five-Stage Native Rule-Based BGC Detection Architecture and Benchmark Design 
========================================================================================================================
  STAGE 1  PROFILE LIBRARY : 34 Pfam seed alignments (PKS, NRPS, terpene, RiPP, siderophore, tailoring) -> Plan-7 profile HMMs
                             (Henikoff position-based weights, 50% match-column rule, Laplace/Dirichlet-style regularisation)
------------------------------------------------------------------------------------------------------------------------
  STAGE 2  GENE CALLING    : 6-frame ORF graph + dicodon Markov coding model + start-codon/RBS scoring (prokaryote model)
------------------------------------------------------------------------------------------------------------------------
  STAGE 3  DOMAIN SCAN     : MSV prefilter (>= 7 bits) -> full Plan-7 forward/backward domain definition
                             Hit accepted only if best domain bit-score >= the Pfam curated gathering (GA) cut-off
------------------------------------------------------------------------------------------------------------------------
  STAGE 4  CLUSTER CALLING : core-gene ARCHITECTURE rules          support window  S(i) = sum of w(j), j in [i-4, i+4]
                             PKS-I    = ketosynthase fused to acyltransferase / reductase / dehydratase
                             NRPS     = condensation domain, or adenylation fused to a carrier protein
                             NRPS-like (v2) = lone adenylation gene + carrier-protein gene within 3 genes
                             seed if core gene has S(i) >= 1.8 (NRPS-like: >= 1.0); extend while gap <= 5 kb; pad 2 kb
------------------------------------------------------------------------------------------------------------------------
  STAGE 5  BENCHMARK       : compare against GECCO v0.11.0 (CRF over the full Pfam library; default posterior cut-off 0.8,
                             sensitive 0.3) and against experimentally curated MIBiG 4.0 loci on the same contigs
========================================================================================================================
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner โ€ข Native 4-Stage 6-Frame ORF Graph, Dicodon Hexamer Markov & 28-Bin Shine-Dalgarno RBS EM Engine โ€ข Native 4-Stage Plan-7 Profile HMM, Forward-Backward Log-Odds, Optimal Accuracy & Null2 Bias Engine

2. Benchmark Design: Five Genomes and Experimentally Curated Ground Truth

We selected five complete genomes that carry MIBiG 4.0 loci with genomic coordinates (Zdouc et al., 2025): Streptomyces coelicolor A3(2), Bacillus velezensis FZB42, Pseudomonas aeruginosa PAO1, Escherichia coli K-12 MG1655 (a near-negative control whose only curated locus is enterobactin), and chromosome II of the fungus Aspergillus nidulans. The fungal genome is deliberately outside the regime of both tools: GECCO was trained on bacterial clusters, and our gene caller uses a prokaryotic model that cannot reconstruct spliced multi-exon genes. After collapsing records with identical coordinates, the five genomes contribute 33 curated loci.

Two cautions apply to every number below. First, 32 of the 33 MIBiG records for these loci carry the curation flag questionable and several have unknown completeness, so coordinates are approximate. Second, MIBiG is far from exhaustive: these genomes contain additional real clusters that are simply not curated. The fraction of predicted clusters that overlap a MIBiG locus is therefore a lower bound on precision, not precision itself. GECCO was run at its documented default posterior cut-off (0.8) and at a permissive cut-off (0.3) to show how its call count scales.

Figure 2. Five-Genome Benchmark Telemetry: Native Rule Engine vs. GECCO v0.11.0 (MIBiG 4.0 Ground Truth) 
========================================================================================================================
  FIVE-GENOME BENCHMARK (27.50 Mb, 33 MIBiG 4.0 loci): NATIVE RULE ENGINE vs. GECCO v0.11.0
========================================================================================================================
  Genome      Organism                              Mb   MIBiG  Native(v2/v1)  GECCO-0.8  GECCO-0.3  Clusters N/G.8/G.3
------------------------------------------------------------------------------------------------------------------------
  AL645882.2  Streptomyces coelicolor A3(2)         8.67     15        7/7         14         15     12/46/73
  CP000560.1  Bacillus velezensis FZB42             3.92      6        4/4          5          5      9/10/12
  AE004091.2  Pseudomonas aeruginosa PAO1           6.26      6        1/1          2          4       3/9/18
  U00096.3    Escherichia coli K-12 MG1655          4.64      1        1/0          1          1        1/2/7
  BN001302.1  Aspergillus nidulans FGSC A4 chr II   4.01      5        2/2          5          5      4/41/45
------------------------------------------------------------------------------------------------------------------------
  TOTAL       5 genomes (4 bacterial, 1 fungal)    27.50     33      15/14         27         30   29/108/155
========================================================================================================================
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner

Across all five genomes the rule engine overlapped 15 of 33 curated loci and GECCO at its default cut-off overlapped 27 of 33 (13 of 28 versus 22 of 28 in the four bacterial genomes). GECCO reached 30 of 33 at the permissive cut-off, at the cost of 155 clusters. The two tools differ most in how many clusters they report: 29 for the rule engine against 108 for GECCO (0.8), of which 14 and 24 respectively overlap a curated locus (Table 1). Because MIBiG is incomplete, the unconfirmed clusters in either list may include real but uncurated BGCs; the data cannot say which.

Table 1. Per-Genome Benchmark Results Against MIBiG 4.0 (Hits = Curated Loci Overlapped by at Least One Predicted Cluster)
Genome Organism Kingdom Genome Size (Mb) Genes Called MIBiG Loci Native Hits (rules v2) Native Hits (rules v1) GECCO Hits (cut-off 0.8) GECCO Hits (cut-off 0.3) Native Clusters GECCO Clusters (0.8) GECCO Clusters (0.3) Native Clusters Overlapping MIBiG GECCO (0.8) Clusters Overlapping MIBiG Base-Pair Jaccard (Native vs GECCO 0.8)
AL645882.2 Streptomyces coelicolor A3(2) Prokaryota 8.67 7738 15 7 7 14 15 12 46 73 6 13 0.288
CP000560.1 Bacillus velezensis FZB42 Prokaryota 3.92 3741 6 4 4 5 5 9 10 12 4 4 0.825
AE004091.2 Pseudomonas aeruginosa PAO1 Prokaryota 6.26 5681 6 1 1 2 4 3 9 18 1 2 0.370
U00096.3 Escherichia coli K-12 MG1655 Prokaryota 4.64 4319 1 1 0 1 1 1 2 7 1 1 0.424
BN001302.1 Aspergillus nidulans FGSC A4 chr II Eukaryota 4.01 5094 5 2 2 5 5 4 41 45 2 4 0.068
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner

3. Sensitivity Depends on Cluster Class

Breaking the 33 loci down by chemistry shows where the rule engine’s gaps come from (Table 2). Its definitions cover PKS, NRPS, terpene and RiPP cores, so loci annotated as saccharide or other (a catch-all category for chemistry that does not fit the named classes) cannot be seeded by design. Within NRPS and hybrid NRPS/PKS clusters, the engine overlapped 11 of 13 loci against 13 of 13 for GECCO. PKS-only loci are the weakest class for the rule engine (2 of 7), partly because the fungal PKS loci sit on a genome where gene calling itself fails.

Table 2. Curated-Locus Recall by Biosynthetic Class (All Five Genomes)
MIBiG Cluster Class Loci Native Hits GECCO Hits (0.8) GECCO Hits (0.3) Class Has Core Definition In Native Engine
Nonribosomal peptide synthetase (NRPS) 11 9 11 11 Yes
Polyketide synthase (PKS) 7 2 6 7 Yes
Hybrid NRPS/PKS 2 2 2 2 Yes
Terpene 3 1 2 3 Yes
Ribosomal peptide (RiPP) 2 0 2 2 Yes
Saccharide 1 0 0 1 No
Other / uncategorised 7 1 4 4 No
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner

4. A Worked Failure: Enterobactin and the Stand-Alone Adenylation Pattern

The E. coli K-12 enterobactin locus (MIBiG BGC0002476) is the clearest single illustration of how a rule gap, rather than a weak search, produced a miss. Enterobactin is assembled from a stand-alone adenylation enzyme (EntE, 536 residues), a stand-alone carrier protein (EntB, 285 residues) and a multi-domain NRPS (EntF) that carries a condensation domain (Raymond et al., 2003). The native engine found the EntE and EntB domain signatures with high confidence, but under the first rule set a lone adenylation gene was not a core gene, so no cluster was called (Figure 3).

We then added an NRPS-like rule (a lone adenylation gene with a carrier-protein gene within three genes). This was introduced after examining this miss, so the enterobactin recovery is a diagnosis-and-repair result, not an independent test. The rule added 3 clusters across the five genomes (26 to 29); only the enterobactin cluster overlaps a curated locus, while the other 2 (in Bacillus velezensis and Aspergillus nidulans) do not overlap any MIBiG record and so cannot be confirmed or rejected here. The total of curated loci overlapped therefore moved from 14 to 15 of 33. The same panel also shows a sensitivity gap: GECCO assigns a reductase (Pfam KR) domain to the third gene of the block at E = 4 × 10−12, which the native profile misses (Section 5).

Figure 3. Enterobactin Locus in E. coli K-12: Curated Boundaries, Both Tools, and Per-Gene Domain Evidence 
========================================================================================================================
  FOCAL LOCUS: ESCHERICHIA COLI K-12 ENTEROBACTIN BIOSYNTHETIC LOCUS (MIBiG BGC0002476; U00096.3:619384-629713)
========================================================================================================================
  Coordinate (kb): 618       620       622       624       626       628       630     631
                   |.........|.........|.........|.........|.........|.........|.........|
  MIBiG curated  :            ###############################################################################
  GECCO (0.8)    : ============================================================================================================================================================   609459-629713 (NRP)
  Native (v2)    : ----------------------------------------------------------------------------------------------------   607459-633999 (NRPS+NRPS-like)
------------------------------------------------------------------------------------------------------------------------
  Genes with a signature domain in the window (native bit-score | GECCO e-value):
   626070-627680 +  536 aa  native: PF00501:126.1,PF13193:40.9   |  GECCO: PF00501 E=3e-77, PF13193 E=2e-12
   627694-628551 +  285 aa  native: PF00550:27.8   |  GECCO: PF00550 E=4e-8
   628551-629297 +   native: (no hit)   |  GECCO: PF08659 E=4e-12   <- sensitivity gap (reductase family)
------------------------------------------------------------------------------------------------------------------------
  Rules v1 : lone adenylation gene (PF00501) has no core architecture  -> no cluster called at this locus
  Rules v2 : adenylation gene + carrier-protein gene within 3 genes   -> NRPS-like cluster; cluster also holds the NRPS gene
========================================================================================================================
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner

5. Domain-Level Concordance: High Precision, Moderate Recall

To separate profile quality from cluster-calling logic, we compared the per-gene domain calls of the native engine with the calls GECCO makes for the same 34 Pfam families on the same genes (genes matched by stop-codon coordinate). Of 1013 native calls, 983 were also found by GECCO (0.970 of native calls confirmed). Of 1490 GECCO calls, 983 were recovered by the native engine (recall 0.660). The native engine is therefore rarely wrong but misses a third of GECCO’s calls (Table 3).

Two effects probably contribute, and the data here cannot apportion them. First, GECCO keeps a domain hit when its p-value is below 10−9 (its default), a single statistical cut-off, whereas the native engine applies each family’s curated Pfam gathering threshold; for broad families such as the short-chain-reductase-like KR domain (PF08659) the two criteria can differ widely, and that family shows the largest gap. Second, the native profiles are built from the small seed alignments (as few as 5–8 sequences for Pfam families PF00195, PF02797 and PF00698) with a simplified prior, whereas the released Pfam models use hmmbuild’s mixture priors and entropy-based sequence weighting (Eddy, 2011; Mistry et al., 2021). The type III PKS families, which have the smallest seeds, are recovered least well (Table 3), and the one type III PKS locus in this benchmark (flaviolin, MIBiG BGC0002127) is missed.

Table 3. Per-Family Domain-Call Concordance: Native Engine vs. GECCO (21 of 34 Families Shown; Five Genomes Pooled)
Pfam Family Name Seed Sequences Model Length (aa) Curated GA Cut-off (bits) Native Calls GECCO Calls Shared Calls Native Recall vs GECCO
PF00109 ketoacyl-synt 75 253 20.8 57 61 55 0.90
PF02801 Ketoacyl-synt_C 75 119 20.8 52 53 52 0.98
PF00698 Acyl_transf_1 8 320 32.7 18 28 18 0.64
PF08659 KR 42 180 22.6 42 181 41 0.23
PF14765 PS-DH 59 168 23.7 16 25 16 0.64
PF00550 PP-binding 150 67 20.8 89 89 87 0.98
PF00501 AMP-binding 136 385 29.9 106 115 106 0.92
PF00668 Condensation 21 457 30.6 29 41 29 0.71
PF00975 Thioesterase 16 231 26.4 20 31 20 0.65
PF13193 AMP-binding_C 222 76 24.5 72 77 70 0.91
PF00195 Chal_sti_synt_N 5 225 27.0 1 5 1 0.20
PF02797 Chal_sti_synt_C 6 151 27.0 2 4 2 0.50
PF05147 LANC_like 45 352 23.2 3 3 3 1.00
PF04738 Lant_dehyd 45 651 28.0 2 2 2 1.00
PF02624 YcaO 181 325 25.7 3 3 3 1.00
PF04183 IucA_IucC 160 240 25.0 4 4 4 1.00
PF00067 p450 50 463 22.8 31 45 31 0.69
PF01494 FAD_binding_3 19 357 22.0 34 71 30 0.42
PF08241 Methyltransf_11 141 96 21.2 109 118 94 0.80
PF13489 Methyltransf_23 40 166 27.0 31 100 31 0.31
PF00155 Aminotran_1_2 48 365 19.7 59 114 59 0.52
๐Ÿ”ง Generated by: Native 4-Stage Plan-7 Profile HMM, Forward-Backward Log-Odds, Optimal Accuracy & Null2 Bias Engine
Figure 4. Most Conserved 60-Column Window of the Adenylation-Domain (PF00501) Pfam Seed Alignment Used to Build the Native Profile 
========================================================================================================================
  FOCAL ALIGNMENT: AMP-BINDING (ADENYLATION) DOMAIN, MOST CONSERVED 60-COLUMN WINDOW OF THE PFAM SEED (PF00501; 136 sequences)
  Alignment columns 43-102 of 722; first six seed sequences shown
========================================================================================================================
  Column index    : .........1.........2.........3.........4.........5.........6   (tick = every 10 columns)
  P95227_MYCTU/6  : TFSELDEAAHAVANGL--LA--K-GV-RAGDG--VAILARNHRWFVIANYGAARVGARII
  Q19878_CAEEL/1  : TYAEFNAHCNRYANYFQ--G--L-GY-RSGDV--VALYMENSVEFVAAWMGLAKIGVVTA
  M2UU11_COCH5/8  : SYKEFSEAYTRVANWLI-DE--L-DV-QVGEM--VAIDGGNSAEHLMLWLALDAIGAATS
  YDAB_BACSU/8-3  : TYKGYRKRINQLANAML--Q--K-GI-QKGDR--VALLCKNGHPASTVMFAALEIGAVVV
  MENE_ARATH/13-  : TGREFVDGVLSLAAGLI-RL----GL-RNGDV--VSIAAFNSDLFLEWLLAVALVGGVVA
  FAC8_MYCTU/54-  : TGGQLADRISQYIQAFE--A--L-GA-GTGVA--VGLLSLNRPEVLMIIGAGQARGYRRT
  Seed consensus  : TYRELDERANRLAAGLR--A--L-GV-GPGDR--VAILLPNSPEWVVALLACLKAGAVYV
  Conservation    : :: ..      .:  :      . :    ..   :.    .  .   .  .   .:.  .   (* >=90% , : >=60% , . >=40% identity)
  Entropy (bits)  : โ–ƒโ–ƒโ–†โ–„โ–„โ–‡โ–‡โ–†โ–…โ–‡โ–†โ–…โ–‚โ–‡โ–‡โ–‚โ–‡โ–‚โ–‡โ–‡โ–โ–ƒโ–†โ–‚โ–‚โ–…โ–‚โ–‡โ–†โ–…โ–…โ–‡โ–„โ–„โ–ƒโ–…โ–„โ–‡โ–†โ–†โ–…โ–†โ–†โ–…โ–†โ–†โ–†โ–†โ–‡โ–…โ–ƒโ–…โ–†โ–†โ–…โ–ƒโ–…โ–…โ–‡โ–…   (taller = more variable)
========================================================================================================================
๐Ÿ”ง Generated by: CRV Science Labs Native Bioinformatics Suite

6. Audit of Every Locus the Rule Engine Did Not Recover

Table 4 lists all 20 curated loci covered by less than half of their length by the native clusters. Of these, 4 are in the fungal genome (out of regime for the gene caller), 8 are bacterial loci in classes that have no core definition (saccharide or other), and 8 are bacterial loci in classes the engine is meant to detect. That last group is the actionable one: it is where profile sensitivity and rule design should be improved. GECCO’s coverage of the same loci is shown for context; it too misses several of the saccharide and other-class loci.

Table 4. Curated Loci with Under 50% Native Coverage, With Probable Cause
Genome MIBiG Locus Class Locus Length (bp) Native Coverage (fraction) GECCO 0.8 Coverage (fraction) Interpretation
AL645882.2 BGC0000325 NRPS 22626 0.00 1.00 Core NRPS architecture not recovered at the curated cut-offs, or cluster support below the seeding threshold.
AL645882.2 BGC0002127 PKS 3014 0.00 1.00 Type III-like or small cluster whose seed-built domain profile scored below the curated Pfam gathering cut-off.
AL645882.2 BGC0000910 other 1420 0.00 1.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AL645882.2 BGC0000940 other 7328 0.00 1.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AL645882.2 BGC0000660 terpene 2468 0.00 1.00 Terpene locus not recognised by the two terpene architectures in the engine (terpene synthase, or paired cyclase domains).
AL645882.2 BGC0001181 terpene 2181 0.00 0.00 Terpene locus not recognised by the two terpene architectures in the engine (terpene synthase, or paired cyclase domains).
AL645882.2 BGC0000849 other 1710 0.05 1.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AL645882.2 BGC0000551 ribosomal 6728 0.00 1.00 Core NRPS architecture not recovered at the curated cut-offs, or cluster support below the seeding threshold.
AL645882.2 BGC0002128 other 8672 0.00 0.87 Cluster class outside the 34 signature families (no core definition for this chemistry).
CP000560.1 BGC0000616 ribosomal 4173 0.00 1.00 Core NRPS architecture not recovered at the curated cut-offs, or cluster support below the seeding threshold.
CP000560.1 BGC0001184 other 7298 0.00 0.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AE004091.2 BGC0000922 PKS 13136 0.00 0.00 Type III-like or small cluster whose seed-built domain profile scored below the curated Pfam gathering cut-off.
AE004091.2 BGC0002037 NRPS 16597 0.00 1.00 Core NRPS architecture not recovered at the curated cut-offs, or cluster support below the seeding threshold.
AE004091.2 BGC0002645 other 6095 0.00 0.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AE004091.2 BGC0000725 saccharide 16652 0.00 0.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
AE004091.2 BGC0002489 other 5082 0.00 0.00 Cluster class outside the 34 signature families (no core definition for this chemistry).
BN001302.1 BGC0001515 NRPS 64035 0.40 1.00 Fungal gene structure: spliced multi-exon genes are fragmented by a bacterial-style gene caller, so domain evidence is split or missing.
BN001302.1 BGC0002758 PKS 30659 0.00 1.00 Fungal gene structure: spliced multi-exon genes are fragmented by a bacterial-style gene caller, so domain evidence is split or missing.
BN001302.1 BGC0001121 PKS 9908 0.00 1.00 Fungal gene structure: spliced multi-exon genes are fragmented by a bacterial-style gene caller, so domain evidence is split or missing.
BN001302.1 BGC0000107 PKS 6819 0.00 0.88 Fungal gene structure: spliced multi-exon genes are fragmented by a bacterial-style gene caller, so domain evidence is split or missing.
๐Ÿ”ง Generated by: Native Biosynthetic Gene Cluster (MIBiG / ARTS) & Amphipathic AMP Biophysics Miner

7. Runtime

The native engine was not faster. Its total wall-clock time across the five genomes was 163.7 s, of which 100.9 s was gene calling and 62.7 s was the 34-family domain scan; GECCO at its default cut-off took 127.4 s in total, including a scan against the full Pfam library with a compiled search library. The ratio is a property of these two implementations on one Windows machine rather than of the algorithms, but it does mean the rule engine’s benefit here is selectivity, not speed.

Discussion & Key Takeaways

1. Selectivity versus sensitivity. On this panel the rule engine overlapped fewer curated loci than GECCO (15 versus 27 of 33) but reported far fewer clusters (29 versus 108), and a larger share of its clusters overlap a curated locus (14 of 29 versus 24 of 108). Because MIBiG is incomplete, this is a lower bound on precision, and the unconfirmed clusters from either tool may include genuine uncurated BGCs. The result points to a trade-off worth testing further; it does not show one approach to be better.

2. Failures are mostly explicable and partly fixable. Of the 20 missed loci, 12 lie outside the engine’s declared scope (fungal gene structure, or chemistry without a core definition). The enterobactin miss was a rule gap that a single added rule repaired, although that repair was made after seeing the failure. The remaining in-scope misses are consistent with profiles built from small seed alignments and a stricter cut-off than GECCO uses.

3. Signal Rating Assessment (2๐Ÿงช Technical Validation). The benchmark spans five genomes (27.5 Mb) and 33 curated loci, with one near-negative control and one deliberately out-of-regime fungal genome. It is a reproducible technical benchmark rather than a discovery: n = 5 genomes is too few to estimate how performance varies across bacterial clades, MIBiG labels are mostly flagged questionable, the rule revision was informed by the same data it was scored on, and GECCO was used without retraining. We therefore rate it as a technical validation and not as a comparative synthesis.

4. Candidate directions for downstream genome mining. Combining a high-selectivity architecture filter with a high-sensitivity learned detector is a testable strategy for ranking clusters before expensive chemistry or sequencing work. Any such prioritisation is a computational hypothesis for follow-up in the laboratory, not evidence that a cluster produces a compound.

References & Primary Sources

  1. [1] Carroll LM, Larralde M, Fleck JS, Ponnudurai R, Milanese A, Cappio E, Zeller G (2021). Accurate de novo identification of biosynthetic gene clusters with GECCO. bioRxiv (preprint). https://doi.org/10.1101/2021.05.03.442509
  2. [2] Zdouc MM, Blin K, Louwen NLL, et al. (2025). MIBiG 4.0: advancing biosynthetic gene cluster curation through global collaboration. Nucleic Acids Research 53(D1):D678โ€“D690. https://doi.org/10.1093/nar/gkae1115
  3. [3] Mistry J, Chuguransky S, Williams L, et al. (2021). Pfam: The protein families database in 2021. Nucleic Acids Research 49(D1):D412โ€“D419. https://doi.org/10.1093/nar/gkaa913
  4. [4] Eddy SR (2011). Accelerated profile HMM searches. PLoS Computational Biology 7(10):e1002195. https://doi.org/10.1371/journal.pcbi.1002195
  5. [5] Fischbach MA, Walsh CT (2006). Assembly-line enzymology for polyketide and nonribosomal peptide antibiotics: logic, machinery, and mechanisms. Chemical Reviews 106(8):3468โ€“3496. https://doi.org/10.1021/cr0503097
  6. [6] Raymond KN, Dertz EA, Kim SS (2003). Enterobactin: an archetype for microbial iron transport. Proceedings of the National Academy of Sciences 100(7):3584โ€“3588. https://doi.org/10.1073/pnas.0630018100
  7. [7] Henikoff S, Henikoff JG (1994). Position-based sequence weights. Journal of Molecular Biology 243(4):574โ€“578. https://doi.org/10.1016/0022-2836(94)90032-9
  8. [8] White BP (2026). Calibrating a Biosynthetic Gene Cluster Detector: Released Profiles, Added Marker Families and a Reserve-Genome Test Against GECCO. CRV Science Labs, Lab Discovery #60 (follow-up to this article). https://crvscience.com/labs/discoveries/bgc-reserve-genomes