Locus heterogeneity, not polygenicity, explains how gentamicin resistance transfers across Klebsiella pneumoniae lineages: evidence from a five-outcome lineage-held-out evaluation

Background Resistance architecture is usually treated as a dichotomy: one dominant acquired determinant, or a distributed polygenic basis. That conflates two mechanistically different situations, several determinants acting together in one isolate versus several alternative determinants of which any one suffices, and the difference is invisible to discrimination metrics. Methods We evaluated gentamicin resistance in 4,001 Klebsiella pneumoniae isolates with laboratory-confirmed BV-BRC phenotypes under a lineage-held-out design, assigning whole MLST sequence types to exactly one partition so no clone appears in both training and test. Because discrimination cannot separate architectures, we decomposed the fitted model three ways: a union ladder over candidate resistance loci, a mechanism-attribution census, and full-model AUROC recomputed within each cell of the best single indicator. The same framework was applied unchanged to four further outcomes: meropenem, imipenem, ceftazidime and dual meropenem+ceftazidime resistance (n = 400 to 3,317). Results Gentamicin’s 709-feature model beat the best single-gene indicator (AUROC 0.9875 versus 0.9443), which would ordinarily indicate polygenic signal. Decomposition showed otherwise. Widening the resistance-locus union made prediction worse (0.9463 to 0.8482), admitting susceptible isolates carrying co-located streptomycin genes. One locus family, aac(3), explained 90.0% of held-out resistant isolates. Decisively, the full model’s advantage lived almost entirely among aac(3)-negative isolates, separating 24 resistant from 357 susceptible at AUROC 0.9215, with nothing left to refine among aac(3)-positives. This is locus heterogeneity: several alternative, individually sufficient mechanisms, each resistant isolate explained by exactly one. Across all five outcomes the lineage-only baseline scored 0.802 to 0.927 on random splits and collapsed to exactly 0.500 when held out, while mechanism features retained 0.834 to 0.988 and beat that baseline everywhere (p = 2.3e-6 to 2.6e-92). Adding lineage identity on top of mechanism was significant in only one of five outcomes, gentamicin, where mechanism-only was better; pooling the five is not defensible, and no combined directional claim is made. Conclusions Locus heterogeneity is a third resistance architecture, separable from additive polygenicity by a within-cell test that discrimination metrics cannot perform, and it requires feature sets resolving individual genes. The framework’s scope is resistance caused by discrete acquired genes; whether an additive or regulatory signature behaves the same way remains untested, and the attempt to test it was blocked by breakpoint-era label incomparability rather than lack of effort.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24267397
Primary Topic
Antibiotic Resistance in Bacteria
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Locus heterogeneity, not polygenicity, explains how gentamicin resistance transfers across Klebsiella pneumoniae lineages: evidence from a five-outcome lineage-held-out evaluation

Nandha Saravanan
1 links
Zenodo (CERN European Organization for Nuclear Research)
Antibiotic Resistance in Bacteria
preprint

Locus heterogeneity, not polygenicity, explains how gentamicin resistance transfers across Klebsiella pneumoniae lineages: evidence from a five-outcome lineage-held-out evaluation

Nandha Saravanan
preprint en

Abstract

Background Resistance architecture is usually treated as a dichotomy: one dominant acquired determinant, or a distributed polygenic basis. That conflates two mechanistically different situations, several determinants acting together in one isolate versus several alternative determinants of which any one suffices, and the difference is invisible to discrimination metrics. Methods We evaluated gentamicin resistance in 4,001 Klebsiella pneumoniae isolates with laboratory-confirmed BV-BRC phenotypes under a lineage-held-out design, assigning whole MLST sequence types to exactly one partition so no clone appears in both training and test. Because discrimination cannot separate architectures, we decomposed the fitted model three ways: a union ladder over candidate resistance loci, a mechanism-attribution census, and full-model AUROC recomputed within each cell of the best single indicator. The same framework was applied unchanged to four further outcomes: meropenem, imipenem, ceftazidime and dual meropenem+ceftazidime resistance (n = 400 to 3,317). Results Gentamicin’s 709-feature model beat the best single-gene indicator (AUROC 0.9875 versus 0.9443), which would ordinarily indicate polygenic signal. Decomposition showed otherwise. Widening the resistance-locus union made prediction worse (0.9463 to 0.8482), admitting susceptible isolates carrying co-located streptomycin genes. One locus family, aac(3), explained 90.0% of held-out resistant isolates. Decisively, the full model’s advantage lived almost entirely among aac(3)-negative isolates, separating 24 resistant from 357 susceptible at AUROC 0.9215, with nothing left to refine among aac(3)-positives. This is locus heterogeneity: several alternative, individually sufficient mechanisms, each resistant isolate explained by exactly one. Across all five outcomes the lineage-only baseline scored 0.802 to 0.927 on random splits and collapsed to exactly 0.500 when held out, while mechanism features retained 0.834 to 0.988 and beat that baseline everywhere (p = 2.3e-6 to 2.6e-92). Adding lineage identity on top of mechanism was significant in only one of five outcomes, gentamicin, where mechanism-only was better; pooling the five is not defensible, and no combined directional claim is made. Conclusions Locus heterogeneity is a third resistance architecture, separable from additive polygenicity by a within-cell test that discrimination metrics cannot perform, and it requires feature sets resolving individual genes. The framework’s scope is resistance caused by discrete acquired genes; whether an additive or regulatory signature behaves the same way remains untested, and the attempt to test it was blocked by breakpoint-era label incomparability rather than lack of effort.

Zenodo (CERN European Organization for Nuclear Research)
Good health and well-being
Antibiotic Resistance in Bacteria
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon