Collateral Sensitivity Strongly Connected Components in Real-World Clinical Surveillance Data Are Confounded by Clonal Lineage
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Collateral sensitivity (CS) — acquisition of resistance to one antibiotic inducing hyper-sensitivity to another — offers an evolutionary trap for multidrug-resistant pathogens. A strongly connected component (SCC) in the directed CS graph represents a closed cycle in which every drug is reachable from every other via successive CS edges. Prior evidence for CS SCCs derives from in vitro experiments; whether such structures exist in clinical surveillance data has not been tested.
We mined 104,337 antibiotic susceptibility records from BV-BRC spanning four WHO critical-priority pathogens ( Klebsiella pneumoniae, Escherichia coli, Staphylococcus aureus, Pseudomonas aeruginosa ; 18,821 unique isolates), applying Fisher’s exact test with Benjamini– Hochberg FDR correction to all ordered antibiotic pairs, Tarjan’s algorithm for SCC detection, and permutation testing ( n = 1,000).
A 3-node SCC in K. pneumoniae — imipenem, meropenem, tetracycline — met every criterion: empirical p = 0.001, bidirectional carbapenem–tetracycline edges at OR = 1.81– 1.82 ( q < 0.002, n > 850 per edge), tetracycline-specific relative to tigecycline, and stable across independent year bands.
It does not survive stratification by clonal lineage
Assigning MLST sequence types to 5,276 of 5,382 genomes and pooling the same contingency tables by Cochran–Mantel– Haenszel gives within-lineage odds ratios of 0.93 and 0.95 (CMH p = 0.73 and 0.80), with confidence intervals excluding the unadjusted estimate. The two dominant strata pull in opposite directions (ST307, OR = 1.26; ST258, OR = 0.67). Permuting sequence-type labels 300 times while preserving stratum sizes leaves the odds ratio at a median of 1.76, so the collapse is attributable to lineage specifically rather than to stratification; de novo re-typing from genome sequence agreed with the assignments used in 32 of 33 isolates.
The crude association is therefore confounded by clonal structure: carbapenem-resistant and tetracycline-resistant phenotypes co-occur because they are carried by different successful lineages, not because resistance to one induces susceptibility to the other. We report this as a negative result with a reusable control. Two larger clinical collateral-sensitivity analyses have been published without lineage adjustment, and the permutation test used here distinguishes genuine confounding from stratification artefact at negligible cost.
Importance
Drug-resistant bacterial infections killed an estimated 1.27 million people in 2019, and few new antibiotics are reaching the clinic. One proposed response is to rotate drugs so that bacteria adapting to one become more vulnerable to the next. That only helps if the trade-offs connect into a closed loop, leaving the bacterium no escape. Searching routine hospital testing records — 104,337 results from 18,821 patient isolates of four high-priority pathogens — we found what looked like such a loop in Klebsiella pneumoniae , linking two carbapenems and tetracycline. It was statistically strong, specific to one drug class, and stable over time. It is nonetheless not a trade-off. When isolates are grouped by bacterial lineage, the association disappears: the drugs are not trading off against each other, they are carried by different successful clones. We report this because the same confound applies to any collateral-sensitivity signal read from surveillance data, and because the check that detects it is cheap. Testing a lineage-confounded signal at the bench would cost far more than the analysis that rules it out.
Version 4 note
Versions 1 and 2 of this preprint reported a 3-node collateral-sensitivity strongly connected component in K. pneumoniae as a positive finding. It does not survive stratification by clonal lineage. Version 3 reported that analysis and revised the conclusion accordingly; the earlier versions remain accessible at their own DOIs. The manuscript was withdrawn from journal review on this basis.
Version 4 corrects one sentence in §5. Version 3 stated that whether the two larger published clinical CS analyses survive lineage adjustment was “an open question that the data to answer already exist for”. That is not correct. Neither underlying collection links susceptibility phenotypes to genome assemblies, so neither could have performed the adjustment and neither can be re-tested as it stands. Nothing else has changed and no result is affected.
