Evolutionary Traps in WHO Priority Pathogen Collateral Sensitivity Networks: A Graph-Theoretic Hypothesis and Experimental Validation Design
Abstract
REFUTED 20 August 2026. The hypothesis was that the directed collateral sensitivity graph for WHO priority pathogens contains strongly connected components of size ≥3 defining closed evolutionary traps — resistance cycles with no viable single-step escape — which sequential cycling therapy could exploit to hold a pathogen susceptible indefinitely. It was tested against 104,337 susceptibility records from BV-BRC and a candidate was found: a 3-node component in K. pneumoniae linking imipenem, meropenem and tetracycline. It does not survive stratification by clonal lineage. Within MLST sequence types the association is absent (pooled odds ratios 0.93 and 0.95, Cochran-Mantel-Haenszel p = 0.73 and 0.80, confidence intervals excluding the unadjusted estimates), and the two dominant lineages pull in opposite directions (ST307, 1.26; ST258, 0.67). Carbapenem-resistant and tetracycline-resistant phenotypes co-occur because they are carried by different successful clones, not because resistance to one induces susceptibility to the other. The manuscript reporting the positive finding was withdrawn from journal review, and the experimental validation package below is withdrawn with it: an isogenic experiment is a within-lineage comparison by construction, so the prediction it would have tested has already been evaluated across 473 co-tested isolates and is not supported. The earlier claim of 50,000–200,000 deaths averted per year was conditional on confirmation and does not stand. What survives is methodological, and is reported separately: none of the checks that made this signal convincing — FDR correction, permutation significance, drug-class specificity, temporal stability — distinguishes a trade-off from linkage within a clone, and the controls that do are inexpensive.
0/3 confirmed · 2 withdrawn
For at least one WHO critical-priority pathogen (K. pneumoniae, A. baumannii, P. aeruginosa, or E. coli), the directed collateral sensitivity graph G constructed from GLASS WGS data contains at least one SCC of size ≥3 where every directed edge (A → B) represents OR > 2.0 (p < 0.05, BH-corrected) for susceptibility to drug B conditioned on resistance to drug A.
No single-step resistance mutation accessible from any node in the SCC (frequency > 10⁻⁷) confers simultaneous clinical resistance (MIC > EUCAST breakpoint) to all drugs in the SCC.
A sequential cycling protocol following the SCC edge order maintains pathogen MIC below clinical resistance breakpoints in ≥75% of experimental replicates through passage 30, compared to ≤50% of monotherapy controls reaching resistance by passage 15.
Key Findings
- 1REFUTED: the candidate SCC in K. pneumoniae is confounded by clonal lineage and is absent within MLST sequence types (pooled OR 0.93 and 0.95; CMH p = 0.73 and 0.80)
- 2The two dominant lineages pull in opposite directions (ST307, OR 1.26; ST258, 0.67) — the crude association comes from contrasting clones, not from any within-lineage trade-off
- 3Size-matched random partitions leave the odds ratio at 1.76 against a crude 1.81, so the collapse is lineage-specific rather than an artefact of stratification (0 of 300 permutations reached the observed value)
- 4The drug-class specificity offered as evidence against a clonal explanation weakens under the same stratification — mechanistic plausibility constrains which confounded associations arise, it does not establish that one is unconfounded
- 5Transferable result: FDR correction, permutation significance, class specificity and temporal stability are safeguards against chance, and clonal confounding is not chance. Lineage stratification plus a size-matched permutation control detects it and costs nothing
Source Discoveries
Hypotheses in this paper were sourced from the following AegisMind discoveries on solver.press.
- 59%
252 days
Timeline
1,200
GPU hours
2,800
CPU hours
512 GB
Memory
$185k
Budget (min)
$620k
Budget (full)
Required Datasets
- GLASS WGS database (≥50,000 isolates, 6 WHO critical-priority pathogens)
- 47 published collateral sensitivity studies (standardized to EUCAST 2024 breakpoints)
- Deep mutational scanning (DMS) datasets for key resistance genes (GyrA, OmpF, PBP2, RpoB, OXA family)
- Sequential clinical isolate pairs from GLASS for Phase 5 retrospective analysis
- Isogenic reference strains: PA14 and ATCC clinical references for Phase 4 CRISPR validation
Experimental Protocol
Phase 1 (Weeks 1–8): Mine GLASS WGS database; logistic regression edge estimation with BH correction; phylogenetic confounding via ≥10 PCs; Tarjan's SCC detection; 1,000-permutation null model significance testing. Go/No-Go: ≥1 SCC size ≥3 with p < 0.05.
Phase 2 (Weeks 6–12): Single-step mutation enumeration within 2 steps of each SCC node; QSAR-model frequency estimation. Go/No-Go: no escape variant at frequency > 10⁻⁷.
Phase 3 (Weeks 10–24): Four arms (SCC cycling, monotherapy, random cycling, no treatment), 12 replicates each, 30 serial passages at 10⁶ CFU/mL CAMHB. MIC every 5 passages; WGS at passages 0, 5, 10, 20, 30 (1,440 libraries at 50× coverage). Log-rank test: time-to-resistance Arm A vs. Arm B.
Phase 4 (Weeks 20–30): CRISPR-Cas9 introduction of resistance mutations in top 3 SCC edges; isogenic PA14 or ATCC backgrounds; MIC confirmation.
Phase 5 (Weeks 24–36): Retrospective Cox proportional hazards on GLASS sequential isolate pairs; HR ≤ 0.7 (95% CI excludes 1.0) for SCC-approximating treatment sequences.
Success Criteria
- Phase 1: ≥1 SCC size ≥3, all edges OR > 2.0, empirical p < 0.05
- Phase 2: No simultaneous resistance variant at frequency > 10⁻⁷
- Phase 3: ≥75% replicates susceptible at passage 30 vs. ≤50% monotherapy reaching resistance by passage 15 (log-rank p < 0.01)
- Phase 4: ≥2/3 CRISPR-validated edges produce predicted collateral sensitivity (p < 0.05)
- Phase 5: HR ≤ 0.7, 95% CI excludes 1.0
Failure Criteria
- Phase 1 Go/No-Go fails: no SCC ≥3 identified in GLASS data at OR > 2.0 threshold
- SCC cycling fails to outperform random cycling (log-rank p > 0.10)
- WGS reveals escape mutations in > 30% of lineages by passage 20
- CRISPR validation fails all 3 edges
Abort Checkpoints
- Phase 1 Week 8: Abort if no SCC ≥3 found (empirical p > 0.10)
- Phase 2 Week 12: Abort if escape variant identified at frequency > 10⁻⁵
- Phase 3 Week 16: Abort if no difference in resistance emergence between Arm A and Arm B at passage 15
- Phase 4 Week 26: Scope down to retrospective analysis only if CRISPR efficiency < 30%
Commercial ROI
Withdrawn. This rested on the cycling framework being real; it is not supported by this dataset.
Research ROI
Withdrawn as stated — the 50,000–200,000 deaths-averted figure was conditional on confirmation and does not stand. What the work did yield is methodological and is reported separately: a lineage-stratification and permutation control that distinguishes a genuine collateral-sensitivity trade-off from linkage within a successful clone, at no cost beyond the analysis already being run.
Aggregated EVP Package
This paper is part of the AMR Evolutionary Trap EVP cluster. The aggregated EVP combines evidence from multiple papers targeting related mechanisms, enabling shared experimental infrastructure and compounded validation.
View aggregated EVP →