A natural history of AMR in Klebsiella pneumoniae: Global diversity, predictors, and predictions of evolutionary pathways

By analyzing 47,000 global *Klebsiella pneumoniae* genomes and applying hypercubic transition path sampling, this study maps and predicts the diverse evolutionary pathways of antimicrobial resistance, revealing how drug policies and public health contexts drive both consistent and divergent resistance patterns across countries.

The Gist

Imagine a global game of "Evolutionary Tetris," but instead of falling blocks, we are watching how bacteria like Klebsiella pneumoniae (a common germ that causes serious infections) stack up different "resistance blocks" to survive our antibiotics.

This paper is like a massive, high-tech detective story. The authors didn't just look at which bacteria are resistant today; they tried to figure out the order in which they picked up these resistances over time, and why some countries have different "stacking patterns" than others.

Here is the breakdown of their findings using simple analogies:

1. The Detective Tool: "HyperTraPS"

Imagine you have a pile of 47,000 photos of bacteria from 102 different countries. Each photo shows which "weapons" (drug resistances) the bacteria has.

• The Problem: You can't just look at the photos and guess the history. Did the bacteria get the gun first, then the knife? Or the knife, then the gun?
• The Solution: The authors used a machine learning tool called HyperTraPS. Think of this as a "Time-Travel Simulator." It takes all those photos and runs millions of simulations to figure out the most likely story of how the bacteria evolved. It builds a "roadmap" of evolution, showing the most probable path from a harmless germ to a super-bug.

2. The "Universal Rules" vs. "Local Customs"

The study found two types of evolutionary behaviors:

• The "Universal Rules" (Globally Consistent):
  No matter where you are in the world, bacteria seem to follow a similar script for certain drugs.

  • Analogy: It's like learning to drive. Almost everyone learns the basics (steering, braking) before they try to do a handbrake turn.
  • The Science: Bacteria almost always pick up resistance to older, common drugs (like penicillin-like drugs) first. They save the "heavy artillery" (like Colistin, the last-resort drug) for last. This happens everywhere, from Norway to Tanzania.

• The "Local Customs" (Globally Divergent):
  However, for some specific drugs, the order changes depending on the country.

  • Analogy: Think of it like traffic laws. In the UK, you drive on the left; in the US, you drive on the right. The car is the same, but the rules of the road are different.
  • The Science: Resistance to drugs like Carbapenems (very strong antibiotics) and Fluoroquinolones happens at different times in different places.
    • In Sub-Saharan Africa, bacteria tend to pick up these strong resistances later in their evolution.
    • In Asia, they seem to pick them up earlier.
  • Why? The paper links this to drug policy. If a country uses a specific antibiotic heavily, the bacteria there evolve resistance to it faster. It's an arms race driven by how much the "enemy" (humans) is shooting at them.

3. The "Bubble Map"

To visualize this, the authors created "Bubble Plots."

• Imagine a grid where the X-axis is "Time" (Early to Late) and the Y-axis is "Drug Type."
• A big bubble means "It is very likely this drug resistance appears at this specific time."
• A small or missing bubble means "It's random or unlikely."
• By comparing these maps, they could see that while the "Early" and "Late" drugs are consistent globally, the "Middle" drugs vary wildly by region.

4. The Crystal Ball: Predicting the Future

The coolest part of the paper is the prediction.

• The Test: The authors took brand-new bacteria samples from Tanzania (collected in 2001, 2015, and 2017) that they didn't use to build their model.
• The Prediction: They asked their model: "If we have a bacteria with these resistances, what will it evolve next?"
• The Result: The model guessed correctly most of the time! It successfully predicted which new resistance would appear next, outperforming simple guesses based just on how common a drug is.
• The Application: They built a free online app (a "Crystal Ball" for doctors). A doctor can input a patient's bacteria profile, and the app will say, "Be careful, this bacteria is likely to become resistant to Drug X next." This helps doctors choose the right treatment before the bacteria evolves.

5. The Big Picture

This paper changes how we fight superbugs.

• Old Way: "Oh, this bacteria is resistant to Drug A. Let's try Drug B." (Reactive)
• New Way: "We know the 'roadmap' of evolution. If this bacteria has Drug A, it is 90% likely to get Drug B next. Let's avoid Drug B entirely and use Drug C." (Proactive)

In summary: The authors mapped the "evolutionary DNA" of how bacteria learn to fight drugs. They found that while some rules are the same everywhere, local drug use creates unique evolutionary paths. By understanding these paths, we can predict the future of superbugs and stay one step ahead in the medical arms race.

Technical Summary

1. Problem Statement

Antimicrobial resistance (AMR) in Klebsiella pneumoniae (Kp) is a critical global health threat, causing significant mortality, particularly in sub-Saharan Africa. While large-scale genomic datasets exist describing the current prevalence of resistance genes, there is a lack of understanding regarding the evolutionary pathways by which these resistance traits are acquired.

• The Gap: Traditional evolutionary models (e.g., Mk models) struggle with the high dimensionality of AMR data (often >10 coupled resistance characters) and fail to account for phylogenetic relatedness or complex interactions between traits.
• The Question: In what order are specific resistance features acquired globally? Do these pathways vary by geography or drug policy? Can we predict the next evolutionary step for a specific isolate to inform treatment guidelines?

2. Methodology

The study employs a machine learning approach called HyperTraPS (Hypercubic Transition Path Sampling), an Evolutionary Accumulation Modelling (EvAM) technique, to infer probabilistic evolutionary pathways.

• Data Collection:

  • Training Set: 47,721 K. pneumoniae sensu stricto genomes from 102 countries/territories (sourced from PathogenWatch).
  • Features: Resistance profiles for 22 distinct AMR character classes (genes/mutations) derived using Kleborate. Characters include resistance to aminoglycosides, carbapenems, fluoroquinolones, colistin, etc., categorized by mechanism (e.g., acquired via horizontal gene transfer vs. mutation).
  • Phylogeny: A coarse-grained phylogeny was constructed using LIN-codes (Life Identification Numbers) to group isolates and prevent pseudoreplication, ensuring transitions are treated as independent evolutionary events.
  • Validation Set: Newly sequenced genomes from Tanzania (2001–2002, 2017–2018) and Zanzibar (2015–2016), spanning several decades, used as an independent test set.

• Inference Engine (HyperTraPS):

  • Models the evolutionary space as a hypercube where each node represents a binary state of the 22 resistance characters.
  • Infers a transition matrix describing the probability of moving from one state to another (acquiring a new resistance trait).
  • Accounts for interactions between traits (e.g., acquiring Trait A might increase or decrease the rate of acquiring Trait B).
  • Outputs a posterior probability distribution for the ordering of character acquisition (i.e., the likelihood that Character X is the 1st, 2nd, or 10th trait acquired).

• Analysis Framework:

  • Principal Component Analysis (PCA): Applied to the high-dimensional posterior probability matrices of each country to reduce dimensionality and identify global patterns of variability.
  • Correlation Analysis: Linked PCA projections to Global Burden of Disease (GBD) regions and WHO GLASS antibiotic consumption data (2016–2021).
  • Prospective Validation: Tested the trained models against the newly sequenced Tanzanian/Zanzibar data to see if the model could correctly predict the "next step" in evolutionary transitions observed in the new data.

3. Key Contributions

1. Global Evolutionary Roadmap: The first large-scale inference of the ordered acquisition pathways of 22 AMR traits in K. pneumoniae across 102 countries.
2. Differentiation of Dynamics: Distinguished between "Globally Consistent" pathways (traits acquired in a similar order regardless of location) and "Globally Divergent" pathways (traits whose acquisition order varies significantly by region).
3. Predictive Validation: Demonstrated that evolutionary models trained on historical global data can successfully predict prospective evolutionary dynamics in new, independent datasets (Tanzania/Zanzibar) spanning decades.
4. Policy Correlation: Established quantitative links between specific antibiotic usage policies and the timing of resistance evolution, particularly for carbapenems and fluoroquinolones.
5. Open Tool: Development of a public Shiny app allowing clinicians and researchers to input a specific resistance profile and a country of interest to receive probabilistic predictions of future resistance evolution.

4. Key Results

A. Global vs. Local Pathways

• Globally Consistent (PCA1): Certain traits follow a rigid, universal order.
  • Early acquisition: Resistance to aminoglycosides, sulfonamides, trimethoprim, and basic beta-lactams (e.g., Bla_chr, SHV-m).
  • Late acquisition: Resistance to last-resort drugs like colistin, fosfomycin, tigecycline, and glycopeptides.
  • Observation: Variability in PCA1 largely reflects the precision of the inference (data quantity/quality) rather than biological differences.
• Globally Divergent (PCA2 & PCA3): Specific traits show significant regional variation in acquisition order.
  • PCA2: Driven by carbapenem resistance (Bla_Carb-a), fluoroquinolone resistance (Flq-m), and outer membrane protein mutations.
    • Finding: Sub-Saharan Africa shows significantly later acquisition of carbapenem resistance compared to other regions, correlating with the later introduction of carbapenems in clinical practice there.
  • PCA3: Driven by fluoroquinolone and rifampicin resistance.
    • Finding: South/Southeast/East Asia and Oceania show earlier acquisition of fluoroquinolone resistance compared to Europe and Africa.

B. Drivers of Variability

• Drug Policy: There is a strong correlation between per-capita antibiotic usage and the timing of resistance acquisition for "divergent" traits. Higher usage of carbapenems correlates with earlier acquisition of carbapenem resistance.
• Geography: Public health superregions act as strong predictors. For instance, the "Sub-Saharan Africa" cluster is distinct in its delayed acquisition of last-resort drug resistance.

C. Predictive Power

• Validation: When applied to the newly sequenced Tanzanian isolates, the HyperTraPS model successfully ranked the actual next evolutionary steps higher than a naive model based solely on trait prevalence.
• Zanzibar Case Study: The model correctly predicted that Zanzibar's evolutionary dynamics would align with the broader Sub-Saharan African pattern (late carbapenem resistance), validating the model's ability to generalize to unseen regions.

D. Interactions and Bimodality

• Bimodality: The model identified "bimodal" acquisition patterns for certain traits (e.g., SHV-m), suggesting distinct evolutionary pathways where a trait is either acquired very early or very late, but rarely in the middle.
• Interactions: Evidence of "promoting" interactions (e.g., sulfonamide resistance promoting trimethoprim resistance due to genetic linkage) and "repressing" interactions (e.g., MLS resistance potentially repressing carbapenem resistance in some countries), suggesting potential collateral sensitivity opportunities.

5. Significance

• Clinical Utility: The ability to predict the "next step" in AMR evolution allows for the adaptation of treatment guidelines. If a specific resistance profile is detected, clinicians can anticipate which resistance is likely to evolve next and adjust empiric therapy accordingly.
• Policy Impact: The study provides evidence that antibiotic stewardship policies directly shape the order of resistance evolution, not just the prevalence. This supports targeted interventions in regions where specific resistance pathways are emerging.
• Methodological Advancement: By successfully applying HyperTraPS to a massive, phylogenetically structured dataset, the authors demonstrate that evolutionary accumulation modeling can handle the complexity of real-world AMR data, overcoming limitations of traditional phylogenetic comparative methods.
• Future Directions: The framework is portable to other pathogens and can be refined with continuous-time models (HyperTraPS-CT) to link evolutionary dynamics more precisely to historical policy changes and outbreaks.

In summary, this paper moves beyond static genomic surveillance to a dynamic, predictive understanding of AMR, revealing that while some resistance pathways are universal, others are highly plastic and driven by local drug use and geography, offering a new tool for global health strategy.