Inferring the multi-host fitness landscape of endive necrotic mosaic virus from cross-inoculation experiments

This paper presents a Bayesian approach that links cross-inoculation assay data to Fisher's geometrical model to infer a multi-host fitness landscape for endive necrotic mosaic virus, revealing host permissiveness heterogeneity and phylogenetic concordance to better understand viral adaptation and establishment barriers.

The Gist

Imagine a virus as a traveler trying to find a home in a new city. But this isn't just any city; it's a world made up of five different neighborhoods (the five different plants). Each neighborhood has its own unique rules, architecture, and security systems. To survive and thrive in a neighborhood, the traveler needs to speak the local language and look the part.

This paper is about figuring out the "Map of Compatibility" for a specific plant virus called the Endive Necrotic Mosaic Virus (ENMV). The scientists wanted to answer a big question: If a virus gets good at living in one type of plant, how easy is it for it to jump to a different type of plant?

Here is the story of how they did it, explained simply:

1. The Experiment: The "Cross-Training" Gym

The researchers took a virus and played a game of evolutionary "cross-training."

• Step 1: They took the virus and forced it to live on five different plants (Endive, Lettuce, Wild Salsify, Field Marigold, and Zinnia) for several generations. Think of this as sending the virus to five different gyms to get in shape for specific sports.
• Step 2: Once the virus was "trained" on one plant, they took it and tried to infect every other plant. They did this hundreds of times to see which combinations worked and which failed.

2. The Problem: Too Much Data, Too Little Clarity

They ended up with a massive spreadsheet of "Success" and "Failure."

• Did the virus trained on Lettuce succeed on Zinnia? Yes.
• Did the virus trained on Zinnia succeed on Lettuce? Yes, but barely.
• Did the virus trained on Endive succeed on Field Marigold? No, not a chance.

The data was messy. Some jumps were easy, some were hard, and some were impossible. The scientists needed a way to simplify this chaos into a clear picture.

3. The Solution: The "Fitness Landscape" Map

Instead of just listing numbers, they used a mathematical concept called a Fitness Landscape. Imagine a 3D map where:

• Peaks represent a perfect fit (the virus is happy and reproduces like crazy).
• Valleys represent a bad fit (the virus dies out).
• Distance represents how different two plants are.

Usually, scientists can't see this map because it's too complex. But this team built a Bayesian "GPS" (a fancy computer algorithm) to infer the shape of this map just by looking at the success/failure data.

4. The Three Key Features of the Map

The computer revealed three important things about this viral world:

• A. The Distance (How far apart are the neighborhoods?):
  The map showed that the plants fall into two distinct clusters.

  • Cluster 1: Endive, Lettuce, and Salsify are like neighbors in the same suburb. They are genetically similar. If a virus gets good at one, it's already halfway to being good at the others.
  • Cluster 2: Field Marigold and Zinnia are like a different country entirely. They are far away on the map. Jumping from Cluster 1 to Cluster 2 is a huge, difficult leap.

• B. The "Door Width" (Permissiveness):
  Some neighborhoods have wide, open doors; others have tiny, locked gates.

  • Lettuce and Salsify have wide doors. Even if the virus isn't a perfect match, it can still squeeze in and survive. They are "permissive."
  • Field Marigold has a tiny door. The virus must be a perfect match to get in. If it's even slightly off, it gets rejected.

• C. The "Doorman's Efficiency" (Infection Efficiency):
  This is a subtle but crucial point. Sometimes, even if the virus gets through the door, the building might be hostile.

  • Field Marigold is picky (narrow door), but once the virus gets in, it's actually very easy for it to take over the building.
  • Lettuce has a wide door, but once inside, the virus has to work harder to establish itself.

5. The "Evolutionary Rescue" Concept

The paper also explains how a virus survives a jump it shouldn't be able to make.
Imagine a virus trying to enter a building with a locked door (a difficult host). The main virus tries, fails, and starts to die. But, by pure luck, a mutant (a random variation of the virus) appears that can open the door. This mutant saves the whole group from extinction. This is called Evolutionary Rescue.
The map showed that jumping to the "picky" plants (like Field Marigold) usually requires this kind of lucky rescue mutation, whereas jumping to the "permissive" plants (like Lettuce) is often easy enough that no rescue is needed.

6. Why Does This Matter?

This map isn't just about viruses; it's a tool for prediction.

• For Farmers: If you plant a mix of crops, you might think you are protecting them (dilution effect). But if you plant a "permissive" host that is close to a "difficult" host on the map, you might accidentally create a springboard. The virus gets comfortable on the easy plant, mutates, and then easily jumps to the hard one, causing a massive outbreak.
• For Medicine: The same logic applies to bacteria and antibiotics. If we understand the "fitness landscape" of drugs, we can design treatment schedules that force bacteria into a valley where they can't escape, rather than letting them climb a peak to become resistant.

The Bottom Line

The scientists turned a messy pile of infection data into a clear, visual map of the viral world. They showed us that the risk of a virus jumping to a new host depends on two things: how far apart the hosts are and how strict the new host's rules are. This helps us predict where the next epidemic might come from and how to stop it before it starts.

Technical Summary

1. Problem Statement

Understanding how pathogens adapt to multiple hosts is critical for predicting emergence, evolutionary rescue, and designing disease control strategies. While fitness landscapes are a powerful theoretical framework for describing adaptation, they are rarely inferred from empirical data in multi-environment settings.

• The Challenge: In most biological systems, dense genotype-to-fitness mapping is infeasible. Data is often sparse, consisting of binary outcomes (infection success/failure) across a limited set of host species.
• The Gap: Existing methods often fail to disentangle the geometric constraints of host shifts (phenotypic distance between hosts) from host-specific permissiveness (how tolerant a host is to phenotypic mismatch).
• Goal: To develop a mechanistic, Bayesian approach to infer a multi-host fitness landscape from cross-inoculation data, linking infection probabilities to the geometry of phenotypic optima and the strength of selection.

2. Methodology

The authors propose a novel framework combining Fisher's Geometrical Model (FGM) with Evolutionary Rescue (ER) theory under a Strong Selection Weak Mutation (SSWM) regime.

A. Theoretical Framework

1. Fisher's Geometrical Model (FGM):

   • Viral genotypes are mapped to an $n$-dimensional phenotypic space.
   • Each host $k$ corresponds to a quadratic fitness peak centered at a host-specific optimum $O_k$.
   • The growth rate $\rho_{jk}$ of a viral strain $j$ on host $k$ depends on the squared Euclidean distance between the strain's phenotype $x_j$ and the host optimum $O_k$:
     $$\rho_{jk} = \rho^{\max}_k \left( 1 - \frac{\|x_j - O_k\|^2}{R_k^2} \right)$$
   • $R_k$ represents the characteristic radius (width of the fitness peak), quantifying host permissiveness. Larger $R_k$ implies a broader viable region.

2. Evolutionary Rescue (ER) & Stochastic Dynamics:

   • The model treats the early infection phase as a Feller diffusion (branching process).
   • Successful infection occurs if the viral lineage avoids extinction. This can happen via:
     • Direct Establishment: The dominant strain is already within the viable region ($r_{jk} > 0$).
     • Evolutionary Rescue: The dominant strain is suboptimal or subcritical, but a beneficial mutant arises (either as standing variation in the source or de novo in the target) that establishes.
   • The probability of successful infection ($p_{jk}$) is derived analytically based on the distance between optima ($d_{jk}$), the target radius ($R_k$), mutation rates, and demographic noise.

3. Key Parameters Inferred:

   • Distance Matrix ($M_{jk}$): The Euclidean distances between host-specific phenotypic optima (including the ancestral clonal strain).
   • Permissiveness ($R_k$): The width of the fitness peak for each host.
   • Infection Efficiency ($q_k$): A composite parameter capturing how efficiently phenotypic suitability translates into successful infection (combining inoculum size and demographic stochasticity).
   • Mutation Rate ($U$): The genomic mutation rate.

B. Statistical Inference

• Data Source: Experimental evolution of Endive Necrotic Mosaic Virus (ENMV) on five Asteraceae hosts, followed by a full cross-inoculation design (5 sources $\times$ 5 targets $\times$ 8 lineages).
• Observation Model: A Beta-Binomial model is used to account for overdispersion (lineage-to-lineage heterogeneity) in infection outcomes, rather than a simple Binomial model.
• Inference Algorithm: Bayesian inference using Markov Chain Monte Carlo (MCMC) (Metropolis-Hastings sampler) to sample the joint posterior distribution of parameters.
• Visualization: Posterior distance matrices are embedded into 2D space using Classical Multidimensional Scaling (MDS) with Procrustes alignment to visualize the "phenotypic map" of the host community.

3. Key Results

A. Phenotypic Dimensionality

• Model comparison using WAIC (Widely Applicable Information Criterion) and LOO-CV indicates that $n=2$ (two phenotypic dimensions) provides the best compromise between predictive accuracy and parsimony.
• While $n=1$ captures some structure, $n=2$ significantly improves fit. $n=3$ offers negligible improvement, suggesting the host landscape is effectively planar.

B. Inferred Landscape Geometry

• Clustering: The inferred landscape reveals two distinct clusters of hosts:
  1. Cichorieae Cluster: Endive (CH), Lettuce (LA), and Wild Salsify (SA). These are phenotypically close, indicating high cross-compatibility.
  2. Distant Cluster: Field Marigold (SO) and Zinnia (ZI). These are phenotypically distant from the Cichorieae group, consistent with deeper phylogenetic splits (tribes Calenduleae/Heliantheae vs. Cichorieae).
• Phylogenetic Signal: The geometric distances broadly concord with host phylogeny, validating the approach's ability to recover biological structure from infection data.

C. Host Permissiveness and Asymmetry

• Heterogeneity in $R_k$:
  • Restrictive Hosts: SO and ZI have small radii ($R_k$), requiring a very close phenotypic match for infection.
  • Permissive Hosts: LA and SA have large radii (approaching the prior upper bound), indicating broad viable regions where direct establishment is easy.
• Asymmetry Drivers: The model successfully explains directional asymmetries in cross-inoculation (e.g., CH $\to$ SO fails, but SO $\to$ CH succeeds).
  • This is driven by the joint action of geometry ($d_{jk}$ vs. $R_k$) and efficiency ($q_k$).
  • SO is "narrow but efficient" (small $R_k$, large $q_k$): hard to infect unless matched, but highly successful once matched.
  • LA/SA are "broad but less efficient" (large $R_k$, small $q_k$): easy to enter, but establishment is less efficient even with a match.

D. Evolutionary Rescue Implications

• Adaptation to restrictive hosts (SO, ZI) from permissive sources (CH, LA, SA) typically requires evolutionary rescue (mutation) because the source optima lie outside the target's viable region ($d_{jk} > R_k$).
• Conversely, movement within the Cichorieae cluster rarely requires rescue, as optima lie within each other's viable regions.

4. Key Contributions

1. Methodological Innovation: Developed a parsimonious Bayesian method to infer effective fitness landscapes from sparse, binary multi-environment data, bridging the gap between theoretical FGM and empirical cross-inoculation assays.
2. Mechanistic Interpretation: Successfully separated geometric constraints (distance between optima) from target-specific permissiveness (peak width) and infection efficiency, providing a nuanced view of host shifts.
3. Empirical Application: Applied the method to a comprehensive experimental evolution dataset (ENMV), revealing a landscape structure that aligns with host phylogeny and explains complex directional infection patterns.
4. Generalizability: Demonstrated that the approach is applicable beyond plant viruses to any system where cross-environment performance can be measured (e.g., antibiotic resistance, temperature adaptation).

5. Significance and Implications

• Predicting Emergence: The inferred landscape allows researchers to identify "springboard hosts." Permissive hosts that are geometrically central or close to restrictive hosts can sustain pathogen populations and facilitate jumps to new, difficult-to-infect hosts.
• Disease Management: The framework offers a quantitative basis for designing crop mixtures. By selecting cultivars that create "rugged" landscapes (deep valleys between peaks), agricultural systems can impede pathogen adaptation and reduce the risk of evolutionary rescue.
• Evolutionary Rescue Theory: The study provides empirical validation for SSWM-based rescue approximations in viral systems and highlights the critical role of standing variation versus de novo mutation in host shifts.
• Beyond Metaphor: It moves fitness landscapes from a metaphorical concept to a quantifiable, inferable object that can guide experimental design and evolutionary forecasting.

In summary, this paper provides a robust statistical and mechanistic toolkit to visualize and quantify the "rules of engagement" between pathogens and multi-host communities, offering actionable insights for evolutionary biology and epidemiology.