Epistasis and the changing fitness landscapes of SARS-CoV-2

This study analyzes six years of SARS-CoV-2 genomic data to demonstrate that shifts in mutational fitness costs across evolving variants are largely driven by pairwise epistatic interactions between new mutations and the distinct genetic backgrounds of those variants.

The Gist

The Big Picture: A Viral Game of "Musical Chairs"

Imagine the SARS-CoV-2 virus is a massive game of Musical Chairs. The chairs are the human immune systems and the environment. To win (survive and spread), the virus needs to find a chair that fits it perfectly.

For the last few years, scientists have been watching this game by looking at millions of viral genomes (the virus's instruction manuals). They noticed something strange: The rules of the game keep changing.

A specific move that helps the virus win in one version of the game might make it lose in the next version. This paper tries to figure out why the rules change and how the virus's "genetic background" (the other mutations it already has) affects new mutations.

---

1. The Problem: Why does a "Good" move sometimes become "Bad"?

In the world of viruses, a mutation is just a tiny typo in the instruction manual. Sometimes a typo is helpful (it helps the virus hide from antibodies), sometimes it's harmful (it breaks the virus), and sometimes it doesn't matter.

Scientists used to think: "If a typo is helpful, it will always be helpful, no matter what."

But this paper shows that's not true. It's like playing a video game where a "Super Jump" power-up works great in Level 1, but in Level 2, the floor is made of lava, so "Super Jump" actually makes you fall in and lose.

The Analogy:
Think of the virus as a car.

• Mutation A is like adding a turbocharger.
• In a flat, empty desert (Variant 1), the turbocharger makes the car go super fast. It's a great upgrade!
• But in a narrow, winding mountain road (Variant 2), that same turbocharger makes the car spin out of control and crash. Now, the turbocharger is a terrible idea.

The paper asks: How do we predict when a "turbocharger" (mutation) will be good or bad based on the other parts of the car (the genetic background)?

2. The Solution: The "Genetic Neighborhood"

The authors looked at the massive database of virus sequences and realized that mutations don't happen in isolation. They happen in a neighborhood.

• Epistasis is the fancy scientific word for "neighbors influencing each other."
• If a virus has a mutation at position 100, it changes the environment for position 200.

The Analogy:
Imagine a crowded party.

• If you walk into the party alone, you might be able to dance freely.
• But if you walk in with a group of 50 people who are all dancing in a specific circle, your ability to dance depends entirely on where that circle is.
• If the circle moves (a new variant emerges), your "dance move" (mutation) might suddenly be perfect, or it might get you bumped into the wall.

The researchers found that when the virus evolves into a new "Variant" (like Omicron), it changes the "party layout." This changes how every other mutation feels.

3. The Method: Mapping the "Fitness Landscape"

The scientists built a mathematical model to map this out. They call it a Fitness Landscape.

• Imagine a mountain range:
  • High peaks = Good mutations (the virus is very fit).
  • Deep valleys = Bad mutations (the virus dies out).
  • Flat ground = Neutral mutations (doesn't matter).

In the past, scientists thought the mountain range was static. This paper shows that the mountains move.

When the virus changes its background (e.g., from Delta to Omicron), the whole map shifts. A peak in the Delta map might become a valley in the Omicron map.

How they did it:
They used a model called a Potts Model (think of it as a giant spreadsheet of connections). They asked: "If we change the background at spot X, how does that change the value of a mutation at spot Y?"

They found that:

1. Proximity matters: Mutations that are close to each other in the 3D structure of the virus (like neighbors in a house) influence each other the most.
2. One change ripples out: When the virus picks up one new mutation, it doesn't just change itself; it subtly changes the "value" of about 1 to 3 other spots nearby.

4. The Results: What did they find?

• The "Background" is King: The effect of a mutation depends heavily on what other mutations are already there. A mutation that is deadly in the "Delta" virus might be harmless in "Omicron" because Omicron has a different "background" that buffers the damage.
• Predictability: They found that about 50% of the changes in how mutations behave could be explained just by looking at which other mutations were present.
• The "Ripple Effect": Every time the virus changes its background (a "mismatch"), it reshapes the fitness landscape for a few other spots. It's like dropping a stone in a pond; the ripples change the water level for nearby stones.

5. Why Does This Matter?

This is crucial for the future of pandemic response.

• Predicting the Next Variant: If we know how the "party layout" changes, we can predict which mutations will be dangerous next.
• Vaccine Design: If a mutation is good for the virus now, but bad for the virus later because of how it interacts with other parts, we might be able to design vaccines that force the virus into a "bad" corner.
• Beyond SARS-CoV-2: This method can be used for any virus. It turns the chaotic mess of viral evolution into a map we can read.

Summary in One Sentence

This paper discovered that the virus is like a shapeshifting puzzle: a piece that fits perfectly in one version of the puzzle might not fit in the next, and by understanding how the pieces influence each other (epistasis), we can better predict how the virus will evolve and change its rules.

Technical Summary

1. Problem Statement

Since the emergence of SARS-CoV-2 in late 2019, millions of viral genomes have been sequenced, providing an unprecedented dataset for studying viral evolution. While previous studies successfully estimated the fitness effects of individual mutations by comparing observed mutation counts against neutral expectations, a critical gap remained: how do these fitness effects change as the viral genetic background evolves?

Specifically, the authors address the phenomenon of epistasis, where the fitness effect of a mutation depends on the presence of other mutations in the genome. As SARS-CoV-2 evolved through distinct variants (clades) separated by long evolutionary branches involving up to 50 mutations (e.g., the jump from Delta to Omicron), the "fitness landscape" likely shifted. The paper aims to:

• Quantify how the fitness costs/benefits of mutations change across different SARS-CoV-2 clades.
• Determine if these shifts are linked to specific genetic differences (mismatches) between variants.
• Model these interactions to predict how future mutations might behave in new genetic backgrounds.

2. Methodology

Data Source and Preprocessing

• Data: The study utilizes a pandemic-scale phylogenetic tree (UShER) containing millions of SARS-CoV-2 genomes (up to November 2024).
• Fitness Estimation: Building on the framework by Bloom and Neher (2023), the authors estimate the fitness effect ($\Delta f$) of amino acid mutations. This is done by comparing the number of independent occurrences of a mutation on the phylogenetic tree ($n_{obs}$) against the expected number under neutral evolution ($n_{pred}$).
• Clade Partitioning: The tree is partitioned into distinct clades (e.g., 21J/Delta, 21K/Omicron BA.1, 21L/Omicron BA.2). Fitness effects are estimated separately within each clade to establish a "local" fitness landscape.

Quantifying Epistatic Shifts

To measure how a mutation's effect changes between two clades ($a$ and $b$), the authors define a signed z-score ($z_{ab}$):
$$z_{ab}^i(\mu) = \frac{\Delta f_a^i(\mu) - \Delta f_b^i(\mu)}{\sqrt{s^2_{\Delta f_a} + s^2_{\Delta f_b}}}$$
where $\mu$ is the mutation and $s^2$ represents uncertainty. Large $|z|$ scores indicate significant shifts in fitness effects between clades.

The Epistatic Model (Generalized Potts Model)

The authors propose a pairwise epistatic model to explain these shifts. The fitness effect of a mutation $\sigma_i \to \sigma'_i$ in clade $a$ is modeled as:
$$\Delta f_a^i(\sigma_i \to \sigma'_i) = h_i(\sigma'_i) + \sum_{j \neq i} J_{ij}(\sigma'_i, \sigma_j^a)$$

• $h_i$: Intrinsic (additive) fitness effect.
• $J_{ij}$: Pairwise interaction (coupling) between residue $i$ and residue $j$.
• $\sigma_j^a$: The background amino acid at position $j$ in clade $a$.

The change in fitness effect between two clades ($\Delta \Delta f$) is expressed as the difference in interactions caused by background mismatches:
$$\Delta \Delta f_{ab} = \sum_{j: \sigma_j^a \neq \sigma_j^b} [J_{ij}(\sigma'_i, \sigma_j^b) - J_{ij}(\sigma'_i, \sigma_j^a)]$$

Inference and Regularization

Inferring the interaction matrix $J_{ij}$ is an underdetermined problem (more parameters than constraints). To solve this, the authors minimize an objective function using Elastic Net regularization:

1. Data Fidelity: Minimizes the difference between observed fitness shifts and model predictions.
2. Sparsity ($L_1$): Encourages most $J_{ij}$ to be zero (assuming most residues do not interact).
3. Distance-Based Prior: The regularization strength is modulated by the 3D spatial distance between residues ($d_{ij}$). Interactions between residues close in space are penalized less than distant ones. This utilizes structural data from PDB files and AlphaFold models.

3. Key Contributions

1. Systematic Mapping of Fitness Landscape Shifts: The study provides the first large-scale quantification of how mutational fitness effects change across the major SARS-CoV-2 variants (Delta through Omicron sub-lineages).
2. Identification of Epistatic Hotspots: The authors demonstrate that shifts in fitness effects are not random; they are significantly enriched near positions where the genetic backgrounds of clades differ (mismatches).
3. Development of a Predictive Epistatic Model: They introduce a sparse, distance-regularized pairwise interaction model that successfully explains approximately 50% of the variance in fitness effect shifts between clades.
4. Quantification of Interaction Density: The model suggests that each genetic mismatch between variants alters the fitness effects of approximately 1 to 3 additional positions in the protein.

4. Key Results

• Evidence of Epistasis: Comparing fitness effects between clades (e.g., Delta vs. Omicron BA.1) reveals significant discrepancies (long tails in z-score distributions) that cannot be explained by noise alone.
• Spatial Correlation: Sites with altered fitness costs are statistically closer to background mismatches than expected by chance. For example, the fitness effect of mutation P384L depends heavily on the amino acid at position 408 (Serine vs. Arginine).
• Model Performance:
  • The inferred coupling matrix $J_{ij}$ recapitulates known structural domains, particularly the Receptor Binding Domain (RBD) and N-Terminal Domain (NTD).
  • Key residues identified as interaction hubs include 371, 417, 452, and 486, which are known to be critical for ACE2 binding and immune escape.
  • Validation: The model was validated against independent Deep Mutational Scanning (DMS) data. While the correlation between predicted and experimental shifts is moderate ($r \approx 0.32$), the model successfully captures general trends (e.g., Delta being more tolerant to certain mutations than BA.1).
  • Predictive Power: When tested on novel clades (23I and 24A) not included in the training set, the model reduced the raw fitness discrepancy error by ~50% for 75% of mutations, demonstrating partial predictive capability for new backgrounds.

5. Significance and Implications

• Evolutionary Forecasting: The study demonstrates that fitness landscapes are dynamic, not static. Understanding epistasis is crucial for predicting which mutations will be beneficial in future variants. A mutation that is deleterious in one background might become beneficial in another due to compensatory interactions.
• Beyond DMS: While Deep Mutational Scanning (DMS) is powerful, it is often limited to specific variants (e.g., RBD of a single strain). This method leverages natural evolution data to infer interactions across the entire genome and across the entire pandemic timeline, complementing experimental data.
• Structural Insights: The inferred interactions align with 3D protein structures, validating that the statistical signal from sequence data reflects physical biological constraints.
• Future Applications: The framework can be applied to other rapidly evolving viruses. It highlights that as the virus accumulates mutations, the "rules" of evolution change, necessitating continuous re-evaluation of fitness landscapes to anticipate immune escape and transmissibility.

In summary, this paper bridges the gap between large-scale genomic surveillance and molecular biophysics, showing that the evolutionary trajectory of SARS-CoV-2 is heavily constrained and guided by a complex, shifting network of epistatic interactions.