🧬 biology

Fitness inference tested by in silico population genetics

This paper presents a framework for determining the feasibility of inferring fitness parameters and genotype fitness order from time-stratified, whole-genome population data by testing these methods in simulated populations to identify both viable and non-viable parameter ranges.

Original authors: Hong-Li Zeng, Yu-Han Huang, John Barton, Erik Aurell

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hong-Li Zeng, Yu-Han Huang, John Barton, Erik Aurell

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a detective trying to figure out who the "superstars" are in a massive, chaotic crowd of 1,000 tiny, fast-reproducing organisms. You have a time-lapse video of this crowd evolving over 30 generations. Some organisms have tiny genetic changes (mutations), some swap parts with neighbors (recombination), and some just get lucky or unlucky by pure chance (genetic drift). Your goal? To look at the video and guess which specific genetic combinations are the absolute best at surviving and reproducing. This is called fitness inference.

The paper by Zeng, Huang, Barton, and Aurell asks a simple but tricky question: Can we actually solve this mystery just by watching the time-lapse?

To find out, the authors didn't look at real bugs in a lab; they built a digital playground (a computer simulation) where they knew the exact "score" (fitness) of every single organism from the start. This is like a video game where the developers know the winning strategy, and they are testing if a player can figure it out just by watching the gameplay.

The Two Detective Tools

The researchers tested two different detective methods to see which one could best guess the winners:

The "Marginal Path Likelihood" (MPL) Detective: This method is like a strict accountant. It looks at how often specific genetic changes appear and disappear over time. It assumes the crowd is small enough that random luck (genetic drift) matters a lot. It focuses on the "additive" effects—basically, it thinks the total score is just the sum of individual good or bad moves. It doesn't really care if two moves work well together; it just adds them up.
The "Transient Quasi-Linkage Equilibrium" (tQLE) Detective: This method is more like a social network analyst. It assumes the crowd is so huge that random luck barely matters. It looks for "teamwork" between genes (called epistasis). It asks, "Do these two specific mutations work better when they are friends?" It uses a fancy math formula (Gibbs-Boltzmann distribution) to guess how genes interact.

The Big Test: Who Wins?

The authors ran their digital simulation with different settings to see when each detective shines and when they fail. Here is what they found in their simulations:

When the game is simple (Additive only): If the organisms' success depends only on individual moves (no teamwork), both detectives are great. They can accurately guess the fitness of the top performers. However, if the mutations are very weak and the "noise" (mutation rate) is high, both detectives get confused and can't tell the winners from the losers.
When the game is complex (Teamwork matters): If the success depends on genes working together (epistasis), the tQLE detective is better at spotting the true winners, especially the very top 5%. The MPL detective, which ignores teamwork, starts to lose its way.
When the crowd is small (Genetic Drift is strong): If the population is small (like their simulated 1,000 individuals), random chance plays a huge role. Here, the MPL detective wins. The tQLE detective, which assumes a massive crowd where luck doesn't matter, gets it wrong because it ignores the chaos of the small group.

The "Rank Order" Surprise

Here is the most interesting part: The authors found that even though the two detectives use totally different math and make different assumptions, they often agree on who the top 5% winners are.

In many of their simulations, especially when the genetic signals were strong, both methods pointed to the same "superstar" sequences. This is a big deal because, in the real world, we often don't know the "ground truth" (who actually won). But in their simulated world, they proved that if you just want to know "Who are the top 5%?", you can use either method, and you'll likely get the same answer.

What They Explicitly Rule Out

The paper is very careful to say what they did not prove:

They did not prove that these methods work perfectly for every real-world virus or bacteria. They only tested them in simulations with specific settings (1,000 individuals, 25 genetic sites, 30 generations).
They ruled out the idea that you can always get the exact numerical fitness score (the precise number) for every organism. The methods are better at getting the order (who is #1, #2, #3) than the exact point values.
They ruled out the idea that epistatic fitness (teamwork scores) is passed down perfectly from parent to offspring in a high-recombination environment. Their simulations showed that when organisms swap genes frequently, the "teamwork" scores get scrambled and aren't heritable, while the "individual move" scores are.

The Bottom Line

The authors conclude that fitness inference is possible using these methods, but it depends on the situation.

If you have a small population with lots of random noise, use the MPL method.
If you have a huge population where genes team up, use the tQLE method.
If you just want to find the top 5% winners in a wide range of scenarios, both methods work surprisingly well and usually agree with each other.

This work doesn't solve the mystery of evolution for all time, but it gives scientists a map showing exactly when their detective tools will work and when they will get lost. It turns the "unknown unknowns" of evolution into a solvable puzzle, at least within the bounds of their computer simulations.

Technical Summary: Fitness Inference Tested by In Silico Population Genetics

Problem Statement
The paper addresses the challenge of inferring fitness parameters and the relative fitness order of genotypes from time-series genetic data (whole-genome sequences sampled at multiple time points). While such data is increasingly available for microbes, ancient DNA, and human pathogens, disentangling deterministic selection signals from stochastic noise (genetic drift), mutation, and recombination remains technically difficult. The authors specifically investigate whether it is possible to recover fitness landscapes from population-wide data and delineate the parameter ranges where such inference is feasible versus where it fails.

Methodology
The study employs a competitive evaluation of two distinct fitness inference schemes using in silico simulations generated via the FFPopSim toolbox. The simulations model haploid populations ( $N=1000$ ) with genome length $L=25$ evolving over $T=30$ generations under natural selection, mutation, recombination ( $r=0.5$ ), and genetic drift. Fitness is modeled as a quadratic function of binary genotypes, including additive effects ( $f_i$ ) and pairwise epistatic interactions ( $f_{ij}$ ).

The two methods compared are:

Marginal Path Likelihood (MPL): Based on Kimura's diffusion approximation and the Wright-Fisher model. It assumes finite populations where genetic drift is significant. MPL relies on maximizing the path probability of observed allele frequency trajectories. Crucially, the standard MPL implementation considered here assumes a purely additive fitness landscape and does not explicitly model epistasis, treating non-zero epistatic coefficients as confounding factors.
Transient Quasi-Linkage Equilibrium (tQLE): Based on the assumption that the population state transiently follows a Gibbs-Boltzmann distribution. This method is designed for infinite populations (ignoring genetic drift) and explicitly incorporates both additive and epistatic fitness parameters. It utilizes relaxation equations to infer Potts model parameters ( $h_i, J_{ij}$ ) which are then mapped back to fitness coefficients.

The authors tested these methods across various parameter regimes, varying mutation rates ( $\mu$ ) and the variance of additive ( $\sigma(f_i)$ ) and epistatic ( $\sigma(f_{ij})$ ) fitness effects. The primary evaluation metric was not just the numerical accuracy of inferred coefficients, but the rank order of genotypes by fitness, particularly focusing on the top 5% of the fittest sequences.

Key Results

Additive Landscapes: In purely additive scenarios, both MPL and tQLE accurately inferred fitness effects. Inference accuracy improved when selection and mutation were weaker. However, both methods tended to underestimate the magnitude of selection for the most impactful mutations under strong selection, though they maintained high accuracy in relative ranking.
Epistatic Landscapes:
- Low Epistasis/High Additive Variance: As the additive component of fitness increased, MPL outperformed tQLE in global rank correlation. This is attributed to MPL's ability to handle finite population noise (genetic drift), whereas tQLE assumes an infinite population.
- High Epistasis/Low Additive Variance: When epistatic contributions dominated, tQLE outperformed MPL. MPL, lacking an epistatic term, struggled to correctly order genotypes where interactions were the primary driver of fitness.
- Top-Ranked Sequences: A critical finding is that tQLE remained a superior predictor for the rank order of the top 5% of fittest sequences across a wider range of $\sigma(f_i)$ compared to MPL. This suggests that epistatic interactions play a crucial role in sorting the most fit individuals, even when additive effects are present.
Parameter Recovery: While tQLE could recover epistatic parameters, its accuracy in inferring additive parameters degraded as the variance of additive effects increased. Conversely, MPL's performance in inferring additive effects improved with stronger additive signals.
Replicates: Performance gains from using multiple simulation replicates were substantial for tQLE (saturating around 5 replicates) but minimal for MPL, which was relatively insensitive to the number of replicates.

Significance and Claims
The authors claim that fitness inference from time-stratified, whole-genome data is feasible within specific parameter ranges. The study demonstrates that:

Complementarity: The two methods are complementary. MPL is robust in finite populations with significant genetic drift, while tQLE is superior when epistatic interactions are strong or when the focus is on identifying the absolute fittest genotypes.
Convergence: Despite differing theoretical foundations (finite vs. infinite population assumptions, additive vs. epistatic models), the two methods often agree on fitness rankings over broad, reasonable parameter ranges.
Evaluation Criteria: The paper emphasizes that the choice of evaluation criterion is critical. While numerical recovery of parameters may vary, the ranking of top-performing sequences is a more robust and biologically relevant metric, where the methods show significant agreement.
Heritability of Fitness: The results support the theoretical prediction that additive fitness is heritable, whereas epistatic fitness is not heritable in the presence of high recombination rates, as evidenced by the stability of epistatic fitness distributions in the simulations.

The work provides a framework for determining when fitness inference is possible from real-world data (such as pathogen surveillance or ancient DNA) and highlights settings where inference is likely to fail or require specific methodological choices. The authors conclude that for real data applications, the choice between MPL and tQLE is largely a matter of convenience depending on the specific biological context (e.g., population size and the presence of epistasis).