Counting-based inference of mutant growth rates from… — Plain-Language Explanation

⚕️

This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 which of thousands of different runners are the fastest, but you can't watch them race individually. Instead, you have a giant, crowded stadium where all the runners are mixed together in a single pack. You can only take a quick snapshot of the crowd at the start and a few snapshots later on.

Your goal is to count how many runners of each "team" (genetic variant) are in the crowd at each snapshot. If Team A's numbers are growing faster than Team B's, Team A is the winner.

This paper is about building a better mathematical camera to analyze these snapshots. The authors, Deniz Sezer and Erdal Toprak, argue that the old ways of doing the math are like using a blurry lens or a broken ruler. They propose a new, sharper way to calculate exactly how fast each team is growing, even when the race gets crowded and slows down.

Here is the breakdown of their ideas using simple analogies:

1. The Problem: The "Blurry" Old Lens

In the past, scientists looked at these snapshots and tried to draw a straight line through the dots to see who was winning.

The Issue: This is like trying to predict a marathon runner's speed by only looking at the start and finish line, assuming they ran at a perfectly constant speed the whole time.
The Reality: In a crowded stadium, runners bump into each other. Eventually, the track gets so full that no one can run faster, regardless of how fit they are. The old math didn't account for this "traffic jam" (saturation).
The Noise: Counting thousands of runners in a photo is hard. Sometimes you miss a few, or you count a shadow as a person. The old math treated these counting errors as if they were random static, but in a crowd, if you count one person too many, you must have counted someone else too few. The old math missed this connection.

2. The New Tool: The "Smart" Softmax Camera

The authors introduce a new mathematical trick called the Softmax transformation.

The Analogy: Imagine the runners are not just numbers, but slices of a giant pizza. The whole pizza must always equal 100%. If the pepperoni slice gets bigger, the mushroom slice must get smaller.
The Magic: The old math tried to measure the pepperoni slice in isolation. The new math (Softmax) looks at the entire pizza at once. It understands that the slices are connected. This allows the math to handle the "traffic jam" where the pizza stops growing because the oven is full.

3. Three Ways to Solve the Mystery

The paper compares three different detective methods to find the growth rates:

A. Weighted Least-Squares (The "Ruler" Method)

How it works: You draw a line through the data points and try to make the distance between the line and the points as small as possible.
The Flaw: The authors found that if you use a straight line (linear), your answer changes depending on which runner you pick as your "reference" (like picking a specific runner to measure everyone else against). It's like saying "Runner A is 5 minutes faster than Runner B," but if you pick Runner C as the reference, the math gets messy.
The Fix: They showed that using a curved line (the Softmax curve) is much better. It gives the same answer no matter which runner you use as a reference.

B. Maximum Likelihood (The "Best Guess" Method)

How it works: Instead of just drawing a line, this method asks: "If the runners were actually moving at these specific speeds, how likely is it that we would have seen the exact crowd sizes we photographed?"
The Advantage: It uses the entire history of the race (all the snapshots) at once, rather than just connecting the dots. It naturally handles the "traffic jam" and the counting errors without needing to force a straight line. It's like a super-smart AI that simulates millions of races to see which one looks most like your photos.

C. Variational Bayesian Inference (The "Confidence Meter")

How it works: This is the most advanced method. It doesn't just give you a single number for the speed (e.g., "5 mph"). It gives you a range of confidence (e.g., "5 mph, plus or minus 0.2 mph").
The Analogy: Imagine a weather forecast. The old methods just said "It will rain." This new method says, "There is a 90% chance of rain, and if it does, it will likely be a heavy downpour."
Why it matters: If a runner is very rare in the crowd, it's hard to count them accurately. This method tells you, "We are very unsure about this runner's speed because the data is noisy," whereas the old methods might have given you a confident but wrong answer.

4. The Big Picture: From "Exponential" to "Real Life"

Most old math assumes runners keep speeding up forever (Exponential Growth). But in real life, resources run out, and growth slows down (Logistic or Gompertz Growth).

The Breakthrough: The authors showed that their new "Smart Camera" (the inference framework) can handle any type of race. Whether the runners speed up forever or hit a traffic jam, the math works.
The Future: This means scientists can now use these crowded race experiments to figure out not just who is fast, but why they are fast. They can estimate the tiny, invisible biochemical "engines" inside the cells that make them run, all by just counting them in a crowd.

Summary

The authors took a messy, crowded problem (counting thousands of genetic variants in a petri dish) and built a new mathematical engine to solve it.

They stopped using straight lines and started using curves that respect the "pizza slice" nature of the data.
They moved from simple guessing to "best guess" simulations that use all the data at once.
They added a "confidence meter" so we know how much to trust the results.
They made the system flexible enough to handle real-world limits (like running out of food), not just idealized infinite growth.

This allows scientists to read the "fitness landscape" of life with much higher precision, turning a blurry crowd photo into a crystal-clear movie of evolution in action.

1. Problem Statement

High-throughput pooled competition assays, combined with deep sequencing, allow researchers to track the relative abundances of thousands of genetic variants simultaneously. While identifying "winners" (enriched variants) is common, quantitatively inferring the specific growth rates of all variants is challenging due to:

Counting Noise: Sequencing data consists of discrete read counts subject to stochastic noise.
Compositional Constraints: Variant fractions must sum to one; an increase in one variant's fraction necessitates a decrease in others.
Growth Model Limitations: Most existing methods assume exponential growth and rely on linear regression of log-transformed data. This assumption fails in batch cultures where nutrient depletion leads to saturation (logistic or Gompertz growth).
Reference Dependence: Many current methods (e.g., Enrich2) rely on linear regression of log-ratios against a reference variant, making results sensitive to the choice of reference and prone to bias when reference counts are low or zero.
Uncertainty Quantification: Standard least-squares approaches often fail to provide rigorous uncertainty estimates for growth rates, particularly when data is sparse or noisy.

2. Methodology

The authors propose a unified statistical framework that separates the probabilistic model of counting noise from the deterministic model of variant growth, linked via a softmax transformation.

A. Probabilistic Model: Multinomial Noise & Softmax Reparametrization

Noise Model: Instead of assuming independent Poisson or Negative Binomial distributions for each variant, the authors model sequencing reads as draws from a Multinomial distribution. This inherently respects the compositional constraint (fractions sum to 1) and captures negative covariances between variants.
Softmax Transformation: The variant fractions ( $f_k$ $f_{k}$ ) are expressed as a softmax transformation of the log abundances ( $y_k = \log N_k$ $y_{k} = lo g N_{k}$ ):
$f_k = \frac{e^{y_k}}{\sum_i e^{y_i}}$
This reparametrization is crucial because:
1. It allows the log abundances to be modeled as linear functions of time (for exponential growth) without needing to explicitly estimate the mean population growth rate as an intermediate step.
2. It naturally handles the compositional nature of the data.

B. Inference Approaches

The paper evaluates and compares three inference strategies:

Weighted Least-Squares (WLS):
- Revisits the approach used in tools like Enrich2.
- Derives weights based on the variance of a Dirichlet posterior (Bayesian analysis of fractions).
- Proposes a non-linear softmax fit (fitting the growth model directly to fractions) rather than a linear fit to log-ratios. This avoids the "reference variant" bias and handles zero counts more robustly.
Maximum Likelihood Estimation (MLE):
- Formulates the problem as maximizing the likelihood of the multinomial counts given the growth parameters.
- All-Time Integration: Unlike WLS which aggregates point estimates, MLE integrates information across all time points simultaneously.
- Two-Time Point Case: Derives a closed-form MLE for growth rates when only start ( $t_0$ ) and end ( $t_1$ ) data are available. The resulting estimator depends only on the counts of the variant itself, not a reference variant, resolving ambiguity in uncertainty estimation.
Variational Bayesian Inference (VI):
- Extends MLE to a Bayesian framework to quantify uncertainties.
- Uses Variational Inference with a Gaussian posterior approximation for the log abundances and growth parameters.
- ELBO Optimization: Maximizes the Evidence Lower Bound (ELBO). The authors derive an analytical lower bound (using Jensen's inequality) for closed-form solutions in simple cases and use Monte Carlo sampling for complex cases.
- Parameter Counting: Addresses the challenge of inferring $2K$ parameters (means and variances for $K$ variants) from $K$ observations by showing that means encode relative proportions while variances encode absolute count magnitudes (precision).

C. Generalization to Non-Exponential Growth

The framework is extended to Logistic and Gompertz growth models.
Since analytical solutions for these models are generally unavailable for $K > 1$ , the authors use numerical integration (Euler method) combined with automatic differentiation to compute gradients of the likelihood/ELBO with respect to growth parameters.
This allows the inference of absolute growth rates (not just relative) because the saturation point (carrying capacity) fixes the time scale.

3. Key Contributions

Softmax Parametrization: Demonstrates that using softmax to link multinomial counts to growth dynamics eliminates the need to estimate the mean population growth rate, simplifying the inference of relative and absolute rates.
Superiority of Non-Linear Fitting: Shows that fitting a softmax model to fractions is statistically superior to linear regression of log-ratios, particularly regarding reference-variant independence and handling of zero counts.
Closed-Form Two-Point Estimators: Derives new, reference-free maximum likelihood and variational Bayesian estimators for experiments with only two time points, including explicit formulas for uncertainty (variance).
Unified Framework for Saturation: Successfully extends high-throughput growth rate inference to non-exponential regimes (Logistic/Gompertz) using automatic differentiation, enabling the estimation of absolute kinetic parameters.
Uncertainty Quantification: Provides a rigorous method to estimate uncertainties in growth rates, showing that uncertainty scales inversely with read counts and is sensitive to the number of time points.

4. Results

Synthetic Data Validation: Using synthetic data for 4 and 100 variants under Exponential, Logistic, and Gompertz growth:
- WLS: The proposed softmax-based WLS fit produced results invariant to the choice of reference variant, whereas linear log-ratio fits varied significantly.
- MLE vs. WLS: MLE provided more accurate growth rate estimates by naturally weighting time points based on the multinomial noise model, rather than arbitrary weighting schemes.
- VI Performance: Variational inference successfully recovered true growth rates and provided accurate uncertainty estimates (posterior standard deviations). The analytical lower bound to the ELBO provided closed-form solutions but slightly underestimated uncertainties for highly abundant variants compared to full Monte Carlo VI.
- Scalability: The method scaled efficiently to $K=100$ variants, with optimization times remaining low (seconds to minutes) even with numerical integration of non-linear dynamics.
Two-Time Point Analysis: Confirmed that for $T=2$ , the standard relative growth rate formula is the MLE, but the new estimator (independent of reference) yields better-defined uncertainties.

5. Significance

Beyond Exponential Growth: The ability to incorporate arbitrary growth models (Logistic, Gompertz) allows researchers to analyze pooled competition assays in standard batch cultures (where saturation occurs) without needing specialized turbidostats to maintain exponential growth.
Mechanistic Insight: By decoupling the inference framework from specific growth laws, the authors pave the way for high-throughput inference of microscopic biochemical parameters (e.g., $V_{max}$ , $K_M$ of enzymes) directly from sequencing data.
Robustness: The framework is robust to zero counts and low-abundance variants, which are common in large-scale screens, by properly modeling the compositional noise structure.
Practical Utility: The derivation of closed-form estimators for two-time-point experiments makes the method immediately applicable to the vast majority of existing datasets that only sequence the beginning and end of competitions.

In summary, this paper provides a mathematically rigorous, flexible, and computationally efficient framework for extracting precise growth rates and their uncertainties from pooled sequencing data, moving beyond the limitations of linear regression and exponential-only assumptions.

Counting-based inference of mutant growth rates from pooled sequencing across growth regimes