Phase diagrams for biophysical fitness landscape design

⚕️

This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 gardener trying to grow a specific type of flower. In nature, evolution is like a wild storm: it randomly blows seeds around, and only the ones that survive the wind and rain grow. This is "survival of the fittest," but it's chaotic and hard to predict.

This paper introduces a new way to garden: Fitness Landscape Design (FLD). Instead of letting nature take its course, the scientists want to build a custom "mountain range" where they can decide exactly which flowers grow tall and which stay short.

Here is the breakdown of their discovery, using simple analogies:

1. The Landscape: A Mountain Range of Survival

Think of every possible version of a virus (or protein) as a different hiker standing on a giant, bumpy mountain range.

High peaks represent viruses that are very fit (they reproduce easily).
Deep valleys represent viruses that are weak or dead.
Evolution is the hiker trying to climb to the highest peak.

Usually, the shape of this mountain is fixed by nature. But the authors realized they could reshape the mountain using antibodies (the body's natural defense proteins) as their tools.

2. The Tools: Antibodies as "Gravity Modifiers"

Imagine you have a bag of magnets (antibodies).

If you dump a bunch of magnets on the mountain, they act like heavy weights.
If a virus (hiker) is a "bad fit" for the magnet, the magnet pulls it down into a deep valley (low fitness).
If a virus is a "good fit," it might not be pulled down as much.

By choosing which magnets to use and how many to use, you can sculpt the mountain. You can make a specific virus fall into a pit while letting another one climb a peak. This is Fitness Landscape Design.

3. The Problem: Can We Sculpt Anything We Want?

The big question the paper answers is: "How much control do we actually have?"

Can we make any virus weak while keeping any other virus strong? Or are there limits?

The "Designable" Zone: This is the area where you can successfully sculpt the mountain. You can find an antibody that crushes Virus A but leaves Virus B alone.
The "Undesignable" Zone: This is the area where it's impossible. If two viruses are too similar (like twins), you can't find a magnet that hurts one but not the other. They will always rise or fall together.

4. The Map: The Phase Diagram

The authors created a map (called a Phase Diagram) that shows the border between what is possible and what is impossible.

The Theory: They used math to draw this map, predicting exactly how big the "Designable" zone would be based on how different the viruses are and how many different antibodies they have in their toolbox.
The Analogy: Think of it like a map of a country. The blue area is "Buildable Land" where you can construct your custom fitness landscape. The red area is "Swamp Land" where you can't build anything because the ground is too similar everywhere.

5. The Experiment: Testing the Map in the Real World

For years, this was just a computer simulation. In this paper, they tested it with real data.

They took a massive dataset of 62,000 different antibodies and tested how they interacted with three different flu viruses.
They plotted the real results on their map.
The Result: The real-world dots landed exactly where the math predicted they would. The "Buildable Land" (blue) matched the reality perfectly.

6. Why This Matters

This is a game-changer for fighting diseases like the flu or even cancer.

The "Outlier" Antibodies: The map showed that sometimes, there are rare, super-special antibodies (the "outliers") that can push the boundaries of what's possible. These are the "magic wands" that can suppress a dangerous virus mutant that usually escapes our defenses.
The Future: Instead of waiting for a virus to mutate and escape our vaccines, we can use this "landscape design" to proactively build a mountain where the virus cannot survive. We can engineer the environment so that the virus is forced to evolve into a dead end.

Summary

In short, the authors have drawn a blueprint for a custom evolution simulator. They proved mathematically and experimentally that we can use antibodies to reshape the rules of survival for viruses. If we have the right tools (antibodies) and the right map (the phase diagram), we can design a world where the "bad guys" (viruses) have nowhere to hide.

1. Problem Statement

Evolutionary biology traditionally views populations as moving stochastically toward peaks on a "fitness landscape," driven by natural selection and mutation. While the concept of Fitness Landscape Design (FLD) was recently proposed to actively reshape these landscapes using tunable control parameters (specifically antibody sequences and concentrations), previous evidence for its feasibility was limited to numerical simulations.

The core challenges addressed in this work are:

Lack of Analytical Theory: There was no explicit mathematical framework to predict the boundaries of "designable" fitness assignments (i.e., which combinations of protein fitnesses can be achieved by tuning antibodies).
Experimental Validation: It was unknown whether the theoretical limits of FLD hold true in real biological systems with complex, non-ideal binding affinities.
Quantifying Designability: A rigorous method was needed to define the "designable region" in the phase space of two protein sequences and understand how factors like antibody library size and sequence similarity affect this region.

2. Methodology

The authors employed a three-pronged approach combining statistical mechanics, extreme value theory, and large-scale experimental data analysis.

A. Analytical Theory Development

Biophysical Fitness Model: The authors utilized a statistical mechanics-based model derived from chemical reaction kinetics. Viral fitness $F(s)$ is modeled as a function of host receptor binding ( $\Delta G_H$ ) and antibody binding ( $\Delta G_{Ab}$ ), where antibodies act as a suppressive force.
$F(s) \approx \frac{[H]e^{-\beta\Delta G_H(s)}}{C_0 + [H]e^{-\beta\Delta G_H(s)} + [Ab]e^{-\beta\Delta G_{Ab}(s)}}$
Two-Sequence, One-Antibody Limit: To make the problem analytically tractable, they modeled a system with two target protein sequences ( $s_1, s_2$ ) and a single tunable antibody. They derived the relationship between the fitness of sequence 2 ( $F_2$ ) given a fixed fitness of sequence 1 ( $F_1$ ).
Extreme Value Statistics (EVT): Recognizing that the "designability frontier" is determined by the most selective antibodies in a library (those with the largest difference in binding free energy, $\Delta = \Delta G_{Ab}(s_1) - \Delta G_{Ab}(s_2)$ $Δ = Δ G_{A b} (s_{1}) - Δ G_{A b} (s_{2})$ ), they applied EVT.
- They assumed the binding free energies follow a bivariate Gaussian distribution.
- They used the Random Energy Model (REM) analogy to estimate the ground state energy (the maximum/minimum $\Delta$ ) for a library of size $M$ .
- This yielded closed-form expressions for the upper and lower bounds of the designable region, $F_2^+(F_1)$ and $F_2^-(F_1)$ .

B. Experimental Data Integration

Dataset: The authors utilized a recently published dataset (Ref [27]) containing binding affinity measurements ( $K_d$ ) for 65,536 combinatorial variants of the human CH65 antibody against three influenza H1 glycoprotein antigens (MA90, SI06, and G189E).
Processing: They converted $K_d$ values to binding free energy differences ( $\Delta$ ) and mapped the experimental data onto the theoretical phase diagrams.
Handling Non-Gaussianity: Acknowledging that real experimental data often deviates from ideal Gaussian distributions (due to evolutionary selection and detection limits), they developed a semi-empirical approach using Generalized Pareto Distributions (GPD) to fit the tails of the $\Delta$ distribution. This allowed for more accurate boundary predictions without assuming a perfect Gaussian distribution.

C. Scalability Analysis

They performed a downsampling study, reducing the antibody library to 10% of its size (approx. 6,500 variants) and using GPD tail fitting to predict the phase boundaries of the full dataset. This tested the feasibility of FLD design with limited experimental resources.

3. Key Contributions

Analytical Phase Diagrams: The derivation of explicit analytical equations (Eq. 9) defining the designable region in the fitness phase space. This region represents the set of fitness pairs $(F_1, F_2)$ achievable by varying antibody concentration and sequence.
Codesignability Score (CDS): The introduction of a quantitative metric, the Codesignability Score, defined as the area of the designable region. The authors proved analytically that CDS decreases monotonically with the Pearson correlation ( $r_{12}$ ) between the binding affinities of the two antigens. This mathematically confirms that chemically similar antigens are harder to distinguish (design) independently.
Experimental Validation: The construction of the first experimental FLD phase diagrams using real-world binding affinity data, demonstrating close agreement with theoretical predictions.
Tail-Fitting Scalability: The demonstration that accurate phase boundaries can be predicted from a small subset (10%) of an antibody library by fitting the tails of the distribution using GPD, significantly reducing the experimental burden for future FLD applications.

4. Results

Agreement with Theory: The experimental data points (colored lines in Fig. 3) closely followed the theoretical boundaries derived from the Gaussian approximation and the GPD tail-fitting method.
Designable vs. Undesignable Regions: The phase diagrams clearly delineated a "designable" region (blue) where fitnesses can be independently tuned, and an "undesignable" region (red) where no antibody configuration can achieve the desired fitness pair.
Role of Outliers: The boundaries were found to be determined by "outlier" antibodies—specific variants that bind one antigen significantly more strongly than the other. These outliers define the sharp corners of the designable region.
Impact of Correlation and Library Size:
- Library Size ( $M$ ): Increasing the antibody library size expands the designable region (more degrees of freedom).
- Correlation ( $r_{12}$ ): Higher correlation (chemically similar antigens) shrinks the designable region, making it difficult to suppress one strain without suppressing the other.
Scalability: The downsampling experiments showed that median phase boundaries predicted from 10% of the data were often tighter and more accurate than the Gaussian approximation of the full data, provided the tails were fitted correctly. However, rare outliers in the full dataset were sometimes missed in the 10% samples, leading to slight underestimations of the designable frontier in specific trials.

5. Significance

Foundational Theory: This work moves FLD from a purely numerical concept to a rigorous analytical framework, providing tight bounds on what is physically possible in protein evolution engineering.
Proactive Pandemic Preparedness: The ability to design fitness landscapes allows for the engineering of "proactive" antibodies that can simultaneously suppress wild-type viruses and their escape mutants, effectively trapping viral evolution.
Broader Applications: The principles extend beyond virology to cancer therapeutics (e.g., engineering CAR-T cells to prevent immune escape) and directed evolution (e.g., phage display), offering a toolkit to engineer custom fitness landscapes for laboratory evolution.
Experimental Efficiency: The demonstration that tail-fitting on downsampled datasets can predict full landscape boundaries suggests that FLD can be implemented efficiently in the lab without needing to screen millions of antibody variants, making the approach scalable for practical applications.

In summary, the paper establishes that biophysical fitness landscapes are not just passive backdrops for evolution but can be actively and quantitatively engineered, with the theoretical limits of this engineering now rigorously defined and experimentally verified.