Characterizing Physicochemical Selection in Protein Evolution with Property-Informed Models (PRIME)

The paper introduces PRIME, a novel framework of codon-level maximum likelihood models that explicitly links amino acid substitution patterns to physicochemical properties, thereby revealing the mechanistic biophysical drivers of protein evolution and offering superior interpretability and predictive power compared to traditional rate-based methods.

The Gist

Imagine you are trying to understand why a specific car engine has changed its design over the last 50 years.

The Old Way (Standard Models):
Traditional scientists look at the engine and say, "Wow, this part changed a lot! It must be very important for speed." They count how many times a part was swapped out. If a part changed often, they assume it's under "positive selection" (evolving fast). If it never changed, they assume it's "conserved" (too important to touch).

But this approach has a blind spot. It tells you that something changed, but not why. Did they swap a heavy steel bolt for a light aluminum one to save weight? Did they change the color for style? Did they swap a round gear for a square one? The old models just see "Swap happened" and stop there. They don't understand the physics of the change.

The New Way (PRIME):
This paper introduces PRIME (Property Informed Models of Evolution). Think of PRIME as a detective who doesn't just count the swaps; they look at the properties of the parts being swapped.

PRIME asks:

• Was the new part bigger or smaller? (Volume)
• Is it greasy/oily or water-loving? (Hydrophobicity)
• Is it electrically charged? (Isoelectric Point)
• Does it like to form spirals or flat sheets? (Secondary Structure)

The Three Detectives in the PRIME Team

The authors built three versions of this detective to solve different mysteries:

1. G-PRIME (The Global Detective):

   • Job: Looks at the whole protein (the whole engine) and asks, "What are the general rules for this machine?"
   • Analogy: It's like looking at a whole city and saying, "In this city, everyone drives small, fuel-efficient cars." It gives you the big picture of what the protein generally cares about (e.g., "This protein really hates big, bulky parts").

2. E-PRIME (The Episodic Detective):

   • Job: Looks at specific moments in history (branches of the family tree) to see if the rules changed temporarily.
   • Analogy: Imagine a family that usually drives sedans, but during a specific war, they all switched to heavy trucks for a few years, then went back to sedans. E-PRIME catches that temporary shift. It finds moments where the protein had to break the usual rules to survive a crisis (like a virus trying to escape a vaccine).

3. S-PRIME (The Site-Specific Detective):

   • Job: Zooms in on a single screw or bolt (a single amino acid) to see exactly what property is being guarded or changed.
   • Analogy: This is the most powerful tool. It looks at one specific spot and says, "This spot is allowed to change color, but it is strictly forbidden to change its size." It reveals the hidden "biophysical code" that standard models miss.

What Did They Discover?

By applying these detectives to thousands of proteins (from viruses to humans), they found some fascinating patterns:

• The "Core" is Rigid: The inside of the protein (the engine block) is like a fortress. It is obsessed with size and greasiness. If you try to put a giant part in a small hole, or a wet part in a dry hole, the protein breaks. These rules never change.
• The "Surface" is Flexible: The outside of the protein (the paint job and the bumper) is where the fun happens. This is where proteins tweak their electric charge or spiral shape to adapt.
  • Example: Viruses often change the "charge" on their surface to trick our immune system, while keeping their core structure intact.
• The "Alpha-Helix" is the Playground: The paper found that the tendency to form spirals (alpha-helices) is the most common thing to change. It's like the protein's way of tuning its "knobs" to fit new environments without breaking the engine.

Why Does This Matter?

1. It explains the "Why":
Instead of just saying "This virus is evolving fast," PRIME says, "This virus is evolving fast because it needs to change its electric charge to dodge antibodies, but it must keep its size the same to fit inside the cell."

2. It finds hidden secrets:
Sometimes a part of a protein looks like it's doing nothing (it's not changing fast). Standard models ignore it. But PRIME might find that while the identity of the amino acid changes, its hydrophobicity (greasiness) stays exactly the same. This reveals a "cryptic constraint"—a hidden rule that was invisible before.

3. It connects to AI:
The authors compared their "physics-based" rules to modern AI (like the protein language models used by AlphaFold). They found that the AI, which is a "black box" that just guesses the next letter in a sequence, is actually learning these exact same physical rules (size, charge, greasiness) without being explicitly told to. PRIME helps us open the black box and see why the AI makes its predictions.

The Bottom Line

Think of protein evolution not as a random game of "Hot Potato" where parts are swapped randomly. Instead, it's a highly regulated construction site.

PRIME is the blueprint that finally explains the building codes. It tells us that while the workers (mutations) are constantly changing the bricks, they are strictly following rules about the bricks' weight, texture, and electrical charge. By understanding these rules, we can better predict how viruses will evolve, how drugs might fail, and how life adapts to new challenges.

Technical Summary

1. Problem Statement

Standard probabilistic models of protein evolution (e.g., codon substitution models like MG94) are highly effective at identifying where and when natural selection acts (via metrics like $dN/dS$ or $\omega$). However, they are agnostic to the mechanistic realization of these forces. They treat amino acid substitutions as abstract token exchanges without explaining the underlying biophysical rules (e.g., why a specific change in hydrophobicity or volume is selected for or against).

Existing attempts to integrate physicochemical realism (e.g., Grantham distances, CoRa models) have failed to achieve widespread adoption due to a lack of intuitive appeal, computational scalability, or the inability to significantly improve biologically meaningful inference. Furthermore, while deep learning models (like ESM-2) capture high-order sequence dependencies, they operate as "black boxes," lacking explicit mechanistic interpretability regarding why specific constraints exist.

2. Methodology: The PRIME Framework

The authors introduce PRIME (PRoperty Informed Models of Evolution), a family of codon-level maximum likelihood methods that explicitly model amino acid exchangeability as a function of physicochemical properties.

Core Mathematical Formulation

The model departs from standard substitution rates by defining the nonsynonymous substitution rate multiplier ($\beta_{xy}$) between amino acids $x$ and $y$ as:
$$\beta_{xy} = \alpha \exp\left( \psi - \sum_{i=1}^{D} \lambda_i |x_i - y_i| \right)$$
Where:

• $\alpha$: Synonymous rate multiplier.
• $\psi$: Baseline log nonsynonymous/synonymous rate ratio.
• $D$: Number of modeled physicochemical properties.
• $x_i, y_i$: The value of property $i$ for amino acids $x$ and $y$.
• $\lambda_i$: Importance coefficient for property $i$.
  • $\lambda_i > 0$: Indicates purifying selection (conservation); exchangeability decreases as the difference in property $i$ increases.
  • $\lambda_i < 0$: Indicates diversifying selection; there is a drive to change property $i$.

Three Implementations

1. G-PRIME (Global): Assumes constant $\lambda$ values across the entire alignment to characterize gene-wide constraints.
2. E-PRIME (Episodic): Models $\lambda$ as a random effect varying across phylogenetic branches and sites, allowing for the detection of transient, lineage-specific adaptation.
3. S-PRIME (Site-specific): Estimates independent $\lambda$ vectors for each codon site to resolve fine-grained biophysical architecture.

Physicochemical Properties

The framework utilizes a hierarchical set of properties (2-prop to 5-prop models), including:

• Hydrophobicity (Kyte-Doolittle)
• Volume (Zamyatnin)
• Isoelectric Point (Charge)
• Alpha-Helix Propensity
• Beta-Sheet Propensity

Statistical Testing & Validation

• Hypothesis Testing: Uses Likelihood Ratio Tests (LRTs) with hierarchical False Discovery Rate (FDR) correction (Benjamini-Hochberg followed by Holm-Bonferroni) to identify significant property constraints.
• Simulation: Extensive simulations assess power and False Positive Rates (FPR), identifying the redundancy ratio ($R = \text{substitutions} / \text{unique amino acids}$) as the primary determinant of detection power.
• External Validation: Benchmarked against Deep Mutational Scanning (DMS) data and Protein Language Models (ESM-2) embeddings.

3. Key Contributions

• Mechanistic Interpretability: Transforms abstract evolutionary rates into concrete biophysical rules, identifying which properties drive selection.
• Hierarchical Modeling: Provides a unified framework (Global, Episodic, Site-level) to analyze constraints at different scales.
• Synergy with Rate Heterogeneity: Demonstrates that modeling physicochemical constraints and rate variation ($dN/dS$) are synergistic, not redundant. They capture orthogonal dimensions of evolution.
• Bridge to Deep Learning: Shows that explicit physicochemical weights in PRIME correlate with the principal components of latent spaces in protein language models (ESM-2), validating that these models learn fundamental biophysical rules.

4. Key Results

Model Performance

• Superior Fit: PRIME models (especially the 5-prop version) consistently outperformed standard MG94 and CoRa models across 24 diverse benchmark datasets (viruses, mammals, plants, bacteria), with substantial improvements in AICc scores.
• Synergy: The combination of rate heterogeneity (BUSTED) and physicochemical heterogeneity (PRIME) yielded a mean gain of 723.6 AICc units, significantly higher than the sum of their individual contributions, confirming they capture distinct evolutionary signals.

Biophysical Hierarchy of Constraints

Analysis of 18,944 mammalian genes revealed a distinct hierarchy:

• Rigidly Conserved: Core packing properties (Hydrophobicity and Volume) and Beta-Sheet propensity are under strict purifying selection to maintain structural integrity.
• Adaptive Substrates: Alpha-Helix propensity and Surface Electrostatics (Isoelectric Point) are the primary substrates for adaptive tuning and diversifying selection.
• Case Studies:
  • Influenza HA: PRIME identified Site 226 as "constrained diversification" (changing hydrophobicity/volume while conserving charge/helix propensity), explaining host-switching mechanisms missed by standard metrics.
  • HIV RT: Identified Site 184 as conserving hydrophobicity despite high substitution rates, explaining drug resistance mechanisms.

Detection Power & Information Theory

• Redundancy Ratio ($R$): Detection power is governed by $R$ (substitutions per unique amino acid).
  • $R > 2.0$ is a critical threshold for meaningful inference.
  • Sensitivity exceeds 90% in data-rich alignments ($R \ge 10$).
  • "Toggling" sites (high substitutions, low diversity, e.g., only 2 amino acids) can cause statistical artifacts (inflated FPR) if not corrected, but are biologically informative.

Validation

• DMS Comparison: PRIME showed moderate-to-strong correlation with experimental fitness landscapes (Pearson $r \approx 0.60$), performing comparably to ESM-2 in identifying viable residues but offering superior mechanistic interpretability.
• ESM-2 Alignment: PRIME importance weights correlated with the top principal components of ESM-2 embeddings, suggesting that deep learning models implicitly learn the same biophysical grammar modeled explicitly by PRIME.

5. Significance

• From Phenomenology to Mechanism: PRIME moves evolutionary biology from detecting that selection occurs to explaining how it occurs biochemically. It reveals that adaptation is often "constrained diversification" rather than random exploration.
• Functional Annotation: It can identify "cryptic constraints" in sites that appear neutral by standard $dN/dS$ metrics (e.g., sites where only chemically similar residues are allowed).
• Practical Utility: The framework is implemented in HyPhy, making it accessible for large-scale genomic screens. It provides a tool for predicting the impact of mutations, understanding immune escape, and characterizing the biophysical drivers of protein diversity without relying on opaque deep learning models.
• Limitations & Future Work: The method requires sufficient mutational depth (informational redundancy) and currently assumes site independence (ignoring epistasis). Future iterations may integrate hierarchical Bayesian models to capture co-evolutionary dependencies.

In summary, PRIME provides a robust, interpretable, and statistically sound framework for decoding the "biophysical grammar" of protein evolution, bridging the gap between classical phylogenetics, structural biology, and modern deep learning.