DEX: a consensus-based amino acid exchangeability measure for improved codon substitution modelling

This study evaluates 30 amino acid distance measures across diverse biological datasets to demonstrate that a new consensus-based measure, DEX, derived from deep mutational scanning and experimental exchangeability data, outperforms existing metrics in modeling codon substitution patterns and predicting mean substitution frequencies.

The Gist

Imagine you are trying to predict how a sentence changes when you swap out its words. If you swap "cat" for "dog," the sentence still makes sense. But if you swap "cat" for "pizza," the meaning breaks completely.

In biology, proteins are like sentences made of amino acid "words." When DNA mutates, it swaps these amino acids. Sometimes the swap is harmless (like cat $\to$ dog), and sometimes it destroys the protein's function (like cat $\to$ pizza).

For decades, scientists have tried to build a "dictionary" to predict which swaps are safe and which are dangerous. This paper is about finding the best possible dictionary for this job.

Here is the story of how the authors found the ultimate dictionary, which they named DEX.

1. The Problem: Too Many Dictionaries, No Agreement

Scientists have been trying to measure how "different" two amino acids are for a long time. They've created about 30 different "rulers" or "distance measures" to do this.

• The Old Rulers: Some are based on chemistry (e.g., "How big is this molecule?" or "Is it oily?"). These are like measuring words by their length or number of vowels.
• The New Rulers: Others are based on real-world experiments where scientists actually swapped amino acids in a lab to see what happened. These are like testing the words in a real sentence to see if it still makes sense.

The problem? No one agreed on which ruler was the best. Some said the chemistry ones were great; others swore by the lab-experiment ones. It was like everyone in the room using a different map to find the same destination.

2. The Experiment: A Race to the Finish Line

The authors, Gavin and Louis, decided to settle the debate by running a massive race. They took 30 different amino acid "rulers" and tested them against real-life biological data.

They looked at three very different groups of life:

• Bacteria (Streptococcus): Tiny, fast-evolving microbes.
• Flies (Drosophila): Insects with complex genetics.
• Mammals (Humans and relatives): Large, slow-evolving animals.

They asked a simple question: "Which ruler best predicts how these creatures actually evolve over time?"

3. The Results: The "Lab-Tested" Winners

The race results were clear:

• The Chemistry Rulers (Old School): They did okay, but they were often wrong. They were like guessing how a word fits based only on its spelling, without ever hearing it spoken.
• The Lab-Tested Rulers (New School): These won easily. Specifically, two measures stood out:
  1. EX: An older measure based on lab experiments from 2005.
  2. DMS-EX: A brand-new measure the authors created using massive, modern "Deep Mutational Scanning" data. This is like having a supercomputer test every single possible word swap in a library to see which ones work.

4. The Solution: The "Consensus" Dictionary (DEX)

The authors realized that even the two winners had slight differences. So, they did what a wise judge does: they combined them.

They used a statistical magic trick (called DISTATIS) to blend the two best lab-tested rulers into one super-ruler. They named this new champion DEX (which stands for DISTATIS-based Experimental eXchangeability).

Think of DEX like a "Best of Both Worlds" GPS:
If one map says "turn left" and the other says "turn slightly left," DEX calculates the perfect middle ground that works for almost every driver, in every city, on every road.

5. The Twist: Are Super-Computers Better?

The authors also tested modern AI tools (like RaSP and VespaG) that try to predict if a specific mutation is bad by looking at the 3D shape of the protein.

• For the "Average" Swap: The simple DEX ruler was just as good as the complex AI. Knowing the general "distance" between two amino acids tells you most of what you need to know about the average outcome.
• For the "Disaster" Swaps: The AI tools were much better at spotting the rare, catastrophic mutations that would break a protein completely. It's like DEX is great at telling you if a sentence is "okay," but the AI is needed to spot the one typo that turns a love letter into a breakup note.

The Bottom Line

This paper gives scientists a new, improved tool called DEX.

• Why it matters: It helps us understand how evolution works, how diseases arise from mutations, and how to design better drugs.
• The Analogy: Before, scientists were trying to navigate the world of protein evolution with a mix of old paper maps and untested guesses. Now, they have DEX: a GPS built from the best real-world driving data, blended together to give the most accurate route possible.

In short: DEX is the new gold standard for understanding how life changes its "words" without losing its meaning.

Technical Summary

1. Problem Statement

In molecular evolution, amino acid substitutions between physicochemically similar residues occur more frequently than those between dissimilar pairs. While incorporating amino acid distances into codon substitution models (e.g., $dN/dS$ models) can improve the estimation of selection efficacy, these measures remain underused.

• The Challenge: There is a lack of consensus regarding which amino acid distance measure is most accurate. Hundreds of measures exist, ranging from simple physicochemical properties (e.g., Grantham's distance) to complex empirical matrices (e.g., BLOSUM) and recent deep mutational scanning (DMS) data.
• The Gap: Existing studies have not systematically compared the performance of these diverse measures across different evolutionary lineages or against modern site-specific variant effect predictors. Furthermore, it is unclear whether continuous distance measures outperform binary classifications (radical vs. conservative) or complex deep learning tools for predicting general substitution patterns.

2. Methodology

The authors conducted a comprehensive benchmarking study involving 30 distinct amino acid distance/exchangeability measures.

A. Data Collection and Measure Selection
The study evaluated a wide spectrum of measures:

• Physicochemical/Structural: Grantham's distance, Miyata's distance, Sneath's D, Epstein's coefficient, EMPAR, Xia, and CSW.
• Dimensionality Reduction: Measures derived from principal components of physicochemical properties (Cruciani, FASAGI, Kidera, zScales, VHSE, Atchley, Venkatarajan, and AA-Ontology categories).
• Experimental/Exchangeability: The classic Experimental Exchangeability (EX) measure, DeMaSk, and substitution matrices (BLOSUM62, VTML200).
• Binary Classifications: Radical vs. Conservative (RvC) groupings based on polarity/volume and charge.
• New Measure (DMS-EX): The authors computed a novel measure based on 18 high-quality, independent deep mutational scanning (DMS) datasets from the ProteinGym database. They calculated robust Z-scores for substitutions in buried vs. exposed protein regions, weighted by background amino acid frequencies.

B. Consensus Construction (DEX)
Using the DISTATIS approach (a multivariate statistical method for analyzing multiple distance matrices), the authors generated consensus measures by combining the top-performing individual measures. The primary consensus measure, named DEX, was derived from the combination of the new DMS-EX and the existing EX measure.

C. Evaluation Frameworks
The performance of these measures was tested in two distinct contexts:

1. Codon Substitution Modeling:
   • Data: Alignments from three divergent lineages: Streptococcus (71 species), Drosophila (6 species), and Mammals (190 species).
   • Method: Codon substitution models (M0 and models incorporating amino acid distance matrices via the aaDist option in PAML's CODEML) were fitted to concatenated gene alignments.
   • Metric: Model fit was assessed using the Akaike Information Criterion (AIC).
2. Polymorphism Frequency Prediction:
   • Data: Segregating non-synonymous variants in E. coli (61k+ strains) and human genotypes (gnomAD v4.01).
   • Method: Correlation (Spearman's $\rho$) was calculated between the predicted exchangeability of amino acid pairs and the observed mean minor allele frequencies.
   • Comparison: The authors compared these distance measures against state-of-the-art site-specific predictors: RaSP (structure-based) and VespaG (sequence-based/LLM).

3. Key Contributions

• Introduction of DMS-EX: A new experimental exchangeability measure derived from a curated set of deep mutational scanning datasets, utilizing robust statistical transformations to account for buried vs. exposed residues.
• Development of DEX: A consensus-based measure created via DISTATIS that integrates the strengths of the new DMS-EX and the established EX measure.
• Systematic Benchmarking: The first large-scale comparison of 30 diverse amino acid distance measures across prokaryotic and eukaryotic lineages, evaluating both evolutionary model fit and polymorphism prediction.
• Re-evaluation of Classic Measures: The authors recalculated Grantham's and Miyata's distances using updated volume and polarity estimates to ensure fair comparison, finding no significant improvement over original values.

4. Key Results

• Model Fit Performance:
  • Models incorporating amino acid distances consistently outperformed the standard M0 model (which assumes equal substitution probabilities).
  • DEX emerged as the top-performing measure overall, achieving the lowest AIC scores across all three lineages (Streptococcus, Drosophila, Mammals).
  • The new DMS-EX and the existing EX measure were the best individual performers, with DEX providing a statistically significant improvement over DMS-EX alone.
  • Interestingly, a consensus measure combining EX and DeMaSk performed exceptionally well for Streptococcus but poorly for eukaryotes, suggesting lineage-specific preferences.
• Polymorphism Prediction:
  • There was a strong positive correlation between amino acid exchangeability and mean allele frequency (deleterious mutations are rarer).
  • In E. coli, DeMaSk correlated best, followed closely by DEX.
  • In humans, VespaG (deep learning) correlated best, followed by RvC (Zhang) and DEX.
  • DEX remained the most consistent performer across both lineages when averaged.
• Site-Specific vs. General Predictors:
  • While general distance measures (like DEX) are excellent at predicting mean substitution frequencies, they are inferior to site-specific deep learning tools (RaSP, VespaG) for identifying individual highly deleterious mutations.
  • RaSP and VespaG showed significant enrichment for rare variants among the top 1% of predicted deleterious mutations, a capability static distance matrices lack.
• Selection Efficacy:
  • Using a codon model that incorporates amino acid distances, the authors inferred parameters for selection strength. They observed a negative association between the non-synonymous rate ($a$) and the penalty for dissimilar substitutions ($b$), particularly in Streptococcus, suggesting stronger purifying selection against radical substitutions in bacteria with large effective population sizes.

5. Significance

• Improved Molecular Evolution Models: The study provides a definitive recommendation for the DEX matrix as the most generalizable and accurate amino acid distance measure for codon substitution models. This allows for more precise inference of selection efficacy ($dN/dS$) and evolutionary dynamics.
• Benchmarking Standard: The authors establish that simple, static amino acid distance measures (like DEX) should serve as the baseline for benchmarking complex site-specific variant effect predictors. Using inferior baselines (like raw Grantham distances) can lead to misleading conclusions about the performance of deep learning tools.
• Biological Insight: The results confirm that continuous measures of amino acid dissimilarity capture evolutionary constraints better than binary (radical/conservative) classifications. Furthermore, the study highlights that while general physicochemical rules explain average substitution patterns, site-specific structural and sequence context is critical for predicting the impact of individual mutations.

In conclusion, this paper bridges the gap between experimental deep mutational scanning data and theoretical molecular evolution models, offering DEX as a robust, consensus-based tool for future evolutionary analyses.