ESMRank reveals a transferable axis of protein mutational constraint from overlapping variant effect assays

This paper introduces ESMRank, a sequence-based predictor that leverages a novel "variant soundness" framework to unify heterogeneous multiplexed variant effect assays into a transferable axis of mutational constraint, thereby achieving superior performance in predicting protein stability, fitness, and pathogenicity across the proteome.

The Gist

The Big Picture: Making Sense of Protein Chaos

Imagine your body is a massive factory, and proteins are the machines that keep it running. Sometimes, a tiny screw is swapped for a different one (a genetic mutation). Most of the time, the machine still works fine. But sometimes, that swap breaks the machine, leading to disease.

Scientists have been trying to predict which screw-swaps will break the machine. They have run thousands of experiments (called MAVEs) to test these swaps. However, there's a huge problem:

• The "Language" Problem: One lab measures "brokenness" on a scale of 1 to 10. Another lab uses 1 to 100. A third lab uses "Good/Bad" instead of numbers.
• The "Noise" Problem: Because the experiments are so different, it's hard to compare them. It's like trying to combine weather reports from different countries where one uses Celsius, one uses Fahrenheit, and one just says "It's raining."

The result? We have a mountain of data, but it's messy and fragmented. We can't easily see the big picture of what makes a protein break.

The Solution: Finding the "Ranking" Signal

The authors of this paper realized that while the numbers are different, the order is usually the same.

The Analogy: The Race Track
Imagine three different judges watching a race.

• Judge A says: "Runner 1 is 100 points faster than Runner 2."
• Judge B says: "Runner 1 is 500 points faster than Runner 2."
• Judge C says: "Runner 1 is 2 minutes faster than Runner 2."

The numbers are totally different. But they all agree on the ranking: Runner 1 is the fastest, and Runner 2 is slower.

The authors created a new method called Variant Soundness. Instead of trying to average the confusing numbers, they looked at the ranking. They asked: "Across all these different experiments, which mutations are consistently at the bottom (bad) and which are at the top (good)?"

By focusing on the order rather than the specific score, they filtered out the noise and found a clear, unified signal.

The Discovery: The "Stability" Axis

Once they cleaned up the data, they found a hidden pattern. They discovered that the biggest reason proteins break is instability.

The Analogy: The Jenga Tower
Think of a protein as a Jenga tower.

• Buried blocks (inside the tower): If you pull a block from the middle, the whole tower collapses. These are "buried" amino acids. The data showed these are extremely sensitive to change.
• Surface blocks (on the outside): If you change a block on the very top, the tower might wobble a bit, but it usually stays standing. These are "surface" amino acids.

The study found that the "bad" mutations are mostly the ones that knock the Jenga tower over (destabilizing the structure). This "stability" signal was so strong that it showed up even in experiments designed to measure other things, like how well a protein binds to a virus.

The New Tool: ESMRank

Using this new understanding, the authors built a super-smart AI called ESMRank.

The Analogy: The Master Chef
Imagine you want to teach a chef how to cook a perfect steak.

• Old way: You give the chef a list of 1,000 recipes with exact temperatures and times (Regression). But if the chef tries to cook a steak in a different pan, the recipe fails.
• New way (ESMRank): You teach the chef the concept of "doneness." You say, "This steak is too rare, this one is perfect, this one is burnt." You teach the AI to rank the outcomes rather than predict a specific number.

ESMRank combines two types of knowledge:

1. The "Language" of Life: It reads the protein sequence like a language (using a tool called ESM-2), understanding how words (amino acids) fit together.
2. Physics: It also knows basic physics, like how heavy or sticky a piece of the protein is.

By learning to rank mutations (Bad vs. Good) instead of guessing a specific score, ESMRank became much better at predicting which mutations break proteins, even for proteins it has never seen before.

Why This Matters: Real-World Impact

The paper tested this tool on Cystic Fibrosis (CF), a disease caused by a broken protein called CFTR.

The Analogy: The Broken Elevator
In CF, the elevator (the protein) is stuck on the ground floor.

• The Problem: Some mutations break the elevator so badly it can't be fixed. Others just jam the doors, which can be fixed with a wrench (medicine).
• The Result: ESMRank could look at a mutation and predict:
  1. How broken the elevator is (folding efficiency).
  2. Whether a specific medicine (like a "corrector" drug) can fix it.

The AI successfully predicted which patients would respond to expensive drugs and which wouldn't, simply by looking at the protein's sequence and its "stability score."

Summary

1. The Problem: We have too many different protein experiments that don't speak the same language.
2. The Fix: We stopped trying to match the numbers and started matching the rankings (who is worse than whom).
3. The Discovery: The biggest factor in breaking proteins is structural stability (keeping the Jenga tower standing).
4. The Tool: They built ESMRank, an AI that learns to rank mutations by stability.
5. The Win: This AI is better than previous tools at predicting disease and even guessing which medicines will work for specific genetic errors, all without needing to be taught about specific diseases first.

It's like turning a pile of confusing, conflicting weather reports into a single, clear map that tells you exactly where the storm is coming from.

Technical Summary

1. Problem Statement

The interpretation of missense variants across the proteome faces two primary challenges:

• Data Heterogeneity: Multiplexed Assays of Variant Effect (MAVEs), such as Deep Mutational Scanning (DMS), generate massive datasets but are intrinsically heterogeneous. They differ in readouts (e.g., stability vs. activity), dynamic ranges, cellular contexts, and scoring conventions. This makes direct aggregation or comparison of scores across different experiments impossible.
• Model Limitations: Current computational predictors often rely on regression to predict absolute effect sizes. However, because experimental scales vary wildly, naïve fine-tuning on pooled MAVE data often fails to generalize. Furthermore, while absolute effect magnitudes are noisy, the relative ordering (ordinal structure) of variant effects within a specific protein is often reproducible across different assays.

The authors aim to develop a principled method to reconcile these heterogeneous assays into a unified representation of mutational constraint and leverage this to build a superior, generalizable predictor of variant effects.

2. Methodology

A. Overlap-Aware Integration: "Variant Soundness"

The authors propose a framework to extract a consensus signal from partially overlapping MAVE datasets without relying on absolute score scales.

• Data Source: They analyzed 1,122 MAVEdb score sets covering >2.1 million mutations across 596 proteins.
• Rank Alignment: Instead of aggregating raw scores, they focus on the ordinal ranking of variants within each protein.
• Reciprocal Rank Fusion (RRF): For variants tested in multiple assays, they apply RRF to align rankings. This creates a consensus metric called "Variant Soundness," which quantifies the consistency with which a variant is ranked across experiments.
• Noise Suppression: This approach suppresses assay-specific noise while preserving the reproducible biological signal. The resulting scores are normalized to a common scale, creating an assay-agnostic representation of mutational tolerance.

B. ESMRank: A Learning-to-Rank Model

Recognizing that the integrated signal is inherently relative (ordinal) rather than absolute, the authors formulated variant effect prediction as a Learning-to-Rank (LTR) problem rather than a regression problem.

• Architecture: ESMRank uses LambdaMART (a gradient-boosted decision tree implementation of pairwise learning-to-rank).
• Input Features (Multimodal):
  1. Deep Features: Embeddings from the ESM-2 protein language model, capturing global sequence context, implicit structural priors, attention-derived residue contacts, and masked marginal probability shifts.
  2. Shallow Features: A curated set of 18 biophysical and structural descriptors (e.g., melting temperature, instability index, solvent accessibility, packing perturbation).
• Training Objective: The model optimizes the discrimination between more and less deleterious substitutions within each protein, preventing information leakage via strict protein-level stratified cross-validation.

3. Key Contributions

1. Unified Mutational Landscape: The authors successfully demonstrated that partial redundancy in MAVEs encodes a reproducible "transferable axis" of mutational constraint. This axis is enriched for structural stability determinants (residue burial, packing perturbation) and domain architecture.
2. Variant Soundness Metric: They introduced a novel, overlap-aware metric that harmonizes heterogeneous assays, revealing coherent gradients of constraint that align with biophysical principles (e.g., buried residues are less tolerant than surface residues).
3. ESMRank Predictor: They developed a state-of-the-art, sequence-based predictor that outperforms existing stability and fitness models by aligning its learning objective with the intrinsic ordinal structure of experimental data.
4. Mechanistic Interpretability: The study shows that the learned constraint landscape is not just a statistical artifact but reflects biological reality, stratifying genes by disease mechanisms (e.g., Gain-of-Function vs. Haploinsufficiency) and predicting pharmacological responsiveness without explicit clinical supervision.

4. Key Results

A. Biological Structure of the Integrated Signal

• Biophysical Correlates: The integrated "Variant Soundness" scores strongly correlate with structural features. Buried residues show lower tolerance than surface residues. Hydrophobic-to-polar/charged substitutions are highly deleterious in cores but less so on surfaces.
• Domain Architecture: Proteins were clustered into communities based on mutational response patterns. The most tolerant community contained long, disordered proteins with small metal-binding domains (e.g., zinc fingers), while the least tolerant contained compact, $\beta$-rich folds.
• Clinical Relevance: Pathogenic variants from ClinVar are significantly enriched at the deleterious end of the integrated axis compared to benign variants.

B. Benchmarking Performance

• Human Domainome: ESMRank achieved a median Spearman correlation ($\rho$) of 0.62 on a dataset of ~560k variants, significantly outperforming ThermoMPNN ($\rho$ = 0.46) and other stability predictors.
• ProteinGym: In zero-shot settings (excluding training proteins), ESMRank achieved the highest mean Spearman correlation (0.63) on stability assays, surpassing sequence-based, structure-based, and hybrid methods.
• Robustness: Performance remained high even under strict homology filtering (<25% identity) and across diverse contexts (conserved vs. variable, buried vs. exposed).
• Kinetics: ESMRank correlated well with independent folding/unfolding rates (VariBench), validating its biophysical grounding.

C. Clinical and Mechanistic Insights

• Pathogenicity Stratification: ESMRank provided sharper separation between pathogenic and benign variants than $\Delta\Delta G$-based methods, particularly at exposed residues where thermodynamic proxies often fail.
• Disease Mechanism: Genes associated with different disease mechanisms showed distinct global tolerance profiles: Gain-of-Function (GOF) genes were most tolerant, followed by Dominant-Negative (DN), Autosomal Recessive (AR), and Haploinsufficiency (HI) genes (most constrained).
• CFTR Case Study:
  • ESMRank scores correlated strongly with CFTR folding efficiency, channel activity, and pharmacological rescue (response to correctors elexacaftor/tezacaftor and potentiator ivacaftor).
  • Variants predicted to be less destabilizing (higher ESMRank scores) were more likely to respond to pharmacological correction.
  • ESMRank outperformed AlphaMissense and ThermoMPNN in classifying therapeutically responsive variants (AUC = 0.83 vs. 0.80).

5. Significance

This work establishes experimental overlap as a scalable statistical resource for extracting transferable biological signals. By shifting the paradigm from predicting absolute effect sizes to learning ordinal relationships, the authors created a model (ESMRank) that is:

• Generalizable: It performs robustly across diverse proteins and structural contexts without needing protein-specific training data.
• Mechanistically Interpretable: It captures stability-mediated constraints that are fundamental to protein biology, linking sequence variation to folding, function, and drug response.
• Clinically Actionable: It provides a framework for prioritizing variants and anticipating therapeutic responsiveness in genetically heterogeneous disorders, particularly those driven by protein instability.

The study suggests that the "noise" of heterogeneous assays can be transformed into a coherent signal through rank-based integration, offering a new pathway for building next-generation variant effect predictors.