Deconvolving mutation effects on protein stability and function with disentangled protein language models

The paper introduces DETANGO, a deep learning framework that disentangles the intertwined effects of protein stability and function by removing stability-related components from protein language model predictions, thereby accurately identifying stable-but-inactive variants and pinpointing functionally critical residues to advance rational protein engineering.

The Gist

The Big Picture: The "Double-Edged Sword" of Protein Mutations

Imagine a protein as a complex, high-performance car. For this car to work, two things must happen simultaneously:

1. The Engine Must Run (Stability): The car needs to be sturdy enough to hold together. If the frame is too flimsy, the car falls apart.
2. The Car Must Drive (Function): The car needs to steer, accelerate, and brake correctly to get you where you need to go.

For millions of years, evolution has been the "mechanic" tweaking these cars. Sometimes, a mechanic changes a part (a mutation) to make the car faster (better function), but it makes the frame slightly weaker (less stable). Other times, they reinforce the frame, but the car becomes sluggish.

The Problem:
Scientists have powerful tools called Protein Language Models (pLMs) (like a super-smart AI that has read every car manual ever written). These AIs can look at a car and say, "If you change this bolt, the car will break."

But here's the catch: The AI is often vague. It says, "This car will break," but it doesn't tell you why.

• Did the car break because the frame collapsed? (Stability issue)
• Did the car break because the steering wheel was removed? (Function issue)

This is a huge problem for engineers. If you want to fix a broken car, you need to know which part is broken. If you think the frame is the problem, you might reinforce the frame, but if the real issue was the steering wheel, you've wasted your time.

The Solution: Meet DETANGO

The researchers created a new tool called DETANGO (which sounds like "detangling" a knot).

Think of the AI's prediction as a smoothie that mixes two flavors: Stability and Function.

• Old AI: "This smoothie tastes bad." (You don't know if it's the sour lemon or the bitter kale).
• DETANGO: It acts like a magical juicer that separates the smoothie back into its ingredients. It tells you: "Okay, the lemon (stability) is fine, but the kale (function) is the problem."

How DETANGO Works (The "Subtraction" Trick)

DETANGO uses a clever math trick to separate the flavors:

1. The Total Score: It takes the AI's original "bad smoothie" score (how likely the mutation is to break the protein).
2. The Stability Score: It calculates how much the mutation hurts the structural "frame" of the protein (using physics-based tools like FoldX).
3. The Magic Subtraction: It subtracts the Stability Score from the Total Score.
   • Result: What's left is the Function Score.

If the result is still very "bad" even after removing the stability issues, it means the mutation broke the "steering wheel," not the frame.

What Did They Discover?

By using this "separation" tool, the team found some amazing things:

1. The "Stable but Inactive" (SBI) Variants
They found many mutations that were like a perfectly built car with no engine. The car (protein) was structurally sound and didn't fall apart, but it couldn't do its job.

• Why this matters: Before, scientists might have ignored these because the car looked "fine." DETANGO spots them immediately, telling us, "Hey, this part is crucial for driving, even if it's not holding the car together."

2. Finding the "Secret Spots" (Ligand Binding)
Proteins often have specific pockets where they grab onto other molecules (like a hand grabbing a ball). These are called ligand-binding sites.

• DETANGO found these spots with incredible accuracy, even for things like DNA, RNA, and tiny metal ions. It's like having a map that highlights exactly where the driver's seat and the gas pedal are, even if you've never seen the car before.

3. The "Hidden Switches" (Allostery)
Some proteins have "remote controls." You can press a button on the back of the car (a distant part of the protein), and it changes how the engine runs in the front. These are called allosteric sites.

• DETANGO found these hidden switches in the KRAS protein (a famous protein involved in cancer). This is huge because these hidden switches are often the best targets for new drugs.

4. The "Family Resemblance" (Protein Families)
The researchers looked at families of proteins (like cousins). They found that while all cousins share a basic face (stable structure), they have different expressions (specific functions).

• DETANGO showed exactly which parts of the face changed to make one cousin a "smiler" and another a "frowner," helping us understand how evolution fine-tunes proteins for different jobs.

Why Should You Care?

This isn't just about abstract science; it's about better medicine and engineering.

• Drug Design: If you are trying to design a drug to stop a virus, you need to know exactly which part of the virus protein to hit. DETANGO helps you find the "Achilles' heel" (the functional part) rather than just the "armor" (the stable part).
• Fixing Genetic Diseases: Many diseases happen because a protein stops working. DETANGO helps doctors understand why it stopped working. Is the protein falling apart, or is it just broken? The treatment might be different for each.
• Building Better Proteins: If we want to build a new enzyme to eat plastic or clean oil spills, we need to know which parts to tweak to make it work without making it fall apart. DETANGO gives us the blueprint.

The Bottom Line

DETANGO is like a high-tech mechanic's diagnostic tool. It stops us from guessing whether a protein mutation is a structural failure or a functional glitch. By separating these two effects, it gives scientists a clear, mechanical understanding of how life works at the molecular level, paving the way for smarter drugs and better bio-engineering.

Technical Summary

1. Problem Statement

Proteins evolve under coupled constraints: they must maintain structural stability to remain folded while retaining the conformational plasticity required for specific functions (e.g., catalysis, ligand binding, allostery).

• The Challenge: Current methods for predicting mutation effects (such as Protein Language Models or pLMs) provide an "evolutionary plausibility" score. However, this score conflates two distinct selective pressures:
  1. Stability-driven pressure: Mutations that destabilize the protein fold.
  2. Function-driven pressure: Mutations that disrupt specific functional mechanisms (e.g., active sites) while leaving the fold intact.
• The Gap: Existing approaches cannot distinguish between a mutation that causes loss of function because the protein unfolds (destabilization) and a mutation that causes loss of function because it directly disrupts a functional mechanism while the protein remains stable. This obscures the identification of Stable-but-Inactive (SBI) variants and limits the ability to rationally engineer proteins or understand pathogenicity.

2. Methodology: DETANGO Framework

The authors propose DETANGO, a deep learning framework that "reprograms" a pre-trained pLM (specifically ESM-1v) to explicitly disentangle these effects.

Core Mathematical Formulation

The authors posit a multiplicative factorization of the evolutionary probability of a sequence $x$:
$$p_e(x) = p_s(x) \cdot p_f(x)$$
Where:

• $p_e(x)$: Overall evolutionary probability (captured by pLMs).
• $p_s(x)$: Structural probability (stability constraints).
• $p_f(x)$: Functional probability (constraints not explained by stability).

Taking the log-likelihood difference between a mutant ($x_{MT}$) and wild-type ($x_{WT}$), the Functional Plausibility ($f$) is defined as:
$$f(x_{MT}) = e(x_{MT}) - s(x_{MT})$$
Where $e$ is the evolutionary plausibility from the pLM, and $s$ is the structural plausibility derived from stability measurements.

Architecture and Training

1. Input: A wild-type protein sequence ($x_{WT}$) and a pre-trained pLM (ESM-1v).
2. Representation Decomposition: The internal residue embeddings from ESM-1v are decomposed into two latent components:
   • Stability Representation ($h^{(s)}$): Obtained via a projection network (MLP) trained to predict stability measurements (e.g., $\Delta\Delta G$ from FoldX or experimental cellular abundance).
   • Function Representation ($h^{(f)}$): Calculated by subtracting the stability representation from the original evolutionary embedding ($h^{(f)} = h^{(e)} - h^{(s)}$).
3. Prediction Heads:
   • A Functional Predictor estimates the functional plausibility score based on $h^{(f)}$.
   • A Structural Predictor estimates the stability score based on $h^{(s)}$ (or directly uses experimental/computational stability data).
4. Loss Function: The model is trained to minimize the reconstruction error between the original pLM evolutionary score and the sum of the predicted structural and functional scores. This ensures the disentanglement remains faithful to the original evolutionary signal.
   • Note: DETANGO requires no residue-level functional annotations for training; it only needs stability data (computational or experimental).

3. Key Contributions

• Novel Disentanglement Strategy: Unlike post-hoc heuristics that simply subtract stability scores from pLM scores, DETANGO learns a representation-level disentanglement, adapting to local sequence contexts.
• Identification of SBI Variants: The framework specifically identifies "Stable-but-Inactive" variants—mutations that preserve the fold but destroy function.
• Unsupervised Functional Mapping: It can identify functional sites (ligand binding, catalysis, allostery) using only sequence and structure, without requiring prior knowledge of functional sites.
• Scalability: The method is computationally efficient and applicable to large-scale proteome analysis and protein families.

4. Results and Benchmarks

A. Classification of Stable-but-Inactive (SBI) Variants

• Benchmark: Tested on 11 proteins with multidimensional Deep Mutational Scanning (DMS) data (functional readout + cellular abundance/stability).
• Performance: DETANGO significantly outperformed baselines (ESM-1v, ESM-2, FoldX, ESM-IF1, and logistic regression models) in classifying SBI variants (measured by AUROC, AUPRC, and nDCG).
• Insight: It successfully identified residues where mutations were deleterious to function but neutral to stability, which standard pLMs misclassified as general deleterious mutations.

B. Identification of Functional Sites

• Human Domainome: Applied to 408 human proteins. DETANGO assigned significantly higher "function scores" to residues annotated as functional in the Conserved Domains Database (CDD) compared to non-functional residues.
• Comparison: It outperformed sequence conservation baselines (SeqIC) and structure-based models (FoldX, ESM-IF1). Notably, it identified 100% of CDD-annotated zinc-binding sites correctly.
• Energetic Frustration: Residues with high DETANGO scores were enriched for local energetic frustration, consistent with the theory that functional sites often reside in locally frustrated regions.

C. Ligand Binding and Cryptic Pockets

• Ligand Binding Sites (LBS): Evaluated on BioLiP2 (24,252 proteins). DETANGO accurately identified binding sites for DNA, RNA, peptides, metal ions, and small molecules.
• Cryptic Pockets: Tested on the PocketMiner dataset. DETANGO successfully predicted cryptic binding pockets (sites invisible in the apo-structure but visible in the holo-structure) using only the apo-structure as input. For example, it identified the ADP-binding segment in CooC1 that only forms upon conformational change.

D. Allosteric Landscapes

• KRAS Case Study: Applied to KRAS variants interacting with six different partners. DETANGO (using experimental stability data) correlated strongly with experimental binding affinity changes ($\Delta\Delta G_b$) and outperformed ESM-1v in identifying allosteric sites.
• Generalization: Validated on 17 proteins with known allosteric sites, showing robust generalization beyond the training context.

E. Protein Families

• Globins and RAS Superfamily: Applied to mammalian globins and the human RAS superfamily. DETANGO revealed both shared functional patterns (e.g., conserved G-box motifs in RAS) and divergent patterns (e.g., specificity-determining positions distinguishing $\alpha$ and $\beta$ globins or RAB vs. RHO families), offering insights into evolutionary adaptation.

5. Significance and Impact

• Mechanistic Interpretability: DETANGO moves beyond "black box" predictions of deleteriousness to provide a mechanistic explanation: Is the mutation bad because it breaks the fold, or because it breaks the function?
• Protein Engineering: By identifying positions where stability can be traded for function (locally frustrated regions), DETANGO guides rational protein engineering strategies (e.g., "stabilize-then-diversify").
• Drug Discovery: The ability to identify cryptic pockets and allosteric sites without prior structural knowledge of the bound state offers a new avenue for targeting "undruggable" proteins.
• Evolutionary Biology: It provides a principled framework to separate evolutionary constraints driven by stability from those driven by function, refining our understanding of how protein sequences evolve.

In summary, DETANGO represents a significant advancement in computational biology by leveraging the representational power of pLMs to explicitly deconvolve the intertwined evolutionary pressures of stability and function, enabling more precise identification of functional residues and mutation mechanisms.