Signatures of Electron-hole Hopping in Myoglobin Peroxidase Activity Revealed by Deep Mutational Learning

By combining deep mutational scanning with protein language models, researchers identified and experimentally validated myoglobin variants with significantly enhanced peroxidase activity, demonstrating that strategic aromatic substitutions facilitate long-range electron-hole hopping to improve catalytic efficiency.

The Gist

The Big Idea: Turning a Storage Tank into a Power Tool

Imagine Myoglobin as a tiny, spherical oxygen storage tank found in your muscles. Its main job is to hold onto oxygen and release it when your body needs a burst of energy (like when you're sprinting).

However, scientists discovered that this tank has a hidden superpower: it can also act like a chemical power tool (a peroxidase). It can break down harmful substances and clean up toxic chemicals. But in its natural state, this power tool is a bit sluggish and clunky.

The goal of this research was to answer a simple question: Can we redesign this storage tank to make it a much more powerful chemical tool?

The Problem: Too Many Possibilities

Myoglobin is made of a chain of building blocks called amino acids. To make it better, you could swap out one of these blocks for a different one. But there are millions of possible combinations. Trying them all one by one would take a human lifetime.

The researchers needed a way to test thousands of versions at once and then use a "smart guesser" (Artificial Intelligence) to find the best ones.

The Solution: The "Yeast Factory" and the "Fluorescent Tag"

1. The Yeast Factory (EP-Seq)
The team used yeast cells as tiny factories. They programmed each yeast cell to display a different version of the Myoglobin protein on its surface.

• They created a library of 6,000 different Myoglobin variants.
• To test them, they added a special chemical (Tyramide) that glows red when the Myoglobin does its job.
• The Analogy: Imagine a massive dance floor where 6,000 dancers (yeast cells) are wearing different colored shirts (Myoglobin variants). When the music starts (the chemical reaction), the dancers who are the best at the dance glow the brightest.
• Using a high-speed camera (Flow Cytometry), they sorted the brightest dancers into a "winner's circle." This told them which Myoglobin versions were the most active.

2. The AI Coach (Deep Mutational Learning)
Once they had the data on which 6,000 versions worked best, they didn't stop there. They fed this data into a Machine Learning model (a type of AI).

• The Analogy: Think of the AI as a master coach who has watched 6,000 practice games. The coach learns the rules of the game: "Oh, I see that whenever we put a specific type of player (an aromatic amino acid like Tryptophan) in a specific spot, the team scores more points."
• The AI then looked at the entire universe of possible Myoglobin combinations (over 4 million double-mutant versions) and predicted which ones would be the champions, even ones they hadn't seen before.

The Discovery: The "Electron Relay"

The AI predicted that the best way to boost the Myoglobin's power was to add Tryptophan and Tyrosine residues (special amino acids) to the surface of the protein.

The "Hole-Hopping" Analogy:
Think of the Myoglobin's active center (where the chemistry happens) as a locked vault deep inside a fortress.

• The Old Way: To get energy out of the vault, you have to walk all the way through the dark, narrow tunnels. It's slow and difficult.
• The New Way (Hole-Hopping): The researchers added "relay stations" (the Tryptophan/Tyrosine residues) on the surface of the fortress.
• Now, the energy (electrons) can hop from station to station, like a game of "Red Light, Green Light" or passing a hot potato, until it reaches the vault. This makes the process incredibly fast and efficient.

The Results: From Theory to Reality

The researchers took the top 20 predictions from the AI and built them in the lab.

• The Result: Every single one of the 20 AI-predicted variants was better than the original Myoglobin.
• The Champion: They found a "Super Variant" (Q92W/F107W) that was nearly 5 times more efficient at breaking down chemicals than the original.
• Real-World Check: They tested these new proteins not just on yeast, but as free-floating proteins in a test tube (soluble format). The results were the same: the new designs worked perfectly outside the yeast cell, proving the method is reliable.

Why Does This Matter?

This study is a blueprint for the future of enzyme engineering.

1. Speed: Instead of guessing and checking for years, we can use AI to find the best designs in days.
2. Versatility: This method could be used to design better enzymes for cleaning up industrial waste, breaking down plastics, or creating new medicines.
3. Understanding Life: It helps us understand how nature builds "electron highways" inside proteins to protect cells from damage.

In short: The researchers built a massive testing ground, taught a computer to spot the winners, and used that computer to design a super-powered version of a natural protein that works 5 times better than the original. They proved that by adding a few "relay stations" to the protein's surface, you can turn a slow chemical reaction into a lightning-fast one.

Technical Summary

1. Problem Statement

Myoglobin (Mb), traditionally known for oxygen storage, possesses latent peroxidase activity mediated by high-valent iron(IV)-oxo species (Compound I/II). While this activity can be beneficial for detoxification, it is often suppressed in physiological conditions. Researchers aim to engineer Mb to enhance this peroxidase activity for applications like dye decolorization and antibiotic degradation.

However, two major challenges hinder the rational design of such enzymes:

1. Mechanistic Uncertainty: It is unclear how specific sequence variations introduce or alter catalytic mechanisms, particularly regarding long-range electron transfer (hole-hopping) pathways mediated by aromatic residues (Tyrosine and Tryptophan).
2. Experimental Limitations: Traditional high-throughput screening methods often conflate improved enzymatic activity with increased protein expression levels. Furthermore, linking genotype to phenotype for complex enzymatic reactions in a massively parallel manner is technically difficult.

2. Methodology

The authors employed a multi-stage workflow combining high-throughput experimental screening with deep learning:

A. Experimental Platform: Enzyme Proximity Sequencing (EP-Seq)

• System: Yeast surface display of a barcoded library of human myoglobin variants (>6,000 variants).
• Assay: A single-cell tyramide proximity labeling technique.
  • Peroxidase activity on the yeast surface oxidizes a fluorescent tyramide-Alexa Fluor 594 substrate.
  • The resulting radical covalently labels the cell surface, with fluorescence intensity proportional to enzymatic activity.
  • Normalization: Cells are dual-stained with antibodies against a C-terminal His-tag to measure expression levels.
• Sorting: Fluorescence-Activated Cell Sorting (FACS) separates variants based on expression-normalized activity (activity per unit of expression) into four bins.
• Sequencing: Next-Generation Sequencing (Illumina) quantifies variant abundance in each bin to calculate fitness scores.
• Validation: Selected variants were expressed as soluble proteins in E. coli and tested using Michaelis-Menten kinetics and decoupled labeling assays to confirm that activity gains were not artifacts of the yeast display format.

B. Computational Framework: Deep Mutational Learning (DML)

• Data Curation: A dataset of ~4,769 unique variants (filtered for low-confidence and high-order mutants) was used for training.
• Model Architecture:
  • Embeddings: Protein sequences were encoded using pre-trained protein language models (PLMs), specifically ESM3 and ProtTrans.
  • Regressor: A Multi-Layer Perceptron (MLP) was trained to predict fitness scores from sequence embeddings.
  • Ensemble: 20 models were trained using 5-fold cross-validation and different random seeds to ensure robustness.
• In Silico Screening: The trained ensemble was applied to screen a virtual library of >4.25 million double mutants to identify high-performing candidates.

C. Mechanistic Analysis

• Tools like eMap and EHPath were used to model electron/hole transfer pathways from the heme cofactor to the protein surface, analyzing the role of aromatic residues in facilitating long-range electron transfer.

3. Key Contributions

1. High-Throughput Fitness Landscape: Generated a comprehensive map of human myoglobin peroxidase activity, distinguishing between stability defects and catalytic defects.
2. Deep Mutational Learning Pipeline: Successfully demonstrated a DML workflow where experimental data trains a model to predict the activity of unseen double mutants with high accuracy.
3. Mechanistic Insight (Hole-Hopping): Provided experimental evidence that introducing surface-exposed aromatic residues (specifically Tryptophan and Tyrosine) creates "hole-hopping" pathways that enhance peroxidase activity for bulky substrates.
4. Generalizability: Validated that activity trends observed in yeast surface display translate directly to soluble enzymes, confirming the platform's utility for biocatalyst design.

4. Key Results

• DMS Findings:
  • ~88% of missense mutations reduced peroxidase activity, indicating stricter constraints on catalysis than on protein folding stability.
  • Mutations at critical heme-coordinating residues (His94, His65) were highly deleterious.
  • Aromatic Substitutions: Introducing Tryptophan (Trp) or Tyrosine (Tyr) at surface-exposed positions (e.g., Q92, A72, F107) frequently enhanced activity, while removing native aromatic residues was often deleterious.
• Machine Learning Predictions:
  • The ESM3-based MLP model achieved high correlation ($R^2$) with experimental data.
  • Perfect Success Rate: All 20 experimentally tested double mutants selected from the top 0.2% of ML predictions showed improved peroxidase activity compared to Wild Type (WT).
• Soluble Validation:
  • Three top candidates (Var4, Var9, Var14) expressed in E. coli confirmed the yeast-display results.
  • The double mutant Q92W/F107W exhibited a 4.9-fold increase in catalytic efficiency (reaction velocity) compared to WT when using the bulky substrate tyramide.
  • Kinetic analysis showed that these variants significantly improved activity for bulky substrates (tyramide, Reactive Blue 19) but showed similar activity to WT for smaller substrates (guaiacol), supporting the "surface relay" hypothesis.
• Epistasis: While single Trp mutations were beneficial, their combination showed non-additive (negative epistatic) effects in some cases, though Q92W/F107W remained the top performer.

5. Significance

• Enzyme Engineering: This study establishes a robust framework for engineering redox enzymes by leveraging deep mutational learning to navigate vast sequence spaces.
• Mechanistic Understanding: It validates the "hole-hopping" hypothesis in myoglobin, demonstrating that surface-exposed aromatic residues can act as relay stations to shuttle oxidative equivalents to bulky substrates that cannot access the heme pocket directly.
• Biomedical Applications: The findings offer insights into engineering Hemoglobin-based Oxygen Carriers (HBOCs) to tune redox activity and prevent oxidative damage, as well as designing industrial biocatalysts for dye decolorization and pollutant degradation.
• Methodological Advance: The EP-Seq platform successfully decouples expression from activity, solving a long-standing bottleneck in high-throughput enzyme screening.