Protein Language Modeling and Evolutionary Analysis Reveal an N-terminal Determinant of Functional Divergence in Cytochrome P450s from Sophora. tonkinensis

By integrating the protein language model ESM-2 with evolutionary analyses, this study identifies a conserved N-terminal functional fingerprint in *Sophora tonkinensis* cytochrome P450s and proposes a functional-adaptive decoupling model where a stable targeting core coexists with diversifying interfaces driven by alternating phases of positive selection and neutral drift.

The Gist

The Big Picture: The "Chef" and the "Recipe Card"

Imagine a plant called Sophora tonkinensis (a type of medicinal plant) as a giant, bustling kitchen. Inside this kitchen, there are hundreds of specialized chefs called Cytochrome P450s.

These chefs are incredibly important. They take basic ingredients (simple molecules) and turn them into fancy, powerful dishes (medicines like anti-inflammatory compounds). If these chefs didn't exist, the plant couldn't make its medicine, and we couldn't use it to cure diseases.

For a long time, scientists thought that to understand how these chefs evolved to make different dishes, they just needed to look at the chefs' main bodies (their core structures). They thought, "If two chefs look alike, they probably cook the same thing."

This paper says: "Not quite."

The researchers discovered that the secret to what these chefs cook isn't in their main bodies, but in the name tag and apron they wear at the very front (the N-terminus). By changing this small tag, a chef can suddenly be assigned to a completely different station in the kitchen, cooking a totally different dish, even if their body hasn't changed much.

---

The New Tool: The "Super-Intelligent Librarian"

To figure this out, the scientists used a new, high-tech tool called ESM-2.

Think of traditional science as a librarian who only sorts books by their cover color or the author's last name. If two books have the same author, the librarian assumes they are about the same topic.

ESM-2 is like a Super-Intelligent Librarian who has read every book in the universe. It doesn't just look at the cover; it understands the meaning of the words inside. It can tell you that two books by the same author are actually about completely different topics, or that two books by different authors are actually about the same topic.

Using this "Super Librarian," the scientists looked at the plant's chefs and realized:

• Old Way: "These two chefs are cousins (genetically close), so they must cook the same thing."
• New Way (ESM-2): "Wait, even though they are cousins, they are actually cooking totally different dishes!"

The Discovery: The "Magic Switch" on the Apron

The researchers zoomed in to find out why these cousins were cooking different things. They found the answer in the N-terminus—the very first 100 letters of the chef's "recipe card" (the protein sequence).

They discovered three key things:

1. The "Address Label" Theory (The Localization Switch)

Imagine the chef's apron has a tiny address label on it.

• If the label says "Kitchen A," the chef goes to the stove to cook spicy food.
• If the label says "Kitchen B," the chef goes to the fridge to make cold desserts.

The study found that the plant changes this "address label" (the N-terminus) very quickly. By slightly altering the first few inches of the apron, the plant can instantly send a chef to a new part of the cell (like the Endoplasmic Reticulum vs. the Mitochondria).

• Why this matters: Once the chef is in a new room, they are surrounded by different ingredients. They naturally start making new medicines without needing to rebuild their entire body. It's like moving a baker from a bakery to a chocolate factory; they start making chocolate cakes just because the ingredients are there!

2. The "Stable Core" vs. The "Playground"

The scientists found a fascinating split in how these chefs evolve:

• The "Fingerprint" (The Stable Core): There is a specific set of letters in the apron (positions 41–50) that acts like a fingerprint. This part is very strict. It tells the cell, "This chef belongs to the 'Spicy Food' team." This part doesn't change much because if it does, the chef gets confused and the plant stops working.
• The "Playground" (The Adaptive Zone): The very first part of the apron (positions 1–40) is a playground. This is where the plant experiments. It tries out new letters, deletes old ones, and swaps things around. This is where the "evolutionary magic" happens, allowing the plant to adapt to new threats or environments.

The Big Surprise: The "Fingerprint" (what makes the chef unique) and the "Playground" (where the changes happen) are almost never the same spots.

• Analogy: Imagine a car. The "Fingerprint" is the engine model (which defines if it's a Ferrari or a Ford). The "Playground" is the paint job and the stickers. You can change the paint (adaptation) a thousand times, but as long as the engine model stays the same, it's still the same type of car. The plant keeps the engine stable while painting the car in new colors to survive.

Why This Matters for Us

1. Better Medicine: By understanding exactly which "address label" makes a chef produce a specific medicine, scientists can engineer plants to make more of the good stuff or new medicines we haven't discovered yet.
2. A New Way to Read Biology: This paper shows that we can't just look at family trees (who is related to whom) to understand what an organism does. We have to look at the "functional code" hidden in the small details. It's like realizing that to understand a person's job, you shouldn't just look at their family tree, but at the specific badge they wear on their chest.

Summary in One Sentence

This study discovered that plant enzymes evolve new abilities not by rebuilding their whole bodies, but by swapping out a tiny "address label" at the front, which sends them to a new chemical workshop to invent new medicines.

Technical Summary

1. Problem Statement

Cytochrome P450s (P450s) are a massive enzyme superfamily responsible for the biosynthesis of specialized plant metabolites (e.g., alkaloids, terpenes). While they are crucial for ecological resilience and pharmacology, understanding their functional divergence is challenging due to:

• High Sequence Diversity: Traditional phylogenetic methods based on homology often fail to resolve fine-scale functional shifts, especially among closely related paralogs.
• Limitations of Selection Models: Standard codon-based selection models (e.g., Hyphy) can identify sites under positive selection but cannot determine which specific residues are causative for functional identity versus those merely undergoing lineage-specific adaptation.
• The "Black Box" of N-termini: The N-terminal signal sequences are known to be critical for subcellular localization, yet their specific evolutionary drivers and mechanisms in driving functional diversification remain largely unresolved.

2. Methodology

The study employs an integrative framework combining Protein Language Models (PLMs), Evolutionary Genetics, and Interpretable Machine Learning.

• Data Source: 345 non-redundant P450 proteins from the medicinal plant Sophora tonkinensis.
• Protein Language Modeling (ESM-2):
  • Sequences were projected into a high-dimensional semantic space using the ESM-2 (650M) model.
  • Hierarchical clustering of ESM-2 embeddings was used to define "functional clusters" independent of phylogeny.
  • Decoupling Analysis: The congruence between ESM-2 functional clusters and traditional phylogenetic (monophyletic) clusters was quantified using V-measure and Adjusted Mutual Information (AMI).
• Evolutionary Analysis:
  • Positive Selection: The Mixed Effects Model of Evolution (MEME) in Hyphy was used to detect sites under episodic diversifying selection across the protein family.
  • Synteny & Phylogeny: Standard tools (IQ-TREE, MCscanX, JCVI) were used to reconstruct phylogenetic trees and analyze gene duplication/synteny.
• Interpretable Machine Learning (The Core Innovation):
  • Region Ablation: To identify the minimal functional determinant, classifiers (Linear SVM, Random Forest, Logistic Regression) were trained on various sequence segments: N-terminal (1–100 aa), shuffled N-terminal (composition control), C-terminal, middle, random fragments, and full-length sequences.
  • Length-Scan: Systematic testing of N-terminal segments (40–100 aa) to find the shortest length sufficient for functional classification.
  • Positional Importance: Using Linear SVM decision weights projected back to the original ESM-2 feature space to identify specific "key residues" driving classification.
  • Validation: Key residues identified by ML were cross-referenced with positively selected sites using a hypergeometric test to determine overlap.

3. Key Results

A. Functional-Phylogenetic Decoupling

• ESM-2 clustering revealed a moderate but significant decoupling (V-measure = 0.576) between phylogenetic ancestry and functional similarity.
• 13 phylogenetically adjacent sister pairs were assigned to distinct functional clusters by ESM-2, indicating recent functional divergence that traditional trees missed.
• Sequence alignment of these divergent pairs revealed that the most dramatic variations (deletions, insertions, substitutions) were concentrated in the N-terminal region, while catalytic cores remained conserved.

B. The N-terminus as a Functional Determinant

• Sufficiency: An N-terminal segment of 100 amino acids achieved classification accuracy statistically indistinguishable from the full-length sequence.
• Specificity: The N-terminal signal was significantly more predictive than shuffled N-termini (preserving composition but destroying order), C-termini, or random fragments. This confirms the signal relies on specific sequence order, not just amino acid composition.
• Hierarchical Architecture: A length-scan identified three functional zones:
  1. Strong Divergence Zone (1–47 aa): Contains adaptive "hotspots" but is insufficient alone for classification.
  2. Transition Zone (48–55 aa): Where accuracy rapidly converges to full-sequence levels.
  3. Context Zone (56–100 aa): Stabilizes the functional identity.
• Minimal Determinant: A 51-amino-acid segment was identified as the minimal sufficient determinant for functional classification.

C. The "Functional-Adaptive Decoupling" Model

• Key Residues: ML identified 20 robust key positions (the "functional fingerprint"), with a dense cluster at positions 41–50.
• Minimal Overlap with Positive Selection: Crucially, there was minimal overlap (Jaccard index = 0.085, p = 0.932) between the ML-identified key functional residues and sites under positive selection.
  • Interpretation: The residues defining the stable functional class (the fingerprint) are distinct from the residues undergoing rapid, lineage-specific adaptation.

D. Evolutionary Mechanism: The "Localization Switch"

• The study proposes that N-terminal variations act as an evolutionary "localization switch."
• Changes in the N-terminal signal (length, charge, hydrophobicity) may reroute P450s to different subcellular compartments (e.g., ER vs. plastids).
• This spatial segregation exposes the enzyme to different substrate pools and redox partners, driving functional divergence without requiring mutations in the conserved catalytic core.
• Once a new localization is established, the N-terminal sequence becomes a stable "functional fingerprint" under purifying selection, while the surrounding interface remains a playground for adaptive evolution.

4. Key Contributions

1. Methodological Framework: Established a novel pipeline ("PLM-based functional clustering → evolutionary validation → key region mapping") to bridge genotype to metabolic phenotype in non-model species.
2. Discovery of N-terminal Determinants: Demonstrated that the N-terminus is not merely a passive anchor but a compact, hierarchical module that disproportionately encodes functional identity.
3. Resolution of the Neutralist-Selectionist Debate: Provided evidence for a "functional-adaptive decoupling" where stable functional codes (fingerprint residues) and dynamic adaptive interfaces (positively selected sites) operate on distinct evolutionary layers.
4. Hypothesis Generation: Proposed the "Localization Switch" hypothesis, offering a mechanistic explanation for how P450s rapidly diversify functionally via subcellular re-targeting.

5. Significance

• For Evolutionary Biology: This work reconciles neutralist and selectionist perspectives by showing that functional innovation can be driven by a two-step cycle: initial positive selection driving innovation in a specific module (N-terminus), followed by neutral/purifying evolution that stabilizes the new functional identity.
• For Plant Metabolism: It provides a precise target (the N-terminal 51-aa module) for metabolic engineering in medicinal plants like Sophora tonkinensis to optimize the production of bioactive compounds (e.g., anti-inflammatory alkaloids).
• For Protein Science: It validates the utility of Protein Language Models (ESM-2) in uncovering cryptic functional relationships and structural determinants that are invisible to traditional homology-based methods.