Transferable excited-state dynamics enable screening of fluorescent protein chromophores

This paper introduces X-MACE, a transferable machine-learning potential combined with curvature-driven surface hopping, to efficiently screen fluorescent protein chromophores and reveal how steric crowding and conjugation extension govern their excited-state dynamics and fluorescence properties.

The Gist

Imagine you are a chef trying to invent the perfect glowing dessert. You know that the secret ingredient is a special molecule (the "chromophore") that makes the dessert light up when you shine a flashlight on it. But here's the problem: every time you tweak the recipe by adding a pinch of this or a dash of that, you have to bake the whole cake, wait for it to cool, and then test it in the dark. If you want to test 200 different recipes, it would take you years of baking and waiting.

This is exactly the challenge scientists face with Green Fluorescent Protein (GFP), the molecule that makes jellyfish glow and helps doctors see inside our cells. They want to design better, brighter, and more stable versions, but testing them one by one is too slow and expensive.

The "Magic Recipe Book" (X-MACE)

In this paper, the researchers (Rhyan, Sophia, and Julia) created a super-smart AI chef called X-MACE.

Think of X-MACE as a "Magic Recipe Book" that has already tasted thousands of different glowing molecules. It knows the general rules of how these molecules behave. Usually, to teach a computer to predict how a new molecule will behave, you'd have to feed it millions of examples of that specific molecule. That's like teaching a chef to bake a new cake by making 10,000 of them first.

But X-MACE is different. It's transferable.

• The Analogy: Imagine a master chef who knows the physics of baking so well that if you give them a new type of flour, they only need to taste three or four cookies made with it to figure out how to bake the perfect cake.
• The Result: The team only needed to run about 100 tiny simulations for each new molecule variant. The AI then used its "general knowledge" to predict how the other 193 variants would behave, saving them years of computer time.

The Two Rules of the Glow

Once the AI started "baking" (simulating) these 193 different glowing molecules, it discovered two simple rules that control whether a molecule glows brightly or just flickers out:

1. The "Crowded Dance Floor" Rule (Steric Crowding)

• What happened: When the researchers added bulky, heavy groups to a specific spot on the molecule (the phenolate ring, specifically position R4), it was like putting a giant, clumsy dancer on a small dance floor.
• The Effect: The molecule got "crowded." This forced it to twist and turn violently to find space.
• The Result: This twisting made the molecule lose its energy quickly by spinning around (a process called isomerization). Instead of glowing, it just got hot and stopped. Crowding kills the glow.

2. The "Super-Sticky Glue" Rule (Conjugation Extension)

• What happened: When they added a long, chain-like tail to the other side of the molecule (the imidazolinone ring, position R5), it was like adding a super-strong glue that held the molecule flat.
• The Effect: This "glue" (extended conjugation) made the molecule very stiff and resistant to twisting. It refused to dance around.
• The Result: Because it couldn't twist and lose energy that way, it had to get rid of its energy by glowing. Stiffness creates a brighter, longer-lasting glow.

Why This Matters

Before this study, scientists were like explorers guessing their way through a dark forest, testing one path at a time. Now, they have a GPS.

• For Doctors: This means we can quickly design new fluorescent proteins that are brighter and don't fade away as fast. This helps in taking clearer pictures of cancer cells or tracking viruses inside the body.
• For Science: It proves that we don't need to simulate every single atom from scratch for every new molecule. We can use a "pre-trained" brain and just give it a tiny nudge to understand new systems.

The Bottom Line

The researchers built a smart AI that learns the "rules of the game" for glowing molecules. They found that if you want a molecule to glow, you should keep it stiff and flat (add long tails) and avoid crowding it (don't add bulky groups in the wrong spots). This allows them to screen hundreds of potential new "glow-in-the-dark" proteins in a few days, a task that used to take decades.

Technical Summary

1. Problem Statement

The rational design of fluorescent proteins (FPs), such as Green Fluorescent Protein (GFP), requires a deep understanding of the photophysical mechanisms governing their chromophores (specifically the HBDI⁻ anion). The relaxation of excited states in these systems involves competing radiative (fluorescence) and non-radiative pathways (photoisomerization via twisted intramolecular charge transfer).

• The Bottleneck: Accurately simulating these nonadiabatic excited-state dynamics is computationally prohibitive using conventional quantum chemistry methods (e.g., CASSCF, CASPT2). Achieving statistically significant results for a single molecule can take years of compute time; scaling this to screen hundreds of variants is practically impossible.
• The Gap: While Machine Learning Potentials (MLPs) have accelerated ground-state simulations, existing excited-state MLPs typically require massive, molecule-specific training datasets (hundreds of thousands of points), preventing systematic screening across diverse chemical spaces. There is a lack of transferable models that can learn general relationships from diverse data and adapt to new systems with minimal data.

2. Methodology: X-MACE and Workflow

The authors introduce X-MACE, a transferable machine-learning potential designed specifically for excited-state dynamics, integrated into a two-stage screening workflow.

A. The X-MACE Model

• Architecture: Based on the Message Passing Atomic Cluster Expansion (MACE), a high-order equivariant neural network. X-MACE extends MACE to predict multiple potential energy surfaces (PES), forces, and oscillator strengths simultaneously.
• Transferability Strategy:
  1. Pre-training: A foundational model is trained on a diverse dataset of 12,000 chromophores (generated via metadynamics and quantum calculations) to learn general structure-property relationships.
  2. Fine-tuning: For a specific target system (e.g., a specific HBDI⁻ derivative), the pre-trained model is fine-tuned on a very small reference dataset (<100 geometries). This allows the model to specialize to the specific conformational space of the derivative while retaining the general chemical knowledge from pre-training.
• Quantum Reference: The training data is generated using ADC(2) (Algebraic Diagrammatic Construction to second order) with the cc-pVDZ basis set. This method was selected over CASSCF/CASPT2 because it accurately reproduces the S₁ torsional PES shape without requiring manual active space selection, making it a robust "black-box" method for large-scale screening.

B. Simulation Workflow

1. Dataset Generation: 193 HBDI⁻ variants were created by systematically substituting positions R1–R4 on the phenolate ring and R5 on the imidazolinone ring.
2. Fine-tuning: Each variant was fine-tuned using ~50–100 reference geometries sampled along torsional coordinates and via Wigner sampling.
3. Dynamics: The fine-tuned X-MACE models were interfaced with the SHARC (Surface Hopping Arbitrary Coupling) program to perform nonadiabatic molecular dynamics.
   • Settings: 7 ps trajectories, 0.5 fs time steps, curvature-driven time-derivative couplings (kTDC).
   • Scale: 19,300 total trajectories were run across 193 chromophores.

3. Key Contributions

• First Transferable Excited-State Foundation Model: Demonstrated that a single pre-trained model can be adapted to diverse excited-state systems with fewer than 100 data points per variant, a drastic reduction from the hundreds of thousands required by previous methods.
• Data-Efficient Screening: Established a workflow that scales excited-state dynamics from static single-point calculations to statistically robust trajectory ensembles for hundreds of molecules.
• Mechanistic Design Rules: Uncovered specific structure-property relationships governing GFP chromophore fluorescence that were previously difficult to quantify systematically.

4. Key Results and Findings

The screening of 193 HBDI⁻ derivatives revealed two competing design principles governing excited-state lifetimes and photoisomerization yields:

A. Steric Crowding (Phenolate Ring: R1–R4)

• Observation: Substitutions at the R4 position (closest to the imidazolinone ring) introduce significant steric repulsion.
• Effect: This steric crowding lowers the torsional activation barrier, facilitating rapid access to twisted geometries (conical intersections).
• Outcome: R4-substituted variants exhibit the shortest lifetimes (~1.3 ps) and near-zero activation barriers, leading to rapid non-radiative decay and high photoisomerization yields (low fluorescence). R1 substitutions have a similar but weaker effect. R2 and R3 substitutions have minimal steric impact, preserving longer lifetimes.

B. Conjugation Extension (Imidazolinone Ring: R5)

• Observation: Introducing $\pi$-connected tail substituents at the R5 position extends the conjugation of the chromophore.
• Effect: Extended conjugation delocalizes electron density, stabilizing the planar excited-state configuration. It increases the energetic penalty for torsion (raising the barrier to the conical intersection).
• Outcome: Highly conjugated variants show longer S₁ lifetimes, suppressed non-radiative decay, and significantly enhanced fluorescence. The population remains trapped in the S₁ state longer, favoring radiative relaxation.

C. Representative Case Studies

• R4-CN (Cyano): Steric clash accelerates torsion; rapid decay to S₀.
• R5-WYG (Extended Tail) + R2/R4-CN: The electronic stabilization from the R5 tail counteracts the steric acceleration. Even with a CN group at R4, the extended conjugation prevents full isomerization, leading to a "trapped" S₁ population (~30%) and prolonged fluorescence compared to the R4-only variant.

5. Significance and Impact

• Accelerated Material Discovery: This workflow transforms excited-state dynamics from a theoretical bottleneck into a practical design engine. It enables the screening of thousands of chromophore variants to identify candidates with optimized brightness and photostability before wet-lab synthesis.
• Interpretable Design Rules: The study provides clear, actionable guidelines for protein engineering:
  • To increase fluorescence: Extend conjugation (R5) and avoid steric crowding at R4.
  • To promote photoisomerization (e.g., for photoswitches): Introduce steric bulk at R4.
• Broader Applicability: The "pre-train on diverse data + fine-tune on specific coordinates" strategy offers a general template for photochemical design in other chromophore families, potentially revolutionizing the development of bioimaging tools, sensors, and optogenetic actuators.

In conclusion, the paper demonstrates that transferable machine learning potentials can overcome the computational cost barrier of nonadiabatic dynamics, providing a scalable route to understanding and engineering the photophysics of complex biological systems.