Transferable excited-state dynamics enable screening of fluorescent protein chromophores

Rhyan Barrett †, Sophia Wesely †, Julia Westermayr · 2026

Read Transferable excited-state dynamics enable screening of fluorescent protein chromophores by Rhyan Barrett †, Sophia Wesely †, Julia Westermayr (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-12699.txt.

Introduction

Rhyan Barrett and two colleagues screen fluorescent protein chromophores with a transferable model. Standard excited-state simulations cost too much for broad screening. Their X-MACE potential predicts energy maps, forces, and strengths. Hopping guided by curvature then enables efficient pathway screening. Green fluorescent protein variants show how small changes reshape relaxation. Tuning one pretrained model on fewer than 100 geometries per variant yields good dynamics. Crowding on the ring lowers twist barriers and speeds twisted intersections. Longer chains stabilize planar states, curb dark decay, and prolong fluorescence.

Catalog record

CREATORRhyan Barrett †, Sophia Wesely †, Julia Westermayr
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41071
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.12699

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Transferable excited-state dynamics enable screening of fluorescent protein chromophores Rhyan Barrett †, Sophia Wesely †, Julia Westermayr Abstract Transferable excited-state dynamics offer a route to efficient screening of photophysical behavior across molecular systems, but conventional nonadiabatic simulations remain prohibitively expensive. Here we introduce X-MACE, a transferable machine-learning potential for excited-state dynamics that predicts multiple potential energy surfaces, forces and oscillator strengths, and combine it with curvature-driven surface hopping to enable data-efficient screening of photochemical pathways. We apply this framework to fluorescent chromophores as an example application, using green fluorescent protein chromophore variants to demonstrate how subtle structural modifications reshape excited-state relaxation, lifetimes and photoisomerization yields. Fine-tuning a single pretrained model with fewer than 100 reference geometries per derivative yields accurate dynamics across a chemically diverse set of analogues. The screening reveals two governing design principles: steric crowding on the phenolate ring lowers the torsional barrier and accelerates access to twisted conical intersections, whereas conjugation extension stabilizes planar excited-state configurations, suppresses non-radiative decay and prolongs fluorescence. More broadly, this workflow provides a general framework for scalable excited-state screening and interpretable design of photophysical properties. I Introduction The ability to visualize cellular processes using fluorescent proteins has revolutionized our understanding of life at the molecular scale. A major advance came with the discovery and development of the green fluorescent protein (GFP), recognized by the Nobel Prize in Chemistry in 2008 30 , which fluoresces upon light excitation to enable real-time tracking of dynamics in living systems. This breakthrough launched a new era of cellular imaging that has transformed disease research, diagnostics and therapies. 20 ; 12 ; 5 Clarifying photochemical mechanisms in fluorescent proteins can thus guide rational optimization and design of fluorescent proteins, improving brightness, photostability, maturation, and spectral properties thereof. At the centre of GFPs fluorescence behavior is the p-hydroxybenzylidene-imidazolinone anion (HBDI - ), the chromophore of GFP (see Figure 1a). It is known that, upon photoexcitation, its relaxation occurs through competing radiative and non-radiative pathways, where the dominant non-radiative channel is commonly associated with twisted intramolecular charge transfer that can drive partial photoisomerisation. 19 This twisting can proceed via rotation about either the imidazolinone dihedral or the phenolate dihedral (Fig. 1 a), and is coupled to pyramidalisation at the methine-bridge carbon, providing access to efficient S_{1}/S_{0} decay. 19 ; 18 Consistent with this picture, HBDI - exhibits a pronounced temperature dependence: fluorescence is enhanced at low temperature and suppressed as temperature increases, indicating increased access to non-radiative relaxation channels at higher internal energies. 23 ; 32 This behaviour suggests that the S_{1}/S_{0} conical intersection seam is not readily reachable from the equilibrium excited-state minimum, but instead requires substantial structural distortion, specifically large-amplitude torsion about either ring toward roughly 90 degrees, to reach torsional conical intersections as identified in prior electronic structure analyses. 18 ; 23 Mechanistically, the excitation from non-bonding to the anti-bonding orbital permits bridge torsion, while sufficient kinetic energy is needed to attain the distorted geometries, directly linking non-radiative yield to temperature; accordingly, improving GFP performance can be framed as shifting the balance toward radiative decay by restricting these torsional deactivation routes, thereby enhancing fluorescence efficiency for bioimaging applications. A great advantage of fluorescent proteins such as GFP is the existence of a diverse family of variants, whose excited-state properties can be tuned by subtle structural changes to their chromophore and surrounding protein environments. 24 ; 27 Consequently, many studies focus on the effects of small variations in the chromophore on fluorescent properties. 6 ; 1 However, current studies are limited to a few individual molecules due to the time-consuming nature of experiments and design processes. Theoretical modelling could overcome this challenge by screening many chromophores prior to any wet lab work, but this process demands accurate theoretical non-adiabatic excited-state dynamics simulations that are prohibitively expensive: even for a single HBDI - derivative, the required sampling of different conformations to achieve statistically significant dynamics can consume years of compute time, and extending this to a substituted series of chromophores rapidly scales to hundreds of years on high-performance resources. Although recently emerged machine learning potentials can alleviate these costs, most existing excited-state approaches still require large, molecule-specific training datasets for each chromophore to be simulated, making systematic screening across diverse HBDI - variants difficult in practice. Foundational models offer a promising route toward screening, as they can learn from large, heterogeneous datasets and transfer this knowledge to new systems with little to no additional data. While such models have already transformed ground-state simulations, they have not yet been applied to photodynamics screening. Here, we introduce the first step towards a foundational model for excited state dynamics and a general framework for screening excited-state dynamics of hundreds of chromophores. Applied to a diverse set of HBDI - derivatives, the approach enables efficient comparative simulations and provides mechanistic insights into key factors governing their dynamics. By systematically simulating many variants, we show that increased conjugation leads to substantially longer lifetimes and sustained S_{1} population, resulting in a larger fraction of radiative de-excitation and consequently, enhanced fluorescence potency of the HBDI - system. II Results and Discussion II.1 Machine Learning Workflow To efficiently assess how structural changes influence photophysical and photochemical behaviour across a range of chemical systems, a two-stage machine learning workflow for excited-state modelling and screening (Fig. 1 c) was developed. As illustrated, the first step requires an excited-state surrogate model that needs to be trained on a chemically diverse collection of chromophores to learn general relationships between molecular structure and excited-state properties. Therefore, we extend the message-passing atomic cluster expansion model (MACE) 4 , which has previously shown high generalizability between systems, to multiple excited state potentials, forces, and oscillator strengths, and call it X-MACE. For training, we first compile a dataset of 12,000 chromophores with details on the dataset and quantum-chemical calculations provided in the methods section.The resulting X-MACE model is then adapted to a specific chromophore system using fine-tuning on a small reference dataset that is, ideally, representative of the relevant conformational space of each chromophore. In this work, we fine tune the model on a family of HBDI - derivatives with systematic substitutions that are illustrated in Fig. 1 b. These derivatives are generated by combining single substitutions at different positions on the phenolate ring of HBDI - (panel a) with substitutions on the imidazolinone ring, yielding a broad set of variants. To construct the fine-tuning dataset, we sample geometries for each HBDI - derivative along the relevant torsional and pyramidalisation coordinates and generate additional geometries via Wigner sampling. This step specializes the model to enable accurate predictions for the target system while retaining the transferable knowledge learned during pretraining (Fig. 1 c middle). An amount of approximately 50-100 data points per system is found to be sufficient, while earlier excited-state machine learning models usually required several hundred thousand of data points for nonadiabatic molecular dynamics. 2 ; 33 After fine-tuning, the model is used as a fast screening engine: predicted oscillator strengths can be used to generate absorption spectra and define initial conditions for excited-state dynamics simulations, while the learned excited-state potential energy surfaces and corresponding forces enable nonadiabatic molecular dynamics (see Fig. 1 c right). For photodynamics, X-MACE is interfaced with the SHARC

Read the complete text