【Reading guide】
Intramolecular charge transfer (ICT) is one of the most important photophysical mechanisms in organic fluorophores. Among ICT processes, TICT (Twisted Intramolecular Charge Transfer) and PICT (Planar Intramolecular Charge Transfer) represent two highly representative yet frequently confused mechanisms. Although their ground-state structures appear remarkably similar, their excited-state conformations and emission behaviors diverge dramatically. This “similar structures but opposite properties” paradox has long hindered the rational design of fluorescent molecules, making probe development costly, time-consuming, and difficult to scale to large molecular libraries. To address this challenge, the authors Prof. Jie Dong and Prof. Wenbin Zeng from the Xiangya School of Pharmaceutical Sciences, Central South University employed interpretable artificial intelligence to unveil the deep chemical structural essence distinguishing TICT and PICT fluorophores at a systematic level. They further proposed AI-guided design rules for intelligent fluorophore development, significantly improving design efficiency. The key highlights of the study include: (1) Constructing the first comprehensive TICT and PICT fluorophore dataset, covering molecules from nearly a decade of research. (2) Using interpretable algorithms to successfully identify the key factors that critically influence TICT and PICT mechanisms. (3) Releasing an easy-to-use decision tree only based on simple molecular descriptors and fingerprints, ensuring a fast decision and modification when designing TICT and PICT molecules. (4) Proposing the first AI-guided structural design rules for TICT and PICT fluorophores. (5) Conducting both experimental tests and quantitative calculations which confirmed the potential of the approach for the efficient and reliable discovery of TICT and PICT fluorophore candidates.

【Abstract】
D-π-A type fluorescent materials are crucial tools in life sciences and medicine, with their development hinging on a precise understanding of fluorophore mechanisms, particularly twisted intramolecular charge transfer (TICT) and planar intramolecular charge transfer (PICT) processes. These fluorophores exhibit unique charge transfer properties, making them highly valuable in organic optoelectronics, fluorescent probes, and sensors. However, despite their growing applications, the structural essence of TICT and PICT fluorophores remains poorly understood. This often results in molecules with similar structures displaying charge transfer modes that contradict design expectations, substantially hindering the application of TICT and PICT fluorescent probes. In this study, we meticulously designed various computational strategies based on interpretable machine learning to thoroughly deconstruct the chemical structural essence of TICT and PICT fluorophores. Utilizing the first real-world TICT and PICT dataset, we constructed predictive models that balance both interpretability and accuracy (area under the receiver operating characteristic curve = 0.846) using a range of algorithms, including deep learning. We established artificial intelligence (AI)-guided rules comprising 5 structural factors—electron-donating group strength, electron-withdrawing group strength, alkyl cyclization, steric hindrance, and solvent–solute interactions—that influence TICT and PICT. These rules provide obvious guidance for probe design based on molecular rigidity and charge transfer driving forces. Compared to community-suggested rules, the AI-guided rules achieved an over 20% improvement in accuracy in a controlled evaluation. By applying these rules, we successfully synthesized and validated several representative fluorophores that are challenging to distinguish using chemical intuition alone. Both quantitative calculations and experimental results confirmed the accuracy of the model and the practicality of the AI-guided rules. This novel approach is expected to establish a novel paradigm for exploring ideal TICT and PICT molecules, offering a robust framework for future research and application in fluorescent materials.

【Methods and Strategies】
The study establishes an innovative strategy that deeply integrates interpretable artificial intelligence with quantitative calculations to systematically decode the chemical structural essence of TICT and PICT fluorophores (Fig. 1). Unlike traditional quantum chemical methods, leveraging AI to analyze the intrinsic nature of TICT and PICT requires simultaneously addressing several key challenges: achieving high predictive accuracy, maintaining computational efficiency, and, most importantly, enabling the model to truly understand how structural variations influence TICT and PICT behaviors. To accomplish this, the research team compiled the first accurately labeled dataset of TICT and PICT fluorophores by systematically curating large-scale literature from 2012 to 2022. During model construction, the team sequentially evaluated 12 traditional machine learning algorithms and a deep learning model, progressing from simple to complex architectures to strike an optimal balance between predictive performance and model interpretability. By employing multiple model interpretation methods, the authors distilled five AI-guided design rules that reveal the key structural factors governing TICT and PICT behaviors, particularly from the perspectives of molecular rigidity and charge-transfer driving force. The study further introduces a more intuitive and practical visualization tool that demonstrates the step-by-step decision-making process for distinguishing TICT from PICT molecules to promote broader real-world applications. Compared with probes designed based on traditional chemical intuition, those designed using AI-guided rules exhibit significantly improved structural design accuracy. To validate the accuracy of the proposed guiding rules, the research team selected entirely new compounds whose behavior could not be readily predicted by empirical intuition. The rules were first examined from a computational perspective through excited-state potential energy surface scanning and electron–hole analysis, followed by the synthesis of the compounds and rational validation through wet-lab experiments.

【Main Results】
1.Construction of the TICT and PICT dataset
Accurately distinguishing TICT from PICT molecules remains highly challenging beyond experimental methods, leading to a long-standing scarcity of reliable data. To ensure robust model training and generalization, the research team systematically screened more than 10,000 publications from the past decade (2012-2022). Using the excited-state potential energy surface (PES) scanning results reported in the literature as the criterion for classification, the team manually examined and identified potential TICT and PICT molecules one by one. The initial collection contained more than 3,000 candidate structures. These candidates were then converted into InChIKey format using ChemDes and PyBioMed, strictly deduplicated, and further split and retained when structural entries had inconsistent labels. Ultimately, the finalized dataset consisted of 511 TICT molecules and 369 PICT molecules. To enable the models to effectively capture key factors influencing TICT and PICT mechanisms, each molecule was further computed into 12 categories of molecular representations, including 2D descriptors, three semi-empirical 3D descriptors, and eight types of molecular fingerprints, comprehensively covering essential chemical information ranging from molecular topology, conformation, and electronic structure to substructural fragments.

2.Accurate prediction of TICT and PICT probes
Based on this high-quality dataset, the authors constructed multiple models by combining 12 machine learning algorithms with the deep learning framework MolMapNet. For both 2D descriptors and eight molecular fingerprint types, they conducted 50 rounds of repeated training with different dataset partitions, selecting the best-performing algorithm in each round. Ultimately, the XGBoost model based on AtomPair fingerprints achieved the best overall predictive performance, with a test accuracy of 0.784 and an AUC of 0.846. Among the semi-empirical 3D descriptor models, AM1, PM3, and MNDO all showed strong performance, with AM1 achieving the highest test accuracy of 0.730. In comparison, MolMapNet exhibited slightly weaker generalization due to the limited dataset size, with an accuracy of 0.716, although its multi-descriptor fusion approach still offers valuable complementary
interpretability (Fig. 2).

3.Interpretable contribution of important structures to TICT and PICT
The contribution analysis of important structures pinpoints the key structures that most influence the predictions of TICT and PICT. The team conducted a multidimensional interpretability analysis of the optimal models from four perspectives: 2D descriptors, 3D descriptors, MACCS fingerprints, and deep learning multimodal features. While the feature importance may not be entirely consistent due to differences between the selected algorithms, the most important features always have relatively high scores. All four interpretability approaches converged on the conclusion that conformational rigidity, molecular steric hindrance, electronic effects, and solvent-related factors collectively define the fundamental differences between TICT and PICT (Fig. 3).

4.Practical simplified model based on interpretable descriptors
To enhance usability in practical applications, the team developed a more concise and operational lightweight model based on the preceding interpretability analysis. They first examined the correlations between the top 20 most important features and the true labels, where the highest correlation coefficient reached 0.67. However, the distribution of the five most critical descriptors indicated that although each individual feature showed some association with TICT and PICT mechanisms, none was sufficient for accurate discrimination on its own. Building on this, the researchers constructed two simplified decision-tree models using MACCS fingerprints and 2D+3D descriptors, respectively. Although the decision-tree models perform below the optimal XGBoost model, both achieve a favorable balance between interpretability and predictive ability, serving as practical tools for researchers to quickly assess TICT and PICT behaviors (Fig. 4).

5.Deconstruction of the structural essence of TICT and PICT
To deeply explore the mechanisms of TICT and PICT, we deconstructed the structural essence of TICT and PICT from the perspective of 36 core molecular descriptors. The results showed that the most influential descriptors were closely associated with alkyl cyclization, steric effects, and solvent-solute interactions. When these factors are strengthened, the system tends to favor PICT, whereas weaker values shift the equilibrium toward TICT. More importantly, the team found that descriptors related to charge and energy are equally critical. When EDG/EWG strength increases, the compound tends to exhibit TICT, and the reverse favors PICT. Based on these insights, the study distilled five AI-derived structural rules: EDG strength, EWG strength, alkyl cyclization, steric hindrance, and solvent-solute interactions (Fig. 5). Among these, EDG and EWG strength together account for approximately 56% of the total contribution, serving as the dominant factors governing charge-transfer dynamics, whereas cyclization, steric effects, and solvent interactions influence the TICT/PICT mechanism by modulating conformational freedom and environmental stabilization.

6.Validation of AI-guided rules in TICT and PICT probe design
To validate the practicality of the AI-guided rules, the study designed 2 TICT and PICT molecular libraries: one based on community-suggested rules and the other utilizing AI-guided rules. The team conducted a 3-level systematic evaluation to assess the differences between the 2 design strategies. In the first validation stage, the best-performing model was used to evaluate both molecular libraries. The AI-guided library reached a design accuracy of 90.96%, significantly higher than the 67.47% obtained using traditional chemical intuition (Fig. 6). To further assess the real-world applicability of the AI-derived rules, two previously unreported and mechanistically ambiguous molecules (Compound 1 and Compound 2) were selected for additional validation. In the second stage, excited-state PES scanning and electron–hole analysis showed that Compound 1 displays clear TICT characteristics, whereas Compound 2 remains strictly PICT. The third stage of experimental validation confirmed that Compound 1 exhibited strong viscosity sensitivity, while Compound 2 showed stable PICT emission with no viscosity response. The three validation layers demonstrate that the AI-derived design rules can reliably guide the precise design of TICT and PICT fluorophores (Fig. 7).

【Conclusion】
In conclusion, this work proposes a novel AI-based approach to deconstructing the structural essence of TICT and PICT without relying on extensive computational cost, thereby facilitating the rational design of the corresponding fluorescent probes. The authors first collected a high-quality dataset of TICT and PICT molecules and established a series of machine learning models to distinguish between the two. The best model achieved an accuracy of 0.784 on the test set. More importantly, benefiting from the model’s excellent interpretability, five AI-guided design rules based on molecular rigidity and charge-transfer driving forces were uncovered: EDG strength, EWG strength, alkyl cyclization, steric hindrance, and solvent–solute interactions. A more intuitive and practical tool was also created to make these rules more accessible for researchers in real-world design scenarios. Compared with fluorophores designed using community-suggested rules, those designed using the AI-guided rules showed a marked improvement: the design accuracy increased from 67.47% to 90.96%. These results were fully validated through both quantum chemical calculations and wet-lab experiments. Therefore, this study not only effectively deconstructs the structural essence of TICT and PICT but also represents a meaningful step toward the intelligent design of their molecular probes.

The complete study is accessible via DOI:10.34133/research.1021