A modular artificial intelligence framework to facilitate fluorophore design

Data collection and processing

In our previous study, a database (SMFluo1) focusing on near infrared fluorophores was constructed, containing five widely used fluorescent scaffolds²⁹. The limitation in data volume of SMFluo1 makes it difficult to meet the requirements of deep learning algorithms, particularly those based on graph neural networks (GNN), including GCN and Attentive FP³⁴, leading to moderately good prediction accuracy. In addition, to evaluate a fluorophore for bioimaging, four key photophysical parameters (λ_abs, λ_em, Φ_PL, and ε_max) are needed, where λ_abs and λ_em are related to the penetration depth and Φ_PL×ε_max indicates the brightness¹⁵. Of note, other factors including the blinking, thermal stability, photobleaching, and labeling specificity should be taken into consideration when designing probes for bioimaging, but parameters for these properties were not included in FluoDB due to limited access⁶. Taking these into consideration, data collection was carried out as follows (Fig. 2a): (i) Literature survey via searching the name of fluorescent scaffolds on PubMed; (ii) Retrieval of experimental data from various open-source databases^{23,35,36,37,38,39,40} and supplement with four photophysical parameters & solvent information from the original literature. These data were processed after the combination (see “Data processing” in Methods).

a General pipeline for data collection and processing to construct the new database, FluoDB, followed by systematic analysis and statistics of FluoDB to visualize the optical properties of various fluorophores in different solvents. b UMAP (Uniform Manifold Approximation and Projection) of different databases (Deep4Chem³⁶, DyeAgg³⁸, ChemFluor²³, ChemDataExtrator (CDEx)³⁵, and SMFluo1²⁹) using Morgan fingerprints. The UMAP algorithm was applied with a neighborhood size of 10 and a minimum distance of 0.3. The number of unique compounds in each database is listed in the bracket. c Distribution of various fluorescent scaffolds in different databases. Source data are provided as a Source data file.

Most fluorescent compounds are derived from some basic scaffolds and they may share common optical characteristics; thus, we categorized the fluorophores into twelve classic fluorescent scaffolds and four non-classical scaffolds (Fig. S1; detailed skeletal structures for 728 subgroups are shown in Table S1). FluoDB, a new database containing 35,528 unique fluorophores and 55,169 fluorophore-solvent pairs, was therefore constructed with SMILES of fluorophores/solvents, category of fluorescent scaffolds, experimental photophysical data, and original reference.

Compared to representative open-source databases (e.g., Deep4Chem³⁶, DyeAgg³⁸, ChemFluor²³, ChemDataExtrator (CDEx)³⁵, and SMFluo1²⁹), FluoDB gets improved in the number of molecules and the richness in optical information (Fig. 2b and Fig. S2), exhibiting much higher molecular diversity according to the data distribution and structural analysis (Fig. S3 and Tables S2–3). In addition, different scaffolds distribute relatively even in both FluoDB and Deep4Chem, while the data of each category is largely enriched in FluoDB (Fig. 2c and Table S4).

Data analysis with FluoDB

With FluoDB, the correlations between different parameters were investigated (Fig. S4) and indicated an obvious positive correlation between λ_abs and λ_em. In addition, molecular weight (MW) also has a certain positive correlation with λ_abs, λ_em, and ε_max, which is consistent with the scatter plot analysis (Fig. S5) and experimental results (i.e., introducing large π bridging moieties or strong electron acceptors/donors is commonly used for longer absorption and emission wavelengths⁴¹, and these modifications often increase the MW). External factors such as the surrounding solvent are reported to influence the optical properties of certain fluorophores⁴². To investigate this in a large scope, fluorophores with experimental data available for different solvents (≥5) were selected from FluoDB. The variance distribution of each photophysical parameter for selected molecules in different solvents is shown in Fig. S6, where these parameters do vary along the change of solvent type, underscoring the importance of solvents when predicting optical properties.

As mentioned earlier, we divided the fluorescent scaffolds into 16 types (Fig. S1 and Table S1), and any input fluorophore can be quickly classified accordingly to explore potential commonalities from the same group. As indicated in Figs. S7–22, a discrepancy was found in λ_abs and λ_em distribution with different scaffolds, where most groups centered in the UV-Vis range, while BODIPY, porphyrin, and squaraine lie in longer wavelength (above 550 nm)^43,44. In addition, larger wavelength tunability was found from acridine, naphthalimide, coumarin, and cyanine. Besides, Δλ (Stokes shift, Δλ = λ_em − λ_abs) of BODIPY, porphyrin, and squaraine is relatively small (~25 nm). Statistical analysis of such large-scale data is indicative for choosing the ideal fluorescent scaffold to start with, and we also prepared a toolkit where users can search for fluorophores with desired similarity as the molecule of interest from the database.

Workflow and prediction performance of FLSF

With FluoDB, open-source prediction models, including GBRT²³, SMFluo²⁹, UVVisML⁴⁵, SchNet²⁶, and ABT-MPNN⁴⁶, were tested. Data in FluoDB-Lite (SMILES in the mixture/complex form were removed from FluoDB) was divided randomly in a ratio of 7:1:2 for training, validation, and testing, respectively (Table S5). As shown in Table 1 and Table S6, ABT-MPNN, a general molecular property prediction model based on an atom-bond Transformer, performed best in predicting λ_abs and λ_em, highlighting the advantage of combining Transformers with GNN for molecular representation. However, the introduction of attention mechanisms in ABT-MPNN led to 10 times slower training than UVVisML (a Directed MPNN, D-MPNN), despite the improved MAE for λ_abs and λ_em by 9.18% and 4.86%, respectively. For practicability, it is desirable to replace the attention mechanism in ABT-MPNN with a new way to speed up the training while maintaining the prediction accuracy.

As all fluorophores in FluoDB were classified into 16 core scaffolds and distribution discrepancy in optical properties among them was observed (Figs. S7–21), a special molecular fingerprint—fluoroscaffold (a 728-dimensional digital fingerprint encoded by 728 fluorescent-scaffold subgroups listed in Table S1), fused with the current feature extraction method based on MPNN, was designed for better molecular representation of fluorophores. A new prediction model (FLSF) was constructed based on it (Fig. 3a). As shown in Table 2, FLSF predicted well for different fluorophores, especially for BODIPY-based compounds (the largest proportion in FluoDB) with MAE of 6.44 nm/7.37 nm for λ_abs/λ_em. For non-classical scaffolds (i.e., [6 + 5], [6 + 6], 6-n-5, 6-n-6), FLSF also has a good performance, promising for dealing with novel fluorophores. Overall, FLSF performs well at predicting λ_abs and λ_em (R² = 0.94) and needs improvement at Φ_PL and ε_max (R² ≈ 0.6; Fig. 3b). Then we conducted benchmark tests of FLSF and summarized the results in Table 1 and Table S6 for direct comparison with reported SOTA models. Obvious improvements were seen in the prediction accuracy of λ_abs, λ_em, and ε_max by FLSF than ABT-MPNN (the same of Φ_PL), at a much faster speed, indicating its great potential for high-throughput screening of candidate fluorophores. To check whether FLSF can capture the solvent effect, a multi-solvent test set (fluorophores with experimental data available for ≥4 different solvents) together with the control test set (fluorophores in the same solvent) was selected and the prediction performance of FLSF was compared with other baseline models. As shown in Tables S7–8, FLSF has the best prediction performance on the multi-solvent test set. To our delight, FLSF can also predict λ_abs and λ_em of fluorophores showing solvatochromism with high accuracy (Tables S9–10), further supporting its potency to capture solvent effects.

a The model architecture of FLSF. A domain-knowledge-derived fingerprint based on the 728 fluorescent-scaffold subgroups (called fluoroscaffold) is fused with a message-passing neural network (MPNN) for the feature extraction of the input fluorophore. The feature extraction of the solvent molecule is based on MPNN. The feature vectors of both the fluorophore and the solvent are input together to output a prediction of the property of interest. MLP: multilayer perceptron. b The overall prediction performance of FLSF for different photophysical parameters. λ_abs: maximum absorption wavelength; λ_em: maximum emission wavelength; Φ_PL: photoluminescence quantum yield; ε_max: molar absorption coefficient. c Comparison between FLSF (red points) and TD-DFT (time-dependent density functional theory) calculations (gray points) for λ_abs (left) and λ_em (right) prediction. MAE mean absolute error, R² the coefficient of determination. Source data are provided as a Source data file.

Time-dependent density functional theory (TD-DFT) used to be the most widely used tool for predicting optical properties⁴⁷. However, such traditional theoretical calculations require high computational and time costs⁴⁸, faced with insufficient accuracy in predicting λ_abs and λ_em, much less in parameters like Φ_PL involved in various radiation and non-radiation processes⁴⁹. For direct comparison with FLSF, we collected 162 fluorophore-solvent pairs from FluoDB (Table S11) and used TD-DFT to calculate their λ_abs and λ_em (Fig. 3c and Table S12). The MAE of FLSF decreased by more than 0.2 eV for predicting λ_abs and λ_em than TD-DFT. Of note, FLSF can provide all prediction results in less than one second, while the average calculation time of TD-DFT exceeds 200 CPU hours in the current test set.

Interpretability analysis of FLSF

The interpretability of a model, illustrating how it makes decisions and achieves related results, helps to verify the reliability of the model and excavate valuable information from the data. First, we analyzed the interpretability of FLSF from the molecule-level perspective⁵⁰. The embedding vectors from three states of FLSF were studied, namely, the state only treated by D-MPNN without fluoroscaffold integration, the state with fluoroscaffold integration but before solvent incorporation, and the state after solvent incorporation. According to the 2D-PCA (two-dimensional principal component analysis) dimension reduction distribution diagram (Fig. 4a), there is a clear difference in the distribution between short-wavelength and long-wavelength fluorophores in λ_abs and λ_em prediction tasks, indicating that FLSF can effectively identify their structural features with different wavelengths. Interestingly, the integration of fluoroscaffold makes this difference more significant, and the data distribution dispersion is further improved after the introduction of solvent, highlighting the importance of scaffold information for the prediction and implying that FLSF is sensitive in capturing subtle differences caused by solvents.

a FLSF embedding interpretability through PCA (Principal Component Analysis). 2D-PCA plots of the molecular embeddings at different stages: (Left) before integrating fluoroscaffold information, (Center) after integrating fluoroscaffold information, and (Right) after further integration of solvent information. Each dot is colored by experimental values. PC1: Principal Component 1; PC2: Principal Component 2. The intensity of the color scale represents the magnitude of the experimental values of the molecular parameters, with darker colors indicating higher experimental values. Source data are provided as a Source data file. λ_abs: maximum absorption wavelength; λ_em: maximum emission wavelength; Φ_PL: photoluminescence quantum yield; ε_max: molar absorption coefficient. b Summary of reported structural modification strategies for coumarin-based fluorophores (left) and atomic contributions learned by FLSF (right). The color bar values represent the normalized difference in the predicted values before and after masking specific atoms. EDG electron-donating group, EWG electron-withdrawing group, Exp. experimental, Pred. predicted.

Subsequently, the explicability analysis of FLSF at the atom-level perspective was conducted⁵⁰. To be specific, each atom in the fluorophore was masked, and the prediction values (e.g., λ_em) before and after masking were compared to reveal the attribution of each atom. Coumarin was taken as the example since it has been derived to cover a wide range of wavelengths, providing invaluable SPR information (Fig. 4b, left) for validating the reliability of FLSF^51,52. With a classic D-π-A structure, the introduction of electron-donating groups (EDG) on the phenyl ring and electron-withdrawing groups (EWG) on the lactone ring can effectively achieve redshift of coumarin according to experimental experience. Representative examples in Fig. 4b (right) demonstrate that FLSF has grasped such rules. In addition, researchers found that the replacement of ketone with imine at position 2 can also produce redshift⁵³ (e.g., compound e-g), and FLSF has also mastered it. Of note, although coumarin derivatives with substitution other than oxygen at position 1 are not recorded in FluoDB, FLSF can indicate the contribution of oxygen at this position to the redshift, which is also supported by recent experimental results⁵⁴. It implies that FLSF has good generalization ability/reliability and may provide new structural modification suggestions for fluorophore design.

Construction of FLAME for fluorophore design

While large databases and various property prediction models have significantly advanced our knowledge of the SPR of certain molecules, a gap exists in their direct applications for molecular design. Therefore, we aimed to build a multifunctional software package, FLAME, to meet the practical needs of researchers for novel fluorophore design by integrating the database, prediction models, and molecule generators into one framework (Fig. 5a). FLAME provides six open-source fluorophore databases, including FluoDB (Fig. 5b). Users can input the molecule of interest to search for related information of existing molecules in the database, as well as to train the model with different databases for illustrating the data impact. Meanwhile, FLAME offers six open-source prediction models (i.e., FLSF, UVVisML, ABT-MPNN, SchNet, SMFluo, and GBRT), which can be combined with the above datasets to meet various requirements from different users. In-parallel comparison between different combinations also helps to identify the best settings for specific parameter prediction (Table 3).

a The framework of FLAME to facilitate the fluorophore design. FLAME, assembled from the latest databases and prediction models, is feasible for various applications, including virtual screening, molecular generation, and structural optimization. SMILES: Simplified Molecular Input Line Entry System; MLP: multilayer perceptron; λ_abs: maximum absorption wavelength; λ_em: maximum emission wavelength; Φ_PL: photoluminescence quantum yield; ε_max: molar absorption coefficient. b The basic workflow of FLAME for various applications, including database search, photophysical property prediction, and creating unreported molecules with predicted optical properties by integrating different fluorophore databases, prediction models, and molecule generators. Databases including Deep4Chem³⁶, DyeAgg³⁸, ChemFluor²³, CDEx³⁵, SMFluo1²⁹, and our FluoDB; prediction models including previously reported UVVisML⁴⁵, ABT-MPNN⁴⁶, SchNet²⁶, SMFluo²⁹, GBRT²³, and our FLSF (FLuorescence prediction with fluoroScaFfold-driven model).

To provide structure-new compounds with predicted optical properties directly, a newly reported open-source generative AI framework, Reinvent 4³³, was introduced. As a scoring tool embedded in FLAME for molecular design, the speed of training and predicting is critical for the prediction model. FLSF proven good at these two aspects was coupled with Reinvent 4 herein (users can make their own choice). With the help of FLAME, both de novo molecular generation and structural modifications can be achieved. For example, if users are interested in the development of novel BODIPY derivatives, they can set the desired photophysical parameters (single or multiple parameters) with FLAME. Then, newly generated molecules (not recorded in FLAME’s built-in database) belonging to BODIPY with predicted properties will be screened out. Alternatively, users can input a parent structure of interest with desired parameters into FLAME to obtain optimized structures. Of note, both processes used to be highly dependent on specialized knowledge and years of experience, while FLAME is promising to think out of the box and offer fluorophore candidates more efficiently.

Experimental evaluation

With increasing interest in coumarin-based fluorescent probes due to their excellent biocompatibility, good structural flexibility, and tunable fluorescence⁵², FLAME is employed to guide the development of novel coumarin derivatives for concept proof (Fig. 6a and Fig. S23). Four optical parameters (λ_abs, λ_em, Φ_PL, and ε_max) were set as scoring targets. We trained the generative model and sampled one million molecules, with a focus on coumarin-type compounds during screening. From the virtual library generated by FLAME, 3,4-oxazole-fused coumarins attracted our attention due to their structural novelty and synthesizability. A variety of oxazole-containing dyes were reported to possess attractive photophysical properties, such as high fluorescence quantum yields^55,56,57, while the fluorescence properties of 3,4-oxazole-fused coumarins have not been reported yet.

a Generation of unreported heterocyclic-fused coumarins with predicted optical properties by FLAME. FLSF: FLuorescence prediction with fluoroScaFfold-driven model; Reinvent 4³³: a newly reported open-source generative AI framework. b The new strategy for one-pot synthesis of 3,4-oxazole-fused coumarins. c Absorption (up) and emission (down) spectra of 3h and 3o recorded in different solvents. H₂O: water; DMSO: dimethyl sulfoxide; EtOH: ethanol; DCM: dichloromethane. Source data are provided as a Source data file. d Confocal fluorescence images of living HeLa cells treated with different concentrations of 3o. The cell imaging was performed three times with similar results.

Available strategies for the synthesis of this scaffold include (a) heating of 7-N,N-dimethylamino-4-hydroxycoumarin in the presence of nitromethane and DABCO⁵⁸, (b) synthesis from 4-hydroxy-3-nitrocoumarin and benzyl alcohol under gold nanoparticle or FeCl₃ catalysis⁵⁹, and (c) synthesis from 4-hydroxy-3-nitrocoumarin and acids in the presence of triphenylphosphine and phosphorus pentoxide under microwave irradiation⁶⁰ (Fig. S24). The lack of structural diversity on the phenyl ring using reported strategies, together with the demand for simple and efficient synthetic procedures to construct diverse 3,4-oxazole-fused coumarins from readily available starting materials, drives us to develop new synthetic methodology. Inspired by our previous work in isocyanide chemistry^61,62,63, we proposed a one-pot approach to synthesize 3,4-oxazole-fused coumarins from ethyl isocyanoacetates and phenyl salicylates promoted by base (Fig. 6b and Fig. S25). Under the optimized conditions (Table S13), 16 oxazole-fused coumarins were synthesized successfully (Figs. S26–27), carrying electron-donating or electron-withdrawing substituents on the phenyl ring. With these compounds in hand, their optical properties were evaluated (Figs. S28–29). Consistent with the prediction result from FLSF, the introduction of an amino group at the 6- or 7-position of the coumarin scaffold (i.e., 3h, 3o) led to a redshift in λ_abs and an increase in Φ_PL (Table S14). Then, solvent effects on these two fluorophores were investigated (Fig. 6c and Table S15). As expected, the solvent polarity has a significant impact on their absorption/emission wavelengths, and 3o liberated much stronger emission than 3h in all solvents, which was selected for bioimaging. Brilliant fluorescence was observed in HeLa cells after 30-min incubation (Fig. 6d), indicating its potential for live-cell imaging.