Skip to main content
NCBI home page
JACS Au logoLink to JACS Au
. 2021 Mar 17;1(4):427–438. doi: 10.1021/jacsau.1c00035

Deep Learning Optical Spectroscopy Based on Experimental Database: Potential Applications to Molecular Design

Joonyoung F Joung 1, Minhi Han 1, Jinhyo Hwang 1, Minseok Jeong 1, Dong Hoon Choi 1, Sungnam Park 1,*
PMCID: PMC8395663  PMID: 34467305

Abstract

graphic file with name au1c00035_0008.jpg

Accurate and reliable prediction of the optical and photophysical properties of organic compounds is important in various research fields. Here, we developed deep learning (DL) optical spectroscopy using a DL model and experimental database to predict seven optical and photophysical properties of organic compounds, namely, the absorption peak position and bandwidth, extinction coefficient, emission peak position and bandwidth, photoluminescence quantum yield (PLQY), and emission lifetime. Our DL model included the chromophore–solvent interaction to account for the effect of local environments on the optical and photophysical properties of organic compounds and was trained using an experimental database of 30 094 chromophore/solvent combinations. Our DL optical spectroscopy made it possible to reliably and quickly predict the aforementioned properties of organic compounds in solution, gas phase, film, and powder with the root mean squared errors of 26.6 and 28.0 nm for absorption and emission peak positions, 603 and 532 cm–1 for absorption and emission bandwidths, and 0.209, 0.371, and 0.262 for the logarithm of the extinction coefficient, PLQY, and emission lifetime, respectively. Finally, we demonstrated how a blue emitter with desired optical and photophysical properties could be efficiently virtually screened and developed by DL optical spectroscopy. DL optical spectroscopy can be efficiently used for developing chromophores and fluorophores in various research areas.

Keywords: deep learning, optical property, photophysical property, fluorophore, chromophore, chromophore-solvent interaction, experimental database

Introduction

The optical and photophysical properties of chromophores and fluorophores are important for industrial and academic applications in a variety of research fields such as, organic solar cells,1 organic light-emitting diodes (OLEDs),2 bioimaging,3 organic sensors,4 and organic dyes. Substantial efforts have been made to design and synthesize new colorful materials based on various chromophores and fluorophores for various purposes. To develop new molecules with desired properties, knowledge of the chemical and physical properties of previously studied molecules is required. In particular, the optical properties of molecules, such as absorption and emission peak positions and bandwidths, extinction coefficients, photoluminescence quantum yield (PLQY), and emission lifetime, are important for the development of chromophores and fluorophores. For given molecules, the aforementioned optical properties are readily measured using various spectroscopic instruments, such as spectrophotometers, spectrofluorometers with integrating spheres, and time-resolved fluorescence (TRF) spectrometers. It is also desirable that the optical properties of newly designed chromophores and fluorophores are accurately estimated in advance to reduce development costs. To estimate the optical properties of molecules, empirical rules or theories, such as the Woodward–Fieser rule,5 first-principles,6,7 and time-dependent density functional theory (TD-DFT) calculations8−10 have been developed. Such methods have limitations resulting from the limited range of molecules to which they can be applied and from the lack of practical theories to calculate specific properties. While these rules and theories have been used to successfully predict many chemical and physical properties, more accurate, facile, and computationally inexpensive methods are still needed.

In this study, we have developed a deep learning (DL) method to reliably and quickly predict the optical and photophysical properties of organic compounds. Our DL method includes chromophore–solvent interactions to account for the effect of environment on the optical and photophysical properties. Our DL model was trained using an experimental database of 30 094 chromophore/solvent combinations consisting of 11 392 unique organic chromophores in 369 solvents or solid states.11,12 Our DL method is found to readily predict seven optical properties such as the first absorption peak position and bandwidth, extinction coefficient, emission peak position and bandwidth, photoluminescence quantum yield, and emission lifetime. The root mean squared errors of the predicted values were found to be 26.6 and 28.0 nm for absorption and emission peak positions, 603 and 532 cm–1 for absorption and emission bandwidths, and 0.209, 0.371, and 0.262 for the logarithm of the extinction coefficient, PLQY, and emission lifetime, respectively. Using these predicted optical properties of organic compounds, the electronic absorption and emission spectra are shown to be readily reconstructed, and the colors of organic molecules associated with the absorption and emission are estimated in Commission Internationale de l’Eclairage (CIE) color coordinates. Finally, we demonstrate that the DL model can be effectively utilized to prescreen newly designed molecules to find target molecules that exhibit the desired optical and photophysical properties.

Results and Discussion

Experimental Optical Properties of Organic Compounds

The absorption properties of a chromophore in a solvent are characterized by the first absorption peak position (λabs), bandwidth (σabs), and extinction coefficient (ε), whereas the emission properties of a fluorophore are characterized by the emission peak position (λemi), bandwidth (σemi), and photoluminescence quantum yield (PLQY, Φ). In addition, the emission (excited-state) lifetime (τ) is an intrinsic property of fluorophores that is importantly considered in fluorescence lifetime imaging and dynamic quenching. The experimental database of these optical and photophysical properties of organic compounds is prerequisite for our DL method. As shown in Figure 1, the experimental database includes 11 392 unique organic chromophores in 369 solvents or solid states, yielding 30 094 chromophore/solvent combinations. Briefly, the experimental database was built by collecting seven optical and photophysical properties of organic compounds from the literature and has been described in detail elsewhere.11,12 Not all seven optical and photophysical properties were available for each chromophore/solvent pair in the database. A variety of chromophore structures are included in the database; the frequencies of specific core structures are summarized in Figure S1. For example, more than 1300 BF2-group-containing boron-dipyrromethene (BODIPY) derivatives are found in the database.

Figure 1.

Figure 1

Open in a new tab

Database of the optical properties of organic compounds. Histograms of (a) first absorption peak position (λabs), (b) bandwidth in full width at half-maximum (σabs), (c) extinction coefficient in logarithm (log ε), (d) emission peak position (λemi), (e) bandwidth in full width at half-maximum (σemi), (f) photoluminescence quantum yield (Φ), (g) lifetime (τ), (h) molecular weights of chromophores, and (i) solvents (CH2Cl2 dichloromethane, CH3CN acetonitrile, Tol toluene, CHCl3 chloroform, THF tetrahydrofuran, MeOH methanol, EtOH ethanol, DMSO dimethyl sulfoxide, CH cyclohexane, and DMF N,N-dimethylformamide). The number of data points (N) and the number of chromophores (mol.) are included in each graph.11,12

DL Model for Prediction of Optical and Photophysical Properties

DL based on big data has attracted substantial attention because it can potentially solve problems via direct learning from data without well-defined rules, empirical laws, or theories.13,14 In chemistry, DL has proven promising for predicting various properties of molecules and materials,15−29 optimizing reactions and retrosynthesis,30−32 and designing new molecules including drugs33,34 and materials.35−37 Additionally, DL has been effectively used in computational chemistry.38,39 Nakata and Shimazaki reported a large-scale electronic structure database (PubChemQC)40 and recently, using the PubChemQC database, Lee and co-workers reported a random forest model to predict the highest oscillator strength and the corresponding excitation energy of molecules.26 Lin and co-workers used a DL method to predict the bandgap of configurationally hybridized graphene and boron nitride,21 and Chang and co-workers used a tuplewise material representation to predict the band gap of organic–inorganic perovskite, 2D materials, and binary and ternary inorganic materials.27 Jiang and co-workers reported DL models that could predict infrared and ultraviolet absorption spectra from the conformations of a molecule using a theoretical database that was generated by molecular dynamics simulations and DFT calculations.23,28 Rinke and co-workers reported a DL method to predict molecular excitation spectra based on the QM7b and QM9 data sets of small organic molecules containing up to 17 C, N, O, S, and halogen atoms.24 Li and co-workers reported a machine learning method using the Morgan fingerprint and an experimental database of 4300 molecules to predict three properties such as absorption and emission wavelengths and PLQY.29

In this work, we developed a DL model to predict the optical and photophysical properties as described in Figure 2. The molecular structure of a chromophore is represented by a 150 × 150 adjacency matrix (Achm) and a 150 × 43 feature matrix (Xchm) (Figure S2). The adjacency matrix includes the connectivity information on atoms in the chromophore, and the feature matrix (Xchm) includes the identity of the atoms, the number of hydrogen and heavy atoms bonded to them, aromaticity, hybridization, ring structure, and formal charge. The graph convolutional network (GCN) is the simplest version of the message-passing neural network, which is effective for graph structures such as molecular structures.16,25,41−44 A single graph convolution of the adjacency matrix (Achm) and feature matrix of the lth layer (Hchml) updates the atom’s features, producing a new feature matrix for the l + 1th layer, Hchm, which contains all the information related to the nearest-neighboring atoms in the chromophore. The final feature matrix of GCN layers, Hchm(f), is reduced to a row vector, zchm, by summing all elements to secure permutation invariance. The row vector, zchm(0), passes through multilayer perceptrons (MLPs) to produce a chemical space vector, zchm, for the chromophore. Similarly, a chemical space vector, zsol(f), of the solvent is obtained. The chemical space vectors for the chromophore and solvent, zchm and zsol(f), are concatenated and pass through the MLPs to construct an interaction vector. Finally, the interaction vector passes through the MLPs to obtain the seven optical properties. Note that the interaction vector in the DL model is used to account for the chromophore–solvent interaction.

Figure 2.

Figure 2

Open in a new tab

Our deep learning (DL) model. The interaction vector is used to account for the chromophore–solvent interaction for predicting the optical and photophysical properties of the chromophore.

It should be highlighted that our DL model includes the chromophore–solvent interaction for accurate estimation of the optical and photophysical properties of organic compounds in various solvents. Details of the DL model and training process are provided in the Supporting Information and Methods. The performance of our DL model is presented in Figure S4. Overall, the DL model is reasonably well trained using the experimental database and the square of the Pearson correlation coefficient (R2) ranges from 0.93 to 0.71 for the test sets (Table S2). It should be noted that σabs, σemi, Φ, and τ for a chromophore cannot be well predicted by currently available theoretical methods because they are highly dependent on many factors, including solute–solvent interactions, density of energy states, and relaxation pathways, which cannot be fully considered in the theoretical methods. However, in our DL model, these properties are shown to be directly learned from the experimental database and reliably predicted, which is a great advantage of the DL method.

Performances of Our DL Model

The DL model trained with the experimental database can be used to achieve DL optical spectroscopy such that seven optical properties for any given molecule are easily predicted. In this section, several interesting examples from the results of our DL model will be discussed.

Structural Diversity of Chromophores

As shown in Table 1a, our DL model predicts the optical properties of widely used core structures of chromophores and fluorophores45−48 such as Nile red,49 Prodan,50 methylene blue,51 and l-tyrosine.52 In addition, we compared the λabs and λemi values predicted by our DL model and TD-DFT calculations53 for 131 combinations involving 53 chromophores (Figure S5 and Table S6). The RMSEs of λabs and λemi predicted by our model are 17.0 and 22.5 nm, respectively, which are 2–3.5 times smaller than those obtained by TD-DFT calculations (46.1 and 81.9 nm, respectively). The biggest advantage of DL method is that the information hidden in the experimental database is directly decoded by the DL model. As a result, the DL model can predict the properties (e.g., σabs, σemi, Φ, and τ) that are traditionally incomputable by theoretical methods. Compared to theory-based TD-DFT calculations, another advantage of the DL method is its low computational cost. The TD-DFT calculations required 236 min per molecule on average to compute the ground- and excited-state geometries, λabs and λemi. Our DL model required only 20 s to simultaneously predict the seven optical properties of 131 chromophore/solvent combinations on the same computer (see Methods for computational details).

Table 1. Prediction of Optical Properties Using Our DL Modela.

graphic file with name au1c00035_0006.jpg

graphic file with name au1c00035_0007.jpg

Open in a new tab
a

(a) Various optical properties of chromophores with different molecular structures. (b) BODIPY derivatives (in red) with different moieties (in green) at the meso position in dichloromethane. (c) Three protonation states of 7-amino-2-naphthol in water and their experimentally measured and predicted values.

For donor–acceptor (D–A) type molecules, the DL model is found to recognize the core moiety that has a more significant effect on the overall optical properties. Table 1b shows D–A type molecules with BODIPY and other moieties.54−58 Our DL model confirms that the emission properties (λemi) of the molecules are mainly determined by the BODIPY moiety. For example, for the anthracene (donor)–BODIPY (acceptor) structure, the emission peak position of BODIPY is longer than that of anthracene, and thus, BODIPY determines the emission peak position of the molecule according to Kasha’s rule.59

In addition, we further examined whether the DL model could distinguish the optical properties of chromophores with similar structures. Ninety-seven chromophores in dichloromethane were selected from the experimental database (Table S7), and their optical properties were obtained by the DL model and compared with the experimental values (Figure S6). The DL model recognizes the core structure of a given molecule that determines the optical properties and the effect of structural modification of the core structures on the optical properties such as changes in the conjugation length and substituents of the core structures. The diverse changes in the core structures were successfully accounted for in the prediction of the optical properties.

Lastly, a change in protonation state is considered the smallest possible modification to a molecular structure which can dramatically alter its optical properties; this is because protonation/deprotonation changes the chromophore charge distribution and resultantly the chromophore–solvent interactions. Fifty-nine molecules with two or more protonation states were identified in the database (Figure S7). The effects of protonation/deprotonation in acid–base reaction on the optical properties are well predicted by the DL model. For example, the λabs and λemi of the three protonation states of 7-amino-2-naphthol in Table 1c are well distinguished.60

Effects of Solvent Polarity

The optical properties are significantly influenced by the solvent polarity known as solvatochromic shifts including bathochromic shift (red shift) and hypsochromic shift (blue shift). In the DL model, chromophore–solvent interactions are included, allowing reliable prediction of solvatochromic shifts. The solvatochromic shifts of Reichardt’s dye61 (Betaine 30, Figure 3a) in 334 solvent molecules are well predicted by the DL model. The DL model is found to quite accurately predict the λabs of Betaine 30 in solvents with similar dielectric constants (εr): 1,2-dichloroethane (εr = 10.36, λabsexp = 692 nm, λabs = 692 nm), 1,1-dichloroethane (εr = 10.46, λabsexp = 726 nm, λabs = 701 nm), and 2,2-dichloropropane (εr = 11.63, λabsexp = 754 nm, λabs = 715 nm). These results indicate that the chromophore–solvent interactions in the DL model were sufficiently well implemented.

Figure 3.

Figure 3

Open in a new tab

Solvent effects on optical properties predicted by our DL model. (a) Experimental vs predicted λabs values of Reichardt’s dye (Betaine 30) in 334 solvents. The λabs of Reichardt’s dye exhibits a hypsochromic (blue) shift from 1000 to 450 nm with increasing solvent polarity. (b) Experimental vs predicted λemi values of dopants in host matrices (films). (c) Experimental vs predicted Φ values of molecules exhibiting aggregation induced emission in solutions or in solid states.

Organic fluorophores can also be incorporated as dopants in solid-state host matrices. Such systems are widely used in OLEDs.62 Our database includes 48 host molecules and 361 fluorophore(dopant)–host systems developed for OLEDs. The matrix effects on the emission properties of the dopants are well predicted (Figures 3b and S8). Estimation of the PLQY of dopants in host matrices is critical for designing dopant molecules, but the PLQYs of molecules in solutions or in films are challenging to estimate theoretically. The DL model could be effectively used to prescreen candidate dopant molecules by accurately estimating their PLQY.

In solid states, nonradiative processes associated with molecular rotations and vibrations are inactive and relatively high fluorescence quantum yields can be achieved. Aggregation-induced emission (AIE) has recently become an active research topic,63 and we investigated whether our DL model could explain AIE. In our DL model, the effects of local environments are implemented using the adjacency and feature matrices of the solvents (Asol and Xsol), which can be treated as either the molecule itself for one-component solids (i.e., molecular solids) or as the host matrix for solid solutions. Therefore, the optical properties of chromophores in solids are able to be trained and predicted by the DL model. Note that amorphous and crystalline phase solid states are not distinguished in our database. Our database contains 730 examples of chromophores with AIE properties (Figure S9 and Table S3), and their Φ are indeed predicted to be larger in aggregates (molecular solids) than in solutions as shown in Figure 3c. In short, the effects of solvent and local environments on the optical properties of chromophores were successfully implemented to account for both solvatochromic shifts and the AIE effect.

Deep Learning Optical Spectroscopy

For a molecule, absorption and emission properties, which are quantified using spectroscopic measurements, are characterized by their position (λabs/λemi), bandwidth (σabs/σemi), and amplitude (ε/Φ). For any given organic compound, these optical properties are readily predicted by our DL model and can be used to simulate the absorption and emission spectra as described in Methods, making DL optical spectroscopy possible. Furthermore, the daylight and emission colors of a molecule in solution and films under the daylight and UV-lamp can be estimated based on the absorption and emission spectra, respectively.

The absorption and emission spectra of coumarin 153 in ethanol were predicted using the DL model and TD-DFT calculations and are compared in Figure 4a.64 The daylight and emission colors of coumarin 153 in ethanol are successfully reproduced by the DL model. In contrast, TD-DFT calculations estimated the daylight color well but not the emission color. For a molecule, λabs, λemi, ε, and the radiative rate constant can be estimated by theoretical calculations, but σabs, σemi, and Φ are highly challenging to predict using currently available theoretical methods. The performance of the DL model was further examined using the thermally activated delayed fluorescence (TADF) dye, BPCP-2CPC, in the host molecule, C-2PC.65 While the optical properties of BPCP-2CPC in toluene are included in our database, those of BPCP-2CPC in C-2PC are not. The emission spectrum and a photograph of BPCP-2CPC in C-2PC are presented in Figure 4b. The emission spectrum of BPCP-2CPC in C-2PC was reasonably well predicted by the DL model, and the emission color is well represented in the CIE 1931 xy chromaticity space (Figure S10). Figures 4c, d and S11 show the photographs of chromophores in various environments along with the emission colors predicted by the DL model.66,67 The solvent-polarity-dependent colors are well reproduced. Accordingly, our DL optical spectroscopy is successfully implemented to predict the optical properties of chromophores in solutions, films, and powders.

Figure 4.

Figure 4

Open in a new tab

DL optical spectroscopy. Experimentally measured and DL predicted absorption (black) and emission (red) spectra. (a) Coumarin 153 in ethanol. (b) BPCP-2CPC (molecule) in C-2PC (host). The bandwidths in fwhm for the calculated spectra are set to 5370 cm–1 (the default value in GaussView 6). (c) Photograph of (E,E,E)-2-(4-diphenylaminostyryl)-4,6-bis(4-methoxystyryl)pyrimidine in several solvents. The colors below the photograph are those predicted using our DL model [Reproduced with permission from ref (66). Copyright 2018 American Chemical Society]. (d) Photographs of solid state emission. The colors below the photograph are those predicted using our DL model [Reprinted with permission from ref (67). Copyright 2016 Published by Elsevier Ltd.].

Development of a Target Chromophore by Virtual Screening via DL Optical Spectroscopy

DL optical spectroscopy could be effectively used to select target chromophores with desired optical and photophysical properties from many predesigned candidate molecules as shown in Figure 5a. Many potentially synthesizable molecular structures are predesigned, their seven optical properties are predicted by DL optical spectroscopy, a target molecule with desired properties is selected and synthesized, and finally the optical and photophysical properties of the synthesized target molecule are characterized by many spectroscopic methods and whether the optical and photophysical properties match those desired is evaluated. In this way, new target molecules optimized for a given purpose can be efficiently developed using DL optical spectroscopy.

Figure 5.

Figure 5

Open in a new tab

Development of a target chromophore by virtual screening. (a) Overview for development of a new chromophore by DL optical spectroscopy. (b) Molecular structures of 3 molecules based on a DOBNA core with carbazole (1), phenoxazine (2), and diphenylamine moieties (3), and the optical properties predicted by DL optical spectroscopy. (c) Absorption and emission spectra of compound 1 in toluene. σabs cannot be experimentally measured because the S1 transition is overlapped with higher electronic transitions. (d) Time-resolved fluorescence signal of compound 1 in toluene.

This section demonstrates how DL optical spectroscopy can be practically used in the development of a new blue fluorophore with a donor (D)–acceptor (A) type. Figure 5b shows three organic molecules with a DOBNA backbone which could be readily predesigned. And then their optical properties in toluene were predicted by DL optical spectroscopy as summarized in Table S4. Depending on the donor unit such as carbazole (1), phenoxazine (2), and diphenylamine moieties (3) as shown in Figure 5b, the predesigned molecules are predicted to exhibit quite different emission colors, very narrow emission bandwidths, and decent PLQYs. Note that DOBNA core structures were shown to exhibit such narrow bandwidths and high PLQYs.68−70 Among three compounds in Figure 5b, Compound 1 is a fluorophore with a deep blue color (λemi = 428 nm and CIE y < 0.1), narrow emission bandwidth (σemi = 29 nm), and decent PLQY (Φ = ∼0.7). As a desired deep blue emitter, compound 1 was synthesized.

As shown in Figure 5c and d, the optical properties of compound 1 were experimentally measured in toluene, which is a typical solvent for optical measurements of organic compounds, and are summarized in Table S5. The optical properties of compound 1 in toluene are found to agree reasonably well with the values predicted by DL optical spectroscopy within the RMSEs although the experimentally determined values are somewhat slightly smaller than the predicted values. The λabs and λemi in the absorption and emission spectra are ∼30 nm blue-shifted compared to the predicted values. But the emission bandwidth (σemi = 22 nm), PLQY (Φ = 0.543), and lifetime (τ = 5.1 ns) are reasonably well predicted. Furthermore, the CIE coordinates are found to be (0.157, 0.020), which is close to the predicted values (0.153, 0.020).

Conclusions

We have successfully developed a DL method based on an experimental database to predict the optical properties of chromophores or fluorophores in various environments. Our DL model was trained using an experimental database of 30 094 chromophore/solvent combinations consisting of 11 392 unique organic chromophores in 369 solvents or solid states.11,12 Seven optical properties of each organic molecule can be reliably estimated using the DL model and used to reconstruct the absorption and emission spectra and predict the real-world color of the molecule, achieving DL optical spectroscopy.

DL optical spectroscopy can be effectively used to prescreen whether newly designed chromophores and fluorophores exhibit desired optical properties before they are actually synthesized, which would reduce development costs significantly. We successfully demonstrated a strategy to developing a new deep blue emitter with desired optical properties using virtual screening by DL optical spectroscopy. By utilizing DL optical spectroscopy, the development of organic molecules for electronic and optoelectronic applications will be boosted.

Our DL method represents a framework that could be generally applicable to predict other chemical and physical properties of organic compounds. The experimental database of 30 094 chromophore/solvent combinations is currently in use but can be continuously extended to include a broader range of core structures. Therefore, the prediction accuracy of optical properties and the coverage of molecular structures by our DL optical spectroscopy will be continuously improved in the future. Recently, DL models that generate valid molecular structures based on a variational autoencoder (VAE) and a generative adversarial network (GAN) have been reported,71−76 and by combining these generative DL models, our DL optical spectroscopy is expected to be used more efficiently in the development of new fluorophores and chromophores. Lastly, using our DL model we developed a web-based interface, “Deep4Chem”,77 to predict the optical properties of organic compounds as well as the electronic absorption and emission peaks and the corresponding original colors which the materials would have, associated with absorption and emission peaks, in CIE color coordinates.

Methods

Training Our DL Model

Our DL model was trained using an experimental database. In the training process, the loss function was defined by the total mean squared error which was obtained by summing the mean squared errors of seven optical and photophysical properties with the same weight. Note that τ and Φ were taken logarithms before normalization because their values reported in the literature were ranged on many orders of magnitude. In the case of Φ = 0 in the database, 10–5 was added to Φ. The absorption and emission bandwidths (σabs and σemi) in nanometers in the database were converted to those in wavenumbers by using the following equation, σj(cm–1) = 107/[λj – σj(nm)/2] – 107/[λj + σj(nm)/2] where λj is the absorption and emission peak positions in nanometers. In addition, all molecular properties were normalized to follow the standard normal distribution so that the loss of each property was on the same scale. A total of 30 094 chromophore/solvent combinations are randomly divided in a ratio of 0.8:0.1:0.1 for training, validation, and testing. For optimizing the hyper-parameters such as the number and dimension of GCN layers and MLPs, training and validation sets were used. No evidence of overfitting was found during the training over 10 000 epochs, as shown in Figure S3. After optimization, the DL model was trained by using training and validation sets. The training results are displayed in Figure S4 and Tables S1 and S2.

Quantum Chemical Calculations

All quantum chemical calculations were performed using the density functional theory (DFT) and time-dependent DFT (TD-DFT) methods at the B3LYP level with a 6-31G(d) basis set as implemented in the Gaussian 16 package.53 For all molecules in Table S5, the integral equation formalism for the polarizable continuum model (IEF-PCM) was used for solvation. The absorption and emission spectra of molecules were obtained by optimizing the electronic ground state and excited state, respectively, and subsequently calculating the TD-DFT transition energies. Using a computer with an Intel CPU (i7-6700) and 16 GB of RAM, TD-DFT calculations took an average of 236 min per molecule. On the other hand, our DL model took 20 s for 131 molecule–solvent pairs on the same computer.

Reconstruction of Absorption and Emission Spectra

Absorption and emission spectra, Sj(ω), were calculated as a single peak using the normalized Gaussian function and predicted values by our DL model

graphic file with name au1c00035_m001.jpg 1

where λj is the absorption or emission peak position in nm and σj is the bandwidth. The vibronic features and asymmetric nature of the spectroscopic peaks were not reflected in eq 1. The absorption intensity is directly related to the extinction coefficient by Beer’s law (A = εbc); thus, the absorption spectrum can be approximated as εSabs(ω). Similarly, the emission intensity is proportional to the absorption intensity and PLQY, and thus, the emission spectrum can be obtained as εΦSemi(ω). The product of PLQY and extinction coefficient (i.e., Φε) is known as the brightness. Using the optical and photophysical properties (λabs, λemi, σabs, σemi, ε, and Φ) predicted by the DL model, the absorption or emission spectra of a molecule were readily calculated. Furthermore, the CIE 1931 color space was computed using the calculated absorption and emission spectra. The CIE Standard Illuminant D65 was used for a daylight color.

Acknowledgments

This study was supported by grants from the National Research Foundation of Korea (NRF) funded by the Korean government (No. 2019R1A6A1A11044070) and Korea University-Future Research Grant (KU-FRG).

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.1c00035.

  • The details of our deep learning model, prediction by our deep learning model, and synthetic procedure of compound 1 (PDF)

Author Contributions

† J.F.J. and M.H. contributed equally to this work.

Author Contributions

S.P. and D.H.C. conceived the research. J.F.J., M.H., and M.J. conducted research under the supervision of S.P. J.F.J., M.H., and M.J. built the experimental database. J.H. synthesized and characterized compound 1. J.F.J. and M.H. developed the deep learning algorithm and codes, and the web-based Deep4Chem. All authors analyzed the data. J.F.J., M.H., and S.P. wrote the manuscript.

The authors declare no competing financial interest.

Supplementary Material

au1c00035_si_001.pdf (2.4MB, pdf)

References

  1. Burlingame Q.; Huang X.; Liu X.; Jeong C.; Coburn C.; Forrest S. R. Intrinsically stable organic solar cells under high-intensity illumination. Nature 2019, 573 (7774), 394. 10.1038/s41586-019-1544-1. [DOI] [PubMed] [Google Scholar]
  2. Uoyama H.; Goushi K.; Shizu K.; Nomura H.; Adachi C. Highly efficient organic light-emitting diodes from delayed fluorescence. Nature 2012, 492 (7428), 234. 10.1038/nature11687. [DOI] [PubMed] [Google Scholar]
  3. Li B.; Zhao M.; Feng L.; Dou C.; Ding S.; Zhou G.; Lu L.; Zhang H.; Chen F.; Li X.; et al. Organic NIR-II molecule with long blood half-life for in vivo dynamic vascular imaging. Nat. Commun. 2020, 11 (1), 3102. 10.1038/s41467-020-16924-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Zhang Y.; Yang H.; Ma H.; Bian G.; Zang Q.; Sun J.; Zhang C.; An Z.; Wong W. Y. Excitation Wavelength Dependent Fluorescence of an ESIPT Triazole Derivative for Amine Sensing and Anti-Counterfeiting Applications. Angew. Chem. 2019, 131 (26), 8865. 10.1002/ange.201902890. [DOI] [PubMed] [Google Scholar]
  5. Woodward R. B. Structure and the Absorption Spectra of α,β-Unsaturated Ketones. J. Am. Chem. Soc. 1941, 63 (4), 1123. 10.1021/ja01849a066. [DOI] [Google Scholar]
  6. Roothaan C. C. J. New Developments in Molecular Orbital Theory. Rev. Mod. Phys. 1951, 23 (2), 69. 10.1103/RevModPhys.23.69. [DOI] [Google Scholar]
  7. McWeeny R.; Diercksen G. Self-Consistent Perturbation Theory. II. Extension to Open Shells. J. Chem. Phys. 1968, 49 (11), 4852. 10.1063/1.1669970. [DOI] [Google Scholar]
  8. Hohenberg P.; Kohn W. Inhomogeneous Electron Gas. Phys. Rev. 1964, 136 (3B), B864. 10.1103/PhysRev.136.B864. [DOI] [Google Scholar]
  9. Bauernschmitt R.; Ahlrichs R. Treatment of electronic excitations within the adiabatic approximation of time dependent density functional theory. Chem. Phys. Lett. 1996, 256 (4–5), 454. 10.1016/0009-2614(96)00440-X. [DOI] [Google Scholar]
  10. Adamo C.; Jacquemin D. The calculations of excited-state properties with Time-Dependent Density Functional Theory. Chem. Soc. Rev. 2013, 42 (3), 845. 10.1039/C2CS35394F. [DOI] [PubMed] [Google Scholar]
  11. Joung J. F.; Han M.; Jeong M.; Park S. Figshare 2020, 10.6084/m9.figshare.12045567.v2. [DOI] [Google Scholar]
  12. Joung J. F.; Han M.; Jeong M.; Park S. Experimental database of optical properties of organic compounds. Sci. Data 2020, 7 (1), 295. 10.1038/s41597-020-00634-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Butler K. T.; Davies D. W.; Cartwright H.; Isayev O.; Walsh A. Machine learning for molecular and materials science. Nature 2018, 559 (7715), 547. 10.1038/s41586-018-0337-2. [DOI] [PubMed] [Google Scholar]
  14. Mater A. C.; Coote M. L. Deep Learning in Chemistry. J. Chem. Inf. Model. 2019, 59 (6), 2545. 10.1021/acs.jcim.9b00266. [DOI] [PubMed] [Google Scholar]
  15. Montavon G.; Rupp M.; Gobre V.; Vazquez-Mayagoitia A.; Hansen K.; Tkatchenko A.; Müller K.-R.; Anatole von Lilienfeld O. Machine learning of molecular electronic properties in chemical compound space. New J. Phys. 2013, 15 (9), 095003. 10.1088/1367-2630/15/9/095003. [DOI] [Google Scholar]
  16. Coley C. W.; Barzilay R.; Green W. H.; Jaakkola T. S.; Jensen K. F. Convolutional Embedding of Attributed Molecular Graphs for Physical Property Prediction. J. Chem. Inf. Model. 2017, 57 (8), 1757. 10.1021/acs.jcim.6b00601. [DOI] [PubMed] [Google Scholar]
  17. Musil F.; De S.; Yang J.; Campbell J. E.; Day G. M.; Ceriotti M. Machine learning for the structure-energy-property landscapes of molecular crystals. Chem. Sci. 2018, 9 (5), 1289. 10.1039/C7SC04665K. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Jinich A.; Sanchez-Lengeling B.; Ren H.; Harman R.; Aspuru-Guzik A. A Mixed Quantum Chemistry/Machine Learning Approach for the Fast and Accurate Prediction of Biochemical Redox Potentials and Its Large-Scale Application to 315000 Redox Reactions. ACS Cent. Sci. 2019, 5 (7), 1199. 10.1021/acscentsci.9b00297. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Yamada H.; Liu C.; Wu S.; Koyama Y.; Ju S.; Shiomi J.; Morikawa J.; Yoshida R. Predicting Materials Properties with Little Data Using Shotgun Transfer Learning. ACS Cent. Sci. 2019, 5 (10), 1717. 10.1021/acscentsci.9b00804. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Afzal M. A. F.; Sonpal A.; Haghighatlari M.; Schultz A. J.; Hachmann J. A deep neural network model for packing density predictions and its application in the study of 1.5 million organic molecules. Chem. Sci. 2019, 10 (36), 8374. 10.1039/C9SC02677K. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Dong Y.; Wu C.; Zhang C.; Liu Y.; Cheng J.; Lin J. Bandgap prediction by deep learning in configurationally hybridized graphene and boron nitride. npj Comput. Mater. 2019, 5 (1), 26. 10.1038/s41524-019-0165-4. [DOI] [Google Scholar]
  22. So S.; Mun J.; Rho J. Simultaneous Inverse Design of Materials and Structures via Deep Learning: Demonstration of Dipole Resonance Engineering Using Core-Shell Nanoparticles. ACS Appl. Mater. Interfaces 2019, 11 (27), 24264. 10.1021/acsami.9b05857. [DOI] [PubMed] [Google Scholar]
  23. Ye S.; Hu W.; Li X.; Zhang J.; Zhong K.; Zhang G.; Luo Y.; Mukamel S.; Jiang J. A neural network protocol for electronic excitations of N-methylacetamide. Proc. Natl. Acad. Sci. U. S. A. 2019, 116 (24), 11612. 10.1073/pnas.1821044116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Ghosh K.; Stuke A.; Todorovic M.; Jorgensen P. B.; Schmidt M. N.; Vehtari A.; Rinke P. Deep Learning Spectroscopy: Neural Networks for Molecular Excitation Spectra. Adv. Sci. 2019, 6 (9), 1801367. 10.1002/advs.201801367. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Back S.; Yoon J.; Tian N.; Zhong W.; Tran K.; Ulissi Z. W. Convolutional Neural Network of Atomic Surface Structures To Predict Binding Energies for High-Throughput Screening of Catalysts. J. Phys. Chem. Lett. 2019, 10 (15), 4401. 10.1021/acs.jpclett.9b01428. [DOI] [PubMed] [Google Scholar]
  26. Kang B.; Seok C.; Lee J. Prediction of Molecular Electronic Transitions Using Random Forests. J. Chem. Inf. Model. 2020, 60 (12), 5984. 10.1021/acs.jcim.0c00698. [DOI] [PubMed] [Google Scholar]
  27. Na G. S.; Jang S.; Lee Y. L.; Chang H. Tuplewise Material Representation Based Machine Learning for Accurate Band Gap Prediction. J. Phys. Chem. A 2020, 124 (50), 10616. 10.1021/acs.jpca.0c07802. [DOI] [PubMed] [Google Scholar]
  28. Ye S.; Zhong K.; Zhang J.; Hu W.; Hirst J. D.; Zhang G.; Mukamel S.; Jiang J. A Machine Learning Protocol for Predicting Protein Infrared Spectra. J. Am. Chem. Soc. 2020, 142 (45), 19071. 10.1021/jacs.0c06530. [DOI] [PubMed] [Google Scholar]
  29. Li B.; Liu R.; Bai H.; Ju C.-W. Can Machine Learning Be More Accurate Than TD-DFT? Prediction of Emission Wavelengths and Quantum Yields of Organic Fluorescent Materials. ChemRxiv 2020, 10.26434/chemrxiv.12111060.v1. [DOI] [PubMed] [Google Scholar]
  30. Gao H.; Struble T. J.; Coley C. W.; Wang Y.; Green W. H.; Jensen K. F. Using Machine Learning To Predict Suitable Conditions for Organic Reactions. ACS Cent. Sci. 2018, 4 (11), 1465. 10.1021/acscentsci.8b00357. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Segler M. H. S.; Preuss M.; Waller M. P. Planning chemical syntheses with deep neural networks and symbolic AI. Nature 2018, 555 (7698), 604. 10.1038/nature25978. [DOI] [PubMed] [Google Scholar]
  32. Voznyy O.; Levina L.; Fan J. Z.; Askerka M.; Jain A.; Choi M. J.; Ouellette O.; Todorovic P.; Sagar L. K.; Sargent E. H. Machine Learning Accelerates Discovery of Optimal Colloidal Quantum Dot Synthesis. ACS Nano 2019, 13 (10), 11122. 10.1021/acsnano.9b03864. [DOI] [PubMed] [Google Scholar]
  33. Torng W.; Altman R. B. Graph Convolutional Neural Networks for Predicting Drug-Target Interactions. J. Chem. Inf. Model. 2019, 59 (10), 4131. 10.1021/acs.jcim.9b00628. [DOI] [PubMed] [Google Scholar]
  34. Lim J.; Hwang S.-Y.; Moon S.; Kim S.; Kim W. Y. Scaffold-based molecular design with a graph generative model. Chem. Sci. 2020, 11 (4), 1153. 10.1039/C9SC04503A. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Jha D.; Ward L.; Paul A.; Liao W. K.; Choudhary A.; Wolverton C.; Agrawal A. ElemNet: Deep Learning the Chemistry of Materials From Only Elemental Composition. Sci. Rep. 2018, 8 (1), 17593. 10.1038/s41598-018-35934-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Sahu H.; Yang F.; Ye X.; Ma J.; Fang W.; Ma H. Designing promising molecules for organic solar cells via machine learning assisted virtual screening. J. Mater. Chem. A 2019, 7 (29), 17480. 10.1039/C9TA04097H. [DOI] [Google Scholar]
  37. Kim B.; Lee S.; Kim J. Inverse design of porous materials using artificial neural networks. Sci. Adv. 2020, 6 (1), eaax9324. 10.1126/sciadv.aax9324. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Hansen K.; Biegler F.; Ramakrishnan R.; Pronobis W.; von Lilienfeld O. A.; Muller K. R.; Tkatchenko A. Machine Learning Predictions of Molecular Properties: Accurate Many-Body Potentials and Nonlocality in Chemical Space. J. Phys. Chem. Lett. 2015, 6 (12), 2326. 10.1021/acs.jpclett.5b00831. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Hermann J.; Schätzle Z.; Noé F. Deep-neural-network solution of the electronic Schrödinger equation. Nat. Chem. 2020, 12 (10), 891. 10.1038/s41557-020-0544-y. [DOI] [PubMed] [Google Scholar]
  40. Nakata M.; Shimazaki T. PubChemQC Project: A Large-Scale First-Principles Electronic Structure Database for Data-Driven Chemistry. J. Chem. Inf. Model. 2017, 57 (6), 1300. 10.1021/acs.jcim.7b00083. [DOI] [PubMed] [Google Scholar]
  41. Gilmer J.; Schoenholz S. S.; Riley P. F.; Vinyals O.; Dahl G. E. Neural Message Passing for Quantum Chemistry. arXiv.org 2017, 1704.01212v2. [Google Scholar]
  42. Ryu S.; Lim J.; Hong S. H.; Kim W. Y. Deeply learning molecular structure-property relationships using attention- and gate-augmented graph convolutional network. arXiv.org 2018, 1805.10988. [Google Scholar]
  43. Wang X.; Li Z.; Jiang M.; Wang S.; Zhang S.; Wei Z. Molecule Property Prediction Based on Spatial Graph Embedding. J. Chem. Inf. Model. 2019, 59 (9), 3817. 10.1021/acs.jcim.9b00410. [DOI] [PubMed] [Google Scholar]
  44. Li X.; Yan X.; Gu Q.; Zhou H.; Wu D.; Xu J. DeepChemStable: Chemical Stability Prediction with an Attention-Based Graph Convolution Network. J. Chem. Inf. Model. 2019, 59 (3), 1044. 10.1021/acs.jcim.8b00672. [DOI] [PubMed] [Google Scholar]
  45. Joung J. F.; Kim S.; Park S. Cationic Effect on the Equilibria and Kinetics of the Excited-State Proton Transfer Reaction of a Photoacid in Aqueous Solutions. J. Phys. Chem. B 2018, 122 (19), 5087. 10.1021/acs.jpcb.8b00588. [DOI] [PubMed] [Google Scholar]
  46. Niko Y.; Hiroshige Y.; Kawauchi S.; Konishi G.-i. Fundamental photoluminescence properties of pyrene carbonyl compounds through absolute fluorescence quantum yield measurement and density functional theory. Tetrahedron 2012, 68 (31), 6177. 10.1016/j.tet.2012.05.072. [DOI] [Google Scholar]
  47. Ciubini B.; Visentin S.; Serpe L.; Canaparo R.; Fin A.; Barbero N. Design and synthesis of symmetrical pentamethine cyanine dyes as NIR photosensitizers for PDT. Dyes Pigm. 2019, 160, 806. 10.1016/j.dyepig.2018.09.009. [DOI] [Google Scholar]
  48. Khopkar S.; Jachak M.; Shankarling G. Novel A(2)-D-A(1)-D-A(2) type NIR absorbing symmetrical squaraines based on 2, 3, 3, 8-tetramethyl-3H-pyrrolo [3, 2-h] quinoline: Synthesis, photophysical, electrochemical, thermal properties and photostability study. Spectrochim. Acta, Part A 2019, 211, 114. 10.1016/j.saa.2018.11.061. [DOI] [PubMed] [Google Scholar]
  49. Cser A.; Nagy K.; Biczók L. Fluorescence lifetime of Nile Red as a probe for the hydrogen bonding strength with its microenvironment. Chem. Phys. Lett. 2002, 360 (5–6), 473. 10.1016/S0009-2614(02)00784-4. [DOI] [Google Scholar]
  50. Niko Y.; Kawauchi S.; Konishi G. Solvatochromic pyrene analogues of Prodan exhibiting extremely high fluorescence quantum yields in apolar and polar solvents. Chem. - Eur. J. 2013, 19 (30), 9760. 10.1002/chem.201301020. [DOI] [PubMed] [Google Scholar]
  51. Santin L. R. R.; dos Santos S. C.; Novo D. L. R.; Bianchini D.; Gerola A. P.; Braga G.; Caetano W.; Moreira L. M.; Bastos E. L.; Romani A. P.; et al. Study between solvatochromism and steady-state and time-resolved fluorescence measurements of the Methylene blue in binary mixtures. Dyes Pigm. 2015, 119, 12. 10.1016/j.dyepig.2015.03.004. [DOI] [Google Scholar]
  52. Wünsch U. J.; Murphy K. R.; Stedmon C. A. Fluorescence Quantum Yields of Natural Organic Matter and Organic Compounds: Implications for the Fluorescence-based Interpretation of Organic Matter Composition. Front. Mar. Sci. 2015, 10.3389/fmars.2015.00098. [DOI] [Google Scholar]
  53. Frisch M. J.; Trucks G. W.; Schlegel H. B.; Scuseria G. E.; Robb M. A.; Cheeseman J. R.; Scalmani G.; Barone V.; Petersson G. A.; Nakatsuji H.; Li X.; Caricato M.; Marenich A. V.; Bloino J.; Janesko B. G.; Gomperts R.; Mennucci B.; Hratchian H. P.; Ortiz J. V.; Izmaylov A. F.; Sonnenberg J. L.; Williams-Young D.; Ding F.; Lipparini F.; Egidi F.; Goings J.; Peng B.; Petrone A.; Henderson T.; Ranasinghe D.; Zakrzewski V. G.; Gao J.; Rega N.; Zheng G.; Liang W.; Hada M.; Ehara M.; Toyota K.; Fukuda R.; Hasegawa J.; Ishida M.; Nakajima T.; Honda Y.; Kitao O.; Nakai H.; Vreven T.; Throssell K.; Montgomery J. A. Jr.; Peralta J. E.; Ogliaro F.; Bearpark M. J.; Heyd J. J.; Brothers E. N.; Kudin K. N.; Staroverov V. N.; Keith T. A.; Kobayashi R.; Normand J.; Raghavachari K.; Rendell A. P.; Burant J. C.; Iyengar S. S.; Tomasi J.; Cossi M.; Millam J. M.; Klene M.; Adamo C.; Cammi R.; Ochterski J. W.; Martin R. L.; Morokuma K.; Farkas O.; Foresman J. B.; Fox D. J.. Gaussian 16; Gaussian, Inc.: Wallingford, CT, 2016. [Google Scholar]
  54. Sun H.; Dong X.; Liu S.; Zhao Q.; Mou X.; Yang H. Y.; Huang W. Excellent BODIPY Dye Containing Dimesitylboryl Groups as PeT-Based Fluorescent Probes for Fluoride. J. Phys. Chem. C 2011, 115 (40), 19947. 10.1021/jp206396v. [DOI] [Google Scholar]
  55. Li Z.; Lv X.; Chen Y.; Fu W.-F. Synthesis, structures and photophysical properties of boron–fluorine derivatives based on pyridine/1,8-naphthyridine. Dyes Pigm. 2014, 105, 157. 10.1016/j.dyepig.2014.01.022. [DOI] [Google Scholar]
  56. Zhang S.; Liu X.; Yuan W.; Zheng W.; Li H.; Li C.; Sun Y.; Wang Y.; Yang Y.; Li Y.; et al. New aryl substituted pyridylimidazo[1,2-a]pyridine-BODIPY conjugates: Emission color tuning from green to NIR. Dyes Pigm. 2018, 159, 406. 10.1016/j.dyepig.2018.04.070. [DOI] [Google Scholar]
  57. Filatov M. A.; Karuthedath S.; Polestshuk P. M.; Callaghan S.; Flanagan K. J.; Telitchko M.; Wiesner T.; Laquai F.; Senge M. O. Control of triplet state generation in heavy atom-free BODIPY-anthracene dyads by media polarity and structural factors. Phys. Chem. Chem. Phys. 2018, 20 (12), 8016. 10.1039/C7CP08472B. [DOI] [PubMed] [Google Scholar]
  58. Zhang X. F.; Zhang G. Q.; Zhu J. Methylated Unsymmetric BODIPY Compounds: Synthesis, High Fluorescence Quantum Yield and Long Fluorescence Time. J. Fluoresc. 2019, 29 (2), 407. 10.1007/s10895-019-02349-5. [DOI] [PubMed] [Google Scholar]
  59. Kasha M. Characterization of electronic transitions in complex molecules. Discuss. Faraday Soc. 1950, 9, 14. 10.1039/df9500900014. [DOI] [Google Scholar]
  60. Cotter L. F.; Brown P. J.; Nelson R. C.; Takematsu K. Divergent Hammett Plots of the Ground- and Excited-State Proton Transfer Reactions of 7-Substituted-2-Naphthol Compounds. J. Phys. Chem. B 2019, 123 (19), 4301. 10.1021/acs.jpcb.9b01295. [DOI] [PubMed] [Google Scholar]
  61. Reichardt C. Solvatochromic Dyes as Solvent Polarity Indicators. Chem. Rev. 1994, 94 (8), 2319. 10.1021/cr00032a005. [DOI] [Google Scholar]
  62. Nakanotani H.; Higuchi T.; Furukawa T.; Masui K.; Morimoto K.; Numata M.; Tanaka H.; Sagara Y.; Yasuda T.; Adachi C. High-efficiency organic light-emitting diodes with fluorescent emitters. Nat. Commun. 2014, 5, 4016. 10.1038/ncomms5016. [DOI] [PubMed] [Google Scholar]
  63. Hong Y.; Lam J. W.; Tang B. Z. Aggregation-induced emission. Chem. Soc. Rev. 2011, 40 (11), 5361. 10.1039/c1cs15113d. [DOI] [PubMed] [Google Scholar]
  64. Jones G.; Jackson W. R.; Choi C. Y.; Bergmark W. R. Solvent effects on emission yield and lifetime for coumarin laser dyes. Requirements for a rotatory decay mechanism. J. Phys. Chem. 1985, 89 (2), 294. 10.1021/j100248a024. [DOI] [Google Scholar]
  65. Kim H. J.; Kim S. K.; Godumala M.; Yoon J.; Kim C. Y.; Jeong J. E.; Woo H. Y.; Kwon J. H.; Cho M. J.; Choi D. H. Novel molecular triad exhibiting aggregation-induced emission and thermally activated fluorescence for efficient non-doped organic light-emitting diodes. Chem. Commun. 2019, 55 (64), 9475. 10.1039/C9CC05391C. [DOI] [PubMed] [Google Scholar]
  66. Feckova M.; le Poul P.; Guen F. R.; Roisnel T.; Pytela O.; Klikar M.; Bures F.; Achelle S. 2,4-Distyryl- and 2,4,6-Tristyrylpyrimidines: Synthesis and Photophysical Properties. J. Org. Chem. 2018, 83 (19), 11712. 10.1021/acs.joc.8b01653. [DOI] [PubMed] [Google Scholar]
  67. Wen W.; Shi Z.-F.; Cao X.-P.; Xu N.-S. Triphenylethylene-based fluorophores: Facile preparation and full-color emission in both solution and solid states. Dyes Pigm. 2016, 132, 282. 10.1016/j.dyepig.2016.04.014. [DOI] [Google Scholar]
  68. Hirai H.; Nakajima K.; Nakatsuka S.; Shiren K.; Ni J.; Nomura S.; Ikuta T.; Hatakeyama T. One-Step Borylation of 1,3-Diaryloxybenzenes Towards Efficient Materials for Organic Light-Emitting Diodes. Angew. Chem., Int. Ed. 2015, 54 (46), 13581. 10.1002/anie.201506335. [DOI] [PubMed] [Google Scholar]
  69. Kondo Y.; Yoshiura K.; Kitera S.; Nishi H.; Oda S.; Gotoh H.; Sasada Y.; Yanai M.; Hatakeyama T. Narrowband deep-blue organic light-emitting diode featuring an organoboron-based emitter. Nat. Photonics 2019, 13 (10), 678. 10.1038/s41566-019-0476-5. [DOI] [Google Scholar]
  70. Madayanad Suresh S.; Hall D.; Beljonne D.; Olivier Y.; Zysman-Colman E. Multiresonant Thermally Activated Delayed Fluorescence Emitters Based on Heteroatom-Doped Nanographenes: Recent Advances and Prospects for Organic Light-Emitting Diodes. Adv. Funct. Mater. 2020, 30 (33), 1908677. 10.1002/adfm.201908677. [DOI] [Google Scholar]
  71. Sanchez-Lengeling B.; Aspuru-Guzik A. Inverse molecular design using machine learning: Generative models for matter engineering. Science 2018, 361 (6400), 360. 10.1126/science.aat2663. [DOI] [PubMed] [Google Scholar]
  72. Popova M.; Isayev O.; Tropsha A. Deep reinforcement learning for de novo drug design. Sci. Adv. 2018, 4 (7), eaap7885. 10.1126/sciadv.aap7885. [DOI] [PMC free article] [PubMed] [Google Scholar]
  73. Lim J.; Ryu S.; Kim J. W.; Kim W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminf. 2018, 10 (1), 31. 10.1186/s13321-018-0286-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  74. Elton D. C.; Boukouvalas Z.; Fuge M. D.; Chung P. W. Deep learning for molecular design—a review of the state of the art. Mol. Syst. Des. Eng. 2019, 4 (4), 828. 10.1039/C9ME00039A. [DOI] [Google Scholar]
  75. Hong S. H.; Ryu S.; Lim J.; Kim W. Y. Molecular Generative Model Based on an Adversarially Regularized Autoencoder. J. Chem. Inf. Model. 2020, 60 (1), 29. 10.1021/acs.jcim.9b00694. [DOI] [PubMed] [Google Scholar]
  76. Krenn M.; Häse F.; Nigam A.; Friederich P.; Aspuru-Guzik A. Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Mach. Learn.: Sci. Technol. 2020, 1 (4), 045024. 10.1088/2632-2153/aba947. [DOI] [Google Scholar]
  77. Joung J. F.; Han M.; Park S.. Deep4Chem, 2020; http://deep4chem.korea.ac.kr/.

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

au1c00035_si_001.pdf (2.4MB, pdf)

Articles from JACS Au are provided here courtesy of American Chemical Society

RESOURCES