Discovering New Color Developers for Thermochromic Microcapsules Using a Variational Autoencoder

doi:10.21203/rs.3.rs-3400210/v1

Download PDF

Research Article

Discovering New Color Developers for Thermochromic Microcapsules Using a Variational Autoencoder

https://doi.org/10.21203/rs.3.rs-3400210/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

In our pursuit of safer alternatives to Bisphenol A (BPA) for use as color developers in thermochromic microcapsules, our goal was to mitigate endocrine disruption without compromising the properties of BPA-based microcapsules. We began by scrutinizing the effect of 11 BPA derivatives on temperatures at which microcapsules change color (color-changing temperatures). The color-changing temperatures were determined using sigmoid fitting of the color density versus temperature plot, leading to four regression models connecting these temperatures to the color developer structures. To assess endocrine-disrupting potential and toxicity, we adopted machine learning models from the Open QSAR Application (OPERA). Concurrently, using atom-wise tokenization, we trained a variational autoencoder on SMILES data of drug-like molecules. With this approach, we have achieved a six-fold speed increase in training with 20% fewer parameters than conventional character-wise tokenization. After transfer learning with potential color developer data, this model generated new SMILES data, which were subsequently evaluated for their properties. In the end, we have obtained a compilation of SMILES predicted to be effective and safer replacements for BPA.

thermochromic microcapsule

color developer

BPA-free

chemical databases

atom-wise tokenization

variational autoencoder (VAE)

Thermochromic microcapsules are materials that change color in response to thermal changes in their surrounding environment. This behavior results from the phase change of their internal solvent, which leads to the formation or breaking of hydrogen bonds between the leuco dye and the color developer. The color of leuco dye appears when it forms a hydrogen bond with the color developer and disappears when this bond breaks.[1] Owing to their unique ability, thermochromic microcapsules have found applications in various fields, such as wood coatings[2], cement,[3] road paving materials,[4] and functional textile coatings.[5, 6] However, it is concerning that Bisphenol A (BPA) is frequently used as a color developer for leuco dye in these microcapsules.[7–9]

In the 1930s, BPA was hypothesized to have estrogenic properties.[10] Building on this initial hypothesis, there are many insightful review articles that have extensively discussed numerous studies reporting on the toxicity of BPA.[11–13] According to these reviews, while assessing the risk of low-dose exposure is challenging, they conclude that potential adverse effects on the endocrine and nervous systems are deeply concerning. Recently, due to stricter regulations concerning BPA, thermochromic products have been formulated using alternative chemicals like Bisphenol S or Bisphenol F.[14] Nevertheless, the safety of these alternatives remains uncertain. The toxicity of substances included in 'BPA-free' products has not been fully elucidated, and both computational and experimental results suggest they might be as toxic as BPA.[15–17]

Although further safety assessments are ongoing, Pergafast® 201 is considered a successful replacement for BPA in the fields of polymer production and thermal paper.[18–22] However, applying it to thermochromic microcapsules is not straightforward. The primary determinants of the color-changing temperature range are the melting and freezing points of the solvent in the core along with the complicated interactions between the solvent, dye, and developer.[23] Therefore, even with the same solvent, the temperature range can shift when different color developers and additives are used, especially in comparison to when BPA is employed as the color developer. Consequently, our research aims not only to identify safe color developers that can be microencapsulated but also to ensure that they do not shift the desired color-changing temperature range of the product.

To address this challenge, we adopted data-driven strategies, providing an effective method to identify suitable new materials. A straightforward approach involves filtering from existing chemical databases. Specifically, we have collected the Simplified Molecular Input Line Entry System (SMILES)[24] of molecules provided by the open-source chemical databases. Beyond simple filtering, we turned our attention to the variational autoencoder (VAE), a leading framework for generative models. Gómez-Bombarelli et al. highlighted how VAEs use continuous molecular representations for automated chemical design.[25] In a related context, there have been successful cases in chemistry where transfer learning is introduced to address the challenges of deep learning with limited data.[26] Based on these findings, we decided to apply deep learning to our relatively scarce color developer-like data. Consequently, we trained a VAE on numerous molecular structures from chemical databases, yielding potential BPA replacement candidates for use as developers of thermochromic microcapsules.

The discovery of a new molecule does not guarantee its safety for use. Experimentally determining the non-toxicity of every generated chemical structure is daunting, both in terms of time and resources. Recognizing this challenge, we leaned on existing models for toxicity assessments. Notably, Mansouri et al. reported useful collaborative models such as CERAPP,[27] CoMPARA,[28] and CATMoS,[29] designed to forecast the estrogenic and androgenic activities, and the overall toxicity of specific compounds. Their efforts are integrated into the open-source application known as OPERA (OPEn structure-activity/property Relationship App),[30] which was used as an invaluable tool for our research. Through OPERA, we could assess the endocrine disruption potential of molecules that are expected to act as color developers.

Before the experimental application of the BPA replacements into our final products, we need to estimate their qualities. Despite limited experimental data, we developed regression models that predict the qualities of new materials when applied. Using these equations, we could prioritize potential candidates worth testing first. In this study, we introduce a series of processes such as finding new candidates among known data, training VAE model, generating new chemical structures using the model, and identifying molecular structures that meet the requirements. These procedures are illustrated in Fig. 1.

2.1 Measuring Thermochromic Microcapsule Properties

We prepared 11 thermochromic microcapsules with different developers shown in Fig. 2. All microcapsules contain methyl stearate:ethyl stearate(weight ratio 4:1) as a solvent for core and 2'-anilino-6'-(dibutylamino)-3'-methyl-3H-spiro[2-benzofuran-1,9'-xanthen]-3-one (ODB-2 as a common name, CAS RN. 89331-94-2) as a dye. The shell of the microcapsules is made of melamine, urea, and formaldehyde. The detailed microcapsule preparation methods can be found elsewhere in references.[31–33] The prepared microcapsules are a concentrated slurry in which they are dispersed in an aqueous phase. The mean diameters of all prepared microcapsules are controlled in the range of 3.0 to 5.0 µm.

Each thermochromic microcapsule slurry was coated on paper having a gram per square meter (gsm) of 70 g with a thickness of 20 µm. For analyzing the colorimetric properties of specimens, the coated papers were cut into 25 mm $\times$ 45 mm pieces and placed on sample panels. The temperature of the panel was increased from 15℃ to 50℃ for 18 minutes and decreased from 50℃ to 15℃ for 35 min. The temperatures of the most discolored and most colored states are assumed to be 50℃ and 15℃, respectively. Then, the videos of color changes caused by temperature changes were recorded by a camera.

A color space is a specific organization or model of colors. Color spaces provide a systematic way to describe the colors we see and capture with devices like cameras and scanners, and then display with screens and printers. In this context, the RGB color space of images obtained with a camera was converted to ${L}^{*}{a}^{*}{b}^{*}$ color space using OpenCV.[34] It is crucial to recognize that within the context of color perception, the Euclidean distance in color spaces does not always correlate with human visual experience. In simpler terms, the direct distance between two colors in the space might not genuinely represent the perceived difference between them. This gap in perception underscores the need for more sophisticated methods to quantify color discrepancies. The CIEDE2000[35] formula addresses this by providing an approximation more in line with how humans discern color variations, making it an optimal choice for our study.

The normalized color density is obtained by applying min-max normalization to the ΔE. The normalized color density versus temperature is parameterized using a seven-parameter (a-g) asymmetric sigmoid function. For a color change according to temperature ($T$), the function $f\left(T\right)$ is formulated as:

$$f\left(T\right)=a+\frac{f-a}{{\left(1+{\left(\frac{c}{\text{e}\text{x}\text{p}\left(g\left(T-b\right)\right)}\right)}^{d}\right)}^{e}}$$

The discoloration start (DS) and discoloration end (DE) temperature are determined by the temperature at the point where the color is different by 3 (ΔE = 3) from the most colored state and the most discolored state in the discoloration (heating) curve, respectively. On the other hand, the coloration start (CS) and coloration end (CE) temperature were determined by the temperature at the point where the color is different by 3 (ΔE = 3) from the most discolored state and the most colored state in the coloration (cooling) curve, respectively. This is because it is widely accepted that humans begin to perceive in color at a ΔE value of 3. The definition of the four temperatures on the color density curve is described in Fig. 3.

2.2 Predicting Color-Changing Temperatures

We create four quantitative structure–property relationships (QSPR) that predict the color-changing temperatures such as DS, DE, CS, and CE, depending on the molecular structure of the color developer. In essence, QSPR models use molecular descriptors to quantitatively forecast certain properties or behaviors of the substance. These models are used to predict if the performance of a microcapsule containing a new color developer and the dye 2-phenyl amino-3-methyl-6-dibutyl amino fluoran (ODB-2) will be similar to its performance when using BPA.

After removing highly correlated descriptors, we get 856 descriptors from the PaDEL.[36] PaDEL-Descriptor, crafted by Chun Wei Yap, is a tool tailored for the calculation of molecular descriptors and fingerprints, primarily leveraging The Chemistry Development Kit. At its essence, a molecular descriptor translates the chemical specifics of a molecule into a numeric value, serving as an aid to predict its various properties or behaviors. Then, linear regression models with two descriptors are established to explain our experimental results by using the genetic function approximation method.[37] Linear regression models for predicting color-changing temperature are as follows:

$$\widehat{T}={a}_{0}+\sum _{k=1}^{2}{a}_{k}{x}_{k}$$

where, $\widehat{T}$ is a predicted color-changing temperature, ${a}_{0}$ is a bias and ${a}_{k}$ and ${x}_{k}$ are coefficients and features corresponding to each other.

2.3 Preprocessing Data

To apply the data-driven methodology, SMILES[24] data of molecules were obtained from ChEMBL,[38, 39] ZINC,[40] and PubChem[41]. It is necessary to filter molecules that can show the role of a developer by selecting only molecules that satisfy several conditions. The conditions are as follows: (1) containing at least an oxygen element; (2) having an aromatic ring or more; (3) not an ionic compound; (4) not a salt; (5) having two or more hydroxyl groups bonded to aromatic rings; (6) having no hydroxyl groups attached to aliphatic carbon chains; (7) containing no carboxylic acid or aldehyde; (8) having no double bond or triple bond; (9) containing no isotopes; (10) non-bicyclic compound; (11) having logP calculated by RDKit[42] between 2.68 (logP of BPF) to 5.74 (logP of BPM).

Based on SMILES syntax, conditions (1), (2), (3), and (4) determine whether specific characters such as ‘O’, ‘c’ (small letter means aromaticity), '+', '-]' and '.' are included in the SMILES string. Here, It is important to note that ‘-]’ and ‘-‘ are different, with ‘-‘ representing a single covalent bond. When filtering data under the conditions of (1) and (2), it can be quickly checked whether it contains aromatic ring and oxygen, which are necessary conditions for color developers. This can significantly reduce the pool of searches before directly looking for the hydroxyl group, as it does not have to consider the graph of the molecule. When data are filtered with conditions (1) and (2), it can be quickly confirmed whether the color developer for leuco dye corresponds to an aromatic compound containing a hydroxyl group according to the rules of SMILES. Conditions (3) and (4) are easy criteria for selecting molecules that are not highly polar because the solvent molecule that controls the discoloration temperature has a long alkyl chain. A substance that can act as a developer empirically has hydroxyl groups with $\text{p}{K}_{a}$ in the range of 9 to 11. Aliphatic alcohol is a too weak acid to develop dyes with $\text{p}{K}_{a}$ of 15 or higher. Additionally, substances like carboxylic acid, which have small $\text{p}{K}_{a}$ values, do not exhibit the discoloration. Therefore, a phenolic hydroxyl group is ideal. Molecules that survived at this range are further filtered with conditions (5), (6), and (7) and then substances that do not meet conditions (8), (9), and (10) are excluded in consideration of synthesis difficulty. Finally, the condition (11) is applied to predict the solubility.

2.4 Training Generative Models

Atom-wise tokenization. To train a VAE with SMILES data, the data must be transformed into numerical vectors. They are in the form of one-hot vectors that digitalize categorical data as 0 or 1. In the process of converting the SMILES string into a one-hot encoded vector, which is the input form of the neural network.

We use atom-wise tokenization[43, 44] to preprocess SMILES data for the sake of clarity in meaning and to reduce the length of the sequence to be trained. Unlike character-wise tokenization, which breaks down two-letter elements like chlorine (Cl) into individual characters (C and l), atom-wise tokenization keeps each element in its original form. Furthermore, the sequence of character tokens is longer, so that the training time of neural networks would increase. When a stereochemical expression is included, even if "[C@@H]" refers to a single chiral carbon atom, it is divided into six characters, such as "[", "C", "@", "@", "H", and "]". On the other hand, atom-wise tokenization, it makes just a single “[C@@H]” token, which treats all the characters enclosed in square brackets (‘[’ and ‘]’) as an atomic token. Thus, this superior atom-wise tokenization produces tokens with explicit meaning and can also reduce the length of sequences for training. In addition, according to SMILES rules, ‘[’ and ‘]’ must always appear as a pair, and when either one of the two is missing or the order is changed when generating SMILES, it becomes invalid, so this invalidity is prevented in advance by using atom-wise tokenization.

Variational autoencoder. The VAE introduced by Kingma et al. can be used as a generative model.[45] In chemistry, it is well known from the work of Gómez-Bombarelli et al. that new SMILES can be obtained by adding noise to the latent representation of an encoded compound.[25] We use this method applying the atom-wise tokenization method mentioned above.

To compare the difference depending on tokenization, we train the identical data on drug-like compounds like in reference.[25] After converting SMILES data into sequences of tokens, it is encoded as one-hot vectors. As the output of a variational layer sampling the mean and variance, we get continuous representation of compounds. Encoder produces a 180-dimensional continuous representation from variational layer sampling the mean and variance after operations of one-hot vectors by three 1D convolutional layers. Each convolution layer has 9, 9, and 10 convolution kernels and their sizes are 9, 9, and 10, respectively. Every convolution process is followed by batch normalization. The decoder is composed of three gated recurrent unit (GRU)[46] layers with hidden dimension of 450. The final layer of the decoder provides the probability distribution of all atomic tokens for the output. The decoder is trained using teacher forcing.[47] The output GRU layer contains an extra input that corresponds to the atomic tokens from the softmax output of the previous time step and the atomic tokens used in the original data.

We adjust the learning rate to 80% if the categorical cross-entropy loss of the validation set does not decrease at least every 3 epochs in the training of the neural network. In addition, an early stop is applied, so we plan to train for a total of 150 epochs, but if the decoding accuracy of the validation set does not improve within 10 epochs during training, we stop learning and obtain the final model.

Training is implemented using TensorFlow[48] 1.15.0, Keras[49] 2.0.6, NumPy[50] 1.19.2, Pandas[51] 0.19.2 version, RDKit[42] 2021.9.4 and the source code is written and executed in Python 3.6.13 environment.

2.5 Generating New Color Developers

We generate new SMILES by adding random noise to the latent vector of each color developer as follows:

$${z}^{{\prime }}=z+\left(X\times \frac{Z}{{‖Z‖}_{2}}\right)$$

where, z is a latent vector, $\text{X}$ is a uniform distribution, $\text{X}\sim\text{u}\text{n}\text{i}\text{f}\text{o}\text{r}\text{m}(0, \text{n}\text{o}\text{i}\text{s}\text{e})$, $Z$ is a random normal distribution, $Z∼N(0, 1)$, and ${‖Z‖}_{2}$ is a L2-norm of $Z$. When ${z}^{{\prime }}$ is fed into the decoder network, the one-hot vector of the new SMILES is decoded. After removing the duplicates from data in training set, we get new SMILES that we have never seen before.

3.1 Thermochromic Property of Microcapsules with 11 Different Color Developers

We investigated the thermochromic properties of microcapsules containing 11 different color developers, as shown in Fig. 2. Using the video analysis technique described in Section 2.1, we observed changes in the color density of these microcapsules with respect to temperature.

Figure 3 demonstrates the sigmoidal relationship between color density and temperature variations during heating and cooling cycle. During the heating phase, the solvent within the microcapsule melts. This action breaks the hydrogen bonds between the dye and color developer, causing the microcapsules to become colorless. Conversely, as the temperature decreases and the solvent solidifies, hydrogen bonds form between the dye and color developer, causing the microcapsules to regain color.

Through this analysis, we determined four color-changing temperatures: DS, DE, CS, and CE, which correspond to the 'thermochromic properties' mentioned above. The corresponding temperature values for these changes, measured in Celsius, are listed in Table 1.

Table 1

Calculated LogP values (unitless) of each developer and color-changing temperatures (in Celsius) of the prepared thermochromic microcapsules. The values enclosed in parentheses below are the predicted values from the regression model presented in Table 2. The last RMSE column numerically presents the difference in the dis-/coloration temperature of each developer compared to when BPA was applied.
* Experiment / (Prediction)
Common name	LogP	DS	DE	CS	CE	RMSE
Bisphenol A	3.4237	29.0	35.6	30.0	24.8	0.0
Bisphenol A	3.4237	(28.7)	(35.4)	(30.1)	(25.0)	(0.2)
Bisphenol AF	4.5085	29.1	35.6	30.3	26.4	0.8
Bisphenol AF	4.5085	(29.4)	(35.5)	(30.4)	(26.3)	(0.8)
Bisphenol AP	4.4521	28.3	38.2	30.8	25.4	1.4
Bisphenol AP	4.4521	(29.6)	(38.6)	(31.5)	(25.4)	(1.7)
Bisphenol E	3.2496	29.8	36.4	31.0	25.8	0.9
Bisphenol E	3.2496	(29.2)	(36.8)	(30.8)	(25.7)	(0.9)
Bisphenol F	2.6886	29.6	36.1	31.0	25.5	0.7
Bisphenol F	2.6886	(29.4)	(37.4)	(31.5)	(25.5)	(1.2)
Bisphenol FL	5.4609	31.1	48.2	34.7	26.4	6.9
Bisphenol FL	5.4609	(30.4)	(47.3)	(33.6)	(26.6)	(6.2)
Bisphenol M	5.7496	21.5	36.3	31.1	22.9	3.9
Bisphenol M	5.7496	(22.1)	(34.9)	(29.9)	(23.0)	(3.6)
Bisphenol P	5.7496	24.5	34.2	29.3	23.6	2.5
Bisphenol P	5.7496	(23.8)	(35.0)	(29.9)	(23.5)	(2.7)
Bisphenol Z	4.348	33.3	42.2	29.7	26.3	4.0
Bisphenol Z	4.348	(33.0)	(41.2)	(30.0)	(26.1)	(3.5)
D-8	3.0122	31.1	48.0	47.7	25.2	10.9
D-8	3.0122	(31.7)	(49.4)	(48.0)	(25.4)	(11.4)
TDP	3.249	29.4	47.4	42.3	25.1	8.5
TDP	3.249	(29.4)	(46.7)	(42.0)	(25.1)	(8.2)

3.2 Predictive Models for Color-Changing Temperatures, DS, DE, CS, and CE

In our study, we utilized PaDEL to examine four temperatures for DS, DE, CS, and CE across 11 distinct thermochromic microcapsules, each associated with a unique color developer. Comprehensive details on each developer are presented in Table 1, which enumerates the calculated LogP values alongside the thermochromic properties of the microcapsules, and emphasizes the predicted values from the regression model delineated in Table 2. Additionally, Table 2 not only portrays the regression models correlating developer structure to color-changing temperature but also juxtaposes the experimentally observed color-changing temperatures with those predicted by the model.

Table 2

The descriptors and parameters of Eq. (2) relating the color developer structure to each color-changing temperature.
	${\varvec{a}}_{0}$	${\varvec{a}}_{1}$	${\varvec{x}}_{1}$	${\varvec{a}}_{2}$	${\varvec{x}}_{2}$	R² (Cross validation)
DS	178.8650682	-81.7272488	piPC4	44.1153953	piPC7	0.964 (0.896)
DE	-216.1975366	-95.3339744	MATS4c	441.4377698	ETA_BetaP_s	0.971 (0.822)
CS	10.2420151	0.2626027	AATS1m	-2.1875432	SM1_Dzs	0.989 (0.974)
CE	28.5930664	-0.0022960	ATSC5m	-0.6227423	GGI1	0.983 (0.971)
piPC4 and piPC7 are defined as molecular multiple path counts of order 4 and 7, respectively.[52, 53] MATS4c is a moran coefficient of lag 4 weighted by gasteiger charge.[54, 55] ETA_BetaP_s is a measure of the electro negative atom count of molecule relative to the molecular size.[56–58] AATS1m is a averaged moreau-broto autocorrelation of lag 1 weighted by mass.[55, 59] SM1_Dzs is a spectral moment of order 1 from Barysz matrix weighted by I-state.[55, 60] ATSC5m is a centered moreau-broto autocorrelation of lag 5 weighted by mass.[55, 59] GGI1 is a topological charge index of order 1.[55, 61]

Now we can predict temperatures by those equations before applying new candidates. We hope there are small differences between the predictions and BPA, but not all four temperatures can be predicted close to the target. To evaluate the priority of the new molecule, the root mean square error (RMSE) of the four predicted values is obtained based on the BPA in Table 1.

$$\text{R}\text{M}\text{S}\text{E}= \sum _{i}^{4}\sqrt{\frac{{\left({T}_{i}-{\widehat{T}}_{i}\right)}^{2}}{4}}$$

where ${T}_{1}$, ${T}_{2}$, ${T}_{3}$, and ${T}_{4}$ correspond to DS, DE, CS, and CE, respectively. ${T}_{i}$ are the four color-changing temperatures based on BPA, and ${\widehat{T}}_{i}$ from Eq. (2), are the predictions of unknown molecules through the models (Table 2).

Although a separate test set should be used to evaluate the performance of the model, owing to the very small dataset, we only conduct cross-validation to estimate the predictive performances of our models. It is worth noting that while a larger dataset would have invariably granted us greater confidence in our predictions and generalizability to unseen data, in the context of our study, cross-validation provided the most robust means to evaluate our models. It is essential to approach the findings with a degree of caution due to constraints in the dataset. Nonetheless, our aim was to provide valuable insights and recommendations based on the data currently available to us.

3.3 Candidates from Databases

We consider as potential color developers those compounds that satisfy all the conditions mentioned above, 260,738, 1,556,222, and 5,466,359 compounds acquired from the open chemical databases ZINC, ChEMBL, and PubChem, respectively. The structures we obtained are processed by OPERA using the CERAPP, CoMPARA, and CATMoS models to predict endocrine disruption and toxicity. As a result, a list of candidate developers is obtained by removing potentially hazardous materials. Through this process, we obtain 287, 1,311 and 29,022 compounds from ChEMBL, PubChem, and ZINC, respectively. Here we remove duplicates, resulting in a total of 30,538 compounds. After applying the color-changing temperature prediction models (Table 2), RMSE is calculated to evaluate the priority. The most experimentally testable 15 molecular structures are presented by RMSE in ascending order in Fig. 4 (a).

3.4 Training Deep Generative Model

Tokenization. We compare the results of both tokenization methods, character, and atom-wise tokenization, using drug-like molecules data of ZINC. This comparison is based on 249,455 molecular structures. As the tokenizer is changed from character-wise to atom-wise, the maximum sequence length is reduced from 120 to 72 and the number of token types is increased from 35 to 62.[25]

Variational autoencoder. We train VAEs by applying two tokenization methods to the drug-like molecule data used in the reference.[25] We improve the performance of neural networks by applying different tokenization methods. There are two ways to tokenize SMILES: character-by-character or atom-by-atom. Character-wise tokenization, which splits a SMILES string into all individual characters, is the simplest way to implement. However, in this case, some characters may have to embed multiple meanings, and the length of the sequence is equal to the number of characters in SMILES. On the other hand, if the atom-wise tokenization method[44] is applied, the number of token types increases, but the length of the sequence to be trained may decrease considerably. Figure 5 shows that when VAE is trained by atomic tokens, a significant improvement can be achieved over the character-wise tokenization method. Short sequence lengths obviously lead to fast learning. Because the sequence that the decoder must predict is shortened, the chances of outputting the wrong token are also reduced. For the same reason, the number of parameters in the model has also been significantly reduced. When using the atomic token, the model has 3.6 million parameters, whereas when using the character token, it has 4.5 million. Briefly, with 20% fewer parameters, a more accurate model was obtained six times faster.

Transfer learning addresses the challenges of deep learning with limited data by leveraging knowledge acquired from previously trained models on larger datasets. This approach not only accelerates the training process but also enhances generalization of the model, enabling it to excel even in the presence of sparse data. Apply the previously obtained 30,538 color developer SMILES to the trained model above. As shown in Fig. 5, it is possible to effectively train the developer structures.

3.5 Generation of New Color Developer Candidates

New molecules were obtained by adding noise to the latent representation of all color developer data. Similarly, in this process, filtering is performed according to the conditions, applied to OPERA to evaluate toxicity, and then the experimental priority is determined by substituting the model in Table 2. Finally, the 15 newly generated color developer structures are summarized in Fig. 4. (b) in ascending order of RMSE. More results are provided as supporting information.

3.6 Analysis of Candidates for Color Developers

According to the results of molecular docking simulations between estrogen receptor (ER) and ligands such as 17β-estradiol (estrogen) and bisphenol derivatives, His524 and Glu353 in the ligand-binding pocket of ER are residues that can form hydrogen bonds with ligands.[17] Thus, the structural features of molecules capable of acting as ER agonists are that they have two hydroxyl groups but are properly spaced by a central hydrophobic core.[62] In estrogen and BPA, two hydroxyl groups are spaced with approximately 10 to 11 Å by the steroid and alkyl backbone, respectively.

Therefore, there are two ways for designing color developers without ER activity: making the distance between the two hydroxyl groups very close or far, and increasing the hydrophilicity of the central structure. The most of molecular structures in Fig. 4 contain not only functional groups, or substituents, consisting of heteroatoms such as nitrogen and chlorine, but also two hydroxyl groups attached to a single aromatic ring. When two hydroxyl groups are substituted on a single benzene, taking p-dihydroxybenzene as an example, the distance between the two groups is within about 7 Å no matter how far they are. In addition to considering their substituents, it is expected that it will be difficult to form hydrogen bonds in the ligand-binding pocket. Also, some heteroatoms of them may increase the polarity locally. The color developers we have tested in Fig. 2 do not contain the features of the candidates, so in the current situation, the candidates seem reasonable for further experimentation.

We have discovered and designed new color developers for thermochromic microcapsules to replace the BPA. Recognizing the need for efficient material discovery based on data-driven approaches, we initiated our research with straightforward regression models tailored for limited datasets. Then, we used the large-scale data provided by the open chemical database and developed a deep generative model. Then, the estrogen/androgen receptor activities of SMILES data, filtered from databases and generated by the model, were evaluated. After all, lists of candidates to replace BPA were obtained in efficient ways.

However, we acknowledge certain limitations in our methodology. Firstly, our QSPR models were constrained by a small dataset, leading to potential biases in our list of candidates based on the model selected. While the accuracy of the model is crucial, the interpretation of the model can shift considerably with the addition of new experimental data. This underscores the pressing need for further experiments with a broader range of color developers. Secondly, the molecules we generate might be overly influenced by the training data. To produce a richer array of molecular SMILES, we must diversify the molecular data utilized in both pre-training and transfer learning.

Nevertheless, the results we obtained reasonably suggest the next experiments based on the current data level, and we believe that data on molecular structures can provide insight to experimental chemists. An efficient cycle in which model updates are performed by additional experiments, followed by experiments again, may contribute to discovering materials with desirable properties.

Supporting Information

50 molecular structures of candidate color developer materials to replace BPA, each filtered from a database or generated by a generative model.

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Author Contributions

Taehoon Kim : Methodology, Software, Formal analysis, Investigation, Data Curation, Writing - Original Draft, Hyein Guk : Methodology, Software, Resources, Data Curation, Writing - Original Draft, Jaehoon Yun : Resources, Jonghwa Kim : Resources, Daejin Kim : Conceptualization, Writing - Review & Editing, Project administration, Seung-Hoon Choi : Supervision, Dong Hyen Chung : Supervision, Writing - Review & Editing

MacLaren DC, White MA (2003) Competition between dye-developer and solvent-developer interactions in a reversible thermochromic system. J Mater Chem 13:1701–1704. https://doi.org/10.1039/b302250a
Zhu X, Liu Y, Li Z, Wang W (2018) Thermochromic microcapsules with highly transparent shells obtained through in-situ polymerization of urea formaldehyde around thermochromic cores for smart wood coatings. Sci Rep 8:. https://doi.org/10.1038/s41598-018-22445-z
Ma Y, Zhu B (2009) Research on the preparation of reversibly thermochromic cement based materials at normal temperature. Cem Concr Res 39:90–94. https://doi.org/10.1016/j.cemconres.2008.10.006
Du P, Chen Z, Zhang H (2019) Rheological and aging behaviors of base and SBS modified asphalt with thermochromic microcapsule. Constr Build Mater 200:1–9. https://doi.org/10.1016/j.conbuildmat.2018.12.083
Zhang Y, Hu Z, Xiang H, et al (2019) Fabrication of visual textile temperature indicators based on reversible thermochromic fibers. Dyes and Pigments 162:705–711. https://doi.org/10.1016/j.dyepig.2018.11.007
Ma X, Zhao S, Wang L, Zhou HJ (2020) Research on the behaviors of extending thermochromic colors for a new thermochromic microcapsule. Journal of the Textile Institute 111:1097–1105. https://doi.org/10.1080/00405000.2019.1684224
Bašnec K, Perše LS, Šumiga B, et al (2018) Relation between colour- and phase changes of a leuco dye-based thermochromic composite. Sci Rep 8:. https://doi.org/10.1038/s41598-018-23789-2
Zhu X, Liu Y, Li Z, Wang W (2018) Thermochromic microcapsules with highly transparent shells obtained through in-situ polymerization of urea formaldehyde around thermochromic cores for smart wood coatings. Sci Rep 8:. https://doi.org/10.1038/s41598-018-22445-z
Zhao W, Yan X (2022) Preparation of Thermochromic Microcapsules of Bisphenol A and Crystal Violet Lactone and Their Effect on Coating Properties. Polymers (Basel) 14:. https://doi.org/10.3390/polym14071393
Edward Charles Dodds, W. Lawson (1938) Molecular structure in relation to oestrogenic activity. Compounds without a phenanthrene nucleus. Proc R Soc Lond B Biol Sci 125:222–232. https://doi.org/10.1098/rspb.1938.0023
Geens T, Aerts D, Berthot C, et al (2012) A review of dietary and non-dietary exposure to bisphenol-A. Food and Chemical Toxicology 50:3725–3740. https://doi.org/10.1016/j.fct.2012.07.059
Chapin RE, Adams J, Boekelheide K, et al (2008) NTP-CERHR expert panel report on the reproductive and developmental toxicity of bisphenol A. Birth Defects Res B Dev Reprod Toxicol 83:157–395
Rosenfeld CS (2017) Neuroendocrine disruption in animal models due to exposure to bisphenol A analogues. Front Neuroendocrinol 47:123–133
Björnsdotter MK, de Boer J, Ballesteros-Gómez A (2017) Bisphenol A and replacements in thermal paper: A review. Chemosphere 182:691–706
Kim SS, Kim JL, Hwang KS, et al (2022) Mechanism of action and neurotoxic effects of chronic exposure to bisphenol F in adult zebrafish. Science of the Total Environment 851:. https://doi.org/10.1016/j.scitotenv.2022.158258
Gao Y, Li A, Zhang W, et al (2022) Assessing the toxicity of bisphenol A and its six alternatives on zebrafish embryo/larvae. Aquatic Toxicology 246:. https://doi.org/10.1016/j.aquatox.2022.106154
Ng HW, Shu M, Luo H, et al (2015) Estrogenic Activity Data Extraction and in Silico Prediction Show the Endocrine Disruption Potential of Bisphenol A Replacement Compounds. Chem Res Toxicol 28:1784–1795. https://doi.org/10.1021/acs.chemrestox.5b00243
Goldinger DM, Demierre A-L, Zoller O, et al (2015) Endocrine activity of alternatives to BPA found in thermal paper in Switzerland. Regulatory Toxicology and Pharmacology 71:453–462. https://doi.org/10.1016/j.yrtph.2015.01.002
Björnsdotter MK, Jonker W, Legradi J, et al (2017) Bisphenol A alternatives in thermal paper from the Netherlands, Spain, Sweden and Norway. Screening and potential toxicity. Science of The Total Environment 601–602:210–221. https://doi.org/10.1016/j.scitotenv.2017.05.171
Björnsdotter MK, de Boer J, Ballesteros-Gómez A (2017) Bisphenol A and replacements in thermal paper: A review. Chemosphere 182:691–706. https://doi.org/10.1016/j.chemosphere.2017.05.070
Eckardt M, Simat TJ (2017) Bisphenol A and alternatives in thermal paper receipts - a German market analysis from 2015 to 2017. Chemosphere 186:1016–1025. https://doi.org/10.1016/j.chemosphere.2017.08.037
Keminer O, Teigeler M, Kohler M, et al (2020) A tiered high-throughput screening approach for evaluation of estrogen and androgen receptor modulation by environmentally relevant bisphenol A substitutes. Science of The Total Environment 717:134743. https://doi.org/10.1016/j.scitotenv.2019.134743
Pu Y, Fang J (2022) Preparation and thermochromic behavior of low-temperature thermochromic microcapsule temperature indicators. Colloids Surf A Physicochem Eng Asp 653:. https://doi.org/10.1016/j.colsurfa.2022.129889
Weininger D (1988) SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J Chem Inf Model 28:31–36. https://doi.org/10.1021/ci00057a005
Gómez-Bombarelli R, Wei JN, Duvenaud D, et al (2018) Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Cent Sci 4:268–276. https://doi.org/10.1021/acscentsci.7b00572
Iovanac NC, Savoie BM (2020) Simpler is Better: How Linear Prediction Tasks Improve Transfer Learning in Chemical Autoencoders. J Phys Chem A 124:3679–3685. https://doi.org/10.1021/acs.jpca.0c00042
Mansouri K, Abdelaziz A, Rybacka A, et al (2016) CERAPP: Collaborative estrogen receptor activity prediction project. Environ Health Perspect 124:1023–1033. https://doi.org/10.1289/ehp.1510267
Mansouri K, Kleinstreuer N, Abdelaziz AM, et al (2020) Compara: Collaborative modeling project for androgen receptor activity. Environ Health Perspect 128:. https://doi.org/10.1289/EHP5580
Mansouri K, Karmaus AL, Fitzpatrick J, et al (2021) CATMoS: Collaborative acute toxicity modeling suite. Environ Health Perspect 129:. https://doi.org/10.1289/EHP8495
Mansouri K, Grulke CM, Judson RS, Williams AJ (2018) OPERA models for predicting physicochemical properties and environmental fate endpoints. J Cheminform 10:. https://doi.org/10.1186/s13321-018-0263-1
Wu Z, Ma X, Zheng X, et al (2014) Synthesis and characterization of thermochromic energy-storage microcapsule and application to fabric. The Journal of The Textile Institute 105:398–405. https://doi.org/10.1080/00405000.2013.814753
Geng X, Li W, Wang Y, et al (2018) Reversible thermochromic microencapsulated phase change materials for thermal energy storage application in thermal protective clothing. Appl Energy 217:281–294. https://doi.org/10.1016/j.apenergy.2018.02.150
Han Y, Yan X, Zhao W (2022) Effect of Thermochromic and Photochromic Microcapsules on the Surface Coating Properties for Metal Substrates. Coatings 12:1642. https://doi.org/10.3390/coatings12111642
Bradski G (2000) The openCV library. Dr Dobb’s Journal: Software Tools for the Professional Programmer 25:120–123
Sharma G, Wu W, Dalal EN (2005) The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations. Color Res Appl 30:21–30. https://doi.org/10.1002/col.20070
Yap CW (2011) PaDEL-descriptor: An open source software to calculate molecular descriptors and fingerprints. J Comput Chem 32:1466–1474. https://doi.org/10.1002/jcc.21707
Rogers D, Hopfinger AJ (1994) Application of Genetic Function Approximation to Quantitative Structure-Activity Relationships and Quantitative Structure-Property Relationships. J Chem Inf Comput Sci 34:854–866. https://doi.org/10.1021/ci00020a020
Davies M, Nowotka M, Papadatos G, et al (2015) ChEMBL web services: Streamlining access to drug discovery data and utilities. Nucleic Acids Res 43:W612–W620. https://doi.org/10.1093/nar/gkv352
Mendez D, Gaulton A, Bento AP, et al (2019) ChEMBL: Towards direct deposition of bioassay data. Nucleic Acids Res 47:D930–D940. https://doi.org/10.1093/nar/gky1075
Sterling T, Irwin JJ (2015) ZINC 15 - Ligand Discovery for Everyone. J Chem Inf Model 55:2324–2337. https://doi.org/10.1021/acs.jcim.5b00559
Kim S, Chen J, Cheng T, et al (2021) PubChem in 2021: New data content and improved web interfaces. Nucleic Acids Res 49:D1388–D1395. https://doi.org/10.1093/nar/gkaa971
Landrum G (2013) RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. Greg Landrum
Schwaller P, Gaudin T, Lányi D, et al (2018) “Found in Translation”: predicting outcomes of complex organic chemistry reactions using neural sequence-to-sequence models. Chem Sci 9:6091–6098. https://doi.org/10.1039/C8SC02339E
Li X, Fourches D (2021) SMILES Pair Encoding: A Data-Driven Substructure Tokenization Algorithm for Deep Learning. J Chem Inf Model 61:1560–1569. https://doi.org/10.1021/acs.jcim.0c01127
Kingma DP, Welling M (2013) Auto-encoding variational bayes. arXiv preprint arXiv:13126114
Chung J, Gulcehre C, Cho K, Bengio Y (2014) Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:14123555
Williams RJ, Zipser D (1989) A Learning Algorithm for Continually Running Fully Recurrent Neural Networks. Neural Comput 1:270–280. https://doi.org/10.1162/neco.1989.1.2.270
Abadi M, Agarwal A, Barham P, et al (2016) Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:160304467
Chollet F (2018) Keras: The python deep learning library. Astrophysics source code library ascl-1806
Harris CR, Millman KJ, van der Walt SJ, et al (2020) Array programming with NumPy. Nature 585:357–362. https://doi.org/10.1038/s41586-020-2649-2
McKinney W (2010) Data Structures for Statistical Computing in Python. In: Proceedings of the 9th Python in Science Conference. pp 56–61
Kessler T, St. John PC, Zhu J, et al (2021) A comparison of computational models for predicting yield sooting index. Proceedings of the Combustion Institute 38:1385–1393. https://doi.org/10.1016/j.proci.2020.07.009
Ruecker G, Ruecker C (1993) Counts of all walks as atomic and molecular descriptors. J Chem Inf Comput Sci 33:683–695. https://doi.org/10.1021/ci00015a005
Moran PAP (1950) Notes on Continuous Stochastic Phenomena. Biometrika 37:17. https://doi.org/10.2307/2332142
Todeschini R, Consonni V (2009) Molecular Descriptors for Chemoinformatics. Wiley
De P, Aher RB, Roy K (2018) Chemometric modeling of larvicidal activity of plant derived compounds against zika virus vector Aedes aegypti : application of ETA indices. RSC Adv 8:4662–4670. https://doi.org/10.1039/C7RA13159C
Roy K, Ghosh G (2004) QSTR with Extended Topochemical Atom Indices. 2. Fish Toxicity of Substituted Benzenes. J Chem Inf Comput Sci 44:559–567. https://doi.org/10.1021/ci0342066
Roy K, Das RN (2011) On some novel extended topochemical atom (ETA) parameters for effective encoding of chemical information and modelling of fundamental physicochemical properties. SAR QSAR Environ Res 22:451–472. https://doi.org/10.1080/1062936X.2011.569900
Moreau G, Broto P (1980) The autocorrelation of a topological structure: a new molecular descriptor. Nouv J Chim 4:359–360
Antanasijević J, Antanasijević D, Pocajt V, et al (2016) A QSPR study on the liquid crystallinity of five-ring bent-core molecules using decision trees, MARS and artificial neural networks. RSC Adv 6:18452–18464. https://doi.org/10.1039/C5RA20775D
Galvez J, Garcia R, Salabert MT, Soler R (1994) Charge Indexes. New Topological Descriptors. J Chem Inf Comput Sci 34:520–525. https://doi.org/10.1021/ci00019a008
Perez P, Pulgar R, Olea-Serrano F, et al (1998) The estrogenicity of bisphenol A-related diphenylalkanes with various substituents at the central carbon and the hydroxy groups. Environ Health Perspect 106:167–174. https://doi.org/10.1289/ehp.98106167

No competing interests reported.

Supplementarymaterial.xlsx

Download PDF

Version 1

posted

You are reading this latest preprint version

Discovering New Color Developers for Thermochromic Microcapsules Using a Variational Autoencoder

Status:

Version 1

Abstract

Figures

1. Introduction

2. Method

2.1 Measuring Thermochromic Microcapsule Properties

2.2 Predicting Color-Changing Temperatures

2.3 Preprocessing Data

2.4 Training Generative Models

2.5 Generating New Color Developers

3. Result and Discussion

3.1 Thermochromic Property of Microcapsules with 11 Different Color Developers

3.2 Predictive Models for Color-Changing Temperatures, DS, DE, CS, and CE

3.3 Candidates from Databases

3.4 Training Deep Generative Model

3.5 Generation of New Color Developer Candidates

3.6 Analysis of Candidates for Color Developers

4. Conclusion

Declarations

Author Contributions

References

Additional Declarations

Supplementary Files

Status:

Version 1