Machine learning-based aie molecule screening method and device

Abstract

The embodiment discloses an AIE molecule screening method and device based on machine learning, a virtual to-be-screened database is constructed, each AIE molecule in the database is subjected to molecular feature processing to obtain a molecular feature matrix, the molecular feature matrix corresponding to each AIE molecule is sequentially input into a wavelength prediction model that has been trained to obtain predicted wavelength information corresponding to each AIE molecule output by the wavelength prediction model, and AIE molecules with wavelengths meeting preset requirements are screened out from the AIE molecules according to the predicted wavelength information of each AIE molecule. The AIE molecule screening method disclosed by the embodiment can screen a large-scale unknown molecular structure space to obtain high-performance fluorescent molecular materials, the method accelerates the development of high-performance organic fluorescent materials, and has the advantages of low calculation cost, high accuracy and high efficiency.

Description

Technical Field

[0001] This application relates to the field of organic fluorescent materials technology, and in particular to an AIE molecular screening method and device based on machine learning. Background Technology

[0002] Due to their excellent biocompatibility and physicochemical properties, aggregation-induced emission (AIE) molecules have played an important role in biomedical fluorescence imaging.

[0003] In existing technologies, there are two different methods for screening AIE molecules at specific wavelengths. One method uses traditional chemical synthesis to screen AIE molecules for specific applications. However, traditional chemical synthesis methods are time-consuming and labor-intensive, and because AIE molecules have complex photophysical properties, their optical properties are highly dependent on solvent polarity, making it difficult to rationally design AIE molecules with the desired properties. The other method is to predict using computational methods, such as time-dependent density functional theory. However, this method suffers from significant errors because different functionals and basis sets can have a huge impact on performance, and its computational cost is high.

[0004] Therefore, existing methods cannot effectively achieve rapid screening of AIE molecules, and existing technologies need to be improved. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide users with a machine learning-based AIE molecule screening method and device, overcoming the shortcomings of the existing methods in that they cannot efficiently achieve rapid screening of AIE molecules.

[0006] The technical solution adopted by this invention to solve the technical problem is as follows:

[0007] This embodiment discloses an AIE molecular screening method based on machine learning, which includes:

[0008] A virtual database to be screened is constructed; wherein, the virtual database to be screened includes multiple AIE molecules, each of which is composed of a collected electron donor, electron acceptor and π bridge combined in the form of molecular docking;

[0009] Each AIE molecule is subjected to molecular characterization processing to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing; wherein, the molecular characterization processing is to encode the atomic feature information of each atom in the AIE molecule.

[0010] The molecular feature matrix corresponding to each AIE molecule is sequentially input into the trained wavelength prediction model to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model; wherein, the wavelength prediction model is trained based on the correspondence between the molecular features of multiple AIE molecules and their wavelengths.

[0011] Based on the predicted wavelength information of each AIE molecule, AIE molecules whose wavelengths meet the preset requirements are selected.

[0012] Optionally, the training method for the wavelength prediction model includes:

[0013] Construct a training dataset, which contains multiple sets of sample AIE molecular data. Each set of sample AIE molecular data includes AIE molecular feature information and wavelength information corresponding to the AIE molecular feature information.

[0014] The AIE molecule data of each group of samples are input into a preset network model to obtain the wavelength prediction value of the sample AIE molecule output by the preset network model.

[0015] Based on the error between the predicted AIE molecular wavelength of the sample and the actual AIE molecular wavelength of the sample, the parameters of the preset network model are corrected.

[0016] Continue performing the steps described above: inputting the sample AIE molecular data of each group into the preset network model, and correcting the parameters of the preset network model according to the wavelength prediction and actual values ​​of the sample AIE molecules, until the training of the preset network model meets the preset conditions, so as to obtain the wavelength prediction model.

[0017] Optionally, the structure of the preset neural network model is a graph convolutional neural network model structure; the step of inputting the sample AIE molecule data of each group into the preset network model to obtain the wavelength prediction value of the sample AIE molecule output by the preset network model includes:

[0018] The AIE molecular data of the sample is processed to obtain the AIE molecular feature matrix of the sample.

[0019] The feature matrix of the sample AIE molecules is input into a graph convolutional neural network model, and the graph convolutional neural network model outputs the wavelength prediction value of the sample AIE molecules.

[0020] Optionally, the step of constructing the virtual database to be filtered includes:

[0021] Multiple substructures with electron donors, electron acceptors, and π bridges were collected;

[0022] Using a molecular space generation algorithm, multiple electron donor, electron acceptor, and / or π-bridge substructures are docked and combined to obtain multiple AIE molecules after docking and combination; wherein, each substructure includes one or more binding sites; the molecular space generation algorithm is based on permutation and combination to establish docking of binding sites for each electron donor, electron acceptor, and / or π-bridge substructure through single bonds.

[0023] A virtual database of molecules to be screened was constructed using multiple AIE molecules obtained after docking and combination.

[0024] Optionally, the molecular feature matrix includes: an atomic feature matrix and an adjacency matrix;

[0025] The step of performing molecular characterization processing on each AIE molecule to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing includes:

[0026] Atomic features of each atom and chemical bond relationships between each adjacent atom were extracted from each AIE molecule.

[0027] The atomic feature matrix is ​​obtained by encoding the atomic features of each atom in each AIE molecule, and the adjacency matrix is ​​obtained by encoding the chemical bond relationships between each adjacent atom.

[0028] Optionally, the wavelength prediction model includes: multiple graph convolutional layers, multiple linear rectifier units, and multiple fully connected layers; the linear rectifier units are disposed between two graph convolutional layers and between two fully connected layers.

[0029] Optionally, the number of graph convolutional layers is k; the number of fully connected layers is J, where k and J are positive integers.

[0030] The step of sequentially inputting the molecular feature matrix corresponding to each AIE molecule into the trained wavelength prediction model to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model includes:

[0031] The atomic feature matrix and adjacency matrix corresponding to each AIE molecule are sequentially and synchronously input into the first graph convolutional layer to obtain the first feature information output by the first graph convolutional layer;

[0032] The first feature information is input into the first linear rectifier unit connected to the first graph convolutional layer to obtain the first linear information output by the first linear rectifier unit;

[0033] The first linear information is input into the second graph convolutional layer to obtain the second feature information output by the second graph convolutional layer.

[0034] The second feature information is input into the second linear rectifier unit connected to the second graph convolutional layer to obtain the second linear information output by the second linear rectifier unit;

[0035] The steps of inputting the feature information output by the graph convolutional layer into the linear rectifier unit and inputting the information output by the linear rectifier unit into the next graph convolutional layer are executed sequentially until the k-th graph convolutional layer is reached, so as to obtain the k-th feature information output by the k-th graph convolutional layer.

[0036] The k-th feature information is sequentially input into the first fully connected layer, the (k+1)-th linear rectifier unit, the second fully connected layer, and so on, until it is input into the J-th fully connected layer, to obtain the predicted wavelength information corresponding to each AIE molecule output by the J-th fully connected layer.

[0037] Optionally, the step of filtering out AIE molecules whose wavelengths meet preset requirements based on the predicted wavelength information of each AIE molecule includes:

[0038] Obtain wavelength range information for preset molecular absorption and emission peaks;

[0039] AIE molecules whose wavelengths fall within the wavelength range information are selected from the predicted wavelength information.

[0040] Secondly, this embodiment discloses an information processing terminal, including a memory and a processor. The memory stores a computer program, wherein the processor executes the computer program to implement the steps of the method.

[0041] Thirdly, this embodiment discloses a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the method described.

[0042] This embodiment discloses a machine learning-based AIE molecule screening method and device. A virtual screening database is constructed, comprising multiple AIE molecules, each consisting of collected electron donors, electron acceptors, and π-bridges combined via molecular docking. Molecular feature processing is performed on each AIE molecule to obtain a molecular feature matrix corresponding to each molecule. This molecular feature processing involves encoding the atomic feature information of each atom in the molecule. The molecular feature matrices corresponding to each AIE molecule are sequentially input into a trained wavelength prediction model, which outputs predicted wavelength information corresponding to each AIE molecule. This wavelength prediction model is trained based on the correspondence between the molecular features of multiple AIE molecules and their wavelengths. Based on the predicted wavelength information of each AIE molecule, AIE molecules whose wavelengths meet preset requirements are screened. This AIE molecule screening method disclosed in this embodiment, through large-scale screening of unknown molecular structures, obtains high-performance fluorescent molecular materials. This method accelerates the development of high-performance organic fluorescent materials and has the advantages of low computational cost, high accuracy, and high efficiency. Attached Figure Description

[0043] Figure 1 This is a flowchart of the AIE molecular screening method in an embodiment of the present invention;

[0044] Figure 2a This is a schematic diagram of the structure of the electron donor in an embodiment of the present invention;

[0045] Figure 2b This is a schematic diagram showing the structure of the electron acceptor in an embodiment of the present invention;

[0046] Figure 2c This is a schematic diagram of the structure of the π-bridge in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram illustrating the principle of molecular docking using the molecular space generation algorithm in an embodiment of the present invention.

[0048] Figure 4 This is a schematic diagram illustrating the principle of molecular feature structure information in a convolutional neural network in an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the wavelength prediction model in an embodiment of the present invention;

[0050] Figure 6 This is a flowchart illustrating the steps of an application embodiment of the AIE molecular screening method in this invention.

[0051] Figure 7aThis is a scatter plot of the absorption peak predictions of the test set by the wavelength prediction model in an embodiment of the present invention.

[0052] Figure 7b This is a scatter plot of the emission peak prediction of the test set by the wavelength prediction model in an embodiment of the present invention. Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] Fluorescent probes, especially organic light emitters, are widely used in biological research due to their lower safety risks and biodegradability. However, traditional organic probes still suffer from inherent drawbacks such as poor water solubility, severe photobleaching, and low stability. These inherent drawbacks significantly hinder the widespread application of organic probes in biomedical research.

[0055] Fortunately, luminescent bioluminescent molecules (AIE molecules) with AIE properties provide a perfect solution. AIE molecules exhibit weak or no emission in their molecular state, but display highly enhanced fluorescence emission in their aggregated state. This characteristic endows AIE molecules with good colloidal stability, greater photobleaching resistance, and highly enhanced functional stability. Therefore, in recent years, AIE molecules have attracted widespread attention among scientists and are considered a better choice than traditional organic probes.

[0056] AIE molecules with different wavelengths can be used in various biological applications, such as in vivo fluorescence imaging, orthogonal monitoring, multicolor fluorescent labeling, and fluorescence resonance energy transfer (FRET) analysis. Furthermore, adjusting the wavelength of AIE molecules can achieve appropriate penetration depths and minimize interference from tissue absorption, scattering, and autofluorescence, resulting in a better signal-to-noise ratio. AIE molecules with specific absorption and emission peaks, in particular, have shown good applicability in these applications, thus researchers have shown great interest in designing AIE molecules with appropriate absorption and emission peaks (λabs, λem). Therefore, a thorough understanding of the structure-property relationship between molecular structure and optical properties is crucial before proceeding with chemical synthesis.

[0057] The challenge of studying the structure-property relationship between molecular structure and optical properties stems from complex AIE mechanisms across various photophysical dimensions, such as restricting intramolecular rotation or vibration, restricting excited-state deformation, and suppressing Kasha's rule. Furthermore, it is noteworthy that the optical properties of molecules are also highly dependent on solvent polarity, making it even more difficult to rationally design molecules with the desired properties.

[0058] While existing computational methods, such as time-dependent density functional theory (TD-DFT), can be used to calculate the absorption and emission peaks of molecules, the predictions are often inaccurate and computationally expensive due to the significant impact of different functionals and basis sets on performance. Furthermore, TD-DFT systematically overestimates absorption and emission energies in many cases. Although some optimization strategies can address these issues, they lead to a significant increase in computational cost. In addition, TD-DFT exhibits significant errors on some frameworks and has extremely high time complexity, making it impractical for large-scale molecule screening. Therefore, a user-friendly method is needed that requires only information on molecular structure and solvent to guide the design and synthesis of AIE molecules.

[0059] To improve the efficiency of data processing, machine learning (ML) is increasingly being applied across various fields, achieving considerable success in areas such as drug design, organic synthesis, and materials chemistry. By scanning large datasets and extracting their molecular features, ML models can predict a wide range of properties without requiring knowledge of the underlying physical or chemical information. Even when the underlying physical or chemical mechanisms are already understood, machine learning can aid in further research, enabling researchers to develop molecules with the expected properties.

[0060] To better address the aforementioned problems in existing technologies, this embodiment discloses a machine learning-based AIE molecule screening method and device. It establishes a database containing sample datasets of known AIE molecules from existing technologies, then trains a preset network model using this database to obtain a trained wavelength prediction model. A molecular space generation algorithm is then used to connect the substructures of electron donors, electron acceptors, and π-bridges into multiple unknown AIE molecules. The trained wavelength prediction model is then used to predict the wavelengths of these unknown AIE molecules, allowing for the selection of suitable AIE molecules. Using the AIE molecule wavelength prediction method provided by this invention, researchers can design new aggregation-induced emission materials with specific wavelengths by leveraging the potential structure-property relationships of molecules, saving experimental and computational time and resources, improving experimental efficiency, and avoiding blind spots. This invention provides a novel molecular design strategy that, compared to traditional trial-and-error chemical synthesis routes, features low computational cost, high accuracy, and high efficiency.

[0061] The following detailed description, in conjunction with the accompanying drawings, illustrates a machine learning-based AIE molecular screening method provided in this embodiment.

[0062] This embodiment discloses an AIE molecular screening method based on machine learning, such as... Figure 1 As shown, it includes:

[0063] Step S1: Construct a virtual database to be screened; wherein the virtual database to be screened includes multiple AIE molecules, each of which is composed of a collected electron donor, electron acceptor and π bridge combined in the form of molecular docking.

[0064] To obtain more AIE molecules that meet the required composition, this step first constructs a virtual screening database. This database stores a large number of unknown molecules, which are constructed based on the docking of known structures. Since the docking of the substructures of the known molecules is random, and based on a large number of known molecular substructures, a huge molecular structure space can be formed. The molecular wavelength range within this molecular structure space is wide, but the wavelengths of the individual unknown molecules are unknown.

[0065] Specifically, the step of constructing the virtual database to be filtered includes:

[0066] Step S11: Collect multiple substructures of electron donors, electron acceptors and π bridges.

[0067] Since AIE molecules are composed of classical electron donors, acceptors, and π-bridges, molecular spatial generation docking algorithms can be used to combine different donors, acceptors, and π-bridges through specific binding sites, generating individual AIE molecules of the Donor-π-Acceptor (DA) and Donor-π-Acceptor-π-Donor (DAD) types. Figures 2a to 2c ,as well as Figure 3 and Figure 4 As shown, Figure 2a List of structures representing electron donors Figure 2b List of structures representing electron acceptors Figure 2c The structure is represented as a list of π-bridges. Figure 3 and Figure 4 The molecular space generation docking algorithm is used to dock the above substructures one by one to form a complete AIE molecule.

[0068] Step S12: Using a molecular space generation algorithm, multiple electron donor, electron acceptor, and / or π-bridge substructures are docked and combined to obtain multiple AIE molecules after docking and combination; wherein, the substructure includes one or more binding sites; the molecular space generation algorithm is based on permutation and combination to establish docking of binding sites for each electron donor, electron acceptor, and / or π-bridge substructure through single bonds.

[0069] Specifically, such as Figure 3As shown, the database to be screened contains multiple Donors (D), multiple Acceptors (A), and multiple π-bridge structures. Each D, A, and π-bridge structure is placed into a substructure pool. Since each D and each A has selectable binding sites, a molecular space generation algorithm is used to connect these D and A according to their binding sites through permutation and combination. In one embodiment, the permutation and combination method involves connecting D1 to A1, A2...An to form multiple AIE molecules such as D1A1, D1A2, D1A3... Then, D2 connects to A1, A2...An to obtain multiple AIE molecules such as D2A1, D2A2, D2A3... Then, D3, D4, D5...Dn connect to A1, A2...An in the same way. That is: connecting D1 to A1 forms the first AIE molecule, connecting D1 to A2 forms the second AIE molecule... connecting D1 to An forms the nth AIE molecule. After D1 is used, D2 is used to connect D1 to form the (n+1)th AIE molecule, D2 is used to connect A2 to form the (n+2)th AIE molecule, and so on, with D2 connecting An to form the 2nth AIE molecule. Alternatively, D can be connected to a π-bridge first and then to A to form a Donor-π-Acceptor (DA) type molecule, or D can be connected to a π-bridge and then to both sides of A to form a Donor-π-Acceptor-π-Donor (DAD) type molecule. Through these docking methods, individual AIE molecules are ultimately obtained.

[0070] Step S13: Construct a virtual database to be screened using multiple AIE molecules obtained after docking and combination.

[0071] Combination Figure 3 The multiple AIE molecules obtained after docking and combination are all saved to a virtual database to be screened, thus obtaining the constructed virtual database to be screened.

[0072] Since each AIE molecule is obtained through random docking, it is an unknown molecule with an unknown wavelength. However, due to the vast molecular space inherent in the AIE molecules during assembly, the wavelength range of the AIE molecules in the virtual screening database is relatively wide, and the number of molecules is large. Based on the increase in the number of substructures with electron donors, electron acceptors, and / or π-bridges, the number of unknown AIE molecules also increases. To achieve better screening results, in one embodiment, the number of molecules in the screening database exceeds 10,000.

[0073] Step S2: Perform molecular characterization processing on each AIE molecule to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing; wherein, the molecular characterization processing is to encode the atomic feature information of each atom in the molecule.

[0074] In this embodiment, learning is based on molecular features to predict the wavelength of each AIE molecule. In the specific data processing, each AIE molecule needs to be molecularly characterized to obtain molecular characteristic information, so that the information obtained after molecular characteristic processing can be transmitted to the wavelength prediction model for wavelength prediction.

[0075] Specifically, after performing molecular characterization on each molecule in this step, the resulting molecular feature matrix includes: an atomic feature matrix and an adjacency matrix; wherein, the atomic feature matrix is ​​a matrix that characterizes the features of the atom itself, and the adjacency matrix is ​​a matrix that characterizes the connection relationship between the atom and its neighboring atoms.

[0076] Specifically, as shown in Table 1, the atomic characteristic matrix includes: atom type, non-hydrogen degree of the atom, degree of the atom, hybridization of the atom, atomic number, implicit valence of the atom, etc.

[0077] Table 1. Atomic Feature Information

[0078] The adjacency matrix represents the connection relationship between atoms. If the adjacency matrix G(i,j)=1, it means that there is a covalent bond connecting atoms i and j. If G(i,j)=0, it means that there is no covalent bond connecting atoms i and j.

[0079] The step of performing molecular characterization processing on each AIE molecule to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing includes:

[0080] Step S21: Extract the atomic characteristics of each atom in each AIE molecule and the chemical bond relationships between each adjacent atom.

[0081] Atomic features of each atom in each AIE molecule, as well as the correlation features between each atom and its neighboring atoms, are extracted separately.

[0082] Step S22: Obtain the atomic feature matrix based on the atomic feature encoding of each atom in each AIE molecule, and obtain the adjacency matrix based on the chemical bond relationship between each adjacent atom.

[0083] The atomic feature matrix and adjacency matrix are obtained by encoding the atomic features of each extracted atom and the correlation features between atoms.

[0084] Combination Figure 4 As shown, data encoding is performed based on the extracted atomic features of each atom and the correlation features between atoms. A feature matrix is ​​obtained by encoding based on the atomic features of each atom and the number of atoms in each AIE molecule. An adjacency matrix is ​​then established based on the chemical bonds between adjacent atoms in each AIE molecule.

[0085] Step S3: Input the molecular feature matrix corresponding to each AIE molecule into the trained wavelength prediction model in sequence to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model; wherein, the wavelength prediction model is trained based on the correspondence between the molecular features of multiple AIE molecules and their wavelengths.

[0086] The molecular feature matrices corresponding to each AIE molecule obtained in step S2 above are sequentially input into the wavelength prediction model to obtain the predicted wavelength values ​​of each AIE molecule output by the wavelength prediction model.

[0087] In one implementation, combined with Figure 5 As shown, the wavelength prediction model includes: multiple graph convolutional layers, multiple linear rectifier units, and multiple fully connected layers; the linear rectifier units are arranged between two graph convolutional layers and between two fully connected layers.

[0088] Specifically, the number of graph convolutional layers is k; the number of fully connected layers is J, where k and J are positive integers.

[0089] The step of sequentially inputting the molecular feature matrix corresponding to each AIE molecule into the trained wavelength prediction model to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model includes:

[0090] The atomic feature matrix and adjacency matrix corresponding to each AIE molecule are sequentially and synchronously input into the first graph convolutional layer to obtain the first feature information output by the first graph convolutional layer;

[0091] The first feature information is input into the first linear rectifier unit connected to the first graph convolutional layer to obtain the first linear information output by the first linear rectifier unit;

[0092] The first linear information is input into the second graph convolutional layer to obtain the second feature information output by the second graph convolutional layer.

[0093] The second feature information is input into the second linear rectifier unit connected to the second graph convolutional layer to obtain the second linear information output by the second linear rectifier unit;

[0094] The steps of inputting the feature information output by the graph convolutional layer into the linear rectifier unit and inputting the information output by the linear rectifier unit into the next graph convolutional layer are executed sequentially until the k-th graph convolutional layer is reached, so as to obtain the k-th feature information output by the k-th graph convolutional layer.

[0095] The k-th feature information is sequentially input into the first fully connected layer, the (k+1)-th linear rectifier unit, the second fully connected layer, and so on, until it is input into the J-th fully connected layer, to obtain the predicted wavelength information corresponding to each AIE molecule output by the J-th fully connected layer.

[0096] Step S4: Select AIE molecules whose wavelengths meet the preset requirements based on the predicted wavelength information of each AIE molecule.

[0097] Once the wavelength prediction model outputs the predicted wavelength values ​​for each unknown AIE molecule, specific wavelength AIE molecules can be selected from them according to requirements.

[0098] Specifically, the training method for the wavelength prediction model includes:

[0099] A training dataset is constructed, which contains multiple sets of sample AIE molecular data. Each set of sample AIE molecular data includes AIE molecular feature information and wavelength information corresponding to the AIE molecular feature information.

[0100] In this embodiment, a training set for training a preset network model is first obtained. The training set contains multiple sets of training samples, each set including sample AIE molecular data and the corresponding wavelength information. Based on the correspondence between the sample AIE molecular data and the corresponding wavelength information, the preset network model is trained to obtain a trained wavelength prediction model.

[0101] Specifically, the AIE molecular data of each group of samples are input into a preset network model to obtain the predicted wavelength of the sample AIE molecules output by the preset network model; the parameters of the preset network model are corrected based on the error between the predicted wavelength of the sample AIE molecules and the actual wavelength of the sample AIE molecules; the above steps of inputting the AIE molecular data of each group of samples into the preset network model and correcting the parameters of the preset network model based on the predicted and actual wavelengths of the sample AIE molecules are continued until the training of the preset network model meets the preset conditions, thus obtaining the wavelength prediction model.

[0102] The structure of the preset network model is a graph convolutional neural network model structure; the step of inputting the sample AIE molecule data of each group into the preset network model to obtain the predicted value of the sample AIE molecule wavelength output by the preset network model includes:

[0103] The AIE molecular data of the sample is processed to obtain the AIE molecular feature matrix of the sample.

[0104] The feature matrix of the sample AIE molecules is input into a graph convolutional neural network model, and the graph convolutional neural network model outputs the predicted wavelength of the sample AIE molecules.

[0105] Specifically, the preset conditions include the loss function value meeting a preset requirement or the training iterations reaching a preset number. The preset requirement can be determined based on the accuracy of the predicted wavelength, which will not be detailed here. The preset number of training iterations can be the maximum number of training iterations for the preset network model, for example, 5000. Therefore, after the preset network model outputs the predicted wavelength and calculates the loss function value, it is determined whether the loss function value meets the preset requirement. If the loss function value meets the preset requirement, training ends. If the loss function value does not meet the preset requirement, it is determined whether the training iterations of the preset network model have reached the predicted number. If not, the network parameters of the preset network model are corrected based on the loss function value. If the preset number of training iterations has been reached, training ends. This method of determining whether the preset network model training has ended using the loss function value and the number of training iterations avoids the preset network model entering an infinite loop due to the loss function value failing to meet the preset requirement.

[0106] In one embodiment, three different evaluation metrics are used to evaluate the model, including mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 This allows for the training of a wavelength prediction model.

[0107] Specifically, the step of selecting AIE molecules whose wavelengths meet preset requirements based on the predicted wavelength information of each AIE molecule includes:

[0108] Obtain wavelength range information for preset molecular absorption and emission peaks;

[0109] AIE molecules whose wavelengths fall within the wavelength range are selected from the predicted wavelength information.

[0110] In the embodiments provided by this invention, machine learning (ML) techniques are applied to screen AIE molecules with expected excitation and emission wavelengths for biomedical deep fluorescence imaging. First, a database of various AIE molecules collected from the literature is established. Then, by extracting molecular features and training various state-of-the-art ML models, the structure-property relationships of AIE molecules are obtained, and the absorption and emission wavelength peaks are predicted. Compared to first-principles calculations, the proposed strategy provides higher accuracy at a lower computational cost.

[0111] The following is combined with Figure 6 As shown, a more detailed explanation of specific application embodiments of this method is provided.

[0112] The steps of the method described in this invention in specific applications include the following:

[0113] Step H1: Constructing the Training Database and Splitting the Dataset: First, a training database containing experimental data for 1245 AIE molecules was created, drawn from literature published over the past 20 years. Each data entry includes the molecular structure, solvent, absorption, and emission peaks of the AIE molecule. If a particular solvated AIE molecule has many absorption and emission peaks, the longest wavelength absorption peak and the highest intensity emission peak are collected. In short, the absorption and emission peaks of most AIE molecules are located in the visible to near-infrared region (400–700 nm). The AIE characteristics of each molecule in the training database are documented in the literature. These AIE molecules include rotor structures or their derivatives, such as triphenylamine (TPA), tetraphenylpyrazine (TPP), tetraphenylethylene (TPE), and hexaphenylthiophene (HPS). After the dataset was constructed, it was divided into training, validation, and test sets in an 8:1:1 ratio.

[0114] Step H2: Perform molecular feature processing on each AIE molecule in the training database to obtain the molecular feature matrix corresponding to each AIE molecule after molecular feature processing.

[0115] Graph Convolutional Neural Networks (GCNs) treat atoms as nodes in a graph and the bonds between atoms as edges. Therefore, compared to descriptor-based machine learning algorithms, they can better learn the relationships between molecular structure and properties, thus providing higher prediction accuracy. Similar to convolution of conventional data such as images and text, GCNs extend this operation to irregular molecular graphs as a natural representation of chemical structure. More specifically, a graph G = (V, E) can be defined as the connections between a set of nodes (V) and a set of edges (E). Naturally, a molecule can also be considered as a graph composed of a set of atoms (nodes) and a set of bonds (edges). Essentially, GCNs aim to learn the representation of each atom by recursively passing messages across the molecular graph, aggregating information from neighboring atoms encoded by atom feature vectors. Before encoding the graph, node features need to be defined. As shown in Table 1, 16 node features were used to represent molecules in this study. Most of these features are encoded using one-hot encoding, except for atomic charge, atomic mass, and the number of free radical electrons, which are encoded as integers due to their additive nature.

[0116] Step H3, Machine Learning Model Development and Validation: The dataset is randomly divided into training set, validation set and test set in a ratio of 8:1:1. The wavelength prediction model is obtained by multiple independent training and validation processes.

[0117] To evaluate the effectiveness of the algorithm, combined with Figure 7a and Figure 7b As shown, the model was evaluated using three different evaluation metrics: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). The final results were obtained through 10 independent training and validation runs. The convolutional neural network model demonstrated good performance in predicting the absorption and emission peaks of AIE molecules, with coefficients of determination as high as 0.99 and 0.97 for absorption and emission peak prediction, respectively. This result indicates that the wavelength prediction model proposed in this embodiment has excellent performance and can be well applied to practical screening applications. Furthermore, Bayesian optimization was used during model training to determine the optimal hyperparameters within a wide parameter range (Table 2). This step is crucial because it has been proven that selecting appropriate hyperparameters can produce better prediction results. Table 3 shows the selected optimal parameters for the model.

[0118] Table 2. Available Hyperparameter Values

[0119] Table 3. Optimal hyperparameters of the model

[0120] Step H4: Construct a virtual database of molecules to be screened and use a machine learning model to predict the molecular wavelengths in the database: (e.g.) Figure 3 As shown, a database containing over 10,000 potential AIE molecules was constructed using a molecular space generation algorithm for screening machine learning models. These AIE molecules consist of classic electron donor, electron acceptor, and π-bridge substructures. Through specific binding sites, a molecular space generation docking algorithm was used to combine different donors, acceptors, and π-bridges to generate individual AIE molecules of the Donor-π-Acceptor (DA) and Donor-π-Acceptor-π-Donor (DAD) types. Furthermore, the compounds in the database constructed in this way have fewer synthetic steps, making them relatively easy to synthesize and potentially addressing the issue of unsynthesizable compounds in the final screening, thus reducing the difficulty of molecular synthesis.

[0121] Step H5: Use the trained machine learning model to predict the absorption and emission peaks of each AIE molecule in the database to be screened, and screen out the required AIE molecules according to the needs.

[0122] The method provided in this invention utilizes a database constructed from AIE molecular data collected from the literature. Through feature extraction, it establishes an efficient wavelength prediction model, offering advantages such as simplicity, convenience, and low cost. Using this wavelength prediction model to predict the wavelengths of aggregation-induced emission molecules, the potential structure-property relationships can help researchers design novel AIE molecules with specific wavelengths, saving experimental and computational time and resources, improving experimental efficiency, and avoiding blind experimentation.

[0123] Based on the method described above, this embodiment also discloses an information processing terminal, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method.

[0124] The processor typically controls the overall operation of the device, such as operations associated with display, telephone calls, data communication, camera operation, and recording. The processor can perform all or part of the steps of the AIE molecular screening method described above. Furthermore, the processor may include one or more modules to facilitate interaction between the processor and other components. For example, the processor may include a multimedia module to facilitate interaction between multimedia components and the processor. In some embodiments, the processor may be a central processing unit (CPU), a microprocessor, or other data processing chip for running program code stored in the memory or processing data, such as performing the training steps of the wavelength prediction model.

[0125] Memory is configured to store various types of data to support the operation of a device. Examples of this data include instructions for any application or method operating on the device, contact data, phonebook data, messages, pictures, videos, etc. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0126] In some embodiments, the memory may be an internal storage unit of the playback device, such as the hard drive or memory of a smart TV. In other embodiments, the memory may be an external storage device of the playback device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the smart TV. Further, the memory may include both internal and external storage units of the playback device. The memory is used to store application software and various types of data installed on the playback device, such as the program code installed on the smart TV. The memory can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory stores a control program for a screening method of AIE molecules. This machine learning-based AIE molecule screening method control program can be executed by a processor to implement the AIE molecule screening method in this embodiment.

[0127] In an exemplary embodiment, the apparatus may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0128] This embodiment also discloses a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the AIE molecule screening method.

[0129] This embodiment discloses a machine learning-based AIE molecule screening method and device. A virtual screening database is constructed, comprising multiple AIE molecules, each composed of collected electron donors, electron acceptors, and π-bridges connected via molecular docking. Molecular feature processing is performed on each AIE molecule to obtain a molecular feature matrix corresponding to each molecule. This molecular feature processing involves encoding the atomic feature information of each atom in the molecule. The molecular feature matrix and adjacency matrix of each AIE molecule are sequentially input into a trained wavelength prediction model, which outputs predicted wavelength information corresponding to each AIE molecule. This wavelength prediction model is trained based on the correspondence between the molecular features of multiple AIE molecules and their wavelengths. Based on the predicted wavelength information of each AIE molecule, AIE molecules with wavelengths meeting preset requirements are screened. This AIE molecule screening method, through large-scale screening of unknown molecular structures, obtains high-performance fluorescent molecular materials. This method accelerates the development of high-performance organic fluorescent materials and has the advantages of low computational cost, high accuracy, and high efficiency.

[0130] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0131] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A machine learning-based AIE molecular screening method, characterized in that, include: A virtual database to be screened is constructed; wherein, the virtual database to be screened includes multiple AIE molecules, each of which is composed of a collected electron donor, electron acceptor and π bridge combined in the form of molecular docking; Each AIE molecule is subjected to molecular characterization processing to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing; wherein, the molecular characterization processing is to encode the atomic feature information of each atom in the AIE molecule. The molecular feature matrix corresponding to each AIE molecule is sequentially input into the trained wavelength prediction model to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model; wherein, the wavelength prediction model is trained based on the correspondence between the molecular features of multiple AIE molecules and their wavelengths. Based on the predicted wavelength information of each AIE molecule, AIE molecules whose wavelengths meet the preset requirements are selected. The wavelength prediction model includes: multiple graph convolutional layers, multiple linear rectifier units, and multiple fully connected layers; the linear rectifier units are disposed between two graph convolutional layers and between two fully connected layers; The number of fully connected layers is J, where k and J are positive integers; The step of sequentially inputting the molecular feature matrix corresponding to each AIE molecule into the trained wavelength prediction model to obtain the predicted wavelength information of each AIE molecule output by the wavelength prediction model includes: The atomic feature matrix and adjacency matrix corresponding to each AIE molecule are sequentially and synchronously input into the first graph convolutional layer to obtain the first feature information output by the first graph convolutional layer; The first feature information is input into the first linear rectifier unit connected to the first graph convolutional layer to obtain the first linear information output by the first linear rectifier unit; The first linear information is input into the second graph convolutional layer to obtain the second feature information output by the second graph convolutional layer. The second feature information is input into the second linear rectifier unit connected to the second graph convolutional layer to obtain the second linear information output by the second linear rectifier unit; The steps of inputting the feature information output by the graph convolutional layer into the linear rectifier unit and inputting the information output by the linear rectifier unit into the next graph convolutional layer are executed sequentially until the k-th graph convolutional layer is reached, so as to obtain the k-th feature information output by the k-th graph convolutional layer. The k-th feature information is sequentially input into the first fully connected layer, the (k+1)-th linear rectifier unit, the second fully connected layer, and so on, until it is input into the J-th fully connected layer, to obtain the predicted wavelength information corresponding to each AIE molecule output by the J-th fully connected layer.

2. The AIE molecular screening method according to claim 1, characterized in that, The training method for the wavelength prediction model includes: Construct a training dataset, which contains multiple sets of sample AIE molecular data. Each set of sample AIE molecular data includes AIE molecular feature information and wavelength information corresponding to the AIE molecular feature information. The AIE molecule data of each group of samples are input into a preset network model to obtain the wavelength prediction value of the sample AIE molecule output by the preset network model. Based on the error between the predicted wavelength value of the sample AIE molecule and the actual wavelength value of the sample AIE molecule, the parameters of the preset network model are corrected until the training of the preset network model meets the preset conditions to obtain the wavelength prediction model.

3. The AIE molecular screening method according to claim 2, characterized in that, The structure of the preset network model is a graph convolutional neural network model structure; the step of inputting the sample AIE molecule data of each group into the preset network model to obtain the wavelength prediction value of the sample AIE molecule output by the preset network model includes: The AIE molecular data of the sample is processed to obtain the AIE molecular feature matrix of the sample. The feature matrix of the sample AIE molecules is input into a graph convolutional neural network model, which outputs the wavelength prediction value of the sample AIE molecules.

4. The AIE molecular screening method according to claim 1, characterized in that, The steps for constructing the virtual database to be filtered include: Multiple substructures with electron donors, electron acceptors, and π bridges were collected; Using a molecular space generation algorithm, multiple electron donor, electron acceptor, and / or π-bridge substructures are docked and combined to obtain multiple AIE molecules after docking and combination; wherein, each substructure includes one or more binding sites; the molecular space generation algorithm is based on permutation and combination to establish docking of binding sites for each electron donor, electron acceptor, and / or π-bridge substructure through single bonds. A virtual database of molecules to be screened was constructed using multiple AIE molecules obtained after docking and combination.

5. The AIE molecular screening method according to claim 4, characterized in that, The molecular feature matrix includes: an atomic feature matrix and an adjacency matrix; The step of performing molecular characterization processing on each AIE molecule to obtain the molecular feature matrix corresponding to each AIE molecule after molecular characterization processing includes: Atomic features of each atom and chemical bond relationships between each adjacent atom were extracted from each AIE molecule. The atomic feature matrix is ​​obtained by encoding the atomic features of each atom in each AIE molecule, and the adjacency matrix is ​​obtained by encoding the chemical bond relationships between each adjacent atom.

6. The AIE molecular screening method according to claim 5, characterized in that, The step of selecting AIE molecules whose wavelengths meet preset requirements based on the predicted wavelength information of each AIE molecule includes: Obtain wavelength range information for preset molecular absorption and emission peaks; AIE molecules whose wavelengths fall within the wavelength range information are selected from the predicted wavelength information.

7. An information processing terminal, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images