A photosensitizer screening method, apparatus, equipment and medium based on deep learning

Abstract

This application discloses a deep learning-based method, apparatus, device, and medium for screening photosensitizers, relating to the field of bioinformatics. The method includes: representing each standardized candidate molecule as a corresponding graph structure; extracting node feature vectors for each atom and edge feature vectors for each chemical bond; determining graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms; performing representation learning on the graph feature data using a directed message passing neural network to obtain molecular representation results; determining predicted values ​​of performance indicators for each candidate molecule based on the molecular representation results; training multiple candidate prediction models; determining the prediction performance results of each candidate prediction model for each performance indicator; selecting a target prediction model from among the candidate prediction models based on the prediction performance results; and using the target prediction model to screen target molecules as photosensitizers from among the candidate molecules. This application can improve screening efficiency.

Description

Technical Field

[0001] This application relates to the field of bioinformatics processing technology, and in particular to a method, apparatus, device and medium for screening photosensitizers based on deep learning. Background Technology

[0002] Photodynamic therapy (PDT) is a promising non-invasive or minimally invasive tumor treatment. It works by using photosensitizers to generate reactive oxygen species (ROS) under specific wavelengths of light, inducing tumor cell damage or death. PDT offers advantages such as high selectivity, repeatability, and fewer adverse reactions, demonstrating significant potential for precision oncology.

[0003] In recent years, the integration of near-infrared fluorescence imaging (NIIR) with photodynamic therapy (PDT) has attracted widespread attention. This strategy offers advantages such as deep tissue penetration, high spatiotemporal resolution, and minimal off-target damage, promising to achieve precise tumor identification, real-time imaging guidance, and efficient treatment. However, the development of high-performance therapeutic optical probes still faces significant bottlenecks. On the one hand, an ideal probe must simultaneously possess long-wavelength absorption / emission, excellent two-photon imaging capabilities, efficient intersystem crossing, and strong ROS generation capabilities, making the synergistic optimization of multiple properties quite challenging. On the other hand, the traditional "design-synthesis-characterization-screening" development model is time-consuming and costly, and the relationship between structure and performance remains unclear, limiting the rapid discovery of highly efficient photosensitizers and optical probes. Summary of the Invention

[0004] In view of this, embodiments of this application provide a photosensitizer screening method, apparatus, device, and medium based on deep learning to efficiently screen for photosensitizers.

[0005] One aspect of this application provides a photosensitizer screening method based on deep learning, the method comprising the following steps:

[0006] Each standardized candidate molecule is represented as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule.

[0007] Extract node feature vectors for each atom and edge feature vectors for each chemical bond. Determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms.

[0008] A directed message-passing neural network is used to perform representation learning on the graph feature data to obtain a first molecular representation of the overall electronic structure and topological features of the candidate molecule.

[0009] The predicted values ​​of the performance indicators of each candidate molecule are determined based on the first molecule representation results;

[0010] Using the predicted values ​​of each performance index as training labels and the corresponding candidate molecules as training samples, multiple candidate prediction models are trained.

[0011] Determine the prediction performance results of each of the trained candidate prediction models for the predicted values ​​of each of the performance metrics;

[0012] The target prediction model is selected from each of the candidate prediction models based on the prediction performance results;

[0013] The target molecule is selected as a photosensitizer from the candidate molecules using the target prediction model.

[0014] In some embodiments, determining the predicted values ​​of the performance indicators of each candidate molecule based on the first molecule representation result includes the following steps:

[0015] Based on the first molecule representation result, the predicted values ​​corresponding to the absorption peak wavelength and the intersystem crossing rate between singlet and triplet states of each candidate molecule are determined.

[0016] Alternatively, explicit features are extracted from each of the candidate molecules as auxiliary features; the first molecule representation result and the auxiliary features are fused to obtain the second molecule representation result; and the predicted values ​​corresponding to the absorption peak wavelength and the singlet and triplet crossover rates of each candidate molecule are determined based on the second molecule representation result.

[0017] In some embodiments, the step of extracting explicit features from each of the candidate molecules as auxiliary features includes the following steps:

[0018] At least one of the RDKit molecular descriptor, MACCS molecular fingerprint, or Morgan molecular fingerprint is extracted from each of the candidate molecules as the auxiliary feature.

[0019] In some embodiments, training multiple candidate prediction models includes the following steps:

[0020] The first molecule representation result and the RDKit molecule descriptor are concatenated to obtain a fused representation;

[0021] The fusion representation is input into the prediction model for training, and a physical constraint loss term is introduced during the training process;

[0022] The physical constraint loss term includes at least one of the following:

[0023] The predicted value falls within the preset physical range constraint;

[0024] Monotonicity constraints apply to a known relationship between structure and performance.

[0025] A priori penalty term that ensures the predicted value conforms to the basic laws of chemistry and physics.

[0026] In some embodiments, the step of using a directed message-passing neural network to perform representation learning on the graph feature data to obtain a first molecular representation result characterizing the overall electronic structure and topological features of the candidate molecule includes the following steps:

[0027] The graph feature data is represented by a directed message passing neural network to obtain the first molecular representation result. The representation learning includes assigning different weights to different atomic-level hidden representations using an attention mechanism, so that the central metal atom, N4 coordination environment, aromatic bond-related atoms, and atoms near the conjugated double bond receive higher weights, and then they are aggregated to form a molecular-level representation.

[0028] In some embodiments, the step of using the target prediction model to screen target molecules from among the candidate molecules as photosensitizers includes the following steps:

[0029] Using the target prediction model, carbole derivatives with peripheral pyridinium substitution, central coordination of a non-metallic phosphorus atom, and two hydroxyl groups were screened from the candidate molecules to be used as photosensitizers.

[0030] In some embodiments, the method further includes a step of performing feature contribution analysis on the target prediction model, specifically including:

[0031] For atomic-level node features, each feature dimension is masked, zeroed, or replaced one by one. The change in the output of the target prediction model before and after the perturbation is calculated, and the change is used as the importance value of the corresponding node feature dimension.

[0032] For each chemical bond-level edge feature, a perturbation is applied, and the change in the predicted value of the target prediction model is calculated to characterize the contribution of the corresponding edge feature to the predicted value.

[0033] Perturbation analysis of nodes and edges was repeatedly performed on multiple test samples, and the importance values ​​were statistically averaged and normalized to obtain the global node feature importance ranking and the global edge feature importance ranking.

[0034] Representative candidate molecules are selected, and the contribution values ​​of atoms and bonds are mapped back to two-dimensional molecular structure diagrams. Different colors or shades of color are used to indicate the degree of importance, so as to show the structural units that play a key role in the target performance.

[0035] Another aspect of this application provides a photosensitizer screening device based on deep learning, the device comprising:

[0036] Graph representation unit, used to represent each standardized candidate molecule as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule;

[0037] The feature extraction unit is used to extract node feature vectors for each atom, extract edge feature vectors for each chemical bond, and determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms.

[0038] A molecular representation unit is used to perform representation learning on the graph feature data using a directed message passing neural network to obtain a first molecular representation result that represents the overall electronic structure features and topological features of the candidate molecule.

[0039] A performance prediction unit is used to determine the predicted value of the performance index of each of the candidate molecules based on the first molecule representation result.

[0040] The model training unit is used to train multiple candidate prediction models by using the predicted values ​​of each of the performance indicators as training labels and the corresponding candidate molecules as training samples.

[0041] A model evaluation unit is used to determine the prediction performance results of each of the trained candidate prediction models for the predicted values ​​of each of the performance metrics.

[0042] A model selection unit is used to select a target prediction model from each of the candidate prediction models based on the prediction performance results.

[0043] A molecular screening unit is used to screen target molecules from the candidate molecules using the target prediction model to select target molecules as photosensitizers.

[0044] Another aspect of this application embodiment provides an electronic device, including a processor and a memory;

[0045] The memory is used to store programs;

[0046] The processor executes the program to implement any of the methods described above.

[0047] Another aspect of this application provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described in any of the above embodiments.

[0048] This application includes at least the following beneficial effects:

[0049] This application represents standardized candidate molecules as corresponding graph structures. Nodes in the graph structure represent atoms in the candidate molecule, and edges represent chemical bonds. Node feature vectors are extracted for each atom, and edge feature vectors are extracted for each chemical bond. Graph feature data is determined based on the node feature vectors, edge feature vectors, and the connections between atoms. A directed message-passing neural network is used to perform representation learning on the graph feature data, obtaining a first molecular representation of the overall electronic structure and topological features of the candidate molecules. Predicted values ​​of performance indicators for each candidate molecule are determined based on the first molecular representation. Multiple candidate prediction models are trained using the predicted values ​​of each performance indicator as training labels and the corresponding candidate molecules as training samples. The prediction performance of each trained candidate prediction model for each performance indicator is determined. A target prediction model is selected from the candidate prediction models based on the prediction performance results. The target prediction model is then used to screen target molecules as photosensitizers from the candidate molecules. This application improves screening efficiency by extracting molecular-level representations and determining predicted values ​​of performance indicators through prediction models, thereby rapidly screening target molecules as photosensitizers based on the predicted values. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a photosensitizer screening method based on deep learning, provided for an embodiment of this application;

[0052] Figure 2 This is a structural block diagram of a photosensitizer screening device based on deep learning, provided in an embodiment of this application. Detailed 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] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows:

[0055] Corrole compounds, as condensed porphyrin derivatives, possess advantages such as easy structural modification, tunable near-infrared absorption, good photostability, and excellent photodynamic activity, making them important candidate systems for optical probes and photosensitizers. However, the mechanisms by which their central metal, N4 coordination environment, conjugated macrocyclic skeleton, and peripheral substituents affect key properties are complex, making it difficult to efficiently obtain optimal molecules with both imaging and therapeutic functions using traditional experimental screening methods.

[0056] Based on this, this application aims to introduce deep learning methods to establish an intelligent screening system for predicting key physical properties of carboxane optical probes, thereby promoting the rational design and rapid discovery of high-performance tumor diagnostic and therapeutic probes. This research will help deepen the understanding of the "structure-performance-effect" relationship of carboxane photosensitizers and provide new candidate molecules and technical pathways for tumor imaging-guided phototherapy (PDT).

[0057] With the development of artificial intelligence, especially deep learning, molecular property prediction based on deep learning is changing the development model of photosensitizers. Graph neural networks can directly extract features from molecular graph structures, providing an effective tool for high-throughput prediction of key photophysical parameters. Existing studies have used them to predict excited-state energy levels, photosensitivity, and molecular structure-activity relationships, achieving good results. However, overall, existing research still has shortcomings: First, it focuses on single-performance prediction, lacking synergistic optimization of imaging and therapeutic functions; second, there are still relatively few deep learning studies on carboxyl photosensitizers, and their structure-activity relationships are still unclear; third, there is a lack of closed-loop research from model prediction to in vitro and in vivo experimental verification, which limits the efficiency of results translation.

[0058] Carbohydrates, as condensed porphyrin derivatives, possess advantages such as easy modification, tunable near-infrared absorption, and good photostability, showing great application potential in optical imaging and photodynamic therapy (PDT). However, the comprehensive influence of the central metal, coordination environment, and peripheral substituents on imaging and photodynamic performance lacks systematic research, making it difficult to rapidly screen high-performance probes using traditional methods. Therefore, developing deep learning-based intelligent screening strategies for carbohydrate molecules is of great significance.

[0059] Overall, related research is gradually shifting from experience-driven to data-driven and intelligent design, but research on the integrated intelligent screening, interpretable analysis, and experimental verification of carboxylic optical probes remains relatively weak. This application focuses on the deep learning-based intelligent screening of carboxylic optical probes and their application in the diagnosis and treatment of tumors such as nasopharyngeal carcinoma, achieving breakthroughs at both the methodological and application levels, and providing new ideas for the development of high-performance therapeutic molecular probes.

[0060] This application aims to conduct research on "the rational design and tumor diagnosis and treatment application of high-performance carbole optical probes empowered by deep learning". The main contents include: First, constructing a carbole molecular dataset and establishing a molecular graph prediction model based on a directed message passing neural network, focusing on two key indicators: the maximum absorption wavelength of the Q-band and the singlet-triplet intersystem crossing rate (kISC). Second, introducing strategies such as molecular fingerprinting, RDKit descriptors, physical constraints, and heterogeneous model fusion on the basis of DMPNN, systematically comparing different models, and selecting the optimal prediction framework. Third, conducting interpretability analysis at the node and edge levels to reveal the effects of structural units such as the central metal, N4 coordination environment, and conjugated macrocyclic framework on Q-band and kISC. Fourth, screening candidate carbole probes based on model predictions and interpretability results, and validating their spectral properties, two-photon fluorescence lifetime imaging performance, in vitro ROS generation ability, tumor cell killing effect, and in vivo anti-tumor effect.

[0061] The focus of this application is to achieve synergistic optimization of multiple indicators of optical probes, including "long-wavelength absorption, efficient intersystem crossing, excellent imaging, and significant therapeutic effects." The challenges lie in high-quality characterization of complex molecular structures, improving model generalization ability under small sample conditions, and effectively connecting model predictions with real biological effects. Regarding these issues, the basic viewpoint of this application is that the discovery of high-performance tumor diagnostic and therapeutic probes should shift from traditional trial-and-error based experience to a new paradigm of "data-driven + mechanism explanation + experimental closed-loop verification." This involves using deep learning to identify the core structural factors that determine key photophysical properties, and then using this identification to guide the precise design and screening of molecules.

[0062] The innovations of this application are mainly reflected in the following aspects: First, it uses Q_band and kISC as dual core indicators, taking into account the needs of tumor imaging and photodynamic therapy, breaking through the limitations of existing research that focuses on a single performance; Second, it constructs a smart screening framework for carbole photosensitizers with DMPNN as the core, integrating feature engineering and heterogeneous models, improving the accuracy of key property predictions. Among them, Q_band prediction performs excellently, and kISC prediction obtains a high coefficient of determination after model fusion; Third, it establishes interpretable analysis methods at the node and bond levels, analyzing the structure-property relationship of carbole molecules from the model level, providing an understandable basis for rational molecular design; Fourth, it completes a closed-loop study from model prediction, candidate molecule screening to cell and animal experimental verification, screening out carbole candidates and confirming their good potential for two-photon fluorescence lifetime imaging and in vivo and in vitro PDT therapy.

[0063] This application proposes a technical route of "data processing—model construction—interpretable analysis—candidate screening—optical characterization—cell experiments—animal experiments". First, molecular map data and key property datasets for carboxyprobes are established; second, DMPNN and its extended models are trained and optimized to obtain high-precision prediction results; third, the most promising molecules are screened based on the interpretation results; finally, the diagnostic and therapeutic performance of the screened probes is systematically validated through fluorescence spectroscopy, two-photon excitation, fluorescence lifetime imaging, DPBF and DCFH-DA detection, live / dead staining, tumor-bearing mouse imaging, and histopathological analysis.

[0064] Reference Figure 1 This application provides a photosensitizer screening method based on deep learning, specifically including the following steps S100~S170:

[0065] S100: Represent each standardized candidate molecule as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule;

[0066] S110: Extract node feature vectors for each atom, extract edge feature vectors for each chemical bond, and determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms;

[0067] S120: Use a directed message passing neural network to perform representation learning on the graph feature data to obtain the first molecular representation result that represents the overall electronic structure features and topological features of the candidate molecule;

[0068] S130: Determine the predicted values ​​of the performance indicators of each candidate molecule based on the first molecule representation results;

[0069] S140: Using the predicted values ​​of each performance index as training labels and the corresponding candidate molecules as training samples, train multiple candidate prediction models.

[0070] S150: Determine the prediction performance results of each of the candidate prediction models after training for the predicted values ​​of each of the performance metrics;

[0071] S160: Select the target prediction model from each of the candidate prediction models based on the prediction performance results;

[0072] S170: Using the target prediction model, a target molecule is selected from each of the candidate molecules as a photosensitizer.

[0073] Optionally, determining the predicted values ​​of the performance indicators of each candidate molecule based on the first molecule representation result includes the following steps:

[0074] Based on the first molecule representation result, the predicted values ​​corresponding to the absorption peak wavelength and the intersystem crossing rate between singlet and triplet states of each candidate molecule are determined.

[0075] Alternatively, explicit features are extracted from each of the candidate molecules as auxiliary features; the first molecule representation result and the auxiliary features are fused to obtain the second molecule representation result; and the predicted values ​​corresponding to the absorption peak wavelength and the singlet and triplet crossover rates of each candidate molecule are determined based on the second molecule representation result.

[0076] Optionally, the step of extracting explicit features from each of the candidate molecules as auxiliary features includes the following steps:

[0077] At least one of the RDKit molecular descriptor, MACCS molecular fingerprint, or Morgan molecular fingerprint is extracted from each of the candidate molecules as the auxiliary feature.

[0078] Optionally, training multiple candidate prediction models includes the following steps:

[0079] The first molecule representation result and the RDKit molecule descriptor are concatenated to obtain a fused representation;

[0080] The fusion representation is input into the prediction model for training, and a physical constraint loss term is introduced during the training process;

[0081] The physical constraint loss term includes at least one of the following:

[0082] The predicted value falls within the preset physical range constraint;

[0083] Monotonicity constraints apply to a known relationship between structure and performance.

[0084] A priori penalty term that ensures the predicted value conforms to the basic laws of chemistry and physics.

[0085] Optionally, the step of using a directed message-passing neural network to perform representation learning on the graph feature data to obtain a first molecular representation of the overall electronic structure and topological features of the candidate molecule includes the following steps:

[0086] The graph feature data is represented by a directed message passing neural network to obtain the first molecular representation result. The representation learning includes assigning different weights to different atomic-level hidden representations using an attention mechanism, so that the central metal atom, N4 coordination environment, aromatic bond-related atoms, and atoms near the conjugated double bond receive higher weights, and then they are aggregated to form a molecular-level representation.

[0087] Optionally, the step of using the target prediction model to screen target molecules from the candidate molecules as photosensitizers includes the following steps:

[0088] Using the target prediction model, carbole derivatives with peripheral pyridinium substitution, central coordination of a non-metallic phosphorus atom, and two hydroxyl groups were screened from the candidate molecules to be used as photosensitizers.

[0089] Optionally, the method further includes a step of performing feature contribution analysis on the target prediction model, specifically including:

[0090] For atomic-level node features, each feature dimension is masked, zeroed, or replaced one by one. The change in the output of the target prediction model before and after the perturbation is calculated, and the change is used as the importance value of the corresponding node feature dimension.

[0091] For each chemical bond-level edge feature, a perturbation is applied, and the change in the predicted value of the target prediction model is calculated to characterize the contribution of the corresponding edge feature to the predicted value.

[0092] Perturbation analysis of nodes and edges was repeatedly performed on multiple test samples, and the importance values ​​were statistically averaged and normalized to obtain the global node feature importance ranking and the global edge feature importance ranking.

[0093] Representative candidate molecules are selected, and the contribution values ​​of atoms and bonds are mapped back to two-dimensional molecular structure diagrams. Different colors or shades of color are used to indicate the degree of importance, so as to show the structural units that play a key role in the target performance.

[0094] The following section will provide a detailed introduction and explanation of the solutions in the embodiments of this application, using specific application examples.

[0095] In recent years, molecular property prediction methods based on graph neural networks have developed rapidly. Among them, Directed Message Passing Neural Networks (DMPNN) can encode molecular graphs into atomic and bond features, and achieve property prediction through directed message passing, graph readout, and feedforward neural networks, providing a new technical path for structure-property modeling of complex carbole systems. This application aims to apply emerging deep learning algorithms to the discovery and optical performance enhancement of novel carbole molecular photosensitizers. Carbole, as a novel porphyrin molecular photosensitizer, has great application potential in photodynamic / photothermal therapy, optical imaging, photodynamic detection, and sensing due to its easily modifiable structure, good metal atom coordination, wide absorption and tunability, high photoelectric conversion performance, and excellent photostability. However, its structural performance applications remain unclear and require further in-depth investigation.

[0096] This application addresses the industry-wide challenges of long development cycles, high costs, and high trial-and-error rates in photosensitizer research and development in the field of photodynamic therapy. It innovatively constructs an intelligent screening strategy based on deep learning, capable of effectively screening near-infrared carboxylic molecular probes with excellent optical properties. By encoding molecular graphs into atomic and bond features, and employing k-step directional message passing, graph output, and a feedforward neural network, 12 strategies, including feature engineering, heterogeneous model fusion, attention mechanisms, and physical constraints, are used to predict optical performance. This application predicts two important indicators of carboxylic molecular probes: the absorption of the optical probe and the singlet-triplet intersystem crossing capability. The light absorption capability is closely related to wavelength, and the selection of the absorption peak wavelength is crucial for the depth of light penetration into the skin; therefore, we strive to generate stronger absorption capabilities in longer wavelength bands. Secondly, the most direct parameter reflecting the photodynamic process of the molecular probe is the intersystem crossing rate, which is calculated from the singlet-triplet energy level difference and spin-orbit coupling coefficient using the Fermi golden rule, and is another important performance characteristic of the photosensitizer. The DMPNN algorithm, which integrates molecular fingerprints, demonstrates superior accuracy in predicting optical performance among numerous algorithms. This application constructs a method for interpretable analysis of node and bond features, exploring key contributing features to the optical performance of the probe. Based on this, this application identifies a gallium carbide molecule with a long emission wavelength and excellent photodynamic properties, and demonstrates through in vitro and in vivo experiments that the predicted carbide molecule exhibits dual-mode imaging capabilities of fluorescence intensity and fluorescence lifetime.

[0097] This application provides a deep learning-based photosensitizer screening method, particularly a high-throughput intelligent screening method for carboxyl-type candidate molecules. This method takes the molecular structure information of candidate molecules as input, and through graph structure encoding, directed message-passing neural network representation, auxiliary feature fusion, regression prediction, multi-model comparison, and screening determination, outputs a set of target molecules that meet preset photosensitizer performance requirements. The method can be used to predict and screen photosensitizer molecules with longer absorption peak wavelengths and higher intersystem crossing rates, thereby improving the efficiency of candidate photosensitizer discovery and reducing the time and experimental costs of traditional "synthesis-testing" screening methods. The method includes the following steps:

[0098] Step S1: Obtain the candidate molecule set and its structural expression information.

[0099] Obtain a set of candidate molecules of the carboxylic acid class to be screened; the candidate molecules can be derived from reported molecular libraries, virtual design molecular libraries, molecular libraries generated by substituent combinations, or existing experimental sample molecular libraries.

[0100] For each candidate molecule, its structural expression information is obtained, which includes at least one of the following: a two-dimensional molecular structure; a SMILES string; or a structural file that can be converted into atomic connection relationships.

[0101] The structural representation information is uniformly transformed into a standard molecule that can be processed by machines, which is then used for subsequent graph structure construction and feature extraction.

[0102] The output of this step is the standardized set of candidate molecules.

[0103] The purpose of this step is to provide a unified data foundation for subsequent graph structure encoding and model input.

[0104] Step S2: Represent the candidate molecules as a graph structure.

[0105] Specifically, each candidate molecule is represented as a graph G=(V,E), where the set of nodes V represents the atoms in the molecule, and the set of edges E represents the chemical bonds in the molecule. Node feature vectors are extracted for each atom, and edge feature vectors are extracted for each chemical bond. These features are then combined with the connections between atoms to form the model input graph data.

[0106] Each atomic node is encoded as a 98-dimensional feature vector, which includes: atomic number, atomic degree, formal charge, number of connected hydrogen atoms, hybridization type, number of lone pairs, hydrogen bond acceptor characteristics, minimum ring size, electronegativity, aromaticity, and atomic mass. The atomic number, atomic degree, formal charge, number of connected hydrogen atoms, hybridization type, number of lone pairs, hydrogen bond acceptor characteristics, and minimum ring size are represented using one-hot encoding, while electronegativity, aromaticity, and atomic mass are represented using numerical scalars.

[0107] Each chemical bond edge is encoded as a 12-dimensional feature vector, which includes: single bond markers, double bond markers, aromatic bond markers, conjugated bond markers, intracyclic bond markers, and bond stereochemical configuration information, wherein the bond stereochemical configuration is represented by one-hot encoding.

[0108] Therefore, each candidate molecule is converted into graph feature data containing node feature matrices, edge feature matrices, and graph connectivity relationships, which is then used as input to the Directed Message Passing Neural Network (DMPNN).

[0109] Step S3: Use the DMPNN model to perform representation learning on the graph feature data to obtain the first molecule representation result.

[0110] The graph feature data obtained in step S2 is input into the Directed Message Passing Neural Network (DMPNN).

[0111] DMPNN performs message passing on directed edges, and its processing includes:

[0112] S31: Construction of the initial state of directed edges.

[0113] For each directed edge, generate the initial hidden state of the directed edge based on the initial atomic features and the edge features of that edge.

[0114] The initial hidden state represents the local chemical environment information corresponding to the directed edge.

[0115] S32: Multi-step directed message passing.

[0116] For each round of message passing, for the current directed edge, collect the hidden state information of all adjacent directed edges that point to its starting atom and do not directly coincide with its direction in the opposite direction, and aggregate them; combine the aggregation result with the existing state of the current directed edge, and update the hidden state of the current directed edge after nonlinear mapping.

[0117] Repeat the above process several times to achieve the propagation of local chemical environment information layer by layer along the molecular graph.

[0118] Through this process, the model can learn: the central metal / nonmetal coordination environment; the electronic structure characteristics of the N4 macrocyclic framework; the distribution of conjugated double bonds and aromatic bonds; and the modulation effect of peripheral substituents on the local electronic environment.

[0119] S33: Atomic layer polymerization.

[0120] After completing multiple rounds of directed edge message passing, the hidden states of the directed edges connected to each atom are aggregated to the corresponding atom to obtain the atomic-level hidden representation.

[0121] S34: The graph is read out, and a molecular-level representation is obtained.

[0122] All atomic-level hidden representations are read out graphically by summing, averaging, or weighted aggregation to obtain the molecular-level vector representation of the candidate molecule, denoted as the first molecular representation result H1.

[0123] The output of this step: The first molecule represents the result H1.

[0124] The purpose of this step is to automatically extract deep features reflecting the structure-property relationship from the molecular diagram, providing core characterization for subsequent performance prediction.

[0125] Note: The data processed by DMPNN consists of node features, edge features, and directed edge connections generated in step S2; the result after DMPNN processing is a molecular-level hidden representation H1, which characterizes the overall electronic structure and topological features of candidate molecules and can be directly used for subsequent performance prediction or further fused with auxiliary features.

[0126] Step S4: Extract explicit molecular descriptors and / or molecular fingerprints to obtain auxiliary features.

[0127] To enhance the model's ability to identify local substituent differences, overall physicochemical properties, and structural fragment patterns, auxiliary features are further extracted from candidate molecules. These auxiliary features include one or more of the following:

[0128] S41: RDKit molecular descriptor.

[0129] Descriptors characterizing the overall physicochemical and topological properties are extracted, including one or more of the following: molecular weight; polar surface area; topological complexity; connectivity index; aromatic ring count; hydrogen bond donor / acceptor count; charge distribution related indicators; and molecular surface area related indicators.

[0130] We obtain the RDKit descriptor vector, denoted as F. R .

[0131] S42: MACCS molecular fingerprint.

[0132] Extracting the MACCS fingerprint to represent the presence of functional groups and local structural fragments yields a binary feature vector, denoted as F. M .

[0133] S43: Morgan molecular fingerprint.

[0134] Morgan fingerprints are extracted based on the local neighborhood topology of candidate molecules; preferably, 256-dimensional, 512-dimensional, or 1024-dimensional Morgan fingerprints are used to describe the distribution of structural fragments in neighborhoods of different radii.

[0135] The Morgan fingerprint vector is obtained, denoted as F. G .

[0136] The output of this step: at least one auxiliary feature, including F R F M F G One or more of them.

[0137] The purpose of this step is to supplement information on physicochemical properties, functional group patterns, and local topological fragments that are not easily highlighted in molecular diagram characterization.

[0138] Step S5: Fuse the first molecule representation result with the auxiliary features to obtain the second molecule representation result.

[0139] The first molecular representation result H1 obtained in step S3 is fused with the auxiliary feature obtained in step S4 to obtain the second molecular representation result H2.

[0140] Preferably, the fusion method includes any of the following:

[0141] Method 1: Splicing and merging.

[0142] Connect H1 and F R F M or F G One or more of the features are directly concatenated along the feature dimension to form a fusion vector:

[0143] H2 = [H1;F];

[0144] Where F represents the selected auxiliary feature vector.

[0145] After splicing, dimensional mapping and nonlinear transformation can be performed through fully connected layers to obtain a fused representation at a unified scale.

[0146] Method 2: Weighted fusion.

[0147] After performing linear transformations on H1 and auxiliary features respectively, they are weighted and combined according to preset weights or learned weights to form a fusion representation.

[0148] Method 3: Attention fusion.

[0149] First, obtain the graph representation and auxiliary feature representation separately, and then assign different weights to different feature channels through an attention allocation mechanism to generate a fused representation.

[0150] The output of this step: The second molecule represents the result H2.

[0151] The purpose of this step is to combine the deep structural representation extracted by the graph neural network with the auxiliary features of the molecular descriptor / molecular fingerprint, thereby taking into account both global structural information and local fragment information and improving prediction accuracy.

[0152] Note: The data fused by the fusion strategy includes the first molecular representation result H1 output by DMPNN and the explicit molecular descriptor / fingerprint vector; the result obtained after fusion is the second molecular representation result H2, which serves as the input for the final regression prediction and can improve the model's ability to represent key performance characteristics.

[0153] Step S6: Construct a performance prediction model and output the target performance prediction value.

[0154] Input the molecular representation obtained in step S3 or step S5 into the performance prediction module, and output the target performance prediction value of the candidate molecule.

[0155] The target performance includes at least:

[0156] Predicted Q-band absorption peak wavelength;

[0157] Predicted value of the singlet-trittite intersystem jump rate kISC.

[0158] Preferably, the performance prediction module includes a feedforward neural network regression head, which includes at least an input layer, a hidden layer, and an output layer.

[0159] When using single-task mode, separate models are built for predicting Q_band and kISC.

[0160] in:

[0161] For models represented using only DMPNN, the input is H1;

[0162] For models using fusion representation, the input is H2.

[0163] The model outputs the following results:

[0164] Q_band prediction result Y Q ;k ISC Prediction result Y K .

[0165] The output of this step: performance prediction results for candidate molecules.

[0166] The purpose of this step is to map the structural characterization into quantifiable key performance indicators of the photosensitizer, providing a basis for subsequent screening and judgment.

[0167] Step S7: Model training and validation.

[0168] To obtain a predictive model that can be used for practical screening, supervised learning training is performed on sample molecules with known performance data.

[0169] Preferably, the sample dataset is divided into a training set, a validation set, and a test set according to a preset ratio; more preferably, it is divided in a ratio of 8:1:1.

[0170] During training: Training set samples are input into the model; model predictions are compared with corresponding ground truth values, and the loss function is calculated; model parameters are updated using gradient backpropagation; model hyperparameters are adjusted based on validation set results; and the model's generalization performance is evaluated on the test set. Preferably, the coefficient of determination R² is used to evaluate the model's prediction performance; if necessary, mean squared error, mean absolute error, and other metrics can also be used for comprehensive evaluation.

[0171] For different model strategies, training and validation are performed separately, including:

[0172] DMPNN;

[0173] DMPNN+RDKit;

[0174] DMPNN+RDKit_PINN;

[0175] DMPNN+MACCSFP;

[0176] DMPNN+MorganFP (256 / 512 / 1024);

[0177] DMPNN+XGB;

[0178] DMPNN+SVM;

[0179] DMPNN+GBDT;

[0180] DMPNN+RF;

[0181] DMPNN+Attention.

[0182] Based on the performance results of the validation and test sets, the optimal model for subsequent screening is determined.

[0183] Preferably, for the Q-band absorption peak wavelength prediction task, DMPNN is selected as the optimal model;

[0184] For the kISC prediction task, the DMPNN and 512-dimensional Morgan fingerprint fusion model is selected as the optimal model.

[0185] The output of this step is: the trained performance prediction model and the result of selecting the optimal model.

[0186] The purpose of this step is to ensure that the models used in subsequent screenings have high predictive accuracy and reliability.

[0187] How model augmentation strategies are applied in the overall process:

[0188] (a) Application steps of Physical Constraint Fusion (PINN) in the overall scheme.

[0189] Physical constraint fusion is applied between steps S5 and S7.

[0190] Specifically, after obtaining the graph representation H1 and the RDKit descriptor FR, the two are first concatenated to obtain a fused representation, and then a physical constraint loss term is introduced during the model training process.

[0191] The physical constraint loss term must include at least one of the following:

[0192] Range constraints ensuring that predicted values ​​fall within a reasonable physical range;

[0193] Monotonicity constraints that constrain certain known structure-performance relationships;

[0194] A priori penalty term that ensures the prediction results conform to basic chemical and physical laws.

[0195] The output of the model after adding physical constraints is still the target performance prediction value, but its prediction results have better physical rationality and small sample generalization ability.

[0196] Its position in the overall plan:

[0197] S2 Graph Structure Construction → S3 DMPNN Representation → S4 RDKit Descriptor Extraction → S5 Fusion → S7 Introducing PINN Constraints During Training → Output Predicted Values.

[0198] (ii) Steps for applying attention mechanisms in the overall scheme.

[0199] The attention mechanism is preferably applied to the graph readout stage in step S3, or to the fusion stage in step S5.

[0200] When applied to the graph readout stage:

[0201] Different weights are assigned to different atomic-level hidden representations, giving higher weights to the central metal atom, N4 coordination environment, aromatic bond-related atoms, and atoms near conjugated double bonds, which are then aggregated to form a molecular-level representation.

[0202] When applied in the fusion phase:

[0203] Different feature channel weights are assigned to the graph representation and the auxiliary feature representation to highlight the input information that is more conducive to the prediction of the target task.

[0204] The result after attention mechanism processing is still a molecular-level representation vector. Its role is to enhance the model's ability to identify key structural units and improve its sensitivity to important atoms / bonds.

[0205] Its position in the overall plan:

[0206] S2 graph structure construction → S3 graph reading with Attention introduced or S5 fusion with Attention introduced → S6 regression prediction.

[0207] (III) Application steps of two-stage heterogeneous model integration in the overall scheme.

[0208] The two-stage heterogeneous model fusion is applied to step S6.

[0209] Specifically, the DMPNN first completes graph structure learning and outputs the first molecular representation result H1 or the second molecular representation result H2. Then, this representation is used as input features and fed into traditional machine learning models, including XGBoost, support vector machine, gradient boosting decision tree or random forest, to complete the final regression fitting.

[0210] The processing flow is as follows:

[0211] Using DMPNN to extract deep features from graph structures;

[0212] Extracting molecular-level vector representations;

[0213] This vector represents the input to XGB / SVM / GBDT / RF;

[0214] The traditional regressor outputs Q_band or k ISC Predicted value.

[0215] Its position in the overall plan:

[0216] S2 Graph Structure Construction → S3 DMPNN Representation → S5 Optional Fusion → S6 Inputting Representation Results into a Two-Stage Regressor → Outputting Predicted Values.

[0217] The purpose of this strategy is to further explore the nonlinear relationships in the output features of DMPNN, forming a heterogeneous prediction framework that combines graph neural networks with traditional machine learning.

[0218] The target molecules obtained from the screening and the screening criteria:

[0219] Based on the model prediction results, comprehensive ranking results, and structural interpretability analysis results, the target molecules selected by the screening are: carbole derivatives with pyridyl substitution on the periphery, non-metallic phosphorus atoms coordinated at the center, and two hydroxyl groups attached.

[0220] The criteria for selection include:

[0221] The prediction results show that the molecule has a high predicted Q-band absorption peak wavelength, indicating that it has good red / near-infrared absorption potential.

[0222] The prediction results show that the molecule has a high kISC prediction value, indicating that it has a strong intersystem crossing ability and potential reactive oxygen species generation ability.

[0223] Interpretability analysis shows that the central coordination environment, N4 conjugated macrocyclic skeleton, and combination of peripheral substituents of this molecule conform to the key favorable structural patterns identified by the model.

[0224] Structures containing hydroxyl and pyridine groups are advantageous for improving molecular polarity, hydrophilicity, and biocompatibility. Therefore, these molecules can be considered as preferred target photosensitizer molecules for subsequent experimental verification.

[0225] Step S8: Perform interpretability analysis on the optimal model.

[0226] To obtain selection criteria and improve the interpretability of model output, feature contribution analysis is performed on the optimal model after model training is completed.

[0227] Specifically, it includes:

[0228] S81: Node feature perturbation analysis.

[0229] For atomic-level node features, one-by-one masking, zeroing, or replacement of a feature dimension is performed, and the change in model output before and after the perturbation is calculated. The change is used as the importance value of the node feature dimension.

[0230] S82: Edge feature perturbation analysis.

[0231] For each chemical bond-level edge feature, a perturbation is applied, and the change in the model's predicted value is calculated to characterize the contribution of that edge feature to the prediction result.

[0232] S83: Global importance statistics.

[0233] Perturbation analysis of nodes and edges was repeatedly performed on multiple samples in the test set, and the obtained importance values ​​were statistically averaged and normalized to obtain the global node feature importance ranking and global edge feature importance ranking.

[0234] S84: Local structure mapping.

[0235] Representative candidate molecules are selected, and the contribution values ​​of atoms and bonds are mapped back to two-dimensional molecular structure diagrams. Different colors or shades of color are used to indicate the degree of importance, so as to show the structural units that play a key role in the target performance.

[0236] Based on the above analysis, we can identify the pairings between Q_band and k. ISC The structural features that play a major role.

[0237] Preferably, the key structures identified by the model include:

[0238] Central metallic / nonmetallic coordination center;

[0239] N4 coordination environment;

[0240] Conjugated macrocyclic framework;

[0241] Aromatic bonds and conjugated double bonds;

[0242] Lone pair electron-related structures;

[0243] Peripheral polyfluoroaryl groups or other substituents that affect electron distribution.

[0244] The outputs of this step are: global feature importance results and local molecular structure contribution maps.

[0245] The purpose of this step is to provide structural technical evidence for "why this molecule was screened out", and it can also be used to guide the design of new candidate molecules.

[0246] In summary, the screening method of this application can be summarized as follows:

[0247] 1. Obtain candidate carboxylic acid molecules and normalize the structural input;

[0248] 2. Represent candidate molecules as graph structures containing atomic nodes and chemical bond edges;

[0249] 3. Use DMPNN to perform multi-step directed message passing on the graph structure to obtain the first molecule representation result;

[0250] 4. Extract the RDKit descriptor, MACCS fingerprint, and / or Morgan fingerprint to obtain auxiliary features;

[0251] 5. The first molecule representation result is fused with the auxiliary features to obtain the second molecule representation result;

[0252] 6. Output the predicted values ​​of Q_band and kISC through the regression prediction module;

[0253] 7. Select the optimal model based on the validation results;

[0254] 8. Screen candidate molecules based on dual-indicator thresholds, comprehensive scores, or constraints;

[0255] 9. Output target photosensitizer molecules;

[0256] 10. Perform interpretability analysis on the optimal model to obtain the selection criteria.

[0257] This method realizes a complete technical closed loop from molecular structure encoding and model prediction to result screening and output, and can be used to quickly screen photosensitizer candidate molecules with excellent light absorption and photodynamic properties.

[0258] Reference Figure 2 This application provides a photosensitizer screening device based on deep learning, comprising:

[0259] Graph representation unit, used to represent each standardized candidate molecule as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule;

[0260] The feature extraction unit is used to extract node feature vectors for each atom, extract edge feature vectors for each chemical bond, and determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms.

[0261] A molecular representation unit is used to perform representation learning on the graph feature data using a directed message passing neural network to obtain a first molecular representation result that represents the overall electronic structure features and topological features of the candidate molecule.

[0262] A performance prediction unit is used to determine the predicted value of the performance index of each of the candidate molecules based on the first molecule representation result.

[0263] The model training unit is used to train multiple candidate prediction models by using the predicted values ​​of each of the performance indicators as training labels and the corresponding candidate molecules as training samples.

[0264] A model evaluation unit is used to determine the prediction performance results of each of the trained candidate prediction models for the predicted values ​​of each of the performance metrics.

[0265] A model selection unit is used to select a target prediction model from each of the candidate prediction models based on the prediction performance results.

[0266] A molecular screening unit is used to screen target molecules from the candidate molecules using the target prediction model to select target molecules as photosensitizers.

[0267] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0268] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0269] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0270] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0271] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0272] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0273] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0274] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0275] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0276] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A deep learning-based photosensitizing agent screening method, characterized by, The method includes the following steps: Each standardized candidate molecule is represented as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule. Extract node feature vectors for each atom and edge feature vectors for each chemical bond. Determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms. A directed message-passing neural network is used to perform representation learning on the graph feature data to obtain a first molecular representation of the overall electronic structure and topological features of the candidate molecule. The predicted values ​​of the performance indicators of each candidate molecule are determined based on the first molecule representation results; Using the predicted values ​​of each performance index as training labels and the corresponding candidate molecules as training samples, multiple candidate prediction models are trained. Determine the prediction performance results of each of the trained candidate prediction models for the predicted values ​​of each of the performance metrics; The target prediction model is selected from each of the candidate prediction models based on the prediction performance results; The target molecule is selected as a photosensitizer from the candidate molecules using the target prediction model.

2. The deep learning-based photosensitizing agent screening method of claim 1, wherein, The step of determining the predicted performance index of each candidate molecule based on the first molecule representation result includes the following steps: Based on the first molecule representation result, the predicted values ​​corresponding to the absorption peak wavelength and the intersystem crossing rate between singlet and triplet states of each candidate molecule are determined. Alternatively, explicit features are extracted from each of the candidate molecules as auxiliary features; the first molecule representation result and the auxiliary features are fused to obtain the second molecule representation result; and the predicted values ​​corresponding to the absorption peak wavelength and the singlet and triplet crossover rates of each candidate molecule are determined based on the second molecule representation result. 3.The deep learning-based photosensitizing agent screening method of claim 2, wherein, The step of extracting explicit features from each of the candidate molecules as auxiliary features includes the following steps: At least one of the RDKit molecular descriptor, MACCS molecular fingerprint, or Morgan molecular fingerprint is extracted from each of the candidate molecules as the auxiliary feature. 4.The method of claim 2, wherein, The training of multiple candidate prediction models includes the following steps: The first molecule representation result and the RDKit molecule descriptor are concatenated to obtain a fused representation; The fusion representation is input into the prediction model for training, and a physical constraint loss term is introduced during the training process; The physical constraint loss term includes at least one of the following: The predicted value falls within the preset physical range constraint; Monotonicity constraints apply to a known relationship between structure and performance. A priori penalty term that ensures the predicted value conforms to the basic laws of chemistry and physics.

5. The deep learning-based photosensitizing agent screening method of claim 1, wherein, The method of using a directed message-passing neural network to perform representation learning on the graph feature data to obtain a first molecular representation of the overall electronic structure and topological features of the candidate molecule includes the following steps: The graph feature data is represented by a directed message passing neural network to obtain the first molecular representation result. The representation learning includes assigning different weights to different atomic-level hidden representations using an attention mechanism, so that the central metal atom, N4 coordination environment, aromatic bond-related atoms, and atoms near the conjugated double bond receive higher weights, and then they are aggregated to form a molecular-level representation.

6. The deep learning-based photosensitizing agent screening method of claim 1, wherein, The process of using the target prediction model to screen target molecules from each candidate molecule as photosensitizers includes the following steps: Using the target prediction model, carbole derivatives with peripheral pyridinium substitution, central coordination of a non-metallic phosphorus atom, and two hydroxyl groups were screened from the candidate molecules to be used as photosensitizers.

7. The deep learning-based photosensitizing agent screening method of claim 1, wherein, The method further includes a step of performing feature contribution analysis on the target prediction model, specifically including: For atomic-level node features, each feature dimension is masked, zeroed, or replaced one by one. The change in the output of the target prediction model before and after the perturbation is calculated, and the change is used as the importance value of the corresponding node feature dimension. For each chemical bond-level edge feature, a perturbation is applied, and the change in the predicted value of the target prediction model is calculated to characterize the contribution of the corresponding edge feature to the predicted value. Perturbation analysis of nodes and edges was repeatedly performed on multiple test samples, and the importance values ​​were statistically averaged and normalized to obtain the global node feature importance ranking and the global edge feature importance ranking. Representative candidate molecules are selected, and the contribution values ​​of atoms and bonds are mapped back to two-dimensional molecular structure diagrams. Different colors or shades of color are used to indicate the degree of importance, so as to show the structural units that play a key role in the target performance.

8. A deep learning-based photosensitizing agent screening device, characterized by, The device includes: Graph representation unit, used to represent each standardized candidate molecule as a corresponding graph structure; wherein, the nodes of the graph structure represent atoms in the candidate molecule, and the edges of the graph structure represent chemical bonds in the candidate molecule; The feature extraction unit is used to extract node feature vectors for each atom, extract edge feature vectors for each chemical bond, and determine graph feature data based on the node feature vectors, edge feature vectors, and the connection relationships between atoms. A molecular representation unit is used to perform representation learning on the graph feature data using a directed message passing neural network to obtain a first molecular representation result that represents the overall electronic structure features and topological features of the candidate molecule. A performance prediction unit is used to determine the predicted value of the performance index of each of the candidate molecules based on the first molecule representation result. The model training unit is used to train multiple candidate prediction models by using the predicted values ​​of each of the performance indicators as training labels and the corresponding candidate molecules as training samples. A model evaluation unit is used to determine the prediction performance results of each of the trained candidate prediction models for the predicted values ​​of each of the performance metrics. A model selection unit is used to select a target prediction model from each of the candidate prediction models based on the prediction performance results. A molecular screening unit is used to screen target molecules from the candidate molecules using the target prediction model to select target molecules as photosensitizers.

9. An electronic device, comprising: The electronic device includes a processor and a memory; The memory is configured to store a program. The processor is configured to execute the program to implement the method in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program, and the program is executed by a processor to implement the method in any one of claims 1 to 7.

Images