A method and system for predicting the molecular characterization of a compound screening for cholesterol efflux activity

Abstract

The application provides a molecular characterization prediction method and system for screening compounds with cholesterol efflux activity, which comprises the following steps: S101, obtaining a molecular structure representation of a compound to be predicted, and constructing a molecular fingerprint and a molecular graph; S102, constructing a cholesterol efflux activity prediction model; S103, inputting the molecular fingerprint into a fingerprint feature extraction branch to obtain a molecular fingerprint representation; S104, inputting the molecular graph into a graph structure feature extraction branch to obtain a molecular graph representation; S105, inputting the molecular fingerprint representation and the molecular graph representation into a feature fusion layer to obtain a fused molecular representation; S106, inputting the fused molecular representation into a classification prediction layer to output a prediction probability of the compound to be predicted belonging to a cholesterol efflux activity category; and S107, pre-setting a classification threshold to convert the prediction probability into a binary classification prediction result of whether the compound has activity. The application realizes feature complementation and information synergy, significantly improves prediction accuracy and model generalization ability, and improves the recognition effect of a small number of active compounds.

Description

Technical Field

[0001] This invention relates to the field of compound activity prediction technology, and in particular to a molecular characterization prediction method and system for screening cholesterol efflux-active compounds. Background Technology

[0002] Cholesterol efflux is the process by which cells transport excess cholesterol to extracellular receptors. It is a crucial initiation step in reverse cholesterol transport and is closely related to lipid metabolism-related diseases such as atherosclerosis. Small molecule compounds that can regulate cholesterol efflux activity can not only serve as tool molecules for studying cholesterol transport mechanisms and lipid metabolism regulation mechanisms, but also have potential value in screening lead compounds.

[0003] Traditional cholesterol efflux activity screening mainly relies on in vitro experimental detection. While the experimental results are relatively reliable, they are often hampered by long cycles, high costs, and limited screening scale. With the development of artificial intelligence and molecular machine learning, computational models can be used to predict the cholesterol efflux activity of small molecules before experiments, helping researchers narrow down the range of candidate compounds in advance and improving the efficiency of subsequent screening experiments.

[0004] However, the task of predicting cholesterol efflux activity still has the following problems.

[0005] First, the publicly available and readily available labeled data on cholesterol efflux activity is limited in scale, and the number of training samples is small. Complex deep models are easily limited by the data scale and it is difficult to stably learn the structure-activity relationship.

[0006] Second, the proportion of samples in the live dataset is not completely balanced. Ordinary classification models may be biased towards the majority class, thus affecting the recognition effect of minority class or boundary samples.

[0007] Third, the scaffold distribution of cholesterol efflux-related small molecules is relatively sparse, and the number of samples in a large number of scaffold groups is limited. The model is prone to instability in its predictions when faced with new chemical frameworks or external candidate molecules. In other words, a model's good performance near the training set does not necessarily mean that it will be reliable when faced with new structures.

[0008] Fourth, existing models typically only output a single predicted probability, lacking structural support and explanation for the predicted results. For screening real candidate molecules, it is necessary not only to determine whether they are active, but also to understand the reasons for that determination.

[0009] Existing methods for predicting molecular activity typically use molecular fingerprints, molecular graph neural networks, or a simple combination of both as input. Molecular fingerprints offer advantages such as high computational efficiency and structural prior stability, but their ability to represent complex topological relationships is limited. Molecular graph neural networks can learn trainable molecular topological representations, but they may suffer from training instability and extrapolation risks in scenarios with small samples and insufficient scaffold coverage. While simply combining the two types of representations allows the use of multiple pieces of information simultaneously, it does not clearly distinguish the functional division between fingerprint and graph representations and cannot adequately address the stability issues in small sample scenarios. Summary of the Invention

[0010] Therefore, the purpose of this invention is to provide a molecular characterization and prediction method and system for screening cholesterol efflux-active compounds, so as to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0011] To achieve the above objectives, the first aspect of the present invention provides a molecular characterization and prediction method for screening cholesterol efflux-active compounds, the method comprising the following steps: S101. Obtain the molecular structure representation of the compound to be predicted, and construct the molecular fingerprint and molecular map based on the molecular structure representation; S102. Construct a cholesterol efflux activity prediction model, including a fingerprint feature extraction branch, a graph structure feature extraction branch, a feature fusion layer, and a classification prediction layer; S103. The molecular fingerprint is fed into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; S104. The molecular graph is fed into the graph structure feature extraction branch to obtain the molecular graph representation; S105. The molecular fingerprint representation and molecular map representation are fed into the feature fusion layer to obtain the fused molecular representation. S106. Input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. S107. Pre-set a classification threshold, and convert the predicted probability into a binary classification prediction result of whether the activity exists based on the classification threshold.

[0012] Furthermore, the molecular fingerprint is represented as follows:

[0013] in, Molecular fingerprint, This refers to the input molecule in molecular structure representation. Indicates the fingerprint radius. Indicates fingerprint dimension; The structure of the molecular diagram includes a set of atomic nodes and a set of chemical bond edges. Each atomic node has initial node characteristics, including atomic type, atomicity, formal charge, hybridization mode, aromaticity, and whether it is located in a ring. Each chemical bond edge has chemical bond characteristics, including bond type, whether it is conjugated, and whether it belongs to a ring structure.

[0014] Furthermore, the fingerprint feature extraction branch uses a multilayer perceptron network to perform a nonlinear mapping on the molecular fingerprint input to obtain a molecular fingerprint representation, as shown below:

[0015] in, and The first Layer weights and biases It is a non-linear activation function. Indicates the number of fingerprint branch layers. For the final molecular fingerprint characterization ， Characterization of the original molecular fingerprint, For the first Hidden molecular fingerprint characterization obtained after layer nonlinear mapping For the process The final molecular fingerprint characterization obtained after layer mapping For the first Characterization of hidden molecular fingerprints from layer output.

[0016] Furthermore, the graph structure feature extraction branch uses a graph neural network to perform message passing, node updating, and graph-level readout on the molecular graph input to obtain a molecular graph representation, specifically including the following steps: S11. Feed the molecular graph into the graph structure feature extraction branch, and use a graph neural network to learn the structure of the molecular graph through a message passing mechanism; S12, Based on the structure of the learned molecular map, in the... In a layered graph neural network, atomic nodes aggregate messages from their neighbor sets and update those atomic nodes, as shown below:

[0017]

[0018] in, For the first The layer consists of atomic nodes The message vector is obtained by aggregating the features of neighboring nodes and edges. For the first Atomic nodes in layers The current hidden representation, For the first Neighbor nodes in the layer The hidden representation, atomic nodes with neighboring nodes Edge features between them For the set of neighboring nodes, For message functions, For the first Atomic nodes in layers The updated hidden representation, This represents the node update function; S13. After multi-layer message passing based on the message passing mechanism, a readout operation is performed on all atomic nodes that have undergone node updates to obtain the molecular graph representation, as shown below:

[0019] in, For molecular diagram characterization, For a set of atomic nodes, For the process After message passing in a layered graph neural network, atomic nodes The final node representation; The first in the molecular diagram Atom nodes, The number of layers in a graph neural network. This is the pooling function.

[0020] Furthermore, step S105 specifically includes the following steps: S41. Molecular fingerprint representation and molecular graph representation are projected onto a latent representation space of the same dimension through a feature fusion layer. S42. The projected molecular fingerprint characterization and molecular map characterization are spliced ​​together and input into the residual correction network to generate residual correction terms. S43. Add the residual correction term to the projected molecular fingerprint characterization to obtain the fused molecular characterization, as follows:

[0021] in, This represents the molecular fingerprint characterization after projection. This represents the molecular diagram characterization after projection. This represents vector concatenation. Represents the residual correction network. This indicates the characterization of fused molecules.

[0022] Furthermore, the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category is expressed as follows:

[0023] in, This represents the sigmoid function. This represents a classification prediction network. This represents the predicted probability.

[0024] Furthermore, the cholesterol efflux activity prediction model is trained using repeated core framework cross-validation, and a binary classification cross-entropy loss is used as the loss function during the training phase, as shown below:

[0025] in, For binary classification, cross-entropy loss, The model represents the first The predicted probability of a sample. Indicates the true label, This represents the total number of samples.

[0026] Furthermore, the classification threshold is determined by the out-of-fold prediction probability and the corresponding true label obtained during the training phase of the cholesterol efflux activity prediction model through repeated core skeleton cross-validation.

[0027] A second aspect of the present invention provides a molecular characterization and prediction system for screening cholesterol efflux-active compounds, characterized in that the system comprises: Molecular Input Construction Module: Used to construct molecular fingerprints and molecular maps based on the molecular structure representation of compounds; Cholesterol efflux activity prediction module: used to build a cholesterol efflux activity prediction model, including fingerprint feature extraction branch, graph structure feature extraction branch, feature fusion layer and classification prediction layer; Fingerprint feature extraction module: used to feed the molecular fingerprint into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; Graph structure feature extraction module: used to feed the molecular graph into the graph structure feature extraction branch to obtain the molecular graph representation; The residual fusion module is used to feed the molecular fingerprint representation and molecular map representation into the feature fusion layer to obtain the fused molecular representation. The classification prediction module is used to input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. Threshold decision module: used to pre-set classification thresholds and convert the predicted probabilities into binary classification prediction results of whether the entity is active based on the classification thresholds.

[0028] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a molecular characterization prediction method and system for screening cholesterol efflux active compounds. By feeding molecular fingerprints into the fingerprint feature extraction branch, stable local structural priors are provided, offering a relatively reliable basis for model judgment. By feeding molecular graphs into the graph structure feature extraction branch, structural connectivity relationships that are difficult to express by fingerprints are further supplemented. By feeding molecular fingerprint and molecular graph representations into the feature fusion layer, the instability caused by the model's complete reliance on graph neural networks is reduced under small sample conditions, while retaining the expressive advantages of graph structure information. By converting the predicted probability into a binary prediction result of whether or not the compound is active based on a classification threshold, the stability and interpretability of cholesterol efflux activity prediction are improved. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic flowchart of a molecular characterization and prediction method for screening cholesterol efflux-active compounds, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall structure of a molecular characterization and prediction system for screening cholesterol efflux-active compounds, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the cholesterol efflux activity prediction model provided in an embodiment of the present invention; Figure 4 The performance of the cholesterol efflux activity prediction model provided in the embodiments of the present invention is compared with that of ablation. Detailed Implementation

[0031] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0032] Reference Figure 1 This invention provides a molecular characterization and prediction method for screening cholesterol efflux-active compounds, the method comprising the following steps: S101. Obtain the molecular structure representation of the compound to be predicted, and construct the molecular fingerprint and molecular map based on the molecular structure representation; S102. Construct a cholesterol efflux activity prediction model, including a fingerprint feature extraction branch, a graph structure feature extraction branch, a feature fusion layer, and a classification prediction layer; S103. The molecular fingerprint is fed into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; S104. The molecular graph is fed into the graph structure feature extraction branch to obtain the molecular graph representation; S105. The molecular fingerprint representation and molecular map representation are fed into the feature fusion layer to obtain the fused molecular representation. S106. Input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. S107. Pre-set a classification threshold, and convert the predicted probability into a binary classification prediction result of whether the activity exists based on the classification threshold.

[0033] In this embodiment, a residual fusion structure in the feature fusion layer is used to take fingerprint representation as the basic structural representation, and supplementary topological correction information is learned through graph structure feature extraction branches. Compared with using molecular fingerprint or molecular graph representation alone, this embodiment can better balance stability and expressive power under small sample conditions. In addition, this embodiment does not simply use a fixed threshold, but presets a classification threshold, which can obtain a more reasonable classification decision boundary under different sample distribution imbalances.

[0034] In step S101, the molecular structure is represented as a SMILES molecular structure representation. The SMILES molecular structure representation is normalized and its structural validity is checked. Molecules that cannot be correctly resolved by cheminformatics tools, have incomplete structures, or have obvious abnormalities are removed or marked.

[0035] For the training data used in the cholesterol efflux activity prediction model, further duplicate molecule checks and label cleanup were performed. For samples with identical structures and consistent labels, a representative sample was retained; for samples with identical structures but conflicting labels, verification was conducted based on source information and experimental conditions, and those that could not be confirmed were not included in the training set.

[0036] After processing, each sample is represented as:

[0037] in, Indicates the first A small molecule, This indicates its cholesterol efflux activity label. This indicates that it is active. Indicated as inactive.

[0038] In step S101, an ECFP molecular fingerprint is constructed based on the molecular structure representation, and the ECFP molecular fingerprint representation is as follows:

[0039] in, For molecular fingerprint input, Input molecule in molecular structure representation Indicates the fingerprint radius. The fingerprint dimension is represented as follows: The fingerprint vector of the ECFP molecular fingerprint is represented as follows:

[0040] Each bit indicates whether the corresponding substructure appears in the molecule. This can be understood as ECFP fingerprints providing a relatively stable "list of local structures" for the model.

[0041] The structure of the molecular graph It is expressed as follows:

[0042] in, For a set of atomic nodes, This is a set of chemical bond edges. Each atomic node has initial node characteristics, including atomic type, atomicity, formal charge, hybridization mode, aromaticity, and whether it is located in a ring. Each chemical bond edge has chemical bond characteristics, including bond type, whether it is conjugated, and whether it belongs to a ring structure. The molecular diagram is used to preserve the atomic connections and topological structure information of the molecule, that is, to allow the cholesterol efflux activity prediction model to see "how the atomic nodes are connected".

[0043] In step S102, based on the processed samples from step S101, a training dataset is given. , means as follows:

[0044] The cholesterol efflux activity prediction model is trained and evaluated using repeated scaffold cross-validation, and a binary cross-entropy loss is used as the loss function during the training phase, as shown below:

[0045] in, For binary classification, cross-entropy loss, The model represents the first The predicted probability of a sample. Indicates the true label, This represents the total number of samples.

[0046] To reduce the risk of overfitting under small sample conditions, weight decay regularization can be further introduced:

[0047] in, Indicates model parameters, This represents the weight decay coefficient. This is the final optimized total loss function during model training.

[0048] In this embodiment, a training dataset, a validation set, and a test set are constructed using scaffold groups as units. The training dataset is used for updating the parameters of the cholesterol efflux activity prediction model, the validation set is used for selecting the cholesterol efflux activity prediction model and determining the threshold, and the test set is only used for final performance evaluation. Furthermore, the cholesterol efflux activity prediction model uses binary cross-entropy loss as the loss function, and can be combined with weight decay, early stopping strategies, and learning rate scheduling strategies to improve training stability.

[0049] In step S103, the fingerprint feature extraction branch uses a multilayer perceptron network to perform a nonlinear mapping on the molecular fingerprint input to obtain a molecular fingerprint representation, as shown below:

[0050] in, and The first Layer weights and biases It is a non-linear activation function. Indicates the number of fingerprint branch layers. To obtain the final molecular fingerprint characterization, Characterization of the original molecular fingerprint, For the first Hidden molecular fingerprint characterization obtained after layer nonlinear mapping For the process The final molecular fingerprint characterization obtained after layer mapping For the first Characterization of hidden molecular fingerprints from layer output.

[0051] In this embodiment, the fingerprint feature extraction branch is mainly used to provide stable local structural priors, which is equivalent to providing a relatively reliable basis for judgment for the cholesterol efflux activity prediction model.

[0052] In step S104, the graph structure feature extraction branch uses a graph neural network to perform message passing, node updating, and graph-level readout on the molecular graph input to obtain a molecular graph representation, specifically including the following steps: S11. The molecular graph is fed into the graph structure feature extraction branch. A graph neural network is used to learn the structure of the molecular graph through a message passing mechanism. The initial features of the nodes are... It is expressed as follows:

[0053] in, Indicates the first Initial node characteristics of each atom.

[0054] S12, Based on the structure of the learned molecular map, in the... In a layered graph neural network, atomic nodes Aggregating messages from the set of neighboring nodes is represented as follows:

[0055] Then, the atomic nodes are updated, as shown below:

[0056] in, For the first The layer consists of atomic nodes The message vector is obtained by aggregating the features of neighboring nodes and edges. For the first Atomic nodes in layers The current hidden representation, For the first Neighbor nodes in the layer The hidden representation, atomic nodes with neighboring nodes Edge features between them For the set of neighboring nodes, For message functions, For the first Atomic nodes in layers The updated hidden representation, This represents the node update function; S13. After multi-layer message passing based on the message passing mechanism, a readout operation is performed on all atomic nodes that have undergone node updates to obtain the molecular graph representation, as shown below:

[0057] in, For molecular diagram characterization, For a set of atomic nodes, For the process After message passing in a layered graph neural network, atomic nodes The final node representation; The first in the molecular diagram Atom nodes, The number of layers in a graph neural network. This is a pooling function used to aggregate all atomic node representations into a molecular-level representation.

[0058] In this embodiment, the graph structure feature extraction branch is used to learn trainable topological structure information of molecules, serving as a supplement to fingerprint representation. Its role is not to completely replace fingerprints, but rather to further supplement structural connectivity relationships that fingerprints struggle to express.

[0059] Step S105 specifically includes the following steps: S41. The molecular fingerprint representation and molecular graph representation are projected onto a latent representation space of the same dimension through a feature fusion layer, as shown below:

[0060] in, This represents the fingerprint pattern after projection. This represents the molecular diagram characterization after projection. This is the linear projection weight matrix corresponding to the molecular fingerprint characterization. The linear projection weight matrix is ​​used to characterize the molecular diagram. This is the bias vector in the projection process of molecular fingerprint characterization. The bias vector is used to characterize the projection process of the molecular diagram.

[0061] S42. The projected molecular fingerprint characterization and molecular map characterization are spliced ​​together and input into the residual correction network to generate residual correction terms, as shown below:

[0062] in, Vector concatenation ， Represents the residual correction network. This represents the residual correction term generated jointly by the graph structure information and the fingerprint information. This represents the molecular fingerprint characterization after projection. This represents the molecular diagram characterization after projection.

[0063] S43. Add the residual correction term to the projected molecular fingerprint characterization to obtain the fused molecular characterization, as follows:

[0064] in, This represents the characterization of the fusion molecule. The structure of the fusion molecule characterization can be understood as follows: and The sum of .

[0065] This embodiment does not simply combine molecular fingerprint representation and molecular graph representation. Instead, it uses the ECFP molecular fingerprint as a relatively stable foundation, and then learns a supplementary correction value through the graph structure feature extraction branch. This reduces the instability caused by the model's complete reliance on graph neural networks under small sample conditions, while retaining the expressive advantages of graph structure information.

[0066] In step S106, the predicted probability that the compound to be predicted belongs to the cholesterol efflux active category (i.e., the active category) is expressed as follows:

[0067] in, This represents the sigmoid function. This represents a classification prediction network. This represents the predicted probability. When... The closer the value is to 1, the higher the likelihood that the model considers the molecule to have cholesterol efflux regulation activity; when... The closer the value is to 0, the higher the likelihood that the model considers the molecule to be inactive.

[0068] In step S107, the classification threshold is determined by the out-of-fold prediction probability and corresponding true label obtained during the training phase of the cholesterol efflux activity prediction model through repeated core backbone cross-validation. Candidate thresholds are scanned, and within the candidate threshold range, at least one of the following evaluation metrics is calculated: Matthews correlation coefficient (MCC), balanced accuracy (BACC), and harmonic mean of precision and recall (F1). The threshold that achieves the best results for the evaluation metrics is selected as the final classification threshold. It is expressed as follows:

[0069] in, For candidate thresholds, It is at least one of the evaluation metrics of MCC, BACC or F1, and is calculated based on the candidate threshold according to the predicted probability and the true label.

[0070] The step of converting the predicted probability into a binary classification prediction result of whether the entity is active based on the classification threshold is expressed as follows:

[0071] in, This indicates that the prediction is that it is active. This indicates that the prediction is that the product is inactive.

[0072] In this embodiment, since the cholesterol efflux activity data has class imbalance and the predicted probability output by the cholesterol efflux activity prediction model is not necessarily naturally calibrated, this embodiment does not use 0.5 as the only classification threshold. Instead, the classification threshold is selected based on the out-of-fold prediction results in the repeated core skeleton cross-validation.

[0073] The molecular characterization and prediction method for screening cholesterol efflux active compounds further includes a structural support analysis step for external candidate molecules. The structural support analysis step includes: to improve the interpretability of the prediction results of external candidate molecules, calculating the maximum structural similarity between the external candidate molecules in the compound to be predicted and the known active molecules in the training dataset, and the maximum structural similarity between the external candidate molecules in the compound to be predicted and the known inactive molecules in the training dataset, and determining the structural support direction of the compound to be predicted based on the difference between the two.

[0074] external candidate molecules Maximum structural similarity with known active molecules in the training dataset , means as follows:

[0075] in, For the set of known active molecules in the training dataset, For the first in the training dataset A known active molecule.

[0076] The maximum structural similarity between the external candidate molecule and the known inactive molecules in the training dataset is expressed as follows:

[0077] in, To provide the set of known inactive molecules in the training dataset, For the first in the training dataset A known inactive molecule.

[0078] And define the structure to support differences :

[0079] in, This represents the Tanimoto similarity based on ECFP molecular fingerprints; when When this occurs, it indicates that external candidate molecules are closer to known active molecules in the structure space of the training dataset; when... When this occurs, it indicates that external candidate molecules are closer to known inactive molecules in the structure space of the training dataset; when... When the value is close to 0, it indicates that the external candidate molecules are located in the structural boundary region that determines the mixture or active and inactive molecules.

[0080] This embodiment can help explain whether the predicted probability of the cholesterol efflux activity prediction model has local structural neighborhood support. In other words, the cholesterol efflux activity prediction model can not only determine whether the molecule is active, but also provide further support for the determination.

[0081] In the aforementioned molecular characterization and prediction method for screening cholesterol efflux-active compounds, to further evaluate the reliability of the prediction probability, a multi-random seed stability analysis step is also included. This involves using cholesterol efflux activity prediction models trained with multiple random seeds to predict external candidate molecules, and statistically analyzing the consistency of prediction results under different random seeds to help determine the stability of the prediction probability. Specifically, this includes the following: For a given external candidate molecule, if it is predicted as the active class under multiple random seeds and its structural neighborhood is closer to that of an active molecule, then the molecule can be considered a high-confidence active candidate molecule. If it is predicted incorrectly under multiple random seeds and its structural neighborhood is inconsistent with the true label orientation, then it can be considered a structural support conflict sample. If the prediction results are inconsistent under different random seeds, then the molecule can be considered a boundary-type candidate molecule.

[0082] Therefore, this invention not only outputs the predicted probability, but also provides prediction stability and structural support explanations, which facilitates subsequent priority ranking of candidate molecules.

[0083] In this embodiment, as Figure 4 As shown, the cholesterol efflux activity prediction model was evaluated using repeated scaffold cross-validation with five random seeds. Experimental results show that the cholesterol efflux activity prediction model achieved relatively stable performance on AUROC, AUPRC, BACC, F1, and MCC.

[0084] Furthermore, compared with baseline models such as CatBoost, XGBoost, LightGBM, MLP-ECFP, and GIN, the advantages of the cholesterol efflux activity prediction model are mainly reflected in fixed threshold classification indicators such as MCC, BACC, and F1.

[0085] Furthermore, ablation experiments showed that using fingerprint feature extraction branch alone, graph structure feature extraction branch alone, or simple fusion method alone were all insufficient to achieve the performance of the complete model, indicating that the residual fusion structure can more effectively integrate fixed molecular fingerprint characterization and molecular graph characterization.

[0086] This invention can provide stable and structurally supported predictions for some external candidate molecules, while also identifying structurally conflicting and boundary-type samples. This demonstrates that this invention can not only predict activity but also provide interpretive support for candidate molecule screening.

[0087] Reference Figure 2 This invention provides a molecular characterization and prediction system for screening cholesterol efflux-active compounds, used to implement any of the above-mentioned molecular characterization and prediction methods for screening cholesterol efflux-active compounds, the system comprising: Molecular Input Construction Module: Used to construct molecular fingerprints and molecular maps based on the molecular structure representation of compounds; Cholesterol efflux activity prediction module: used to build a cholesterol efflux activity prediction model, including fingerprint feature extraction branch, graph structure feature extraction branch, feature fusion layer and classification prediction layer; Fingerprint feature extraction module: used to feed the molecular fingerprint into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; Graph structure feature extraction module: used to feed the molecular graph into the graph structure feature extraction branch to obtain the molecular graph representation; The residual fusion module is used to feed the molecular fingerprint representation and molecular map representation into the feature fusion layer to obtain the fused molecular representation. The classification prediction module is used to input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. Threshold decision module: used to pre-set classification thresholds and convert the predicted probabilities into binary classification prediction results of whether the entity is active based on the classification thresholds.

[0088] Interpretation and Analysis Module: Used to interpret the prediction results of external candidate molecules based on the structure support analysis step and the multi-random seed stability analysis step.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A molecular characterization and prediction method for screening cholesterol efflux-active compounds, characterized in that, The method includes the following steps: S101. Obtain the molecular structure representation of the compound to be predicted, and construct the molecular fingerprint and molecular map based on the molecular structure representation; S102. Construct a cholesterol efflux activity prediction model, including a fingerprint feature extraction branch, a graph structure feature extraction branch, a feature fusion layer, and a classification prediction layer; S103. The molecular fingerprint is fed into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; S104. The molecular graph is fed into the graph structure feature extraction branch to obtain the molecular graph representation; S105. The molecular fingerprint representation and molecular map representation are fed into the feature fusion layer to obtain the fused molecular representation. S106. Input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. S107. Pre-set a classification threshold, and convert the predicted probability into a binary classification prediction result of whether the activity exists based on the classification threshold.

2. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 1, characterized in that, The molecular fingerprint is represented as follows: in, Molecular fingerprint, This refers to the input molecule in molecular structure representation. Indicates the fingerprint radius. Indicates fingerprint dimension; The structure of the molecular diagram includes a set of atomic nodes and a set of chemical bond edges. Each atomic node has initial node characteristics, including atomic type, atomicity, formal charge, hybridization mode, aromaticity, and whether it is located in a ring. Each chemical bond edge has chemical bond characteristics, including bond type, whether it is conjugated, and whether it belongs to a ring structure.

3. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 2, characterized in that, The fingerprint feature extraction branch uses a multilayer perceptron network to perform a nonlinear mapping on the molecular fingerprint input to obtain a molecular fingerprint representation, as shown below: in, and The first Layer weights and biases It is a non-linear activation function. Indicates the number of fingerprint branch layers. To obtain the final molecular fingerprint characterization, Characterization of the original molecular fingerprint, For the first Hidden molecular fingerprint characterization obtained after layer nonlinear mapping For the process The final molecular fingerprint characterization obtained after layer mapping For the first Characterization of hidden molecular fingerprints from layer output.

4. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 3, characterized in that, The graph structure feature extraction branch uses a graph neural network to perform message passing, node updating, and graph-level readout on the molecular graph input to obtain a molecular graph representation, specifically including the following steps: S11. Feed the molecular graph into the graph structure feature extraction branch, and use a graph neural network to learn the structure of the molecular graph through a message passing mechanism; S12, Based on the structure of the learned molecular map, in the... In a layered graph neural network, atomic nodes aggregate messages from their neighbor sets and update those atomic nodes, as shown below: in, For the first The layer consists of atomic nodes The message vector is obtained by aggregating the features of neighboring nodes and edges. For the first Atomic nodes in layers The current hidden representation, For the first Neighbor nodes in the layer The hidden representation, atomic nodes with neighboring nodes Edge features between them For the set of neighboring nodes, For message functions, For the first Atomic nodes in layers The updated hidden representation, This represents the node update function; S13. After multi-layer message passing based on the message passing mechanism, a readout operation is performed on all atomic nodes that have undergone node updates to obtain the molecular graph representation, as shown below: in, For molecular diagram characterization, For a set of atomic nodes, For the process After message passing in a layered graph neural network, atomic nodes The final node representation; The first in the molecular diagram Atom nodes, The number of layers in a graph neural network. This is the pooling function.

5. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 1, characterized in that, Step S105 specifically includes the following steps: S41. Molecular fingerprint representation and molecular graph representation are projected onto a latent representation space of the same dimension through a feature fusion layer. S42. The projected molecular fingerprint characterization and molecular map characterization are spliced ​​together and input into the residual correction network to generate residual correction terms. S43. Add the residual correction term to the projected molecular fingerprint characterization to obtain the fused molecular characterization, as follows: in, This represents the molecular fingerprint characterization after projection. This represents the molecular diagram characterization after projection. This represents vector concatenation. Represents the residual correction network. This indicates the characterization of fused molecules.

6. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 5, characterized in that, The predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category is expressed as follows: in, This represents the sigmoid function. This represents a classification prediction network. This represents the predicted probability.

7. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 1, characterized in that, The cholesterol efflux activity prediction model is trained using repeated core framework cross-validation, and a binary cross-entropy loss is used as the loss function during the training phase, as shown below: in, For binary classification, cross-entropy loss, The model represents the first The predicted probability of a sample. Indicates the true label, This represents the total number of samples.

8. The molecular characterization and prediction method for screening cholesterol efflux-active compounds according to claim 7, characterized in that, The classification threshold is determined by the out-of-fold prediction probability and the corresponding true label obtained during the training phase of the cholesterol efflux activity prediction model through repeated core skeleton cross-validation.

9. A molecular characterization and prediction system for screening cholesterol efflux-active compounds, characterized in that, The system includes: Molecular Input Construction Module: Used to construct molecular fingerprints and molecular maps based on the molecular structure representation of compounds; Cholesterol efflux activity prediction module: used to build a cholesterol efflux activity prediction model, including fingerprint feature extraction branch, graph structure feature extraction branch, feature fusion layer and classification prediction layer; Fingerprint feature extraction module: used to feed the molecular fingerprint into the fingerprint feature extraction branch to obtain the molecular fingerprint characterization; Graph structure feature extraction module: used to feed the molecular graph into the graph structure feature extraction branch to obtain the molecular graph representation; The residual fusion module is used to feed the molecular fingerprint representation and molecular map representation into the feature fusion layer to obtain the fused molecular representation. The classification prediction module is used to input the fused molecular characterization into the classification prediction layer and output the predicted probability that the compound to be predicted belongs to the cholesterol efflux activity category. Threshold decision module: used to pre-set classification thresholds and convert the predicted probabilities into binary classification prediction results of whether the entity is active based on the classification thresholds.

Images