A method for constructing a traditional Chinese medicine extract target prediction model and an application method thereof

Abstract

The application discloses a construction method and application method of a traditional Chinese medicine extract target prediction model, and comprises the following steps: obtaining an original data set of traditional Chinese medicine extract targets; the original data set comprises an LC-MS data set of traditional Chinese medicine extracts, a candidate target protein sequence set and a hit target information set; the LC-MS data is expressed by a two-dimensional matrix R(i, j) to represent the signal intensity corresponding to the retention time i and the mass-to-charge ratio j; the original data set is preprocessed to construct a training sample pair and generate a training data set; a network model is defined, the network model adopts a three-branch structure and comprises an extract image encoder, a protein sequence encoder, a cross-modal fusion and target prediction head; the network model is trained by using the training data set to construct a target prediction model. According to the technical scheme, the target prediction of the traditional Chinese medicine extract can be realized without the need of splitting and analyzing chemical components.

Description

Technical Field

[0001] This invention relates to the field of multimodal deep learning, and more specifically, to a method for constructing and applying a target prediction model for traditional Chinese medicine extracts. Background Technology

[0002] Traditional Chinese medicine (TCM) possesses complex characteristics of "multiple components, multiple targets, and multiple pathways," exhibiting unique advantages in the treatment of inflammation, metabolic diseases, autoimmune diseases, and tumors. However, TCM extracts typically contain a large number of unknown compounds, making it difficult to accurately locate their pharmacological targets. Target prediction often relies on costly and low-throughput screening methods. Commonly used target prediction methods mainly depend on the following approaches: (1) providing structural information for each component and performing target prediction based on molecular docking of the structure. However, the large number of components in TCM extracts, including a significant number of unknown compounds, limits the application of this method; (2) constructing a knowledge base based on literature databases and network pharmacology methods, which cannot predict unreported components; (3) performing target prediction based on single-component machine learning models, but these methods struggle to handle data at the "whole extract" level of TCM and lack expression of synergistic effects between components.

[0003] With the development of high-resolution LC-MS / MS technology, extracts of traditional Chinese medicine can be fully characterized through two-dimensional chromatograms and mass spectrometry fingerprints. However, existing studies mainly focus on peak extraction and fragment ion annotation, without analyzing and applying the potential mapping relationship between the "overall spectrum" and the target sites.

[0004] Therefore, a technical solution is needed to integrate the complex chemical profiles of Chinese herbal medicine extracts with target protein characteristics across modalities, enabling target prediction of Chinese herbal medicine extracts without knowing the specific components, thus breaking through the current methods of target prediction in Chinese herbal medicine research. Summary of the Invention

[0005] To achieve the above objectives, this application provides a method for constructing a target prediction model for traditional Chinese medicine extracts, comprising the following steps: Obtain the original dataset of targets of Chinese herbal extracts; the original dataset includes the LC-MS dataset of Chinese herbal extracts, the candidate target protein sequence set and the hit target information set; the LC-MS data uses a two-dimensional matrix R(i,j) to represent the signal intensity corresponding to retention time i and mass-to-charge ratio j; Preprocess the original dataset to construct training sample pairs. Generate a training dataset; where, To extract a two-dimensional grayscale image, For candidate protein sequences, For binary pairing tags; The network model is defined and adopts a three-branch structure, including: an extract image encoder, a protein sequence encoder, and a cross-modal fusion and target prediction head; The network model is trained using the training dataset to construct a target prediction model.

[0006] Preprocessing includes: LC-MS data of Chinese herbal extracts were converted into two-dimensional grayscale images; The target protein sequence is converted into a sequence index vector to generate a set of candidate target proteins; Generate the target hit vector, including: Let the first... Two-dimensional grayscale images of Chinese herbal extracts are The candidate target protein set is Define the target hit value , indicating extract Does it hit the target protein? When all candidate targets are evaluated simultaneously on the same extract sample, a length of [length missing] is generated. Hit target vector .

[0007] The conversion of LC-MS data of traditional Chinese medicine extracts into two-dimensional grayscale images includes: The signal strength is normalized, and the normalized signal strength is expressed as: ; Mapping the normalized signal intensity to [0,1], and generating a two-dimensional matrix S through outer product, it is represented as: ; A two-dimensional grayscale image is generated by linearly scaling the two-dimensional matrix S to 0~255, as shown below: .

[0008] The extracted image encoder employs a lightweight CNN and a (B,1,H,W) layout to extract local and global features from a single-channel grayscale image, outputting a batch feature vector of the extracted image. ; The protein sequence encoder employs a structure combining embedding and sequence modeling. It captures long-range dependencies for amino acid index sequences of shape (B,L) and outputs batch feature vectors of the protein sequence. ; The cross-modal fusion and target prediction head extracts the image encoder output. and the output of the protein sequence encoder The process involves fusing the data to generate a fusion vector. This fusion vector is then passed through multiple fully connected layers to output the probability of hitting the target point. This probability is further mapped to a confidence score as the prediction result, expressed as follows: ,in, .

[0009] On the other hand, this application provides an application method for a target prediction model of traditional Chinese medicine extracts, which is implemented based on the target prediction model constructed in this invention, including the following steps: The preparation of traditional Chinese medicine extracts and LC-MS / MS data detection were carried out to obtain raw mass spectrometry data files. The raw mass spectrometry data files were preprocessed to generate two-dimensional grayscale images. Prepare candidate target protein sequences and generate protein sequence vectors; Load the target prediction model, input a two-dimensional grayscale image and a protein sequence vector into the target prediction model, and obtain the target prediction results.

[0010] The preprocessing of the raw mass spectrometry data file includes the following steps: Peak extraction, denoising, retention time alignment, and intensity normalization are performed on the raw mass spectrometry data file to export structured peak table data as input sample A; the structured peak table information includes: retention time RT, first-order mass-to-charge ratio m / z, and peak intensity; A feature vector is constructed from input sample A and used as input sample B; Convert the input sample B into a two-dimensional grayscale image.

[0011] The construction of feature vectors includes: Mapping the normalized peak intensity to a one-dimensional feature vector is represented as: Where k is a preset dimension, with a value range of 32 to 1024, preferably 64 or 128; The mapping method supports m / z binning, RT-m / z joint binning, and substructure projection based on fragment ion fingerprints. When using m / z binning, the m / z range is divided into k intervals, and the normalized intensity falling into each interval is accumulated to obtain the corresponding value. Normalize X to 0-1 to obtain the input sample B.

[0012] Converting to a 2D grayscale image includes: The two-dimensional space of RT and m / z is mapped to a grayscale distribution map G_RT-m / z. RT is divided into r intervals and m / z is divided into m intervals. The corresponding normalized intensity is accumulated in the r×m grid and then quantized into grayscale values ​​to form a two-dimensional grayscale image that reflects the overall fingerprint distribution of the sample. This image reflects the overall chemical distribution of the "retention time-mass-charge ratio" two-dimensional space as the fingerprint characterization of the sample.

[0013] The generated protein sequence vector includes: Obtain the amino acid sequences of candidate target protein sets, including: obtaining protein sets from inflammation-related pathways, immune-related pathways, or disease-related networks, and for each protein in the set... Obtain its amino acid sequence SEQ_j or three-dimensional structure STR_j; SEQ_j can be organized in FASTA format, and STR_j can be organized in PDB format, as input sample D; The input sample D is encoded to generate a fixed-dimensional vector, represented as follows: .

[0014] This invention introduces a two-dimensional grayscale image to represent the LC-MS fingerprint of the extract, and combines it with a deep embedding vector of the target protein sequence or structure to achieve target prediction of traditional Chinese medicine extracts without the need to disassemble and analyze chemical components. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the construction steps of a target prediction model for traditional Chinese medicine extracts according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the application method of the target prediction model for traditional Chinese medicine extracts provided in the embodiments of the present invention. Figure 3 This is a schematic diagram of a two-dimensional grayscale image generated by mapping the RT and m / z two-dimensional space of a traditional Chinese medicine extract according to an embodiment of the present invention. Detailed Implementation

[0016] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] The method for constructing a target prediction model for traditional Chinese medicine extracts provided by this invention is as follows: Figure 1 As shown, it includes the following steps: Step S100: Obtain the original dataset of targets of traditional Chinese medicine extracts; the original dataset includes the LC-MS dataset of traditional Chinese medicine extracts, the candidate target protein sequence set, and the target hit information set; 1) Select raw medicinal materials and extract the effective components using ethanol, water, or methanol as solvents to obtain traditional Chinese medicine extracts. Perform liquid chromatography (LC), mass spectrometry (MS), and ionization measurements on the extracts to obtain LC-MS data. LC-MS data are three-dimensional information, and a two-dimensional matrix R(i,j) is used to represent the signal intensity corresponding to retention time i and mass-to-charge ratio j, specifically represented as the retention time axis. Mass-to-charge ratio axis and signal strength matrix ; 2) The candidate target protein sequence set consists of multiple target protein sequence information; When obtaining candidate target protein sequence sets, target protein sequence information is obtained through public databases (such as Uniprot and PDB). Each protein sequence consists of 20 standard amino acids and a group representing any amino acid. Composed of dictionaries, represented as: ; 3) The target hit information set consists of target hit information, which is used to reflect the hit status of the specified Chinese medicine extract on the specified target protein sequence, and is represented by 0 and 1.

[0018] When obtaining target information, known component-target pairs in existing databases, such as TCMSP, HERB, and BATMAN-TCM, are used to generate target information by mapping the known components in the extracts to their corresponding targets through network pharmacology.

[0019] Step S110: Preprocess the original dataset to construct training sample pairs. Generate a training dataset; where, To extract a two-dimensional grayscale image, For candidate protein sequences, For binary pairing tags; In this step, preprocessing includes: 1) Convert the LC-MS data of the Chinese herbal medicine extracts into two-dimensional grayscale images: First, the signal strength is normalized. The normalized signal strength is expressed as: Furthermore, the normalized signal intensity is mapped to [0,1], and a two-dimensional matrix S is generated through outer product, which is represented as: This matrix can reflect the synergistic relationship between peak clusters within the extract; Then, by linearly scaling the two-dimensional matrix S to 0~255, a two-dimensional grayscale image can be generated, represented as: A two-dimensional grayscale image can be considered a single-channel image, and its size is... (Corresponding to the retention time axis and mass-to-charge ratio axis).

[0020] 2) Convert the target protein sequence into a sequence index vector to generate a set of candidate target proteins: Use pre-trained protein language models (such as ESM, ProtBERT) to process protein sequences. Encoded as a 1024-dimensional integer index vector If the protein has structural information, a graph convolutional network (GCN) is used to extract structural features to automatically learn the protein's active site information.

[0021] Integer index vector The length is fixed at L_max; if the length of the mapped vector is less than L_max, it is padded. Supplement the length, otherwise truncate the portion exceeding L_max; the candidate target protein set is represented as follows. .

[0022] 3) Generate the target hit vector, including: Let the first... Two-dimensional grayscale images of Chinese herbal extracts are The candidate target protein set is Define the target hit value , indicating extract Does it hit the target protein? When all candidate targets are evaluated simultaneously on the same extract sample, a length of [length missing] is generated. Hit target vector .

[0023] In this step, based on the extracted two-dimensional grayscale image Candidate protein sequences and the target vector Construct training sample pairs ,in, and As input to the sample, the target vector is hit. These are paired labels. The training sample pairs generated from the original dataset constitute the training dataset.

[0024] Meanwhile, Mixup and contrastive learning are used to generate extended training data, improving the model's ability to generalize to unknown components.

[0025] Step S120: Define the network model. The network model adopts a three-branch structure, including: extract image encoder, protein sequence encoder, and cross-modal fusion and target prediction head; 1) The extracted image encoder uses a lightweight CNN (such as a simplified version of ResNet18 or ConvNext) and a (B,1,H,W) layout to extract local and global features from a single-channel grayscale image (the input 2D grayscale image), and outputs a batch feature vector of the extracted image. .

[0026] 2) The protein sequence encoder employs a structure combining embedding and sequence modeling, such as embedding + 1D convolution, BiLSTM, or a small Transformer. This structure can capture long-range dependencies for amino acid index sequences of shape (B,L) and output batch feature vectors of the protein sequence. .

[0027] 3) The cross-modal fusion and target prediction head extracts the image encoder output... and the output of the protein sequence encoder The process involves fusing the data to generate a fusion vector. This fusion vector is then passed through multiple fully connected layers to output the probability of hitting the target point. This probability is further mapped to a confidence score as the prediction result, expressed as follows: ,in, .

[0028] Step S130: Train the network model using the training dataset to build a target prediction model.

[0029] During training, the Adam optimizer combined with dynamic learning rate decay is used, enabling rapid convergence in the early stages and fine-tuning in the later stages. The loss function employs a weighted combination of cross-entropy loss and contrastive loss; the former supervises the main classification task, while the latter enhances feature discriminative power by bringing positive samples closer and pushing negative samples further away. Batch size can be flexibly selected between 16 and 64, balancing training stability and hardware efficiency.

[0030] On the other hand, this invention provides a method for applying the target prediction model, the specific steps of which are as follows: Figure 2 As shown, it includes: Step S210: Prepare the Chinese herbal extract and perform LC-MS / MS data detection to obtain the raw mass spectrometry data file. Preprocess the raw mass spectrometry data file to generate a two-dimensional grayscale image. The present invention provides a specific embodiment: 30 kg of *Ligusticum striatum* bark was weighed, and 10 times the amount of 70% ethanol was added for the first extraction for 1.5 h. The filtrate was filtered and retained. 8 times the amount of 70% ethanol was added to the residue and the extraction was continued for 1.0 h. The filtrate was filtered again and retained. The filtrates from the two extractions were combined, the ethanol was recovered and concentrated under reduced pressure to 1 g / mL to obtain the Chinese herbal extract.

[0031] Traditional Chinese medicine extracts were analyzed using liquid chromatography-tandem mass spectrometry (LC-MS / MS). Electrospray ionization (ESI) was the preferred ion source for mass spectrometry. MS data were collected in positive and / or negative ion modes. 1 With MS 2 Data. The raw mass spectrometry data files obtained can be in formats such as .raw and .wiff.

[0032] Preprocessing raw mass spectrometry data files includes the following steps: Step S211: Perform peak extraction, noise reduction, retention time alignment, and intensity normalization on the original mass spectrometry data file to export structured peak table data as input sample A. The structured peak table includes: retention time RT (min), first-order mass-to-charge ratio m / z, adduct ion form, MS 2 Fragment ion set, peak intensity, and normalized intensity (Intensity_norm).

[0033] In this embodiment of the invention, some structured peak data of the extract of *Ligustrum lucidum* are shown in Table 1.

[0034] Table 1. Exemplary peak records of YSP-Extract-001 extract from *Cephalotaxus fortunei* bark (LC-MS / MS) (excerpt)

[0035] Step S212: Construct a feature vector from input sample A as input sample B; Specifically, the normalized peak intensities of the structured peak table data are mapped to a one-dimensional feature vector, represented as: Where k is a preset dimension, which can be 32 to 1024, preferably 64 or 128.

[0036] The mapping method uses m / z binning: the m / z range is divided into k intervals, and the normalized intensity falling into each interval is accumulated to obtain the corresponding value. Then, the entire X is normalized to 0 to 1 to obtain the input sample B.

[0037] The mapping method also supports RT-m / z joint binning and substructure projection based on fragment ion fingerprints.

[0038] Step S213: Convert the input sample B into a two-dimensional grayscale image; The transformation method uses an outer product graph construction approach: Let S = X·X T We obtain a k×k matrix; after performing Min-Max normalization on matrix S, we multiply it by 255 and quantize it into grayscale values ​​of 0 to 255, forming a two-dimensional grayscale image G_X. G_X serves as an image representation of the extracted chemical features and can be directly input into the image branch network for representation learning.

[0039] In this invention, the two-dimensional space of RT and m / z is mapped to a grayscale distribution image G_RT-m / z, as shown below. Figure 3 As shown, RT is divided into r intervals and m / z into m intervals. The corresponding normalized intensities are accumulated in the r×m grid and quantized into grayscale values, forming a two-dimensional grayscale image reflecting the overall fingerprint distribution of the sample. This image represents the overall chemical distribution in the "retention time-mass-charge ratio" two-dimensional space as a fingerprint characterization of the sample. Higher brightness in the image indicates a stronger ion signal in that region. Different samples will form different grayscale distribution patterns, which can be used by the model to distinguish and learn their chemical characteristics.

[0040] Step S220: Prepare candidate target protein sequences and generate protein sequence vectors; In this step, the candidate target protein set is represented as follows: Where n can be 10 to 20000, preferably 200 to 2000, the process of generating protein sequence vectors includes: Step S221: Obtain the amino acid sequences of the candidate target protein set: The candidate target protein set data comes from protein sets within inflammation-related pathways, immune-related pathways, or disease-related networks. For each protein in the set... Obtain its amino acid sequence SEQ_j or three-dimensional structure STR_j; the three-dimensional structure can be obtained from a public structure library (such as PDB) or generated by a structure prediction method. SEQ_j can be organized in FASTA format, and STR_j can be organized in PDB format, as input sample D.

[0041] Step S222: Encode the input sample D to generate a fixed-dimensional vector, represented as: ; The encoding methods support amino acid composition frequency, k-mer frequency, and protein sequence vectors or combinations of protein sequence vectors obtained based on pre-trained protein language models.

[0042] In practical applications, the same outer product mapping method can be used to generate a two-dimensional grayscale image of the protein, G_Pj: Let S_j = Y_j·Y_j T S_j is normalized and quantized to a grayscale value of 0–255 to obtain a two-dimensional grayscale image of the protein, G_Pj. The two-dimensional grayscale image G_Pj is used for the visual representation and result display of candidate target protein sequences, facilitating the explanation of the input data format and enhancing interpretability; preferably, it is a protein sequence index vector and its embedded representation, but it is not a necessary input for target prediction.

[0043] Step S230: Load the target prediction model, input a two-dimensional grayscale image and a protein sequence vector into the target prediction model, and obtain the target prediction results.

[0044] In the inference prediction process of the target prediction model, the extract image encoder obtains the extract representation vector from the two-dimensional grayscale image. The protein sequence encoder obtains the protein representation vector from the protein sequence vector. ; In cross-modal fusion and target prediction heads, As a query, As the key and value, attention weights are calculated and a fused representation is obtained. The fusion will then be represented. The matching probability is obtained by inputting the fully connected layer. : Furthermore, Mapped to confidence level: ,in, For the Sigmoid function, .

[0045] In this embodiment of the invention, after inference on the sample YSP-Extract-001 of the extract of the Chinese privet bark, the output Top target results include TNF-α, IL-1β, IL-6 and AKT, as shown in Table 2.

[0046] Table 2 Target prediction results of *Ligustrum lucidum* bark extract

[0047] The present invention provides a method for constructing and applying a target prediction model for traditional Chinese medicine extracts, breaking through the traditional paradigm of traditional Chinese medicine research. For the first time, it uses the full LC-MS spectrum of traditional Chinese medicine as the overall image input model, avoiding complex component decomposition. Through multimodal deep learning, it directly establishes an implicit mapping between the spectrum and the target, achieving intelligent matching of complex mixtures with protein targets. An innovative cross-modal attention mechanism is introduced, which can explain the correlation between key peak clusters and targets, providing a basis for screening active ingredients. The model possesses the ability to predict unknown components and perform multi-target collaborative analysis, significantly improving prediction efficiency and reliability compared to traditional network pharmacology methods, and can greatly reduce R&D costs and accelerate the discovery of traditional Chinese medicine targets.

[0048] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for constructing a target prediction model for traditional Chinese medicine extracts, characterized in that, Includes the following steps: Obtain the original dataset of targets for traditional Chinese medicine extracts; the original dataset includes the LC-MS dataset of traditional Chinese medicine extracts, the candidate target protein sequence set, and the target hit information set; The LC-MS data uses a two-dimensional matrix R(i,j) to represent the signal strength corresponding to retention time i and mass-to-charge ratio j; The original dataset is preprocessed to construct training sample pairs. Generate a training dataset; where, To extract a two-dimensional grayscale image, For candidate protein sequences, For binary pairing tags; The network model is defined and adopts a three-branch structure, including: an extract image encoder, a protein sequence encoder, and a cross-modal fusion and target prediction head; The network model is trained using the training dataset to construct a target prediction model.

2. The method for constructing a target prediction model for traditional Chinese medicine extracts according to claim 1, characterized in that, The preprocessing includes: LC-MS data of Chinese herbal extracts were converted into two-dimensional grayscale images; The target protein sequence is converted into a sequence index vector to generate a set of candidate target proteins; Generate the target hit vector, including: Let the first... Two-dimensional grayscale images of Chinese herbal extracts are The candidate target protein set is Define the target hit value , indicating extract Does it hit the target protein? When all candidate targets are evaluated simultaneously on the same extract sample, a length of [length missing] is generated. Hit target vector .

3. The method for constructing a target prediction model for traditional Chinese medicine extracts according to claim 2, characterized in that, The process of converting LC-MS data of traditional Chinese medicine extracts into two-dimensional grayscale images includes: The signal strength is normalized, and the normalized signal strength is expressed as: ; Mapping the normalized signal intensity to [0,1], and generating a two-dimensional matrix S through outer product, it is represented as: ; The two-dimensional matrix S is linearly scaled to 0~255 to generate a two-dimensional grayscale image, represented as: .

4. The method for constructing a target prediction model for traditional Chinese medicine extracts according to claim 1, characterized in that, The extracted image encoder employs a lightweight CNN and a (B,1,H,W) layout to extract local and global features from a single-channel grayscale image, outputting a batch feature vector of the extracted image. ; The protein sequence encoder employs a structure combining embedding and sequence modeling. It captures long-range dependencies for amino acid index sequences of shape (B,L) and outputs batch feature vectors of the protein sequence. ; The cross-modal fusion and target prediction head extracts the image encoder output. and the output of the protein sequence encoder The process involves fusing the data to generate a fusion vector. This fusion vector is then passed through multiple fully connected layers to output the probability of hitting the target point. This probability is further mapped to a confidence score as the prediction result, expressed as follows: ,in, .

5. A method for applying a target prediction model for traditional Chinese medicine extracts, characterized in that, The target prediction model constructed according to claims 1 to 4 is implemented by including the following steps: The preparation of traditional Chinese medicine extracts and LC-MS / MS data detection were carried out to obtain raw mass spectrometry data files. The raw mass spectrometry data files were preprocessed to generate two-dimensional grayscale images. Prepare candidate target protein sequences and generate protein sequence vectors; Load the target prediction model, input a two-dimensional grayscale image and a protein sequence vector into the target prediction model, and obtain the target prediction result.

6. The application method of the target prediction model for traditional Chinese medicine extracts according to claim 5, characterized in that, The preprocessing of the raw mass spectrometry data file includes the following steps: Peak extraction, denoising, retention time alignment, and intensity normalization are performed on the original mass spectrometry data file to export structured peak table data as input sample A; the information in the structured peak table includes: retention time RT, first-order mass-to-charge ratio m / z, and peak intensity; A feature vector is constructed from input sample A and used as input sample B; Convert the input sample B into a two-dimensional grayscale image.

7. The application method of the target prediction model for traditional Chinese medicine extracts according to claim 6, characterized in that, The construction of the feature vector includes: Mapping the normalized peak intensity to a one-dimensional feature vector is represented as: Where k is a preset dimension, with a value range of 32 to 1024, preferably 64 or 128; The mapping method supports m / z binning, RT-m / z joint binning, and substructure projection based on fragment ion fingerprints. When using m / z binning, the m / z range is divided into k intervals, and the normalized intensity falling into each interval is accumulated to obtain the corresponding value. Normalize X to 0-1 to obtain the input sample B.

8. The application method of the target prediction model for traditional Chinese medicine extracts according to claim 7, characterized in that, The conversion to a two-dimensional grayscale image includes: The two-dimensional space of RT and m / z is mapped to a grayscale distribution map G_RT-m / z. RT is divided into r intervals and m / z is divided into m intervals. The corresponding normalized intensity is accumulated in the r×m grid and then quantized into grayscale values ​​to form a two-dimensional grayscale image that reflects the overall fingerprint distribution of the sample. This image reflects the overall chemical distribution of the "retention time-mass-charge ratio" two-dimensional space as the fingerprint characterization of the sample.

9. The application method of the target prediction model for traditional Chinese medicine extracts according to claim 5, characterized in that, The generated protein sequence vector includes: Obtain the amino acid sequences of candidate target protein sets, including: obtaining protein sets from inflammation-related pathways, immune-related pathways, or disease-related networks, and for each protein in the set... Obtain its amino acid sequence SEQ_j or three-dimensional structure STR_j; SEQ_j can be organized in FASTA format, and STR_j can be organized in PDB format, as input sample D; The input sample D is encoded to generate a fixed-dimensional vector, represented as follows: .

Images