A drug representation method based on deep learning and a storage medium
By employing deep learning methods and utilizing the SMILES dataset of drug molecules and self-generated supervisory signals, the problems of adaptability and high computational cost of existing drug characterization methods are solved, achieving efficient and learnable drug characterization and improving the adaptability and speed of drug screening tasks.
Patent Information
- Application Number
- CN202211503046.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-28
- Publication Date
- 2025-12-19
- Estimated Expiration
- 2042-11-28
AI Technical Summary
Existing drug characterization methods suffer from poor adaptability, high computational cost, and huge training overhead, making it difficult to effectively generate characterization vectors that are adapted to different drug screening tasks.
We employ a deep learning-based drug characterization method, which involves collecting data from the SMILES drug molecule dataset, data augmentation, word segmentation, lexical embedding, designing a convolutional neural network model, pre-training, and generating supervisory signals to reduce computational load and improve adaptability.
It achieves efficient and learnable drug characterization, reduces computational resource requirements, and improves training speed and the speed and accuracy of drug characterization vector generation.
Smart Images
Figure CN115762706B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computer-aided drug design and deep learning, and provides a drug representation method based on deep learning and a storage medium. BACKGROUND
[0002] In the process of drug development and reutilization, various experiments such as drug chemistry experiments, animal experiments and clinical experiments are very time-consuming and expensive processes. With the generation of a large amount of drug experimental data and the development of computer technology, using data-based and computational model-based methods for computer-aided drug design is an important method in the current drug development process, which can perform preliminary screening to reduce the number of experiments, reduce research and development costs, and shorten the research and development cycle. The most important and key technology in computer-aided drug design is drug representation.
[0003] In recent years, drug representation methods mainly include molecular descriptors and self-supervised learning methods to obtain representations, but there are still limitations such as poor adaptability, large computational load and difficulty in training.
[0004] The drug representation method based on molecular descriptors is the most mainstream method at present, which mainly obtains a binary representation vector by calculating the types and quantities of various substructures in the drug molecule. This method is generally used in combination with traditional machine learning methods, and has certain representation ability, but it does not have adaptability. Once the molecular descriptors are selected, the generated drug representation vector is fixed, and it is impossible to learn the most suitable representation vector according to different drug screening tasks, and different molecular descriptors will seriously affect the performance of the screening task.
[0005] The drug representation method based on self-supervised learning mainly includes two types, namely self-supervised representation learning methods based on molecular sequence models and molecular graph models. This kind of method is to pre-train a super large scale model through the masking and restoration strategy, and then use the vector of the pre-trained global special token as the drug representation vector. The pre-training model used in this method has a very large number of parameters (usually hundreds of millions or even tens of billions), so it requires huge data and massive computing resources to train, and there are a series of problems such as slow convergence speed, long training time and high computational complexity of drug representation generation, which consumes a lot of manpower, material resources and time.
[0006] The supervised training convolutional neural network model based on extended connection fingerprints uses molecular fingerprints as a supervision signal to supervise the representation learning of the convolutional neural network model, and different drug representations can be generated according to different drug screening tasks. The method uses a convolutional neural network model and a supervised learning pre-training strategy, which greatly reduces the computational complexity, improves the universality of drug representation, and has the characteristics of easy training and interpretability. However, the technology has the disadvantage of long time-consuming for calculating the generation of supervision labels. SUMMARY
[0007] In view of the above problems, the purpose of the present application is to provide a drug representation method based on deep learning, which solves the problems of poor adaptability, unlearnable and huge training cost of drug representation in the prior art.
[0008] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0009] The present application provides a drug representation method based on deep learning, comprising the following steps:
[0010] 1) Collecting drug molecule SMILES data set; 2) Drug molecule SMILES data enhancement; 3) Drug molecule SMILES word segmentation; 4) Word embedding initialization; 5) Designing a convolutional neural network model; 6) Generating drug molecule fingerprint features; 7) Pre-training the convolutional neural network model; 8) Deriving the pre-training model through iterative calculation; 9) Inputting the molecule into the representation model to obtain static molecular representation, or jointly training dynamic representation to adapt to downstream tasks.
[0011] Step 1: Obtain the drug molecule standard SMILES and drug number construction data set provided by ChEMBL database, wherein 80% of the drugs are randomly selected as the training set, 10% as the validation set, and 10% as the test set;
[0012] Step 2: Read the drug molecule using Rdkit toolkit, and randomly generate 5 drug molecule SMILES to obtain the data enhanced random drug molecule SMILES;
[0013] Step 3: Construct a SMILES tokenizer to perform word segmentation on the drug molecule SMILES to obtain a list of word tokenized drug molecule SMILES;
[0014] Step 4: OneHot encode the drug molecule SMILES list obtained in step 3, and then use the Word2Vec model to obtain the embedding vector of the word token;
[0015] Step 5: Construct a deep learning model based on convolutional neural network to obtain a convolutional neural network model;
[0016] Step 6: Calculate the extended connection molecular fingerprint of each drug molecule;
[0017] Step 7: Use the extended connection fingerprint as a label to supervise the training of a convolutional neural network model;
[0018] Step 8: Save all weights and bias parameters of the convolutional neural network model except the output layer, and export the pre-trained model;
[0019] Step 9: Input the drug molecule SMILES into the pre-trained model to obtain a static drug representation, or jointly train a dynamic drug representation to adapt to downstream tasks.
[0020] In the above technical solution, the specific steps of step 2 are:
[0021] Use the Rdkit toolkit to read the standard SMILES of each drug molecule into a Mol file, and use the Mol file to uniquely represent a drug molecule; randomly initialize the atoms of the node, generate 5 random drug molecule SMILES through the drug molecule Mol file, and obtain the data enhanced random drug molecule SMILES.
[0022] In the above technical solution, the specific steps of step 3 are:
[0023] Divide all the drug molecule SMILES collected in steps 1 and 2 according to characters to obtain a list of divided characters;
[0024] Byte pair encoding is performed on the character list, and the pair of adjacent characters with the highest frequency of occurrence is selected as the unit data, i.e. the token, and the original adjacent characters are replaced with the unit data. Repeat the iteration multiple times until the existing new unit data reaches the vocabulary size V=2000, or the maximum number of iterations M=60000 is reached.
[0025] Save the vocabulary table that has met the iteration stop condition, and jointly use the vocabulary table and the atomic regular expression Token_regex as a SMILES tokenizer, where the atomic regular expression Token_regex is:
[0026] Token_regex=(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\\\\ / |:|~|@|\?|>|\*|\$|\%[0-9]{2}|[0-9])
[0027] Input the drug molecule SMILES collected in steps 1 and 2 into the SMILES tokenizer for tokenization to obtain a list of tokenized drug molecule SMILES.
[0028] In the technical solution, the specific steps of step 4 are:
[0029] The drug molecule SMILES list after tokenization is OneHot encoded to obtain the corresponding OneHot vector.
[0030] Randomly shuffle all OneHot vectors of the drug molecule SMILES list, and send the OneHot vectors in batches into a Word2vec model based on Skipgram and negative sampling for training to obtain the embedding vector of each token.
[0031] In the technical solution, the basic structure of the convolutional neural network model in step 5 adopts a TextCNN model, and the specific design is:
[0032] The TextCNN model receives the embedding vector matrix corresponding to the tokenized drug molecule SMILES list as input, and the embedding vector matrix is stacked according to the token order of the drug molecule SMILES list by the embedding vector obtained in step 4, denoted as Xi represents the k-dimensional embedding vector of word i , The embedding vector matrix of a drug molecule SMILES list with a length of n can be represented as:
[0033]
[0034] wherein, represents a concatenation operation, and the final X 1：n is a k*n embedding vector matrix of the drug molecule SMILES list, in the convolution operation, the convolution kernel obtains the corresponding length feature on the embedding vector matrix of the drug molecule SMILES list through a window with a length of h:
[0035] C i =f(w·X i：i+h-1 +b)
[0036] wherein C i is the i-th feature extracted by the convolution layer, w is the weight matrix in the neural network, is a bias term, f is a nonlinear activation function, and X i：i+h-1 represents the i-th row to the i+h-th row of the embedding vector matrix of the drug molecule SMILES list.
[0037] In the convolution calculation process, the convolution kernel slides in the embedded vector matrix of the drug molecule SMILES list to calculate the corresponding feature map, and different scale convolution kernels are set to extract the information of different length word groups in the word vector matrix; the pooling function used by the TextCNN model is the maximum pooling (1-Max), that is, a maximum feature is selected from the extracted feature vector to replace the entire vector, and then the features are spliced to form a new feature vector.
[0038]
[0039] wherein represents the maximum feature, c j represents each value in the feature vector.
[0040] In the above technical solution, the specific steps of step 6 are:
[0041] For each drug molecule, the Rdkit toolkit is used to calculate the extended connectivity fingerprints (Extended Connectivity Fingerprints), the radius of the extended connectivity fingerprints is set to 3, and the fingerprint length is 1024 bits, so as to obtain a 1024-bit binary molecular fingerprint vector.
[0042] In the above technical solution, the specific steps of step 7 are:
[0043] The output vector length of the convolutional neural network model in step 5 is set to 1024, that is, the length of the extended connectivity fingerprints; the last output of the convolutional neural network model is used as the molecular feature vector, the extended connectivity fingerprints are used as the supervision label, and the difference between the two is calculated by the Cosine distance, and the calculation formula is:
[0044] Loss=1-Cosine(Ecfp,O)
[0045] Wherein Loss represents the difference between the extended connectivity fingerprints and the model output feature vector, fcfp represents the 1024-bit extended connectivity molecular fingerprints, and O represents the output of the model.
[0046] The random gradient descent method is used to optimize the parameters of the convolutional neural network model, so that the difference between the extended connectivity fingerprints and the model output feature vector gradually decreases until the difference no longer changes, that is, the model training is completed.
[0047] In the above technical solution, the specific steps of step 9 are:
[0048] If the drug representation is needed directly, the drug molecule SMILES is input into the pre-trained model to obtain the corresponding molecular representation, and at this time the length of the molecular representation vector is fixed at 1024 bits.
[0049] If the drug representation is needed for subsequent computer-aided drug design tasks, a feedforward fully connected neural network needs to be connected after the existing molecular representation model for feature transformation, and the 1024-bit drug representation vector is transformed to the required dimension, and then the label of the subsequent task is combined for joint learning and training.
[0050] Compared with the prior art, the present application has the beneficial effects of:
[0051] In the preprocessing stage, the present application adopts a learnable word segmentation method, which contains statistical significance and drug chemical semantics, making the drug molecule SMILES word segmentation more efficient and reasonable; in the model design stage, a TextCNN-based model is adopted, which has a small number of trainable parameters and can also learn universal drug representation; a molecular fingerprint is used as a pre-training method for supervision signal, which greatly reduces the computational load, specifically:
[0052] 1. In the preprocessing stage, the present application uses a word segmentation method that takes into account byte pair encoding and chemical atom segmentation, fully considering the statistical rules of marked characters in large-scale drug molecule SMILES data and the semantic information of atoms in chemistry, that is, the word units after segmentation have high statistical frequency and contain chemical atomic semantics, and at the same time, a shorter and easier-to-compute word list can be obtained.
[0053] 2. The present application adopts a convolutional neural network model, and the model design is based on the TextCNN model family, with the convolution kernel length set to 3, 5, 7, 9, 12, and 16, which can capture multi-scale local word combination features; compared with other self-supervised deep learning models, the model parameters are only 2 million, about 1% of other models; the model adopted by the present application greatly reduces the number of parameters, improves the training speed and drug representation vector generation speed.
[0054] 3. The present application adopts a self-generated supervised pre-training method, and the supervision signal is generated by drug molecule calculation, without the need for manual labeling; the supervised pre-training method only needs to calculate the loss value in the last step to generate the representation vector, without the need for masking and restoration training, and the pre-training prediction frequency is reduced by about 98%; the pre-training method adopted by the present application greatly reduces the pre-training prediction frequency and training time, and the results are more intuitive and reasonable. BRIEF DESCRIPTION OF DRAWINGS
[0055] Figure 1 Operation flowchart of data preprocessing process.
[0056] Figure 2 Model architecture diagram of the convolutional neural network in the present application.
[0057] Figure 3Technical flowchart of the present application. DETAILED DESCRIPTION
[0058] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0059] The present application proposes a drug representation method based on deep learning, which is used to obtain drug molecule representation.
[0060] The main process of the present application includes: 1) collection of drug molecule SMILES dataset; 2) drug molecule SMILES data enhancement; 3) drug molecule SMILES segmentation; 4) word embedding initialization; 5) design of convolutional neural network model; 6) generation of drug molecule fingerprint features; 7) pre-training of convolutional neural network model; 8) derivation of pre-training model through iterative calculation; 9) input of molecule into representation model to obtain static molecular representation, or jointly train dynamic representation to adapt to downstream tasks, and the specific implementation steps are as follows:
[0061] I. Collection of drug molecule dataset
[0062] The dataset is constructed from the standard SMILES and drug numbers provided by the ChEMBL database, and there are about 1.8 million drug molecules in this part of the dataset. 80% of the drugs in the dataset are randomly selected as the training set, the training set is used for the pre-training stage, 10% is used as the validation set, and 10% is used as the test set. The validation set and the test set are mainly used to test the performance of the pre-training model to ensure that the model can generate effective drug representation vectors.
[0063] II. Drug molecule data enhancement
[0064] The Rdkit toolkit of python language is used to read the standard SMILES of each drug molecule into a Mol file, and the Mol file uniquely represents the drug molecule.
[0065] In a randomized initial atom manner, 5 random SMILES are generated from the drug molecule Mol file to obtain the data enhanced random SMILES.
[0066] III. Drug molecule SMILES segmentation
[0067] The drug molecule SMILES collected in steps 1 and 2 are all cut according to characters to obtain a list of segmented characters;
[0068] The character list is byte-pair encoded, and the pair of adjacent characters with the highest frequency of occurrence is selected as the unit data, i.e. the word unit, and the original adjacent characters are replaced with the unit data. Repeat the iteration multiple times until the existing new unit data reaches the vocabulary size V=2000, or the maximum number of iterations M=60000 is reached.
[0069] Save the vocabulary table that has met the iteration stop condition, and jointly save the vocabulary table and the atomic regular expression Token_regex, as a SMILES tokenizer, where the atomic regular expression Token_regex is:
[0070] Token_regex = (\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\\\\ / |:|~|@|\?|>|\*|\$|\%[0-9]{2}|[0-9])
[0071] Input the drug molecule SMILES collected in steps 1 and 2 into the SMILES tokenizer for tokenization to obtain a tokenized drug molecule SMILES list.
[0072] Four, token embedding initialization
[0073] OneHot encode the tokenized drug molecule SMILES list to obtain the corresponding OneHot vector;
[0074] Randomly shuffle the OneHot vectors of all drug molecule SMILES lists, and input the OneHot vectors in batches into a Word2vec model based on Skipgram and negative sampling for training to obtain the embedding vector of each token.
[0075] Five, design a convolutional neural network
[0076] The TextCNN model receives the embedding vector matrix corresponding to the tokenized drug molecule SMILES list as input, and the embedding vector matrix is stacked according to the token order of the drug molecule SMILES list from the embedding vector obtained in step 4, denoted as Xi represents the k-dimensional embedding vector of word i . The embedding vector matrix of a drug molecule SMILES list with a length of n can be represented as:
[0077]
[0078] where, represents a concatenation operation, and the final X 1：n is a k×n embedding vector matrix of the drug molecule SMILES list. In the convolution operation, the convolution kernel obtains the corresponding length feature on the embedding vector matrix of the drug molecule SMILES list through a window with a length of h:
[0079] Ci = f(w · X i：i+h-1 + b)
[0080] where C i is the i-th feature extracted by the convolutional layer, w is the weight matrix in the neural network, is the bias term, f is a nonlinear activation function, X i：i+h-1 represents the i-th to i+h rows of the embedded vector matrix of the drug molecule SMILES list;
[0081] During the convolution calculation, this convolution kernel slides in the embedded vector matrix of the drug molecule SMILES list at a fixed length to calculate the corresponding feature map. By setting convolution kernels of different scales, information of different length word groups in the word vector matrix is extracted. The pooling function used by the TextCNN model is the maximum pooling (1-Max). The maximum pooling is to select a maximum feature in the extracted feature vector to replace the entire vector, and then these features are spliced to form a new feature vector.
[0082]
[0083] where represents the maximum feature, c j represents each value in the feature vector.
[0084] Six, drug molecule fingerprint feature generation
[0085] For each drug molecule, the Rdkit toolkit is used to calculate the extended connectivity fingerprints (Extended Connectivity Fingerprints). The radius of the extended connectivity fingerprints is set to 3, and the fingerprint length is 1024 bits. A 1024-bit binary molecular fingerprint vector is obtained as the supervised label signal for subsequent model training.
[0086] Seven, pre-training of convolutional neural network
[0087] The output vector length of the convolutional neural network model in step 5 is set to 1024, i.e. the length of the extended connectivity fingerprints. The last output of the convolutional neural network model is used as the molecular feature vector, and the extended connectivity fingerprints are used as the supervised label. The difference between the two is calculated by the Cosine distance, and the calculation formula is:
[0088] Loss = 1 - Cosine (Ecfp, O)
[0089] where Loss represents the difference between the extended connectivity fingerprints and the model output feature vector, Ecfp represents the 1024-bit extended connectivity molecular fingerprints, and O represents the output of the model.
[0090] The random gradient descent method is used to optimize the parameters of the convolutional neural network model, so that the difference between the extended connection fingerprint and the model output feature vector is gradually reduced until it no longer changes, that is, the model training is completed.
[0091] Eight, through iterative calculation, a pre-trained model is derived
[0092] After the loss of the model almost no longer changes, the model is trained, the random gradient descent method is used to optimize the model parameters, so that the difference between the extended connection fingerprint and the model output feature vector is as small as possible, until it no longer changes, that is, the model training is completed.
[0093] The savemodel function of the deep learning framework is used to save the dictionary of all learnable parameters of the current model and store it as a pth file. In subsequent use, the weights and biases of the model corresponding parameters can be imported.
[0094] Nine, input the molecule into the representation model to obtain the static molecular representation, or jointly train the dynamic representation to adapt to the downstream task
[0095] If the drug representation is needed directly, the drug molecule SMILES is input into the pre-trained model, and the corresponding molecular representation can be obtained. At this time, the length of the molecular representation vector is fixed at 1024 bits;
[0096] If the drug representation is needed for subsequent computer-aided drug design tasks, a feedforward fully connected neural network needs to be connected after the existing molecular representation model for feature transformation. The 1024-bit drug representation vector is transformed to the required dimension, and then the label of the subsequent task is combined for joint learning and training.
Claims
1. A drug representation method based on deep learning, characterized in that, Comprising the following steps: Step 1: Obtain the drug molecule standard SMILES and drug number construction dataset provided in ChEMBL database, wherein 80% of the drugs are randomly selected as the training set, 10% as the validation set, and 10% as the test set; Step 2: Read the drug molecule using the Rdkit toolkit and randomly generate 5 drug molecule SMILES to obtain data-enhanced random drug molecule SMILES; Step 3: Construct a SMILES tokenizer to tokenize the drug molecule SMILES to obtain a list of tokenized drug molecule SMILES; Step 4: Perform OneHot encoding on the drug molecule SMILES list obtained in step 3 and use the Word2Vec model to obtain the embedding vector of the token; Step 5: Construct a deep learning model based on a convolutional neural network to obtain a convolutional neural network model, which receives the embedding vector matrix corresponding to the tokenized drug molecule SMILES list as input and outputs a molecular feature vector; Step 6: Calculate the extended connection fingerprint of each drug molecule; Step 7: Use the extended connection fingerprint as a label to supervise the training of the convolutional neural network model; Step 8: Save all weights and bias parameters of the convolutional neural network model except the output layer and export the pre-trained model; Step 9: Input the drug molecule SMILES into the pre-trained model to obtain a static drug representation, or jointly train a dynamic drug representation to adapt to downstream tasks; The specific steps of step 7 are: Set the output vector length of the convolutional neural network model in step 5 to 1024, which is the length of the extended connection fingerprint; take the last output of the convolutional neural network model as the molecular feature vector, and use the extended connection fingerprint as the supervision label; calculate the difference between the two by Cosine distance, and the formula is: wherein represents the difference between the extended connection fingerprint and the model output molecular feature vector, represents an extended connection fingerprint of 1024 bits, represents the output of the model; Optimize the parameters of the convolutional neural network model using the stochastic gradient descent method to gradually reduce the difference between the extended connection fingerprint and the model output feature vector until the difference no longer changes, i.e., the model training is complete; The specific steps of step 9 are: If you need to directly obtain the drug representation, input the drug molecule SMILES into the pre-trained model to obtain the corresponding molecular representation, and the molecular representation vector length is fixed at 1024 bits; If you need to use the drug representation for subsequent computer-aided drug design tasks, you need to connect a feedforward fully connected neural network after the existing molecular representation model to perform feature transformation, transforming the 1024-bit drug representation vector to the required dimension, and then jointly learning and training with the labels of subsequent tasks.
2. The method of claim 1, wherein the method is based on deep learning. The specific steps of step 2 are: Use the Rdkit toolkit to read each drug molecule standard SMILES into a Mol file, which uniquely represents a drug molecule; randomly initialize the atoms of the node, generate 5 random drug molecule SMILES from the drug molecule Mol file, and obtain data-enhanced random drug molecule SMILES.
3. The method of claim 2, wherein the method is based on deep learning. The specific steps of step 3 are: The drug molecule SMILES collected in steps 1 and 2 is all cut according to characters to obtain a cut character list; The character list is byte pair encoded, and the pair of adjacent characters with the highest occurrence frequency is selected as unit data, i.e. a grapheme, and the original adjacent characters are replaced with the unit data, and the iteration is repeated multiple times until the number of new unit data reaches the size of the vocabulary , or the maximum number of iterations is reached . saving the lexicon that has satisfied the iteration stop condition and the union lexicon and the atomic regular expression together as a SMILES tokenizer, wherein the atomic regular expression is: The drug molecule SMILES collected in steps 1 and 2 is input into a SMILES tokenizer for tokenization to obtain a tokenized drug molecule SMILES list.
4. The method of claim 2, wherein the method is based on deep learning. The specific steps of step 4 are: The tokenized drug molecule SMILES list is OneHot encoded to obtain the corresponding OneHot vector; Randomly shuffle the OneHot vectors of all drug molecule SMILES lists, and send the OneHot vectors in batches into a Word2vec model based on Skipgram and negative sampling for training to obtain the embedding vector of each token.
5. The method of claim 3, wherein the method is based on deep learning. The basic structure of the convolutional neural network model in step 5 adopts a TextCNN model, and the specific design is: The TextCNN model receives an embedding vector matrix corresponding to the tokenized drug molecule SMILES list as input, which is stacked by the embedding vectors obtained in step 4 according to the token order of the drug molecule SMILES list, denoted as represents the dimensional embedding vector, then the embedding vector matrix of a drug molecule SMILES list with a length of is represented as: wherein, denotes a concatenation operation, and the final is embedding vector matrix of the drug molecule SMILES list, in the convolution operation, the convolution kernel acquires the corresponding length of the feature on the embedding vector matrix of the drug molecule SMILES list through a window with a length of in It is the first one extracted from the convolutional layer. One characteristic, It is the weight matrix in a neural network. It is a bias term. It is a non-linear activation function. The first element of the embedding vector matrix representing the list of drug molecule SMILES Arrive at the OK; During convolution calculation, the convolution kernel slides in the embedding vector matrix of the drug molecule SMILES list at a fixed length to obtain the corresponding feature map. By setting convolution kernels of different scales, information of word groups of different lengths in the word vector matrix is extracted; the pooling function used by the TextCNN model is max pooling, which is to select a maximum feature from the extracted feature vector to replace the entire vector, and then concatenate these features to form a new feature vector; wherein represents the maximum feature, represents each value in the feature vector.
6. The method of claim 4, wherein the method is based on deep learning. The specific steps of step 6 are: For each drug molecule, use the Rdkit toolkit to calculate the extended connection fingerprint, set the radius of the extended connection fingerprint to 3, and the fingerprint length to 1024 bits to obtain a 1024-bit binary molecular fingerprint vector.
7. A storage medium, characterized by The processor implements the drug representation method based on deep learning according to any one of claims 1-6 when executing the program in the storage medium.