Drug molecule optimization method based on hidden space diffusion model

Abstract

The invention discloses a drug molecule optimization method based on a hidden space diffusion model. The method comprises the steps of 1, constructing drug molecule data sets for different tasks; 2, constructing and pre-training a molecular SMILES auto-encoder; 4, constructing and training a molecular optimization model based on the hidden space diffusion model; according to the method, new molecules which have relatively high similarity with original molecules, have better attributes and are more diversified can be generated, so that the optimization efficiency of multiple attributes of drug molecules is improved.

Description

Technical Field

[0001] The invention belongs to the fields of molecular optimization, artificial intelligence and bioinformatics, and specifically is a drug molecule optimization method based on a latent space diffusion model. Background Art

[0002] New drug research and development is a long and costly process that usually takes more than 10 years and costs billions of dollars. Discovering lead compounds with new structures, good biological activity and drugability is the basis for innovative drug research and development. Molecular optimization of lead compounds is a key step in new drug research and development. It aims to improve the biological activity, reduce toxicity, enhance chemical stability and improve pharmacokinetic properties of lead compounds through a series of chemical and drug design strategies, while maintaining or increasing molecular features similar to the original lead compounds to ensure the effectiveness and safety of the drugs. Traditional drug molecule design technologies, such as virtual screening and methods based on combinatorial chemistry thinking, are difficult to meet the above needs due to limited chemical space, heavy reliance on chemical expert knowledge and poor synthetic feasibility.

[0003] In recent years, molecular optimization algorithms have made significant progress due to the successful use of deep generative models. This model effectively captures the molecular distribution by learning the probability distribution within the molecular corpus, so that new molecules can be generated by sampling from the established distribution. However, current deep learning methods are prone to overfitting existing data sets, resulting in limited exploration of chemical space and producing molecules with low novelty and diversity. Secondly, these methods often fail to balance the trade-offs between multiple properties. This is exactly the property required for the actual discovery of drug-like compounds, so most of the molecules currently optimized by the model are actually of limited use. In addition, since the grammatical structure of the SMILES representation of drug molecules is very helpful for generating molecules, previous molecular optimization methods based on SMILES representation usually use external attribute predictors to guide the optimization of attributes, which also brings additional costs. In addition, the existing deep learning methods have limited exploration of the molecular chemical space, cannot balance the trade-offs between multiple properties, and it is difficult to generate ideal effective molecules. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a drug molecule optimization method based on a latent space diffusion model, in order to generate new molecules that have a higher similarity to the original molecules and have better properties and more diversity, thereby improving the optimization efficiency of multiple properties of drug molecules.

[0005] The present invention adopts the following technical solutions to solve the technical problems:

[0006] The drug molecule optimization method based on the latent space diffusion model of the present invention is characterized in that it is carried out according to the following steps:

[0007] Step 1: Build drug molecule datasets for different tasks:

[0008] Step 1.1: Build a drug molecule dataset ,in, Indicates drug molecules, N is the total number of drug molecules;

[0009] right Enhanced drug molecule dataset ,in, Indicates the enhanced drug molecules;

[0010] Use the BPE word segmentation algorithm from Extract vocabulary from ,in, Indicates words, Indicates the maximum length of the vocabulary;

[0011] Step 1.2: Construct a paired drug molecule dataset ,in, Represents the first Drug molecules pair, Indicates Lead drug molecules, Indicates target drug molecules, represents the total number of paired drug molecules;

[0012] Step 2: construct a SMILES autoencoder, including an encoder and a decoder, and pre-train the SMILES autoencoder to obtain a SMILES autoencoder with optimal parameters;

[0013] Step 2.1: Build an encoder and use it to extract Latent space feature representation of traditional Chinese medicine molecules and Latent space feature representation of traditional Chinese medicine molecules , thereby constructing the contrast loss function of the encoder , and train the encoder to obtain a pre-trained encoder;

[0014] Step 2.2: Use the pre-trained encoder to Middle Drug molecules Perform feature extraction and generate Optimal latent space feature representation of drug molecules ,in, express No. Dimensional features; Indicates the total number of feature dimensions;

[0015] Step 2.3: Build a decoder and use it to Process and get the reconstructed word vector ;

[0016] Step 2.4: Use formula (2) to construct the total loss function :

[0017] (2)

[0018] In formula (2), represents the logarithmic function, Indicates Drug molecules Center front word vectors, Indicated in Under the conditions The predicted probability of

[0019] Step 2.5: Train the SMILES autoencoder using back propagation and gradient descent, and calculate the total loss function. To update the parameters of the SMILES autoencoder until the total loss function Until convergence, the SMILES autoencoder with optimal parameters is obtained;

[0020] Step 3: construct a Transformer-based latent space diffusion model, including a forward denoising module and a reverse denoising module, and train the latent space diffusion model to obtain a trained latent space diffusion model;

[0021] Step 3.1: Input the SMILES autoencoder with the optimal parameters and process it to get the The latent space feature representation set of drug molecule pairs ,in, express No. dimensional features, and , Indicates Lead drug molecules The latent space feature representation is Indicates Target drug molecules Latent space feature representation of ;

[0022] Step 3.2: The forward noise adding module uses formula (3) to Gradually add noise to get In the Latent space feature representation at time steps , thus obtaining In the The latent space feature representation at time steps is used as the noise that conforms to the Gaussian distribution. :

[0023] (3)

[0024] In formula (3), Indicates that from time step to During the time step, The proportion of data retained, and , Indicates The noise intensity corresponding to the time step is express The latent space feature representation at the initial time step, represents the sampling process of forward noise addition, represents sampling from a standard normal distribution Noise of the same shape; To show obedience;

[0025] Step 3.3: Reverse denoising module time steps and Process and obtain Predicted latent space feature representation at the initial time step ;

[0026] Step 4: Use formula (7) to construct the loss function of the denoising network :

[0027] (7)

[0028] Step 5: Based on , use the gradient descent method to train the latent space diffusion model and calculate the total loss The model parameters are updated until the total loss converges or the maximum number of iterations is reached, and the training is stopped to obtain the optimal latent space diffusion model for generating new target drug molecules;

[0029] Step 6: Generate new molecules using the trained latent space diffusion model;

[0030] Enter a new drug molecule to be optimized , using the trained SMILES encoder Encode and get Latent space feature representation of d , and then randomly generate a Gaussian noise of the same shape , and and Perform splicing to obtain the latent space feature representation after splicing Input the optimal latent space diffusion model to obtain the optimized latent space feature representation Finally, the SMILES decoder with the optimal parameters is used to represent the latent space features Convert into target molecule sequence and output.

[0031] The drug molecule optimization method based on latent space diffusion model described in the present invention is also characterized in that the encoder in step 2.1 includes: a word segmenter, a Bert pre-training model and a linear projection module;

[0032] Step 2.1.1, the word segmenter is used to segment the word according to the preset maximum length L. and Perform word segmentation and sort according to the vocabulary Number the characters after word segmentation to get The encoding mark corresponding to each character in and The encoding mark corresponding to each character in ;in, express The encoding mark corresponding to the lth character in express The encoding mark corresponding to the lth character in;

[0033] Step 2.1.2: and They are respectively input into the Bert pre-training model for processing to obtain The initial latent feature vector and ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions;

[0034] Step 2.1.3: and They are respectively input into the linear projection module for dimensionality reduction and normalization, and we get Latent space feature representation and Latent space feature representation ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions;

[0035] Step 2.1.4: Use formula (1) to construct the encoder's contrast loss function ;

[0036] (1)

[0037] In formula (1), represents the temperature parameter; Indicates the enhanced Drug molecules The latent space features of

[0038] Step 2.1.5: Use the SimCLR method to train the encoder and calculate the contrast loss function To update the encoder parameters until the contrast loss function Until convergence, the trained encoder is obtained.

[0039] Furthermore, the decoder in step 2.3 includes: a linear compression module, a multi-head self-attention layer and a feedforward fully connected layer, and a linear output module;

[0040] Step 2.3.1, the linear compression module includes: a fully connected layer and a normalization layer;

[0041] Will After being input into the fully connected layer for dimensionality reduction, it is then input into the normalization layer for processing to obtain Low-dimensional latent space feature representation ,in, express No. Dimensional features, express The total number of feature dimensions of ;

[0042] Step 2.3.2: After being processed by the multi-head self-attention layer and the feed-forward fully connected layer, we get The output vector ;in, express No. Dimensional features;

[0043] Step 2.3.3: The linear output module contains a linear layer and uses autoregression to transform Mapped into dimensions The one-dimensional fractional vector of The reconstructed word vector ={ },in, express The word vectors, express The length of the word vector.

[0044] Furthermore, the reverse denoising module includes a time encoding block, a position encoding block and a Transformer-based denoising network;

[0045] Step 3.3.1: The time coding block pair time steps to embed and get the embedded time vector ={ },in, Indicates The time vector after the time step embedding, and ={ }, express Middle The temporal characteristics of the dimension;

[0046] Step 3.3.2, set the maximum character number to , and Number the encoding mark corresponding to each character in the Location ID ;

[0047] The position encoding block is Encode and get The processed position embedding vector ={ },in, Represents the position embedding vector No. The location characteristics of the dimension;

[0048] Step 3.3.3, the denoising network comprises a self-attention layer and a feedforward network neural network;

[0049] Will With the embedded time vector , position embedding vector After merging, we get The combined latent space representation , and then input into the self-attention layer, so as to obtain through equations (4)-(6) In the The context embedding vector at time steps ;

[0050] (4)

[0051] (5)

[0052] (6)

[0053] In formula (4) to formula (6), express In the The attention score after adding noise at time steps, express In the The attention weight at time steps is is the attention mechanism, is the normalized exponential function, N is the layer normalization process; is the residual connection;

[0054] Will Input into the feedforward neural network for processing, and get In the The predicted latent space feature representation at time steps , thus obtaining Predicted latent space feature representation at the initial time step .

[0055] The computer device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the molecular optimization method based on the latent space diffusion model, and the processor is configured to execute the program stored in the memory.

[0056] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program is suitable for executing the molecular optimization method based on the latent space diffusion model when being run by a processor.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The present invention is based on a latent space diffusion model for generating new molecules. The model uses the chemical information of the lead molecule to guide the generation of molecules with target properties on the continuous latent space representation of drug molecules, while maintaining a high similarity. The present invention performs well in the final generation effect, with high molecular effectiveness and optimization success rate.

[0059] 2. The present invention is based on the diffusion model and combines the pre-trained encoder-decoder with the diffusion model, which has better effects than the previous deep generative model. By partially adding random noise in the forward diffusion process to disturb the distribution of the molecular data to be optimized, while keeping the latent space of the guiding lead molecule unchanged, and using a neural network with parameters to learn the reverse diffusion process to restore the distribution of the target data, it is efficient, controllable and easy to calculate.

[0060] 3. The present invention can realize end-to-end training and generation tasks. Compared with traditional algorithms, this end-to-end structure not only simplifies the training process of drug molecule optimization, but also improves the overall performance of molecular optimization, thereby making the entire drug development more efficient and providing new support for various complex generation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the process of the present invention;

[0062] Figure 2 A schematic diagram of a pre-trained encoder of the present invention;

[0063] Figure 3 A schematic diagram of a pre-trained SMILES autoencoder of the present invention;

[0064] Figure 4 It is a schematic diagram of the molecular optimization framework of the latent space diffusion model of the present invention; DETAILED DESCRIPTION

[0065] In this embodiment, Figure 1 As shown in Figure 1, a molecular optimization method based on latent space diffusion model is carried out in the following steps:

[0066] Step 1: Construct drug molecule data sets for different tasks. In this embodiment, all data sets are from the ZINC database;

[0067] Step 1.1: Obtain drug molecule dataset ,in, Indicates drug molecules, is the total number of drug molecules;

[0068] Taking advantage of the fact that a single drug molecule can be represented by multiple different SMILES sequences, Enhanced drug molecule dataset ,in, Indicates the enhanced drug molecules;

[0069] Use the BPE word segmentation algorithm from Extract vocabulary from ,in, Indicates words, Indicates the maximum length of the vocabulary;

[0070] Step 1.2: Construct a paired drug molecule dataset ,in, Represents the first Drug molecules pair, Indicates Lead drug molecules, Indicates target drug molecules, Represents the total number of paired drug molecules.

[0071] Step 2: construct a SMILES autoencoder, including an encoder and a decoder, and pre-train the SMILES autoencoder to obtain a SMILES autoencoder with optimal parameters;

[0072] Step 2.1: Figure 2 As shown, the encoder is constructed, including: word segmenter, Bert pre-training model and linear projection module, and is used to extract Latent space feature representation of traditional Chinese medicine molecules and Latent space feature representation of traditional Chinese medicine molecules , thereby constructing the encoder's contrast loss function , and train the encoder to obtain the trained encoder;

[0073] Step 2.1.1, the word segmenter is used to and Perform word segmentation and fill or truncate the character sequence according to the preset maximum length L to obtain a number of characters. Specifically: set the maximum length of each character sequence, and truncate the following characters if the length exceeds this range; if the sequence length is insufficient, use the special symbol [PAD] to fill it to make the length reach the maximum value; and number the characters after word segmentation according to the vocabulary to obtain The encoding mark corresponding to each character in and The encoding mark corresponding to each character in ;in, express The encoding mark corresponding to the lth character in express The encoding mark corresponding to the lth character in .

[0074] Step 2.1.2: and Input them into the Bert pre-training model for processing, and get The initial latent feature vector contains rich contextual information and ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions.

[0075] Step 2.1.3: and They are respectively input into the linear projection module for dimensionality reduction and normalization, and we get Latent space feature representation and Latent space feature representation ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions.

[0076] Step 2.1.4: Use formula (1) to construct the encoder's contrast loss function ;

[0077] (1)

[0078] In formula (1), represents the temperature parameter; Indicates the enhanced Drug molecules The latent space features of

[0079] Step 2.1.5: Use the SimCLR method to train the encoder and calculate the contrast loss function To update the encoder parameters until the contrast loss function The pre-trained encoder is obtained until convergence, so that the encoder learns the syntax and information in the SMILES representation of drug molecules.

[0080] Step 2.2: Use the pre-trained encoder to Middle Drug molecules Perform feature extraction and generate Optimal latent space feature representation of drug molecules ,in, express No. Dimensional features; Indicates the total number of feature dimensions.

[0081] Step 2.3: Figure 3 As shown in Figure 2, a decoder is constructed, including a linear compression module, a multi-head self-attention layer, a feedforward fully connected layer, a linear output module, and Process and get the reconstructed word vector ;

[0082] Step 2.3.1, the linear compression module includes: a fully connected layer and a normalization layer;

[0083] In order to prevent overfitting and make it easier for the diffusion model to learn the feature space of molecules, a fully connected layer is used to transform Perform dimensionality reduction and then input it into the normalization layer for processing to obtain Low-dimensional latent space feature representation ,in, express No. Dimensional features, express The total number of feature dimensions of .

[0084] Step 2.3.2: After being processed by the multi-head self-attention layer and the feed-forward fully connected layer, we get The output vector ;in, express No. Dimensional features;

[0085] Step 2.3.3: The linear output module contains a linear layer and uses autoregression to transform Mapped into dimensions The one-dimensional fractional vector of The reconstructed word vector ={ },in, express The word vectors, express The length of the word vector.

[0086] Step 2.4: Use formula (2) to construct the total loss function :

[0087] (2)

[0088] In formula (2), represents the logarithmic function, Indicates Drug molecules Center front word vectors, Indicated in Under the conditions The predicted probability of .

[0089] Step 2.5: Train the SMILES autoencoder using backpropagation and gradient descent, and calculate the total loss function To update the parameters of the SMILES autoencoder until the total loss function Until convergence, the SMILES autoencoder with optimal parameters is obtained.

[0090] Step 3: Figure 4 As shown, a Transformer-based latent space diffusion model is constructed, including: a forward denoising module and a reverse denoising module, and the latent space diffusion model is trained to obtain a trained latent space diffusion model;

[0091] Step 3.1: Input the SMILES autoencoder with the optimal parameters and process it to get the The latent space feature representation set of drug molecule pairs ,in, express No. dimensional features, and , Indicates Lead drug molecules The latent space feature representation is Indicates Target drug molecules The latent space feature representation of .

[0092] Step 3.2: The forward noise module uses formula (3) to Gradually add noise to get In the Latent space feature representation at time steps , thus obtaining In the The latent space feature representation at time steps is the noise that conforms to the Gaussian distribution. :

[0093] (3)

[0094] In formula (3), Indicates that from time step to During the time step, The proportion of data retained, and , Indicates The noise intensity corresponding to the time step is express The latent space feature representation at the initial time step, represents the sampling process of forward noise addition, represents sampling from a standard normal distribution Noise of the same shape; To express obedience.

[0095] Step 3.3, the reverse denoising module includes a time encoding block, a position encoding block and a Transformer-based denoising network;

[0096] Step 3.3.1: Time coding block pair time steps to embed and get the embedded time vector ={ },in, Indicates The time vector after the time step embedding, and ={ }, express Middle The temporal characteristics of the dimension;

[0097] Step 3.3.2, set the maximum character number to , and Number the encoding mark corresponding to each character in the Location ID ;

[0098] Position encoding block pair Encode and get The processed position embedding vector ={ },in, Represents the position embedding vector No. The location characteristics of the dimension.

[0099] Step 3.3.3, denoising network, including self-attention layer and feedforward network neural network;

[0100] The self-attention mechanism is introduced so that the denoising network can take into account the information of other positions in the entire input sequence when processing the input of each position, and the characters in the SMILES representation of drug molecules have long-range dependencies. In the denoising process, the self-attention mechanism can capture these dependencies and restore the original molecules more accurately.

[0101] Will With the embedded time vector , position embedding vector After merging, we get The combined latent space representation ;

[0102] Then input it into the self-attention layer, and then get it through equations (4)-(6): In the The context embedding vector at time steps ;

[0103] (4)

[0104] (5)

[0105] (6)

[0106] In formula (4) to formula (6), express In the The attention score after adding noise at time steps, express In the The attention weight at time steps is is the attention mechanism, is the normalized exponential function, N is the layer normalization process; is the residual connection;

[0107] Will Input into the feedforward neural network for processing, and get In the The predicted latent space feature representation at time steps , thus obtaining Predicted latent space feature representation at the initial time step ;

[0108] Step 4: Use formula (7) to construct the loss function of the denoising network :

[0109] (7)

[0110] Step 5: Based on , use the gradient descent method to train the latent space diffusion model and calculate the total loss The model parameters are updated until the total loss converges or the maximum number of iterations is reached, and the training is stopped to obtain the optimal latent space diffusion model for generating new target drug molecules.

[0111] Step 6: Generate new molecules using the trained latent space diffusion model;

[0112] Enter a new drug molecule to be optimized , using the trained SMILES encoder Encode and obtain its latent space feature representation , and then randomly generate a Gaussian noise of the same shape , and and Perform splicing to obtain the latent space feature representation after splicing , and then iteratively input into the optimal latent space diffusion model to obtain the optimized latent space feature representation Finally, the SMILES decoder with the optimal parameters is used to represent the latent space features Converted into the target molecular sequence and finally output.

[0113] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0114] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A drug molecule optimization method based on latent space diffusion model, characterized in that Follow these steps: Step 1: Build drug molecule datasets for different tasks: Step 1.1: Build a drug molecule dataset ,in, Indicates drug molecules, N is the total number of drug molecules; right Enhanced drug molecule dataset ,in, Indicates the enhanced drug molecules; Use the BPE word segmentation algorithm from Extract vocabulary from ,in, Indicates words, Indicates the maximum length of the vocabulary; Step 1.2: Construct a paired drug molecule dataset ,in, Represents the first Drug molecules pair, Indicates Lead drug molecules, Indicates target drug molecules, represents the total number of paired drug molecules; Step 2: construct a SMILES autoencoder, including an encoder and a decoder, and pre-train the SMILES autoencoder to obtain a SMILES autoencoder with optimal parameters; Step 2.1: Build an encoder and use it to extract Latent space feature representation of traditional Chinese medicine molecules and Latent space feature representation of traditional Chinese medicine molecules , thereby constructing the contrast loss function of the encoder , and train the encoder to obtain a pre-trained encoder; Step 2.2: Use the pre-trained encoder to Middle Drug molecules Perform feature extraction and generate Optimal latent space feature representation of drug molecules ,in, express No. Dimensional features; Indicates the total number of feature dimensions; Step 2.3: Build a decoder and use it to Process and get the reconstructed word vector ; Step 2.4: Use formula (2) to construct the total loss function : (2) In formula (2), represents the logarithmic function, Indicates Drug molecules Center front word vectors, Indicated in Under the conditions The predicted probability of Step 2.5: Train the SMILES autoencoder using back propagation and gradient descent, and calculate the total loss function. To update the parameters of the SMILES autoencoder until the total loss function Until convergence, the SMILES autoencoder with optimal parameters is obtained; Step 3: construct a Transformer-based latent space diffusion model, including a forward denoising module and a reverse denoising module, and train the latent space diffusion model to obtain a trained latent space diffusion model; Step 3.1: Input the SMILES autoencoder with the optimal parameters and process it to get the The latent space feature representation set of drug molecule pairs ,in, express No. dimensional features, and , Indicates Lead drug molecules The latent space feature representation is Indicates Target drug molecules Latent space feature representation of ; Step 3.2: The forward noise adding module uses formula (3) to Gradually add noise to get In the Latent space feature representation at time steps , thus obtaining In the The latent space feature representation at time steps is used as the noise that conforms to the Gaussian distribution. : (3) In formula (3), Indicates that from time step to During the time step, The proportion of data retained, and , Indicates The noise intensity corresponding to the time step is express The latent space feature representation at the initial time step, represents the sampling process of forward noise addition, represents sampling from a standard normal distribution Noise of the same shape; To show obedience; Step 3.3: The reverse denoising module time steps and Process and obtain Predicted latent space feature representation at the initial time step ; Step 4: Use formula (7) to construct the loss function of the denoising network : (7) Step 5: Based on , use the gradient descent method to train the latent space diffusion model and calculate the total loss The model parameters are updated until the total loss converges or the maximum number of iterations is reached, and the training is stopped to obtain the optimal latent space diffusion model for generating new target drug molecules; Step 6: Generate new molecules using the trained latent space diffusion model; Enter a new drug molecule to be optimized , using the trained SMILES encoder Encode and get Latent space feature representation of d , and then randomly generate a Gaussian noise of the same shape , and and Perform splicing to obtain the latent space feature representation after splicing Input the optimal latent space diffusion model to obtain the optimized latent space feature representation Finally, the SMILES decoder with the optimal parameters is used to represent the latent space features Convert into target molecule sequence and output.

2. A drug molecule optimization method based on latent space diffusion model according to claim 1, characterized in that: The encoder in step 2.1 includes: a word segmenter, a Bert pre-training model and a linear projection module; Step 2.1.1, the word segmenter is used to segment the word according to the preset maximum length L. and Perform word segmentation and sort according to the vocabulary Number the characters after word segmentation to get The encoding mark corresponding to each character in and The encoding mark corresponding to each character in ;in, express The encoding mark corresponding to the lth character in express The encoding mark corresponding to the lth character in; Step 2.1.2: and They are respectively input into the Bert pre-training model for processing to obtain The initial latent feature vector and ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions; Step 2.1.3: and They are respectively input into the linear projection module for dimensionality reduction and normalization, and we get Latent space feature representation and Latent space feature representation ,in, and Respectively and No. Dimensional features; Indicates the total number of feature dimensions; Step 2.1.4: Use formula (1) to construct the encoder's contrast loss function ; (1) In formula (1), represents the temperature parameter; Indicates the enhanced Drug molecules The latent space features of Step 2.1.5: Use the SimCLR method to train the encoder and calculate the contrast loss function To update the encoder parameters until the contrast loss function Until convergence, the trained encoder is obtained.

3. A drug molecule optimization method based on latent space diffusion model according to claim 2, characterized in that: The decoder in step 2.3 includes: a linear compression module, a multi-head self-attention layer and a feedforward fully connected layer, and a linear output module; Step 2.3.1, the linear compression module includes: a fully connected layer and a normalization layer; Will After being input into the fully connected layer for dimensionality reduction, it is then input into the normalization layer for processing to obtain Low-dimensional latent space feature representation ,in, express No. Dimensional features, express The total number of feature dimensions of ; Step 2.3.2: After being processed by the multi-head self-attention layer and the feed-forward fully connected layer, we get The output vector ;in, express No. Dimensional features; Step 2.3.3: The linear output module contains a linear layer and uses autoregression to transform Mapped into dimensions The one-dimensional fractional vector of The reconstructed word vector ={ },in, express The word vectors, express The length of the word vector.

4. A drug molecule optimization method based on latent space diffusion model according to claim 3, characterized in that: The reverse denoising module in step 3.3 includes: a time encoding block, a position encoding block and a Transformer-based denoising network; Step 3.3.1: The time coding block pair time steps to embed and get the embedded time vector ={ },in, Indicates The time vector after the time step embedding, and ={ }, express Middle The temporal characteristics of the dimension; Step 3.3.2, set the maximum character number to , and Number the encoding mark corresponding to each character in the Location ID ; The position encoding block is Encode and get The processed position embedding vector ={ },in, Represents the position embedding vector No. The location characteristics of the dimension; Step 3.3.3, the denoising network comprises a self-attention layer and a feedforward network neural network; Will With the embedded time vector , position embedding vector After merging, we get The combined latent space representation , and then input into the self-attention layer, so as to obtain through equations (4)-(6) In the The context embedding vector at time steps ; (4) (5) (6) In formula (4) to formula (6), express In the The attention score after adding noise at time steps, express In the The attention weight at time steps is is the attention mechanism, is the normalized exponential function, N is the layer normalization process; is the residual connection; Will Input into the feedforward neural network for processing, and get In the The predicted latent space feature representation at time steps , thus obtaining Predicted latent space feature representation at the initial time step .

5. A computer device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the molecular optimization method based on the latent space diffusion model as described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is suitable for executing a molecular optimization method based on a latent space diffusion model as described in any one of claims 1 to 4 when executed by a processor.

Images