A covalent drug de novo design method and related device based on enhanced implicit diffusion model

Abstract

The present invention provides a covalent drug de novo design method and related devices based on an enhanced latent diffusion model, comprising constructing a covalent drug de novo design network and training it to obtain a covalent drug de novo design network model; selecting a protein as a target, sampling a latent vector from a standard normal distribution, and inputting the protein sequence and the latent vector into the covalent drug de novo design network model to obtain a drug molecule that can covalently bind to the selected protein; and using the architecture of the latent diffusion model to implement ligand molecule design to generate covalent drug ligand molecules that can bind to a given protein target and have desired properties, while ensuring that the generated molecules have excellent binding ability to the given protein target and also have excellent drug properties including synthesizability, optimizing the acute toxicity of the generated molecules, improving safety, broadening the dosage design scale of the designed drug molecules, and enhancing their ability to be further developed into marketed drugs.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug molecule design, and in particular to a covalent drug de novo design method based on an enhanced implicit diffusion model and a related device. Background Art

[0002] Most traditional drug molecules are non-covalent. Their interactions with receptors are typically non-covalent, such as hydrogen bonds and dipole interactions, resulting in non-persistent binding to protein pockets. In recent years, computational drug design methods have primarily focused on the design of non-covalent drug molecules. Their goal is to design small ligands that bind to a given protein target. A variety of deep generative models, such as variational autoencoders, generative adversarial networks, and flow models, are used to generate molecular SMILES (simile sequences), molecular graphs, or three-dimensional conformations. These methods aim to learn the distribution of real molecules and sample novel and effective chemical molecules from this distribution. They can also incorporate protein target information as a condition to generate molecules with non-covalent binding capabilities.

[0003] Compared to non-covalent drugs, covalent drugs react with widely existing nucleophilic groups on receptors, such as sulfhydryl, hydroxyl, amino, and imidazole groups, through their electrophilic groups to produce covalent bonds, resulting in very high potency and ligand efficiency, more efficient inhibition, and prolonged pharmacodynamic effects. Due to the characteristics of this covalent bond-linked, long-lasting inhibition, covalent drugs often have relatively higher toxicity. Drug safety issues are one of the important reasons for failure in drug development. Drug molecules with lower toxicity can improve safety while also providing more flexible drug dosage options.

[0004] There are two existing non-covalent design methods. The first, Pocket2Mol, learns the distribution of atoms in the protein pocket environment and the spatial positions of ligand molecules, gradually sampling and generating new molecules from this distribution. The second, DiffSBDD, proposes an autoregressive three-dimensional molecule generation method that models the probability of atomic occurrence within the three-dimensional space of the protein pocket and autoregressively samples atoms to generate the final molecule. Both methods generate ligand molecules based on protein targets, but the algorithms in these methods are not suitable for designing covalent drug molecules and do not fully consider the molecular properties of the generated molecules, including synthesizability and toxicity.

[0005] In summary, the limitations of existing technologies are mainly as follows:

[0006] (1) Most current methods are aimed at designing drug molecules that bind non-covalently and reversibly, and it is impossible to design molecules that covalently bind to a given target from scratch.

[0007] (2) The current algorithm does not fully consider the synthesizability and drug-like properties of the molecules, resulting in the generated molecules being unreasonable or difficult to synthesize;

[0008] (3) Most current methods do not take the safety of the designed molecules into consideration and are unable to optimize the safety of the designed molecules. Summary of the Invention

[0009] In order to solve the problems existing in the prior art, the present invention provides a covalent drug de novo design method and related devices based on an enhanced implicit diffusion model, which adopts an implicit diffusion model architecture to realize the design of ligand molecules and generate covalent drug ligand molecules with desired properties that can bind to a given protein target.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for de novo design of covalent drugs based on an enhanced implicit diffusion model, comprising the following specific steps:

[0011] S1 constructs a covalent drug de novo design network and trains it to obtain a covalent drug de novo design network model;

[0012] S2 selects a protein as a target, samples a latent vector from a standard normal distribution, and inputs the protein sequence and latent vector into a covalent drug de novo design network model to obtain drug molecules that can covalently bind to the selected protein;

[0013] The covalent drug de novo design network includes a variational autoencoder, a protein sequence encoder, and a latent diffusion model. The variational autoencoder is used to encode molecular SMILES into latent vectors and decode the latent vectors into drug molecules. The protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding.

[0014] The implicit diffusion model includes the forward diffusion module, the noise prediction network ε θ and classifier models, the forward diffusion module is used to gradually add noise to the latent vector, and the noise prediction network ε θ It is used to fuse the latent vector and protein target information, and perform reverse denoising under the guidance of the classifier model to predict the latent vector of the drug molecule covalently bound to the protein target;

[0015] The toxicity prediction model is used to score the acute toxicity of the drug molecules covalently bound to the protein target. The implicit diffusion model is updated according to the scoring results to obtain drug molecules covalently bound to the protein target with low acute toxicity.

[0016] Furthermore, in S1, the variational autoencoder includes an encoder that encodes the molecular SMILES into a latent vector and a decoder that decodes the latent vector into a drug molecule. Both the encoder and the decoder use multi-layer gated recurrent units to achieve the mapping between characters and continuous latent vectors. The encoder uses the output of the final state of the GRU as the continuous latent vector representation of the molecule.

[0017] The training objective of the variational autoencoder is a combination of the reconstruction error between the original molecule and the decoded molecule and the KL divergence error that constrains the latent space distribution.

[0018] Furthermore, in S1, the protein sequence encoder uses the ESM-2 model to encode the target protein sequence.

[0019] Furthermore, in S1, the forward diffusion module adds Gaussian noise to the latent vector according to the noise strategy β at each time step to obtain the noisy latent vector, which is specifically expressed as follows:

[0020]

[0021] Among them, q(x t |x t-1 ) is a step in the forward denoising process, which is to add noise to the data at time t-1 to make it the data at time t.

[0022] Furthermore, in S1, the noise prediction network is specifically a fully connected neural network with a multi-head self-attention mechanism. The multi-head self-attention mechanism is used to fuse the noisy molecular latent vector and the protein target sequence encoding. The fused vector data is input into the fully connected neural network, and noise prediction is performed in the fully connected neural network, which outputs the prediction of the noise added to the latent vector.

[0023] The training objective is to minimize the L2 norm between the actual noise and the predicted noise.

[0024] Furthermore, in S1, the classifier model is a binary classifier, which is used to predict whether the noisy latent vector has the corresponding molecular structure, and uses the prediction result as a gradient to guide the noise prediction network to generate drug molecules that can be covalently bound to the selected protein. The training goal is to minimize the predicted label The binary cross entropy loss between y and the true label y.

[0025] Furthermore, in S1, a random forest-based regression prediction model is used as the toxicity prediction model. The toxicity prediction model accepts the molecular SMILES representation and gives a prediction of the acute toxicity median lethal dose of the molecule; the input of the toxicity prediction model is the molecular access system fingerprint, and the training goal is to minimize the mean square error between the median lethal dose value predicted by the toxicity prediction model and the true value.

[0026] Furthermore, in S1, the toxicity prediction model is used to predict the median lethal dose of drug molecules obtained by the implicit diffusion model, and the prediction is used as the reward for each step in the reverse process of T time steps of a sampling process of the noise prediction network in the implicit diffusion model. In each step, the gradient with the reward as the weight is calculated and accumulated as the gradient of the total reward with respect to the strategy parameters. The gradient is used to update the implicit diffusion model to obtain an implicit diffusion model that generates drug molecules with lower acute toxicity.

[0027] The present invention also provides a covalent drug de novo design system based on an enhanced implicit diffusion model, comprising:

[0028] A network construction module is used to construct a covalent drug de novo design network and train it to obtain a covalent drug de novo design network model;

[0029] The covalent drug design module is used to select a protein as a target, sample a latent vector from a standard normal distribution, and input the protein sequence and latent vector into the covalent drug de novo design network model to obtain drug molecules that can covalently bind to the selected protein;

[0030] The covalent drug de novo design network includes a variational autoencoder, a protein sequence encoder, and a latent diffusion model. The variational autoencoder is used to encode molecular SMILES into latent vectors and decode the latent vectors into drug molecules. The protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding.

[0031] The implicit diffusion model includes the forward diffusion module, the noise prediction network ε θ and classifier models, the forward diffusion module is used to gradually add noise to the latent vector, and the noise prediction network ε θ It is used to fuse the latent vector and protein target information, and perform reverse denoising under the guidance of the classifier model to predict the latent vector of the drug molecule covalently bound to the protein target;

[0032] The toxicity prediction model is used to score the acute toxicity of the drug molecules covalently bound to the protein target. The implicit diffusion model is updated according to the scoring results to obtain drug molecules covalently bound to the protein target with low acute toxicity.

[0033] The present invention also provides a terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for de novo design of covalent drugs based on an enhanced implicit diffusion model are implemented.

[0034] Compared with the prior art, the present invention has at least the following beneficial effects:

[0035] The present invention provides a method for de novo design of covalent drugs based on a reinforced latent diffusion model. A de novo design network of covalent drugs is used to realize de novo design of molecules that covalently bind to a given target. The latent diffusion model can capture the complexity of the molecular structure, while the classifier can accurately identify and strengthen the required substructure. The latent diffusion model is used for molecule generation, and the classifier is used for gradient guidance to ensure that the generated molecules have the required covalent binding substructure, so that ligand molecules with the ability to covalently bind to the protein target can be designed de novo for the protein target. The present invention introduces a multi-head self-attention mechanism into the noise prediction network, integrates the molecular latent vector and the protein target sequence encoding, learns the relationship between the protein target and the molecular latent vector, and helps to generate a latent vector that binds tightly to the given target, thereby generating a ligand molecule with high affinity. The present invention combines the policy gradient in reinforcement learning and uses the feedback of the toxicity prediction model as a reward to fine-tune the latent diffusion model, so that the generated molecules have lower acute toxicity.

[0036] The present invention uses the policy gradient method to fine-tune the implicit diffusion model. By calculating the gradient with the toxicity prediction score as the weight and accumulating it as the gradient of the total reward with respect to the policy parameters, the implicit diffusion model is updated. This method can directly optimize the target drug properties (such as acute toxicity), so that the drug molecules generated by the model are more in line with actual requirements.

[0037] The covalent drug de novo design method of the present invention utilizes an implicit diffusion model to rapidly generate a large number of potential drug molecules. Through the dual optimization of a classifier and reinforcement learning, molecules with excellent drug properties can be rapidly screened, greatly improving the accuracy of drug design. In addition, it can clearly understand which molecular structural features are critical for binding to specific protein targets and drug properties, which helps to improve the interpretability and reliability of drug design. In summary, while ensuring that the generated molecules have excellent binding ability to a given protein target, they also ensure that they have excellent drug properties including synthesizability, optimize the acute toxicity of the generated molecules, improve safety, broaden the dosage design scale of the designed drug molecules, and enhance their ability to be further developed into marketed drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 .LDDD framework diagram of the covalent drug de novo design network model of the present invention.

[0039] Figure 2 The classifier guides the generation of molecules with different substructures. In the figure, blue represents molecules with substructures, and orange represents molecules without substructures.

[0040] Figure 3 .Optimize the acute toxicity of the generated molecules, Figure 3a is the LD50 (median lethal dose) value of the generated molecule predicted by the toxicity prediction model. This value is converted into units, and the lower the value, the lower the toxicity of the compound. Figure 3 b is the number of molecules containing various acute toxicity warning structures among the generated molecules.

[0041] Figure 4 .Design of covalent drugs targeting EGFR T790M, Figure 4 a is a co-crystal display of the protein covalently bound to a marketed drug, Neratinib. Figure 4 b is the structural formula of neratinib. The blue part is its covalent warhead acrylamide substructure, which is covalently bonded to the sulfur atom on the thiol group of cysteine ​​797 on the protein in Figure a. The table shows the pharmacological properties of neratinib. Figure 4 The horizontal axis of c is the known covalent inhibitors of EGFR, and the vertical axis is the covalent drug molecules designed by the present invention. The numbers in the table are the similarities between them, and the molecular pairs with higher similarity are displayed in boxes. Figure 4 d is the display of the structure, covalent / non-covalent docking pose, and pharmacological properties of the generated molecule; Figure 4 e is a quantitative comparison of the covalent / non-covalent docking postures of the generated molecule, a prior art solution, and the non-covalent molecule generated by the present invention after manual addition of an acrylamide warhead; Figure 4 f is the comparison of the predicted half-inhibitory concentration (IC50) values ​​of the three molecules mentioned above. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] like Figure 1 As shown, the present invention provides a covalent drug de novo design method based on a reinforced latent diffusion model. The covalent drug de novo design network model LDDD can be used to generate targeting ligand molecules from scratch. LDDD maps molecules between the molecular SMILES representation space and the latent space through encoding and decoding of a variational autoencoder. The protein target sequence is encoded by the ESM-2 encoder, and conditional information is introduced into the noise prediction network. The reverse sampling process is guided by a classifier to generate molecules with specified substructures. Through reinforcement learning, samples are sampled from the reverse sampling generation process, and the model is fine-tuned using policy gradients to generate molecules with lower toxicity. The specific steps are as follows:

[0044] 1. Dataset processing

[0045] This invention involves four different datasets, including: small and medium molecule datasets and molecular datasets with and without selected substructures in the ZINC database, CrossDocked2020 dataset, and mouse oral LD50 molecular dataset in the TOXRIC database.

[0046] Specifically, the small molecule data set in the ZINC database selected molecules that are "on the market" and have a lipid-water partition coefficient less than 5 and greater than -1, and an atomic weight greater than 250 and less than 500 in the ZINC database, and randomly selected 1.2 million small molecules from them to form a training set.

[0047] First, filter out the binding conformations in the CrossDocked2020 dataset with RMSD greater than The data was then clustered using the mmseqs2 method at 30% sequence identity, and 100,000 protein-ligand pairs were randomly selected to form the training dataset, with the remaining 100 proteins used as the test set.

[0048] The ZINC database was screened for molecules with and without the selected substructures as positive and negative samples to construct the training set. For each substructure, the dataset size was the same, with 40,000 molecules as positive samples and 40,000 molecules as negative samples.

[0049] The mouse oral LD50 molecular dataset contains 35,299 molecules and their LD50 data.

[0050] 2. Constructing a covalent drug de novo design network

[0051] The covalent drug de novo design network includes a variational autoencoder model, a protein sequence encoder, and a latent diffusion model. The variational autoencoder model includes an encoder that encodes molecular SMILES into a latent vector and a decoder that decodes the latent vector back into the molecule. The latent diffusion model is used to generate new ligand molecules, as follows:

[0052] (1) Variational Autoencoder

[0053] The present invention encodes molecular SMILES into latent vector representation of the molecule based on a variational autoencoder.

[0054] like Figure 1 As shown in the molecular space part, the variational autoencoder is divided into two parts: encoder and decoder. The encoder receives the SMILES string, encodes the SMILES string into a continuous vector, and then transmits it to the latent diffusion model;

[0055] The SMILES string is a SMILES string close to a single character segmented by the word segmenter, excluding common character groups such as chlorine Cl and bromine Br.

[0056] The decoder is used to receive the latent vector output by the latent diffusion model and decode the latent vector into a SMILES representation of the molecule.

[0057] Among them, both the encoder and the decoder use multi-layer gated recurrent units (GRU) to realize the mapping between characters and continuous latent vectors; the encoder uses the output of the GRU final state as the continuous latent vector representation of the molecule.

[0058] (2) Protein sequence encoder

[0059] like Figure 1 As shown, the protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding. For the encoding of protein information, the present invention adopts the ESM-2 model to encode the target protein sequence. In order to improve the quality of protein sequence feature extraction, the weights of the pre-trained ESM-2 model are used here. The ESM-fold model based on ESM-2 accurately realizes the end-to-end prediction of protein sequence to structure. This pre-trained model has high reliability. In actual use, due to the large number of model parameters and the excessive computing resources required, the present invention adopts a strategy of pre-encoding and then storing and reading.

[0060] (3) Implicit Diffusion Model

[0061] The present invention realizes the generation of new covalent ligand molecules based on the implicit diffusion model, such as Figure 1 As shown, it includes the forward diffusion module, the noise prediction network ε θ , classifier, where the latent diffusion model is a type of generative model that uses a forward diffusion module that gradually adds noise to the data and a noise prediction network ε that learns how to remove noise from the noisy data. θ To learn a classifier for the data distribution, and then generate new data by sampling from the learned distribution.

[0062] 3.1 In the forward diffusion module, the initial data is set to x, and a forward process with a total time step of T is defined. At each time step, Gaussian noise is added to the molecular latent vector according to a predetermined noise strategy β. This forward noise addition process has Markov properties, which can be expressed as follows:

[0063]

[0064] Among them, q(x t |x t-1) is a step in the forward noisy process. Adding noise to the data at time t-1 is equivalent to sampling data from this Gaussian distribution. Due to the Markov property, the noisy data at any time t can be obtained in a single step, resulting in the noisy molecule latent vector.

[0065] 3.2 Noise Prediction Network ε θ

[0066] The noise prediction network is implemented by a fully connected neural network combined with a multi-head self-attention mechanism, which accepts the noisy molecular latent vector z t Protein target sequence encoding π θ (P). A multi-head self-attention mechanism is used to fuse the noisy molecular latent vector and the protein target sequence code. The fused vector data is then fed into a fully connected neural network, where noise prediction is performed to output a prediction of the noise added to the latent vector. The protein target sequence code is then introduced using the multi-head self-attention mechanism to learn the relationship between the protein target and the molecular latent vector, thereby generating a latent vector that is well-matched to the given target.

[0067] Specifically: The fusion of protein target sequence encoding and molecular latent vector adopts a multi-head self-attention mechanism, which has three learnable matrices Then these three matrices will transform the noisy molecular latent vector z t Protein target sequence encoding π θ (P) Mapping:

[0068] Q=W Q Z t ,K=W K π θ (P),V=W V π θ (P)

[0069] Then, the updated latent vector is obtained through attention

[0070]

[0071] Multi-head attention performs multiple independent attention calculations simultaneously and concatenates the results, namely:

[0072] Z=Concat(Attn1,…,Attn h )

[0073] The process of generating data is the reverse process, which is the process of removing noise from noisy data step by step, that is, reversing the forward noise addition process, that is, sampling q(x) from such a distribution at each step. t-1 |x t ).

[0074] This distribution is difficult to find an analytical solution, through a fully connected neural network p θ (x t-1 |x t ), it is also a Gaussian distribution, specifically expressed as follows:

[0075]

[0076] According to the above formula, p θ (x t-1 |x t ) The parameters to be fitted by the network are its mean and variance;

[0077] By setting the prediction from x0 to x t Noise prediction network ε of the added noise θ , noise prediction network ε θ Given a prediction of the noise added to the latent vector, the latent vector z for sampling the previous time step t-1 is calculated from this noise prediction t-1 The mean μ and variance ∑ of the distribution are obtained, thereby obtaining the network p of the reverse process fitting distribution θ (x t-1 |x t );

[0078] 3.2 Classifier model guiding covalent ligand generation

[0079] To generate molecules containing the desired covalent warhead substructure, the present invention uses a classifier to guide the forward diffusion module sampling generation. The classifier is used to classify latent vectors with noise added by the forward process. Specifically, a binary classifier is trained to classify the latent vectors of drug molecules. The binary classifier is constructed by a fully connected neural network and accepts a molecular latent vector z with noise added at a certain time step of the forward process of the latent diffusion model mentioned above. t , predict and output the molecular latent vector z t Whether the corresponding molecule has the corresponding substructure. At each time step of sampling, the gradient of the classifier is calculated This gradient is used to perturb the mean of the inverse distribution.

[0080]

[0081] This perturbation process is designed to guide the sampling process to produce molecules with desired properties.

[0082] 3. Training of the Covalent Drug De Novo Design Network

[0083] To train the variational autoencoder model, the present invention selected small molecules from the ZINC database as a training set. Specifically, molecules in the ZINC database that are "on the market" and have a lipid-water partition coefficient less than 5 and greater than -1, and an atomic weight greater than 250 and less than 500 were selected. From this, 1.2 million small molecules were randomly selected to form the training set.

[0084] To train the noise prediction network, the present invention selected the CrossDocked2020 dataset and used a similar processing method to the DiffSBDD dataset. Specifically, this dataset initially contains 225,000 protein-ligand pairings at different quality levels. First, the binding conformation RMSD greater than 100 was filtered out. The data was then clustered using the mmseqs2 method at 30% sequence identity, and 100,000 protein-ligand pairs were randomly selected to form the training dataset, with the remaining 100 proteins used as the test set.

[0085] To train the classifier model, the present invention selected molecules with and without the selected substructure from the ZINC database as positive and negative samples to construct a training set. For each substructure, the dataset size was the same, with 40,000 molecules as positive samples and 40,000 molecules as negative samples.

[0086] To train the toxicity predictor model, the toxicity data were taken from the mouse oral LD50 molecular dataset in the TOXRIC[6] database, which contains 35,299 molecules and their LD50 data.

[0087] 1. Training of Variational Autoencoders

[0088] For the training of the variational autoencoder model used to map molecules to their continuous latent vector representations, the molecules in the training set are fed into the encoder in batches to be encoded into latent vectors, and then fed into the decoder to decode the molecules. The training objective is a combination of the reconstruction error between the original molecule and the decoded molecule and the KL divergence error of the restricted latent space distribution:

[0089]

[0090] 2. Training of noise prediction network

[0091] First, the molecules in the dataset used by the noise prediction network are encoded into latent vectors using a trained variational autoencoder. These latent vectors are then fed with randomly sampled Gaussian noise at time step t via a forward diffusion module. This latent vector, along with the protein target sequence encoding vector obtained by ESM-2 encoding, is then fed into the noise prediction network to predict the added noise. The training objective is to minimize the L2 norm between the actual noise and the predicted noise, thereby training the network to fit the data distribution:

[0092]

[0093] 3. Training of the classifier model

[0094] First, the molecules in the classifier's dataset are encoded into latent vectors through a trained variational autoencoder. These latent vectors are then input into the classifier after adding Gaussian noise at the randomly sampled time step t. The classifier outputs the predicted labels of whether they have the corresponding substructures. The training goal is to minimize the predicted labels. and the binary cross-entropy loss (BCE) between the true label y.

[0095]

[0096] 4. Construction of toxicity prediction model and training of toxicity prediction model

[0097] The present invention adopts a random forest-based regression prediction model as a toxicity prediction model. It accepts a molecule SMILES representation and gives a prediction of its acute toxicity median lethal dose. During training, the molecular access system fingerprint (MACCS) in the toxicity training data is first extracted. This fingerprint serves as the input of the random forest model for predicting toxicity. The model is trained by minimizing the mean square error between the median lethal dose value predicted by the toxicity prediction model and the true value.

[0098] 5. Fine-tune the training process of the trained latent diffusion model using policy gradient

[0099] The trained latent diffusion model of the present invention is fine-tuned using a trained toxicity prediction model. Specifically, the trained latent diffusion model is used to sample and generate a molecule latent vector. The latent vector is decoded to obtain the molecule. The median lethal dose (LD50) of the decoded molecule is predicted using the toxicity prediction model. This prediction serves as the score of the molecule and also as the reward for each step in the reverse process of T time steps of a sampling process of the noise prediction network in the latent diffusion model. At each step, the gradient weighted by the score is calculated and accumulated as the gradient of the total reward with respect to the policy parameters, thereby updating the latent diffusion model.

[0100] Repeated sampling and this process fine-tunes the training of the latent diffusion model so that the molecules it generates have a lower median lethal dose. Specifically:

[0101] To ensure that the generated molecules have lower acute toxicity, the present invention uses a reinforcement learning policy gradient approach to further fine-tune the trained latent diffusion model. Specifically, the reverse process of the latent diffusion model sampling and generating new molecules is defined as a Markov decision process. A T-step Markov decision process can represent a sequence of states m and actions a:

[0102] τ=(m0,a0,m1,a1,…,m T ,a T )

[0103] Then for the sampling generation process of a latent diffusion model with T time steps, we have

[0104]

[0105]

[0106] Where c is the condition information, p θ (z t-1 |z t ,c) is a step in the sampling process, ρ T is the distribution of the state at the initial time step T. R is the reward function. The reward of a T-step sampling trajectory is determined by the score obtained by the numerator of its final latent vector z0 decoded by the decoder and the score function r. Therefore, the model can be optimized using the policy gradient method to maximize the reward R:

[0107]

[0108] Set [r(x0,c)] as a model for scoring the acute toxicity of molecules, and optimize the model in the above manner to reduce the acute toxicity of the molecules it generates.

[0109] After the above training is completed, a trained covalent drug de novo design network model is obtained, and the covalent drug de novo design network model is used to design covalent drug molecules;

[0110] 4. Generation of covalent ligand molecules

[0111] The present invention utilizes a trained covalent drug de novo design network model to generate molecules containing covalent warheads that covalently bind to a given protein target. The specific generation process is as follows:

[0112] 1. Calculate the latent vector at time t given by the classifier, set the total time step to T; sample a latent vector z from the standard normal distribution T As the starting point of the generation process, a protein is selected as the target, and the sequence of this protein is sent to the protein sequence encoder to obtain the protein target sequence encoding vector π θ (P); let "having covalent warhead" be the category of the classifier;

[0113] 2. At time step t, the hidden vector is recorded as z t ; The latent vector z t and protein target sequence encoding vector π θ (P) is fed into the implicit diffusion model, and the noise prediction module gives the z T The noise prediction is used to calculate the latent vector z for sampling the previous time step t-1. t-1 The mean μ and variance ∑ of the distribution;

[0114] 3. The latent vector z T Feed into the classifier model p Φ , the classifier model gives the gradient of the Caintiq vector for the category “with covalent warheads” Using the mean, variance, and gradient calculated in (2), we can calculate the latent vector z for sampling the previous time step t-1 guided by the classifier. t-1 The specific parameters of the distribution are as follows

[0115]

[0116] Sampling from this distribution yields the latent vector z t-1 ;

[0117] 4. Repeat (2) and (3) until all time steps are completed and t = 0 is reached, obtaining the final latent vector z0. z0 is fed into the decoder module of the variational autoencoder and decoded to obtain the final generated numerator x.

[0118] Preferably, the encoder and decoder of the molecule can be implemented by other autoencoder models, such as Grammar Variational Autoencoder GVAE (Grammar Variational Autoencoder).

[0119] Preferably, the dataset used to train the latent diffusion model can be replaced by other protein-ligand interaction datasets.

[0120] Preferably, the protein sequence encoder can be implemented using other networks, such as long short-term memory networks, convolutional neural networks, etc.

[0121] Preferably, the reinforcement learning algorithm used to fine-tune the latent diffusion model can be replaced by other algorithms such as Q-learning.

[0122] One embodiment of the present invention provides a covalent drug de novo design system based on an enhanced implicit diffusion model, comprising:

[0123] A network construction module is used to construct a covalent drug de novo design network and train it to obtain a covalent drug de novo design network model;

[0124] The covalent drug design module is used to select a protein as a target, sample a latent vector from a standard normal distribution, and input the protein sequence and latent vector into the covalent drug de novo design network model to obtain drug molecules that can covalently bind to the selected protein;

[0125] The covalent drug de novo design network includes a variational autoencoder, a protein sequence encoder, and a latent diffusion model. The variational autoencoder is used to encode molecular SMILES into latent vectors and decode the latent vectors into drug molecules. The protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding.

[0126] The implicit diffusion model includes the forward diffusion module, the noise prediction network ε θ and classifier models, the forward diffusion module is used to gradually add noise to the latent vector, and the noise prediction network ε θ It is used to fuse the latent vector and protein target information, and perform reverse denoising under the guidance of the classifier model to predict the latent vector of the drug molecule covalently bound to the protein target;

[0127] The toxicity prediction model is used to score the acute toxicity of the drug molecules covalently bound to the protein target. The implicit diffusion model is updated according to the scoring results to obtain drug molecules covalently bound to the protein target with low acute toxicity.

[0128] One embodiment of the present invention provides a terminal device comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned method for de novo covalent drug design; alternatively, when the processor executes the computer program, it implements the functions of the modules of the aforementioned system for de novo covalent drug design.

[0129] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0130] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0131] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0132] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0133] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0134] Based on this understanding, the present invention can implement all or part of the processes in the above-described method embodiments by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described method for de novo covalent drug design. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form.

[0135] The computer-readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electric carrier signal, a telecommunications signal, and a software distribution medium. It should be noted that the content of the computer-readable medium may be appropriately increased or decreased based on the requirements of legislation and patent practice within a jurisdiction. For example, in some jurisdictions, based on legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunications signals.

[0136] In order to demonstrate the advantages of the present invention and compare it with the two best existing technologies, the present invention uses each generation technology to generate 100 non-repeating molecules for each of the 100 proteins in the test set, and tests these 10,000 molecules. Table 1 Comparison of ligand generation effects of the three methods

[0137]

[0138] (1) The molecules generated by LDDD are more drug-like and easier to synthesize

[0139] To assess the drug-likeness and synthesizability of generated molecules, we used RDKIT to calculate druggability scores and synthetic accessibility scores. Both scores range from 0 to 1, with a higher score indicating better properties. Table 1 shows that the druggability scores for molecules generated by LDDD are significantly lower than those of the two best existing technologies, indicating that LDDD-generated molecules are more easily synthesized.

[0140] (2) LDDD generates molecules more efficiently

[0141] We calculated the time required to generate 100 molecules using various methods. As shown in Table 1, existing technologies are time-consuming due to their autoregressive generation method in 3D space. LDDD, on the other hand, takes much less time than existing methods and has better generation efficiency.

[0142] (3) LDDD can generate molecules with given substructures.

[0143] We selected four covalent drug warheads as the substructures generated by the classifier, such as Figure 2 Here, s is the scale of the classifier's guidance; larger values ​​indicate stronger guidance. Under the guidance of the classifier, the occurrence of various substructures in the 10,000 molecules improved by varying folds, demonstrating that LDDD can guide the generation of molecules with a given covalent warhead substructure.

[0144] (4) LDDD can optimize the acute toxicity of the generated molecules

[0145] like Figure 3 As shown in a, for the model that has undergone strategy gradient optimization of toxicity, the distribution of the predicted values ​​of the median lethal dose of the molecules generated is more to the left, indicating that there are more molecules with lower acute toxicity; Figure 3 As shown in Figure b, the number of acute toxicity warning structures in the generated molecules after the toxicity optimization model is significantly reduced. This indicates that LDDD can optimize the acute toxicity of the generated molecules.

[0146] (5) LDDD can generate molecules with good covalent binding ability for a given target

[0147] The present invention designs a covalent inhibitor molecule for the human epidermal growth factor protein T790M mutant (Epidermal Growth Factor Receptor (EGFR) T790M, abbreviated as EGFR T790M). EGFR T790M is a known drug on the market, such as Figure 4 a, as shown in 4b. Figure 4 c shows the horizontal axis is the known EGFR covalent inhibitors, the vertical axis is the covalent drug molecules designed by the present invention, the numbers in the table are the similarities between them, indicating that the present invention can generate molecules that are relatively similar to the known inhibitors. Subsequently, 100 molecules were generated for EGFR T790M using Pocket2Mol and the present invention, and acrylamide warheads were manually added to the 100 non-covalent molecules generated by the present invention. After covalent and non-covalent docking of these molecules with the protein, the results were shown in Figure 4 Figure e shows the root mean square error (RMSD) between the two docking poses, the distance between the alpha carbon atom of the carbon-carbon double bond of the acrylamide substructure and the sulfur atom of cysteine ​​797 on EGFRT790M, and the success rate refers to the proportion of the above atomic distances less than 4 angstroms. Figure 4 Panel f shows the predicted half-inhibitory concentration (IC50) of the molecule against the cell lines shown on the horizontal axis. It can be seen that the present invention outperforms the prior art in terms of the metrics shown above. These results demonstrate the present invention's ability to design covalent inhibitor molecules for a given target, embodying its ability to design molecules with better properties that are more easily optimized into drug candidates.

Claims

1. A method for de novo design of covalent drugs based on an enhanced implicit diffusion model, characterized in that: The specific steps are as follows: Constructing a covalent drug de novo design network and training it to obtain a covalent drug de novo design network model; A protein is selected as a target, a latent vector is sampled from a standard normal distribution, and the protein sequence and latent vector are input into a covalent drug de novo design network model to obtain drug molecules that can covalently bind to the selected protein. The covalent drug de novo design network includes a variational autoencoder, a protein sequence encoder, and a latent diffusion model. The variational autoencoder is used to encode molecular SMILES into latent vectors and decode the latent vectors into drug molecules. The protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding. The implicit diffusion model includes a forward diffusion module, a noise prediction network and classifier models, the forward diffusion module is used to gradually add noise to the latent vector, and the noise prediction network It is used to fuse the latent vector and protein target information, and perform reverse denoising under the guidance of the classifier model to predict the latent vector of the drug molecule covalently bound to the protein target; The toxicity prediction model is used to score the acute toxicity of the drug molecules covalently bound to the protein target. The implicit diffusion model is updated according to the scoring results to obtain drug molecules covalently bound to the protein target with low acute toxicity.

2. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: The variational autoencoder includes an encoder that encodes molecular SMILES into latent vectors and a decoder that decodes the latent vectors into drug molecules. Both the encoder and decoder use multi-layer gated recurrent units to achieve the mapping between characters and continuous latent vectors. The encoder uses the output of the GRU final state as the continuous latent vector representation of the molecule. The training objective of the variational autoencoder is a combination of the reconstruction error between the original molecule and the decoded molecule and the KL divergence error that constrains the latent space distribution.

3. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: The protein sequence encoder uses the ESM-2 model to encode the target protein sequence.

4. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: The forward diffusion module follows the noise strategy at each time step Add Gaussian noise to the latent vector to obtain the noisy latent vector, which is specifically expressed as follows: in, It is a step in the forward noise addition process, that is, adding noise to the data at time t-1 to obtain the data at time t.

5. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: The noise prediction network is specifically a fully connected neural network with a multi-head self-attention mechanism. It uses the multi-head self-attention mechanism to fuse the noisy molecular latent vector and the protein target sequence code. The fused vector data is input into the fully connected neural network, where noise prediction is performed and the prediction of the noise added to the latent vector is output. The training objective is to minimize the L2 norm between the actual noise and the predicted noise.

6. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: The classifier model is a binary classifier used to predict whether the noisy latent vector has the corresponding molecular structure, and uses the prediction result as a gradient to guide the noise prediction network to generate drug molecules that can covalently bind to the selected protein. The training goal is to minimize the predicted label The binary cross entropy loss between the true label y and the 7. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 1, characterized in that: A random forest-based regression prediction model was used as the toxicity prediction model. The toxicity prediction model accepted the molecular SMILES representation and gave a prediction of the acute toxicity median lethal dose of the molecule. The input of the toxicity prediction model was the molecular access system fingerprint, and the training objective was to minimize the mean square error between the median lethal dose value predicted by the toxicity prediction model and the true value.

8. The method for de novo design of covalent drugs based on enhanced implicit diffusion model according to claim 7, characterized in that: The toxicity prediction model is used to predict the median lethal dose of drug molecules obtained by the implicit diffusion model, and the prediction is used as the reward for each step in the reverse process of T time steps of a sampling process of the noise prediction network in the implicit diffusion model. The gradient with the reward as the weight is calculated in each step and accumulated as the gradient of the total reward with respect to the strategy parameters. The gradient is used to update the implicit diffusion model to obtain an implicit diffusion model that generates drug molecules with lower acute toxicity.

9. A covalent drug de novo design system based on an enhanced implicit diffusion model, characterized in that: include: A network construction module is used to construct a covalent drug de novo design network and train it to obtain a covalent drug de novo design network model; The covalent drug design module is used to select a protein as a target, sample a latent vector from a standard normal distribution, and input the protein sequence and latent vector into the covalent drug de novo design network model to obtain drug molecules that can covalently bind to the selected protein; The covalent drug de novo design network includes a variational autoencoder, a protein sequence encoder, and a latent diffusion model. The variational autoencoder is used to encode molecular SMILES into latent vectors and decode the latent vectors into drug molecules. The protein sequence encoder is used to encode protein target information to obtain protein target sequence encoding. The implicit diffusion model includes a forward diffusion module, a noise prediction network and classifier models, the forward diffusion module is used to gradually add noise to the latent vector, and the noise prediction network It is used to fuse the latent vector and protein target information, and perform reverse denoising under the guidance of the classifier model to predict the latent vector of the drug molecule covalently bound to the protein target; The toxicity prediction model is used to score the acute toxicity of the drug molecules covalently bound to the protein target. The implicit diffusion model is updated according to the scoring results to obtain drug molecules covalently bound to the protein target with low acute toxicity.

10. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for de novo design of covalent drugs based on the enhanced implicit diffusion model as described in any one of claims 1 to 8 are implemented.

Images