Method for generating molecules with specific properties aiming at multiple targets

Abstract

The invention discloses a method for generating molecules with specific properties aiming at multiple targets, which comprises the following steps: acquiring molecular data, protein targets and corresponding ligand data, and the protein target data are two protein targets which are expected to be generated and can target; constructing a compound-target scoring module based on the protein targets and the ligand data corresponding to the protein targets; the molecular data is input into a deep learning model, the deep learning model is trained, and the deep learning model is constructed by combining a diffusion model with KAN and is optimized based on the compound-target scoring module; and obtaining molecules with specific properties based on the optimized deep learning model. According to the present invention, the multi-target 3D molecule with the specific property can be generated, and the contribution is made to the discovery of the multi-pharmacological compound and the treatment of the multi-gene disease.

Description

Technical Field

[0001] The present invention belongs to the fields of biomedicine, computational chemistry and artificial intelligence, and in particular relates to a method for generating molecules with specific properties for multiple targets. Background Art

[0002] Designing drugs with specific properties is an integral part of modern drug development. Various traditional computational methods, such as virtual screening and combinatorial libraries, are used to search for drug candidates in a vast chemical space, but the effectiveness of these methods is limited by the coverage of compound databases and the diversity of synthetic schemes. Moreover, the number of potential drug-like molecules is estimated to be around 10 60 to 10 100 Exploring this almost infinite chemical space is a daunting task. In addition, there are still many potential drug molecules that have not been recorded in existing databases, which limits the diversity of results from traditional methods. Therefore, methods based on deep learning for drug design have attracted much attention.

[0003] In recent years, deep learning, especially generative models, including recurrent neural networks (RNN), Transformer, flow models, variational autoencoders (VAE), and diffusion models, have become very important tools in drug design. By learning the characteristics and distribution of molecular structures, these models can generate new molecules that are similar to the training data or have specific targets and properties. Some one-dimensional (1D) or two-dimensional (2D) molecular generation methods have achieved acceptable results in terms of chemical structure validity or optimization of physicochemical properties, but they are based on 1D simplified molecular linear input specifications (SMILES) or 2D molecular graph data and cannot effectively utilize three-dimensional (3D) spatial information, which plays a vital role in determining molecular properties. Therefore, some methods for generating molecules based on 3D information have been proposed. For example, some methods use the 3D contextual information of the binding pocket for structure-based ligand design, explicitly considering chemical information including the target protein to generate new ligand molecules within the protein binding pocket.

[0004] However, current molecular generation methods still have certain shortcomings. In particular, most deep learning-based molecular generation methods can only design drugs targeting a single target. However, many diseases, especially polygenic diseases, exhibit complex pathological mechanisms involving multiple genes and biological pathways. Although combination therapy can provide some solutions, polypharmacology approaches have irreplaceable advantages, such as superior pharmacokinetics and safety, lower possible risks of acquired resistance and drug interactions, and simpler treatment regimens. Therefore, drugs targeting a single target may not be sufficient to treat these complex diseases, and a comprehensive multi-target approach is needed for effective intervention and treatment. However, since the multiple properties of candidate drugs are difficult to consider simultaneously, generating high-quality drugs targeting multiple specific targets with desired properties remains a long-term goal of drug discovery. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a method for generating molecules with specific properties for multiple targets to solve the problems existing in the above prior art.

[0006] To achieve the above object, the present invention provides a method for generating molecules with specific properties against multiple targets, comprising:

[0007] Obtaining molecular data, protein target data and corresponding ligand data, wherein the protein target data are two protein targets that the molecules to be generated can target;

[0008] Based on the protein targets and the ligand data of the corresponding protein targets, a compound-target scoring module is constructed;

[0009] Inputting the molecular data into a deep learning model to train the deep learning model, wherein the deep learning model is constructed by combining a diffusion model with a KAN, and optimizing the deep learning model based on the compound-target scoring module;

[0010] Based on the optimized deep learning model, molecules with specific properties are obtained.

[0011] Optionally, obtaining the molecular data includes:

[0012] Get the initial data set;

[0013] Performing feature extraction on the initial data set to obtain feature data;

[0014] Noise is added to the characteristic data to obtain the molecular data.

[0015] Optionally, constructing the compound-target scoring module includes:

[0016] Get the label of the compound-target scoring module;

[0017] The compound-target scoring module is trained using the label of the compound-target scoring module to obtain the compound-target scoring module.

[0018] Optionally, a method for obtaining labels of a compound-target scoring module is:

[0019] Calculating the ligand efficiency as a label of the compound-target scoring module according to the IC50 value in the protein target and the ligand data corresponding to the protein target;

[0020] The calculation method is:

[0021]

[0022] Where N is the number of non-hydrogen atoms and E is the ligand efficiency.

[0023] Optionally, the deep learning model includes: an atomic element type denoising module, an atomic position denoising module and a chemical bond denoising module;

[0024] The atomic element type denoising module is used to restore the atomic element type from the noise;

[0025] The atomic position denoising module is used to recover atomic position information from noise;

[0026] The chemical bond denoising module is used to restore the chemical bond type from the noise.

[0027] Optionally, based on the compound-target scoring module, optimizing the deep learning model includes:

[0028] S1. Scoring the first molecule data using the scoring function in the compound-target scoring module, and retaining molecules with preset scores;

[0029] S2. Using the numerator of the preset score, training the deep learning model to obtain second numerator data;

[0030] S3. Repeat S1-S2 until the score of the molecular data reaches the standard, thereby completing the optimization of the deep learning model.

[0031] Optionally, the scoring function includes: ligand efficiency, SA, MW, LogP and QED.

[0032] Optionally, the ligand efficiency and QED score are calculated as follows:

[0033]

[0034] The SA score is calculated as follows:

[0035]

[0036] Among them, s le is the ligand efficiency predicted by the compound-target scoring module, s qed is the raw QED score, s sa is the raw SA score, μ and σ are the threshold mean and standard deviation used for normalization, S le ' is the normalized score of ligand efficiency, S qed ' is the normalized score of the raw QED score, S sa ' is the standardized score of the raw SA score;

[0037] The MW and LogP scores were calculated as:

[0038]

[0039] Among them, S mw ' is the normalized score of the raw MW score, S logp ' is the normalized score of the raw LogP score, β is the exponential decay rate, b l is the lower limit of the target score interval, b u is the upper limit of the target score interval, s mw is the raw MW fraction, s logp is the raw LogP score.

[0040] Compared with the prior art, the present invention has the following advantages and technical effects:

[0041] The present invention understands the three-dimensional structure of molecules based on the diffusion model, and uses KAN instead of MLP in the diffusion model to improve the performance of the model. Compared with MLP, KAN can achieve higher accuracy with fewer parameters.

[0042] The present invention optimizes the molecules generated by the deep learning model through reinforcement learning technology so that the generated molecules have the ability to target multiple targets and the desired molecular properties. In the reinforcement learning process, ligand efficiency, SA, QED, LogP and MW are used as scoring criteria for generating multi-target compounds with desired properties, where ligand efficiency is predicted using an additional compound-target scoring module.

[0043] Compared with the existing technologies, the present invention shows excellent and competitive performance in de novo molecular testing. More importantly, the present invention can generate 3D molecules with specific properties against multiple targets, contributing to the discovery of polypharmacological compounds and the treatment of polygenic diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0045] Figure 1 A flow chart of a method for generating molecules with specific properties for multiple targets according to an embodiment of the present invention;

[0046] Figure 2 A diagram of docking simulation results of a molecule generated by an embodiment of the present invention and a protein target MEK1;

[0047] Figure 3 A diagram of docking simulation results of a molecule generated by an embodiment of the present invention and a protein target mTOR;

[0048] Figure 4 A distribution graph of the scores of the molecules generated by the embodiments of the present invention for molecular docking with the protein targets MEK1 and mTOR;

[0049] Figure 5 This is a first target attribute distribution diagram of an embodiment of the present invention;

[0050] Figure 6 This is a diagram showing the distribution of the second target attribute according to an embodiment of the present invention;

[0051] Figure 7 This is a diagram showing the distribution of the third target attribute according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0053] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0054] In view of the problem that most molecular generation models in the current prior art can only generate molecules for one target, and thus it is difficult to use a single compound to treat polygenic diseases, the present invention provides the following technical solution. The purpose of the present invention is to use a diffusion model to understand the 3D structure of molecules and generate molecules, and to use reinforcement learning technology to optimize the generated molecules to achieve the ability to target multiple targets and have specific molecular properties, that is, to generate polypharmacological compounds with specific properties.

[0055] The present invention proposes a method for generating molecules with specific properties against multiple targets, such as Figure 1 As shown, the specific steps include:

[0056] Obtaining molecular data, protein target data and corresponding ligand data, wherein the protein target data are two protein targets that the molecules to be generated can target;

[0057] Based on the protein targets and the ligand data of the corresponding protein targets, a compound-target scoring module is constructed;

[0058] Input the molecular data into the deep learning model and train the deep learning model, wherein the deep learning model is constructed by combining the diffusion model with KAN, and the deep learning model is optimized based on the compound-target scoring module;

[0059] Based on the optimized deep learning model, molecules with specific properties are obtained.

[0060] Further, obtaining molecular data includes:

[0061] Get the initial data set;

[0062] Extract features from the initial data set to obtain feature data;

[0063] Add noise to the feature data to obtain molecular data.

[0064] Specifically, molecular data were obtained from the GEOM Drugs dataset for training the diffusion model, and molecules containing only the main elements H, C, N, O, F, P, S, and Cl were retained, and H atoms were removed from these molecules. Molecules involving any other chemical bonds except single bonds, double bonds, triple bonds, and aromatic bonds were removed;

[0065] The BindingDB database was used to obtain the ligand-target binding data for a desired pair of protein targets, and ligands with experimentally measured IC50 values ​​(less than 1 μM) were retained for training the compound-target scoring module.

[0066] Extract corresponding features from molecular data to form the input data of the diffusion model, including the element type of atoms in the molecule, the 3D coordinates of the atoms, and the chemical bonds;

[0067] Based on the diffusion model, noise is gradually added to the data;

[0068] Based on the diffusion model and KAN, a deep learning model was constructed to denoise the noisy data and restore it to the original data;

[0069] The molecules generated by the diffusion model were evaluated using effectiveness, uniqueness, novelty, and diversity metrics, where effectiveness represents the proportion of valid molecules generated, uniqueness represents the proportion of unique molecules among all generated molecules, novelty represents the proportion of molecules that are not in the training set among all generated molecules, and diversity represents the average molecular fingerprint similarity of all generated pairs of molecules;

[0070] Furthermore, the deep learning model includes: an atomic element type denoising module, an atomic position denoising module, and a chemical bond denoising module;

[0071] Atomic element type denoising module, used to recover atomic element types from noise;

[0072] Atomic position denoising module, used to recover atomic position information from noise;

[0073] Chemical bond denoising module, used to recover chemical bond types from noise.

[0074] Furthermore, constructing a compound-target scoring module includes:

[0075] Get the label of the compound-target scoring module;

[0076] The compound-target scoring module is trained using the label of the compound-target scoring module to obtain the compound-target scoring module.

[0077] Specifically, the molecular fingerprint of the ligand data in the compound-target binding data of the protein target obtained from the BindingDB database is calculated;

[0078] Ligand efficiency was calculated using IC50 values ​​from compound-target binding data of protein targets obtained from the BindingDB database as follows:

[0079]

[0080] Where N is the number of non-hydrogen atoms and E is the ligand efficiency;

[0081] Molecular fingerprints were used as feature inputs for the compound-target scoring module, ligand efficiency was used as labels, and five-fold cross validation was used to test the performance of the compound-target scoring module, with a root mean square error (RMSE) of 0.55±0.03.

[0082] Furthermore, based on the compound-target scoring module and combined with molecular properties, the deep learning model is optimized including:

[0083] S1. Scoring the first molecule data using a scoring function in a compound-target scoring module, and retaining molecules with preset scores;

[0084] S2. Using the molecules with preset scores, train the deep learning model to obtain second molecule data;

[0085] S3. Repeat S1-S2 until the score of the molecular data meets the standard and the optimization of the deep learning model is completed.

[0086] Furthermore, the scoring functions include: ligand efficiency, SA, MW, LogP and QED.

[0087] Furthermore, the ligand efficiency and QED score are calculated as:

[0088]

[0089]

[0090] The SA score is calculated as:

[0091]

[0092] Among them, s le is the ligand efficiency predicted by the compound-target scoring module, s qed is the raw QED score, s sa is the raw SA score, μ and σ are the threshold mean and standard deviation used for normalization, S le ' is the normalized score of ligand efficiency, S qed ' is the normalized score of the raw QED score, S sa ' is the standardized score of the raw SA score;

[0093] The MW and LogP scores are calculated as:

[0094]

[0095] Among them, S mw ' is the normalized score of the raw MW score, S logp ' is the normalized score of the raw LogP score, β is the exponential decay rate, b l is the lower limit of the target score interval, b u is the upper limit of the target score range, s mw is the raw MW fraction, s logp is the raw LogP score.

[0096] like Figure 1 As shown, the overall process of a method for generating molecules with specific properties for multiple targets of the present invention includes two parts:

[0097] a) Generate molecules using diffusion model: Build a deep learning model based on the diffusion model, input the molecular data into the diffusion model to obtain the trained deep learning model, and evaluate the model performance using effectiveness, uniqueness, novelty and diversity indicators; b) Optimize molecules using reinforcement learning model: Build a compound-target scoring module based on protein targets and their corresponding ligand data to predict the ligand efficiency of the generated molecules and the corresponding targets, optimize the diffusion model based on reinforcement learning, where the scoring function used in reinforcement learning is the ligand efficiency predicted by the compound-target scoring module and a variety of molecular properties, including SA, MW, LogP and QED, to generate molecules with the desired molecular properties that can target multiple protein targets at the same time.

[0098] In the first part, for generating molecules using a diffusion model, a diffusion model is first constructed. Different from previous diffusion models, the present invention uses KAN to construct the diffusion model, including using a deep learning model composed of KAN to denoise the atomic element type, atomic position and chemical bond of the molecule. KAN can achieve higher accuracy than MLP while using fewer parameters than MLP, thereby improving the performance of the diffusion model used in the present invention.

[0099] Before starting to train the diffusion model, molecular data were obtained from the GEOM Drugs dataset for training the diffusion model. Molecules containing only the main elements H, C, N, O, F, P, S, and Cl were retained, and H atoms were removed from these molecules. Molecules involving any other chemical bonds except single bonds, double bonds, triple bonds, and aromatic bonds were excluded.

[0100] For model training, let M = {E, P, B} describe a 3D molecule with n atoms, where E = (e 1 ,e 2 ,…,e n )∈E n Describes the atomic element type, Describes the 3D position of atoms, B = (b 11 ,b 12 ,…,b nn )∈B n×n Describe chemical bonds. Let t = 0, 1, ..., T represent the index of the diffusion step. The diffusion model is a two-step process based on Markov chains: diffusion process and denoising process. The diffusion process gradually adds Gaussian noise to the data using a predefined noise scheme, while the reverse denoising process gradually removes the noise through a parameterized neural network, and finally recovers the real data by learning how to recover the original data from the noisy data. The forward process is as follows:

[0101]

[0102] where βt ∈[0,1] controls the degree of mixing of data and noise, 1 k and 1 k' is a one-hot vector with 1 in the kth or k'th position, and is the identity matrix. For the atom types and chemical bond types, they are represented using the categorical distributions in the formula. By defining The sampling of data can be derived as:

[0103]

[0104] From the above formula, we can see that if t→∞, the final distribution will be close to the standard Gaussian distribution. or absorption type 1 k The denoising process is also a Markov chain that aims to recover the true data from the diffuse data using a neural network, as shown below:

[0105]

[0106] Where N θ , μ θ and N' θ is a neural network, Σ t Set to beta t . Therefore, the loss function is as follows:

[0107]

[0108] The atomic element types and chemical bonds use the Kullback-Leibler (KL) divergence to measure the difference between the actual distribution and the neural network predicted distribution, λ 1 and λ 2 is a constant.

[0109] After training, the model performance is evaluated using the indicators of Validity, Uniqueness, Novelty, and Diversity. Validity indicates the proportion of valid molecules generated. High validity indicates that the model has a deep understanding of the 3D molecular structure and the generated molecules conform to chemical rules:

[0110]

[0111] Uniqueness represents the proportion of unique molecules among all generated molecules. Uniqueness measures the ability of the model to generate new molecules and explore various compounds:

[0112]

[0113] Novelty indicates the proportion of generated molecules that are not in the training set. Low novelty indicates that the model's generative ability is weak or the model is overfitting:

[0114]

[0115] Diversity represents the average molecular fingerprint similarity calculated using all generated molecular pairs. Diversity represents the ability of the model to generate molecules with different structures:

[0116]

[0117] Where N is the total number of molecules, Sim(m i ,m j ) is the molecular fingerprint similarity between molecules i and j. The present invention is compared with multiple baseline models, including MolDiff and MDM, all of which are trained on the GEOM-Drugs dataset. The results are shown in Table 1, where MDRL is the method of the present invention for generating molecules with specific properties for multiple targets, and MDRL-NRL is the version of the present invention that removes reinforcement learning and uses the diffusion model alone:

[0118] Table 1

[0119]

[0120] MDRL-NRL outperforms all baseline models in terms of effectiveness and uniqueness, which means that KAN can understand the 3D structure of molecules at the same or better level than MLP. At the same time, MDRL-NRL is also comparable to baseline models in terms of diversity and novelty, indicating that MDRL-NRL has the ability to fully explore chemical space. Overall, MDRL-NRL performs comparable to state-of-the-art models in all aspects and is competent for the task of generating polypharmacological compounds. After adding reinforcement learning, the effectiveness and novelty of MDRL remain largely unchanged, indicating that the reinforcement learning component does not significantly affect MDRL's understanding of molecular structure. Although it is noted that the uniqueness and diversity of the generated molecules are reduced, this may be due to the contraction of the chemical space. Through reinforcement learning optimization, the model can generate molecules in a targeted manner, which greatly limits the available chemical space and increases the possibility that the model will generate similar or identical molecules, thereby reducing uniqueness and diversity. Despite the decrease in uniqueness and diversity, the model has obviously learned to identify and exploit chemical space in a targeted manner. Overall, MDRL performs well in the generation ability test.

[0121] In the second part, for the molecules generated by the reinforcement learning optimization model, we first constructed the scoring functions required for reinforcement learning, including ligand efficiency, SA, MW, LogP, and QED. The ligand efficiency requires the construction of a compound-target scoring module, and the ligand efficiency is calculated using the IC50 value in the compound-target binding data obtained from the BindingDB database, as shown below:

[0122]

[0123] Where N is the number of non-hydrogen atoms and E is the ligand efficiency. The molecular fingerprints were calculated using the ligands in the compound-target binding data, and the molecular fingerprints were used as the feature input of the compound-target scoring module. The ligand efficiency was used as the label, and the performance of the compound-target scoring module was tested using a five-fold cross validation, with a root mean square error (RMSE) of 0.55 ± 0.03. The SA, QED, MW, and LogP scores were obtained using the RDKit software package. Next, the above five scores were normalized, where the ligand efficiency score and QED score were calculated as follows:

[0124]

[0125] The SA score calculation method is as follows:

[0126]

[0127] Among them, s le is the ligand efficiency predicted by the compound-target scoring module, s qed is the raw QED score, s sa is the raw SA score, μ and σ are the threshold mean and standard deviation used for normalization, S le ' is the normalized score of ligand efficiency, S qed ' is the normalized score of the raw QED score, S sa ' is the normalized score of the raw SA score. The MW score and LogP score are calculated as follows:

[0128]

[0129] Among them, S mw ' is the normalized score of the raw MW score, S logp ' is the normalized score of the raw LogP score, β is the exponential decay rate, b l is the lower limit of the target score interval, b u is the upper limit of the target score range, s mw is the raw MW fraction, s logpis the raw LogP score. The raw SA score, raw QED score, raw LogP score, and raw MW score are calculated by the RDKit software package, and the total score is the weighted average of the individual scores.

[0130] After defining the above scoring function, we use reinforcement learning to optimize it. The steps are as follows:

[0131] i) generating an initial set of 6000 molecules based on the diffusion model, and evaluating and ranking these molecules using a scoring function;

[0132] ii) retaining the top 4000 scoring molecules for additional training of the diffusion model;

[0133] iii) Generate 2000 molecules using the additional trained diffusion model and combine them with the previous 4000 molecules to form a molecule set of the same size as the initial 6000 molecule set;

[0134] iv) Repeat steps from ii) to iii) to guide the model to achieve the optimization goal.

[0135] After the generation was completed, the generated molecules were evaluated. First, the ability of the generated molecules to bind to the two targets was evaluated. Taking the protein targets MEK1 and mTOR as examples, this example obtained the structures of MEK1 and mTOR from the Protein Data Bank database for subsequent molecular docking. Subsequently, the MDRL-generated molecules were docked with the corresponding two targets using the AutoDock Vina software. Figure 2 and Figure 3 One of the best compounds generated by MDRL for both targets is shown, with docking affinity scores of -8.6 kcal / mol for MEK1 and -9.9 kcal / mol for mTOR. The average vina docking scores for the top 1000 reinforcement learning scoring compounds were calculated, with scores of -7.1 kcal / mol for MEK1 and -7.7 kcal / mol for mTOR. The corresponding distributions are shown in Figure 4 As shown, it can be seen that the highest vina scores of the generated molecules with the protein targets MEK1 and mTOR are -9.9 kcal / mol and -10.4 kcal / mol, respectively, and the densely distributed area of ​​molecules is located in the upper right corner, indicating that most molecules show good vina docking scores for both targets. These results show that the present invention can effectively generate polypharmacological compounds.

[0136] In drug design, molecules are often required to have specific properties. For example, designing molecules that are easier to synthesize often requires molecules to have lower SA values. As early as 1997, Lipinski's rule of five provided an empirical guide for evaluating the drug-likeness of compounds. According to this rule, a viable candidate drug should have a LogP value of no more than 5 and a MW value of less than 500 Daltons. Therefore, in drug design, the ability to accurately control the properties of the generated molecules is crucial. Next, LogP was selected as an example and the ability of the present invention to control this single property was evaluated. Different target ranges were set for the LogP values ​​of the generated molecules, and the distribution of 4000 molecules generated within these ranges was evaluated. On average, 81.0% of the generated molecules met the LogP range requirements, and the average LogP value was well controlled.

[0137] Next, the ability of the present invention to control multiple properties of the generated molecules was further evaluated using four properties simultaneously. Based on the relevant research on drugs that cross the blood-brain barrier, three sets of experiments were designed, where LogP was in the range of 2-4 and MW was less than 450 Daltons. Figure 5 (SA<3, QED>0.75, LogP=3-4, MW=350-450), Figure 6 (SA<2, QED>0.7, LogP=2-3, MW=250-350) and Figure 7 (SA<1, QED>0.65, LogP=2-3, MW=150-250) The results of three sets of experiments are shown respectively. In the third set of experiments, to ensure that LogP is in the range of 2-4, the range is kept unchanged and the target ranges of SA, QED and MW are further changed. For SA and QED, a threshold is selected instead of setting a specific range because drug design often requires molecules that are easier to synthesize or drug candidates with a higher probability of success, which requires a lower SA or a higher QED. Figure 5 and Figure 6 The results of the two experiments show that the present invention exhibits effective control over SA, QED, LogP and MW, with the average percentages of molecules within the control range being 96.9%, 81.7%, 70.8% and 92.3%, respectively. Figure 6 As can be seen in Figure 5 In contrast, the lower SA requirement leads to a lower average SA value for the molecules generated by the model. Figure 5 The target value in is higher than Figure 6 , but did not result in Figure 5 This may be due to the complexity of QED, which depends on multiple molecular properties, including LogP and MW. Figure 6In the experiment shown in , the average LogP was 2.66 and the average MW was 295.58 Daltons, which is more consistent with the characteristics of the drug, resulting in an average QED value of Figure 5 The results highlight the interplay between multiple molecular properties and the importance of controlling multiple properties simultaneously. Figure 7 In the experiment shown, the target ranges for SA, QED, and MW were further changed. For QED and MW, the molecules generated by the model are well controlled within the specified range. For SA, whose values ​​are usually between 1 and 10, setting the target to SA<1 does not produce the corresponding molecules, but rather encourages the SA values ​​of the molecules generated by the model to be as small as possible. As can be seen from the results, Figure 7 The average SA value of the molecules generated in was lower than that in the previous two experiments. Overall, these results show that MDRL is able to generate molecules with a variety of desirable properties.

[0138] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for generating molecules with specific properties against multiple targets, characterized in that: include: Obtaining molecular data, protein target data and corresponding ligand data, wherein the protein target data are two protein targets that the molecules to be generated can target; Based on the protein targets and the ligand data of the corresponding protein targets, a compound-target scoring module is constructed; Inputting the molecular data into a deep learning model to train the deep learning model, wherein the deep learning model is constructed by combining a diffusion model with a KAN, and optimizing the deep learning model based on the compound-target scoring module; Based on the optimized deep learning model, molecules with specific properties are obtained.

2. A method for generating molecules with specific properties for multiple targets according to claim 1, characterized in that: Acquiring the molecular data includes: Get the initial data set; Performing feature extraction on the initial data set to obtain feature data; Noise is added to the feature data to obtain the molecular data.

3. A method for generating molecules with specific properties against multiple targets according to claim 1, characterized in that: Constructing the compound-target scoring module includes: Get the label of the compound-target scoring module; The compound-target scoring module is trained using the label of the compound-target scoring module to obtain the compound-target scoring module.

4. A method for generating molecules with specific properties against multiple targets according to claim 3, characterized in that: The method to obtain the label of the compound-target scoring module is: Calculating the ligand efficiency as a label of the compound-target scoring module according to the IC50 value in the protein target and the ligand data corresponding to the protein target; The calculation method is: Where N is the number of non-hydrogen atoms and E is the ligand efficiency.

5. A method for generating molecules with specific properties against multiple targets according to claim 1, characterized in that: The deep learning model includes: an atomic element type denoising module, an atomic position denoising module and a chemical bond denoising module; The atomic element type denoising module is used to restore the atomic element type from the noise; The atomic position denoising module is used to recover atomic position information from noise; The chemical bond denoising module is used to restore the chemical bond type from the noise.

6. A method for generating molecules with specific properties against multiple targets according to claim 1, characterized in that: Based on the compound-target scoring module, optimizing the deep learning model includes: S1. Scoring the first molecule data using the scoring function in the compound-target scoring module, and retaining molecules with preset scores; S2. Using the numerator of the preset score, training the deep learning model to obtain second numerator data; S3. Repeat S1-S2 until the score of the molecular data reaches the standard, thereby completing the optimization of the deep learning model.

7. A method for generating molecules with specific properties against multiple targets according to claim 6, characterized in that: The scoring functions include: ligand efficiency, SA, MW, LogP and QED.

8. A method for generating molecules with specific properties against multiple targets according to claim 7, characterized in that: The ligand efficiency and QED score are calculated as follows: The SA score is calculated as follows: Among them, s le is the ligand efficiency predicted by the compound-target scoring module, s qed is the raw QED score, s sa is the raw SA score, μ and σ are the threshold mean and standard deviation used for normalization, S le ' is the normalized score of ligand efficiency, S qed ' is the normalized score of the raw QED score, S sa ' is the standardized score of the raw SA score; The MW and LogP scores were calculated as: Among them, S mw ' is the normalized score of the raw MW score, S logp ' is the normalized score of the raw LogP score, β is the exponential decay rate, b l is the lower limit of the target score interval, b u is the upper limit of the target score range, s mw is the raw MW fraction, s logp is the raw LogP score.

Images