Small molecule generation method and device based on cross-modal pre-training and electronic equipment

Abstract

The invention provides a cross-modal pre-training-based small molecule generation method and apparatus, and an electronic device. The method comprises the steps of obtaining a given protein pocket; obtaining a target representation vector of a protein pocket; on the basis of a pre-trained molecule generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecule generation model is obtained by performing training optimization according to a training sample set formed by drug-like small molecule samples and corresponding representation vector samples, the representation vector samples are obtained by encoding through a molecular encoder and reducing modal gaps, and the molecule generation model is used for simulating representation vectors of protein pocket samples corresponding to the drug-like small molecule samples; the molecular encoder is a pre-trained contrast learning model. According to the method, a protein pocket and a ligand small molecule are aligned in a representation space through a pre-trained contrast learning model, and a cross-modal protein-small molecule pairing task is learned in an unsupervised manner by using single-modal drug-like small molecule data, so that the chemical structure and conformation distribution of drug molecules can be accurately captured; according to the method, the dependence on experimental data is avoided, the drug micromolecules which have more reasonable geometric structures and are closer to real drug molecular structure distribution are generated, and the synthesizability of the generated drug micromolecules and the affinity between the generated drug micromolecules and target spots are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of drug design, and in particular to a small molecule generation method, device and electronic equipment based on cross-modal pre-training. Background Art

[0002] Since the great success of drugs based on target structure design represented by HIV protease inhibitors at the end of the last century, structure-based drug design (SBDD) has become a key strategy for new drug research and development. SBDD aims to use the three-dimensional structure and physicochemical characteristics of the target pocket to design molecules that bind tightly to the target pocket. Broadly speaking, SBDD includes virtual screening methods and de novo design methods. Compared with virtual screening, de novo molecular design methods can avoid time-consuming blind searches in a vast chemical space, greatly improve the efficiency of drug discovery, and are currently a hot research direction in drug design.

[0003] At the same time, generative AI models have demonstrated excellent performance in fields such as natural language, images, and videos; in the field of drug design, AI molecular generation models can also play an important role.

[0004] At present, the mainstream molecular generation models are: (1) autoregressive small molecule generation model, but the autoregressive generation method has many limitations, such as exposure bias, error accumulation, inability to consider the entire three-dimensional structure when sampling, unnatural sampling order, poor scalability of sampling schemes when generating large molecules, and inability to estimate the mass of generated molecules; (2) non-autoregressive small molecule generation model based on diffusion model, but this model faces problems such as multimodality of molecular data and sensitivity to atomic coordinate noise. The molecules it generates are prone to illegal structures such as large rings that do not conform to chemical constraints and have poor synthesizability.

[0005] The above-mentioned supervised learning models all require the use of annotated protein pocket-ligand small molecule data to learn the distribution of ligand small molecules for a given protein. In fact, the experimentally measured paired protein pocket-ligand small molecule data is extremely limited. Compared with the huge cross-modal datasets in the field of image and text generation (such as the dataset LAION-5B contains 5.85 billion image-text pairs), the protein database PDBbind commonly used in the field of drug generation has only 22,920 protein-ligand small molecule pairs, and the crossdocked2020 dataset (synthetic data) has only 100,000 molecular pairs, which is far from enough to support the model to accurately capture the distribution of the chemical structure and conformation of drug molecules. This may cause the drug molecules generated by the model to still have large differences from the real drug molecules in the distribution of properties such as bond length, bond angle, and torsion angle, resulting in "illegal" phenomena such as geometric distortion and structural fracture.

[0006] The Shape2Mol model uses a large amount of unlabeled small molecule data for pre-training, which improves the "legitimacy" of generated molecules and provides a new way to get rid of the reliance on labeled paired data. However, since it constructs a molecular generation model from the perspective of shape, it ignores important physical and chemical properties of protein pockets such as hydrophilicity and partial charge, and the generated molecules have poor ability to bind specifically to protein pockets.

[0007] In summary, although AI molecular generation models have great potential in improving the efficiency and effectiveness of early drug discovery, they are limited by the scarcity of training data and the lack of effective pre-training methods. The drug molecules they generate have problems such as poor synthesizability, unreasonable conformation, and low binding affinity to the target.

[0008] Therefore, how to solve the problem that the existing drug molecule generation methods are limited by the scarcity of training samples, resulting in poor synthesizability, unreasonable conformation, and low binding affinity with the target, is an important issue that needs to be urgently solved in the field of drug design. Summary of the invention

[0009] The present invention provides a small molecule generation method, device and electronic device based on cross-modal pre-training, which are used to overcome the defects of existing drug molecule generation methods that are limited by the scarcity of training samples, resulting in poor synthesizability, unreasonable conformation and low binding affinity with targets. It can better unsupervisedly learn the chemical structure and conformational information of small molecules, generate drug small molecules with more reasonable geometric structures and substructure distributions that are closer to real drug molecules, and improve the synthesizability of the generated drug small molecules and their affinity with targets.

[0010] On the one hand, the present invention provides a small molecule generation method based on cross-modal pre-training, comprising: obtaining a given protein pocket; obtaining a target representation vector of the protein pocket; based on a pre-trained molecular generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are encoded by a molecular encoder and the modal gap is reduced to simulate the representation vector of the protein pocket sample corresponding to the drug-like small molecule sample, and the molecular encoder is a pre-trained contrast learning model.

[0011] Furthermore, obtaining the target representation vector of the protein pocket includes: encoding the protein pocket into a corresponding initial representation vector based on a pre-trained protein encoder; collapsing the initial representation vector to obtain the target representation vector; wherein the pre-trained protein encoder is a pre-trained contrastive learning model, and the contrastive learning model is obtained by training and optimizing a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples.

[0012] Furthermore, the initial representation vector is collapsed to obtain the target representation vector, including: determining the protein pocket as the first modal data and obtaining the average value of the first modal data; subtracting the average value of the first modal data from the initial representation vector to obtain the target representation vector.

[0013] Furthermore, training and optimizing the molecular generation model specifically includes: based on a pre-trained molecular encoder, encoding the drug-like small molecule samples into corresponding initial representation vector samples; performing collapse and destruction processing on the initial representation vector samples to obtain representation vector samples; during training, using the representation vector samples and the atomic coordinates and atomic types of the drug-like small molecule samples as model inputs, using the generated small molecules as model outputs, and using the difference between the drug-like small molecule samples and the generated small molecules as training losses, and iteratively optimizing the molecular generation model.

[0014] Furthermore, the initial representation vector sample is collapsed and destroyed to obtain a representation vector sample, including: determining the drug-like small molecule as the second modal data, and obtaining the average value of the second modal data; subtracting the average value of the second modal data from the initial representation vector sample, and adding alignment noise to obtain the representation vector sample.

[0015] Furthermore, the alignment noise is obtained according to a budget of a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples.

[0016] In a second aspect, the present invention also provides a small molecule generation device based on cross-modal pre-training, comprising: a protein pocket acquisition module, used to acquire a given protein pocket; a representation vector acquisition module, used to acquire a target representation vector of the protein pocket; a ligand small molecule generation module, used to generate a ligand small molecule corresponding to the protein pocket according to the target representation vector based on a pre-trained molecular generation model; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are encoded by a molecular encoder and the modal gap is reduced to obtain a representation vector for simulating a protein pocket, and the molecular encoder is a pre-trained contrastive learning model.

[0017] In a third aspect, the present invention also provides an electronic 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 method for generating small molecules based on cross-modal pre-training as described above is implemented.

[0018] In a fourth aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for generating small molecules based on cross-modal pre-training.

[0019] In a fifth aspect, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described methods for generating small molecules based on cross-modal pre-training.

[0020] The small molecule generation method based on cross-modal pre-training provided by the present invention obtains a given protein pocket and a target representation vector of the protein pocket, and then generates a ligand small molecule corresponding to the protein pocket based on a pre-trained molecular generation model and the target representation vector; wherein the molecular generation model is trained and optimized based on a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, and the representation vector samples are encoded by a molecular encoder to simulate the representation vector of the protein pocket sample corresponding to the drug-like small molecule sample, and the molecular encoder is a pre-trained contrast learning model. The method aligns the protein pocket with the ligand small molecule in the representation space through the pre-trained contrast learning model, and C 3 The method introduces the field of molecular generation to reduce the modal gap between the representation vectors of proteins and small molecules, thereby using single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks. It can accurately capture the distribution of chemical structures and conformations of drug molecules, get rid of dependence on experimental data, generate drug small molecules with more reasonable geometric structures and distributions closer to real drug molecule structures, and improve the synthesizability of the generated drug small molecules and their affinity with the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0022] Figure 1 It is a flowchart of a small molecule generation method based on cross-modal pre-training provided in an embodiment of the present invention.

[0023] Figure 2 It is a schematic diagram of the overall process of the small molecule generation method based on cross-modal pre-training provided in an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of a ligand small molecule generated by the molecular generation model provided in an embodiment of the present invention during testing.

[0025] Figure 4 It is a bond length distribution difference diagram between the ligand small molecule generated by the molecular generation model provided in an embodiment of the present invention and the original molecule.

[0026] Figure 5 It is a diagram showing the difference in bond angle distribution between the ligand small molecule generated by the molecular generation model provided in the embodiment of the present invention and the original molecule.

[0027] Figure 6 It is a diagram showing the difference in torsion angle distribution between the ligand small molecule generated by the molecular generation model provided in the embodiment of the present invention and the original molecule.

[0028] Figure 7 It is a schematic diagram of the structure of a small molecule generation device based on cross-modal pre-training provided in an embodiment of the present invention.

[0029] Figure 8 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] It should be noted that the molecular generation model based on supervised learning in the existing technology only uses labeled protein pocket-ligand small molecule data. This type of molecular generation model is built on the experimental data space, and the quality of the generated molecules is seriously dependent on the size of the training set, but the data set available for training is very limited.

[0032] In view of this, the present invention proposes a small molecule generation method based on cross-modal pre-training, specifically, Figure 1 A schematic flow chart of a small molecule generation method based on cross-modal pre-training provided in an embodiment of the present invention is shown.

[0033] like Figure 1As shown, the method includes steps S110-S130, and steps S110-S130 and related steps will be described in detail below.

[0034] S110, get a given protein pocket.

[0035] Protein pockets, also known as protein binding pockets, refer to cavities on the surface or inside of proteins that are suitable for binding to ligands (such as small molecules, drugs or other biological molecules). The shape, position, physicochemical properties and functions of protein pockets are mainly determined by the amino acid residues around them. The dynamic properties of protein pockets, such as flexibility and movement, are crucial for the specific interaction of proteins. These properties allow protein pockets to open, close or adapt, thereby regulating the binding process of ligands and exerting specific protein functions.

[0036] It is easy to understand that the three-dimensional structure of a protein can be obtained from a protein database, and the structure file of the protein can be downloaded, and then various software tools can be used to identify pockets in the protein structure, that is, to obtain protein pockets.

[0037] Based on obtaining the given protein pocket in step S110, further, step S120 is performed.

[0038] S120, obtaining a target representation vector of the protein pocket.

[0039] It is easy to understand that features can be extracted from the structure of the protein pocket, where the features include but are not limited to amino acid type, charge, hydrophobicity, geometry, pocket volume and surface area, etc. Then, the extracted features are mapped to a representation vector, each feature will become an element in the representation vector, and the dimension of the representation vector is equal to the number of extracted features.

[0040] In a specific embodiment, a representation vector of a protein pocket is obtained based on a pre-trained encoder, that is, the protein pocket is used as the input of the pre-trained encoder to obtain an output representation vector. The pre-trained encoder can be a pre-trained contrastive learning model (DrugClip) or other network models that can achieve the same function, which is not specifically limited here.

[0041] It should be noted that after obtaining the preliminary representation vector of the protein pocket, considering the significant modal gap between the preliminary representation vector of the protein pocket and the representation vector of the ligand small molecule, a set preprocessing operation can be performed on the preliminary representation vector of the protein pocket, such as C 3 After the “collapse” process in , the final target representation vector can better help the molecular generation model to generate high-quality ligand small molecules.

[0042] Based on obtaining the target representation vector of the protein pocket in step S120, step S130 is further performed.

[0043] S130, based on a pre-trained molecular generation model, generate a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are encoded by a molecular encoder and the modal gap is reduced to simulate the representation vector of the protein pocket sample corresponding to the drug-like small molecule sample, and the molecular encoder is a pre-trained contrastive learning model.

[0044] It is easy to understand that a molecular generation model is pre-trained, which takes the target representation vector as input and the ligand small molecule corresponding to the protein pocket as output when applied. In particular, the molecular generation model can accurately capture the distribution of the chemical structure and conformation of the drug molecule, and the number distribution of substructures such as bond length, bond angle, torsion angle, etc. of the generated ligand small molecule is very close to the real drug small molecule.

[0045] It is worth mentioning that the molecular generation model proposed in this embodiment is obtained based on cross-modal pre-training. Specifically, protein pockets and ligand small molecules can be regarded as data of two different modalities. Ideally, the representation vectors of data of different modalities can be used interchangeably in the aligned representation space. The aligned representation space can enable cross-modal tasks to be learned from single-modal data.

[0046] Based on this inventive concept, this embodiment transforms the modeling in the data space of protein pocket-ligand small molecule into modeling in the aligned representation space, and uses the pre-trained contrastive learning model DrugClip to transform the protein pocket samples into Encoded as a representation vector , drug-like small molecule samples Encoded as a representation vector ,use simulation It can be used as training sample data for pre-training, thus getting rid of the dependence on scarce experimental data.

[0047] To expand, first collect ligand small molecule samples and their corresponding protein pocket samples, which can be obtained from publicly known protein-ligand complex databases (such as PDBbind). Then use the collected ligand small molecule samples and their corresponding protein pocket samples to pre-train the molecular encoder (DrugClip) to obtain a trained encoder, that is, a pre-trained comparative learning model. Next, use the trained encoder to encode a large number of collected drug-like small molecule samples to generate the corresponding representation vector samples.

[0048] Among them, DrugClip is a comparative learning model for drug design and virtual screening, which focuses on extracting features from the structural information of protein pockets to better predict the binding ability of small molecules to proteins. DrugClip includes graph neural networks and convolutional neural networks to achieve efficient and accurate drug design tasks.

[0049] When training the molecular generation model, the representation vector samples corresponding to the drug-like small molecule samples are used to simulate the representation vectors of the protein pocket samples. The representation vector samples and the original structures of the drug-like small molecules are used together as the training inputs of the molecular generation model to output the generated small molecules.

[0050] It should be noted that, since there is a significant modal gap between the representation vectors of different modalities, there is a gap between the representation vectors of protein pocket samples and the representation vector samples corresponding to drug-like small molecule samples. When using the representation vector samples corresponding to drug-like small molecule samples to simulate the representation vectors of protein pocket samples, it is preferred to introduce C 3 The method is used to reduce the modal gap between the representation vector of protein samples and the representation vector of drug-like small molecules (representation vector samples), so as to use single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks, get rid of the dependence on experimental data, and take into account the chemical structure rationality, synthesizability and affinity between the generated molecules and the target.

[0051] In this embodiment, a given protein pocket is obtained, and a target representation vector of the protein pocket is obtained, and then based on a pre-trained molecular generation model, a ligand small molecule corresponding to the protein pocket is generated according to the target representation vector; wherein the molecular generation model is trained and optimized according to a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, and the representation vector samples are encoded by a molecular encoder to simulate the representation vector of the protein pocket sample corresponding to the drug-like small molecule sample, and the molecular encoder is a pre-trained contrastive learning model. This method aligns the protein pocket with the ligand small molecule in the representation space through a pre-trained contrastive learning model, and C 3 The method introduces the field of molecular generation to reduce the modal gap between the representation vectors of proteins and small molecules, thereby using single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks. It can accurately capture the distribution of chemical structures and conformations of drug molecules, get rid of dependence on experimental data, generate drug small molecules with more reasonable geometric structures and distributions closer to real drug molecule structures, and improve the synthesizability of the generated drug small molecules and their affinity with the target.

[0052] On the basis of the above-mentioned embodiment, the process of obtaining the target representation vector of the protein pocket will be described in detail below.

[0053] Obtaining a target representation vector of a protein pocket includes: encoding the protein pocket into a corresponding initial representation vector based on a pre-trained protein encoder; collapsing the initial representation vector to obtain a target representation vector; wherein the pre-trained protein encoder is a pre-trained contrastive learning model, which is obtained by training and optimizing a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples.

[0054] It is easy to understand that when training the molecular generation model, the protein pocket and the ligand small molecule are aligned in the representation space through the pre-trained contrastive learning model, and the representation vector samples corresponding to the drug-like small molecule samples in the representation space are used to simulate the vector representation of the protein pocket samples, which is used as one of the training inputs of the molecular generation model to output the predicted generated small molecule.

[0055] When the molecular generation model is applied for actual reasoning, its input is the vector of the protein pocket in the representation space, that is, the target representation vector. Specifically, first, the protein pocket is encoded using a pre-trained protein encoder to obtain an initial representation vector. Then, as in the training, in order to reduce the modal gap between the representation vector of the protein pocket and the representation vector of the ligand small molecule, the initial representation vector of the protein pocket needs to be C-encoded. 3 The target representation vector can be obtained by the collapse process in .

[0056] The so-called collapse process refers to reducing the modal gap by subtracting the average value from the representation vector of each mode to eliminate the modal gap described above. In this embodiment, the protein pocket is determined as the first modal data, and the average value of the first modal data is obtained, and then the initial representation vector is Subtract the mean of the first mode data from , we can get the target representation vector .

[0057] Among them, the average value of the first modal data is pre-calculated and stored, and can be obtained by summing up the representation vectors of a large number of collected protein pocket samples dimension by dimension and dividing by the total number of samples, which will not be described in detail here.

[0058] It should be noted that the protein encoder in this embodiment is a pre-trained contrast learning model, which is obtained by training and optimization based on a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples, and is the same as the molecular encoder mentioned above, which will be elaborated in detail in the following embodiments.

[0059] In this embodiment, the protein pocket is encoded into a corresponding initial representation vector based on a pre-trained protein encoder, and the initial representation vector is collapsed to obtain a target representation vector, and then based on the pre-trained molecular generation model, the ligand small molecule corresponding to the protein pocket is generated according to the target representation vector. This method aligns the protein pocket with the ligand small molecule in the representation space through a pre-trained contrast learning model, and introduces the C3 method into the field of molecular generation to reduce the modal gap between the protein representation vector and the small molecule representation vector, so as to use the single-modal drug-like small molecule data to unsupervisedly learn the cross-modal protein-small molecule pairing task, which can accurately capture the distribution of the chemical structure and conformation of the drug molecule, get rid of the dependence on experimental data, generate a drug small molecule with a more reasonable geometric structure and a distribution closer to the real drug molecule structure, and improve the synthesizability of the generated drug small molecule and the affinity with the target.

[0060] On the basis of the above-mentioned embodiments, the process of training and optimizing the molecular generation model will be described in detail below.

[0061] The training and optimization of the molecular generation model specifically includes: based on a pre-trained molecular encoder, encoding the drug-like small molecule samples into corresponding initial representation vector samples; collapsing and destroying the initial representation vector samples to obtain representation vector samples; during training, using the representation vector samples, atomic coordinates and atomic types of drug-like small molecule samples as model inputs, using the generated small molecules as model outputs, and using the differences between drug-like small molecule samples and generated small molecules as training losses, and iteratively optimizing the molecular generation model.

[0062] It is easy to understand that to train a molecular generation model, one first needs to construct a training sample set. Specifically, a large number of drug-like small molecule samples are collected, and the drug-like small molecule samples are encoded using a pre-trained molecular encoder to obtain the corresponding initial representation vector samples.

[0063] Then, due to the large gap between the initial representation vector sample and the representation vector of the protein pocket sample to be simulated Therefore, it is necessary to introduce the "collapse" and "destruction" steps in the C3 method to improve the interchangeability between the representation vectors of protein pocket modal data and the representation vectors of small molecule modal data. The constant vector representing the modal gap between the representation vector of the protein pocket and the representation vector of the ligand small molecule is , The spanned surface is perpendicular, represents the alignment noise, which is a random vector that approximately follows a Gaussian distribution.

[0064] Specifically, the drug-like small molecule or ligand small molecule is determined as the second modal data, and the average value of the second modal data is obtained. Then, the average value of the second modal data is subtracted from the initial representation vector sample, and the alignment noise is added to obtain the representation vector sample.

[0065] Among them, the alignment noise can be estimated on the training sample set PDBBind consisting of annotated protein pocket samples and their corresponding ligand small molecule samples.

[0066] In this way, a training sample set of the molecular generation model can be constructed, including drug-like small molecule samples and their corresponding representation vector samples.

[0067] Next, the constructed training sample set is used to train and optimize the molecular generation model. Specifically, the drug-like small molecule samples (atomic coordinates and atomic types of drug-like small molecules) and their corresponding representation vector samples (representation vectors used to simulate protein pocket samples) are used as model input, the predicted generated small molecules are used as model output, and the difference between the drug-like small molecule samples and the generated small molecules is used as the training loss. The molecular generation model is iteratively optimized to obtain a trained molecular generation model.

[0068] The trained molecular generative model can then be used in the actual inference process.

[0069] In this embodiment, a drug-like small molecule sample is encoded into a corresponding initial representation vector sample based on a pre-trained molecular encoder, and the initial representation vector sample is collapsed and destroyed to obtain a representation vector sample. During training, the representation vector sample and the atomic coordinates and atomic type of the drug-like small molecule sample are used as model input, the generated small molecule is used as model output, and the difference between the drug-like small molecule sample and the generated small molecule is used as the training loss. The molecular generation model is iteratively optimized, and then based on the trained molecular generation model, the ligand small molecule corresponding to the protein pocket is generated according to the target representation vector. This method aligns protein pockets and ligand small molecules in the representation space through a pre-trained contrastive learning model, and introduces the C3 method into the field of molecule generation to reduce the modal gap between the representation vector of the protein and the representation vector of the small molecule, thereby using single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks. It can accurately capture the distribution of the chemical structure and conformation of drug molecules, get rid of dependence on experimental data, generate drug small molecules with more reasonable geometric structures and distributions closer to real drug molecule structures, and improve the synthesizability of the generated drug small molecules and their affinity with the target.

[0070] Additionally, in some embodiments, Figure 2 The overall process diagram of the small molecule generation method based on cross-modal pre-training provided by an embodiment of the present invention is shown. Figure 2 As shown, it includes the pre-training stage (solid line) and the sampling generation stage (dashed line).

[0071] In the pre-training stage, the pre-trained contrastive learning model DrugClip is used to encode drug-like small molecules (samples) into initial representation vector samples, and the C3 method is introduced to reduce the modal gap between the representation vector of the drug-like small molecules and the representation vector of the protein pocket to obtain the representation vector samples. Then, the representation vector samples and the atomic coordinates and atom types of the drug-like small molecules themselves are used as training inputs, and the predicted generated ligand small molecules are used as training outputs. The trained molecular generation model is obtained through iterative optimization.

[0072] In the sampling generation stage, for a given protein pocket, the contrastive learning model DrugClip is used to encode the protein pocket into an initial representation vector, and then the C3 method is introduced to collapse the initial representation vector to obtain the target representation vector. Next, the target representation vector is input into the trained molecular generation model to obtain the generated ligand small molecule.

[0073] It is worth mentioning that the small molecule generation method based on cross-modal pre-training provided in an embodiment of the present invention uses the pre-trained contrastive learning model DrugClip to align protein pockets and ligand small molecules in the representation space, and introduces the C3 method into the field of molecule generation, thereby reducing the modal gap between the representation vector of the protein and the representation vector of the small molecule, thereby using single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks, getting rid of dependence on experimental data, and taking into account the rationality of the chemical structure of the generated ligand small molecule, synthesizability and affinity with the target.

[0074] In some other embodiments, the unlabeled drug-like small molecule sample set used for pre-training the molecule generation model comes from the free commercial database - ZINC database. ZINC database is a small molecule compound database that has undergone series preparation and generated high-quality 3D conformations. It can be directly used for virtual screening such as molecular docking and is one of the essential databases for virtual screening.

[0075] In this example, 5M drug-like small molecules were randomly selected from the ZINC database as a training set, 1k drug-like small molecules as a validation set, and 100 drug-like small molecules as a test set. Compared with the crossdocked2020 dataset with only 100k protein-small molecule pairs, the training scale was greatly expanded.

[0076] Specifically, the contrastive learning model DrugClip is used to encode drug-like small molecule samples to obtain their initial representation vector samples, and the representation vector samples of drug-like small molecules are obtained through collapse (subtracting the mean) and destruction (adding Gaussian noise) operations. The representation vectors of protein pocket samples are used as conditions to simulate the representation vectors of drug-like small molecule samples and are input into the molecular generation model together with the atomic coordinates and atomic types of the drug-like small molecule samples. The molecular generation model is trained to unsupervisedly learn the correspondence between the representation vector samples and the atomic coordinates and atomic types of the drug-like small molecule samples.

[0077] Figure 3 Also shown is a schematic diagram of the ligand small molecule generated by the molecular generation model provided in the embodiment of the present invention during testing. Figure 3 As shown in the figure, the generated ligand small molecules have no "illegal" phenomena such as structural distortion and breakage, and are relatively similar to the original molecules. This shows that the pre-trained molecular generation model has successfully learned the correspondence between the representation vector in the aligned representation space and the atomic type and atomic position representation of the small molecule, and can generate "legal" small molecules based on the representation vector of the original molecule.

[0078] 100 drug-like small molecules were randomly selected from the ZINC database as a test set. The bond length, bond angle, torsion angle and other properties of the original molecules and the ligand small molecules generated according to their representation vectors were tested respectively to verify whether the molecular generation model can better capture the distribution of the chemical structure and conformation of drug molecules.

[0079] Figure 4 The figure shows the difference in bond length distribution between the ligand small molecule (rep2mol) generated by the molecular generation model provided in the embodiment of the present invention and the original molecule (ref), Figure 5 The figure shows the difference in bond angle distribution between the ligand small molecule generated by the molecular generation model provided in the embodiment of the present invention and the original molecule. Figure 6 A diagram showing the difference in torsion angle distribution between the ligand small molecule generated by the molecular generation model provided in an embodiment of the present invention and the original molecule is shown.

[0080] Figure 4-Figure 6 Only the distribution of the four chemical bonds with the largest number of ligand small molecules and the original molecule is shown. Figure 4-Figure 6 It can be seen that the ligand small molecules generated by the molecular generation model provided in the embodiment of the present invention have a distribution of the number of substructures such as bond length, bond angle and torsion angle that are very close to real drug molecules, which shows that the molecular generation model captures the distribution of the chemical structure and conformation of the drug molecules.

[0081] Corresponding to the small molecule generation method based on cross-modal pre-training described in the above embodiments, an embodiment of the present invention also provides a small molecule generation device based on cross-modal pre-training.

[0082] Specifically, Figure 7 A schematic structural diagram of a small molecule generation device based on cross-modal pre-training provided in an embodiment of the present invention is shown.

[0083] like Figure 7 As shown, the device includes: a protein pocket acquisition module 710, used to acquire a given protein pocket; a representation vector acquisition module 720, used to acquire a target representation vector of the protein pocket; a ligand small molecule generation module 730, used to generate a ligand small molecule corresponding to the protein pocket according to the target representation vector based on a pre-trained molecular generation model; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are encoded by a molecular encoder and the modal gap is reduced to obtain the representation vector for simulating the protein pocket, and the molecular encoder is a pre-trained contrast learning model.

[0084] In this embodiment, a protein pocket acquisition module 710 is used to acquire a given protein pocket; a representation vector acquisition module 720 is used to acquire a target representation vector of a protein pocket; a ligand small molecule generation module 730 is used to generate a ligand small molecule corresponding to the protein pocket according to the target representation vector based on a pre-trained molecular generation model; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, and the representation vector samples are obtained by encoding with a molecular encoder and reducing the modal gap, and are used to simulate the representation vector of the protein pocket, and the molecular encoder is a pre-trained contrastive learning model. The device aligns the protein pocket with the ligand small molecule in the representation space through the pre-trained contrastive learning model, and C 3 The method introduces the field of molecular generation to reduce the modal gap between the representation vectors of proteins and small molecules, thereby using single-modal drug-like small molecule data for unsupervised learning of cross-modal protein-small molecule pairing tasks. It can accurately capture the distribution of chemical structures and conformations of drug molecules, get rid of dependence on experimental data, generate drug small molecules with more reasonable geometric structures and distributions closer to real drug molecule structures, and improve the synthesizability of the generated drug small molecules and their affinity with the target.

[0085] It should be noted that the small molecule generation device based on cross-modal pre-training provided in the embodiment of the present invention and the small molecule generation method based on cross-modal pre-training described in the above embodiments can be referred to each other, and will not be described in detail here.

[0086] Figure 8 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute a small molecule generation method based on cross-modal pre-training, the method comprising: obtaining a given protein pocket; obtaining a target representation vector of the protein pocket; based on a pre-trained molecular generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are obtained by encoding with a molecular encoder and reducing the modal gap, and are used to simulate the representation vector of the drug-like small molecule sample corresponding to the protein pocket sample, and the molecular encoder is a pre-trained contrast learning model.

[0087] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0088] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the small molecule generation method based on cross-modal pre-training provided by the above-mentioned methods, the method including: obtaining a given protein pocket; obtaining a target representation vector of the protein pocket; based on a pre-trained molecular generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are obtained by encoding with a molecular encoder and reducing the modal gap, and are used to simulate the representation vector of the drug-like small molecule sample corresponding to the protein pocket sample, and the molecular encoder is a pre-trained comparative learning model.

[0089] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the small molecule generation method based on cross-modal pre-training provided by the above-mentioned methods, the method comprising: obtaining a given protein pocket; obtaining a target representation vector of the protein pocket; based on a pre-trained molecular generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; wherein the molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples, the representation vector samples are obtained by encoding with a molecular encoder and reducing the modal gap, and are used to simulate the representation vector of the protein pocket sample corresponding to the drug-like small molecule sample, and the molecular encoder is a pre-trained comparative learning model.

[0090] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0091] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A small molecule generation method based on cross-modal pre-training, characterized in that: include: Get a given protein pocket; obtaining a target representation vector of the protein pocket; Based on the pre-trained molecular generation model, generating a ligand small molecule corresponding to the protein pocket according to the target representation vector; The molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples. The representation vector samples are encoded by a molecular encoder and the modal gap is reduced to simulate the representation vectors of protein pocket samples corresponding to the drug-like small molecule samples. The molecular encoder is a pre-trained contrastive learning model.

2. The method for generating small molecules by cross-modal pre-training according to claim 1, characterized in that: Obtaining a target representation vector of the protein pocket, comprising: Based on a pre-trained protein encoder, encoding the protein pocket into a corresponding initial representation vector; Performing collapse processing on the initial representation vector to obtain the target representation vector; The pre-trained protein encoder is a pre-trained contrastive learning model, which is obtained by training and optimizing a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples.

3. The method for generating small molecules by cross-modal pre-training according to claim 2, characterized in that: The initial representation vector is collapsed to obtain the target representation vector, comprising: determining the protein pocket as first modal data, and obtaining an average value of the first modal data; The average value of the first modality data is subtracted from the initial representation vector to obtain a target representation vector.

4. The method for generating small molecules by cross-modal pre-training according to any one of claims 1 to 3, characterized in that: Training and optimizing the molecular generation model specifically includes: Based on the pre-trained molecular encoder, the drug-like small molecule samples are encoded into corresponding initial representation vector samples; Performing collapse processing and destruction processing on the initial representation vector sample to obtain a representation vector sample; During training, the atomic coordinates and atomic types of the representation vector samples and drug-like small molecule samples are used as model inputs, the generated small molecules are used as model outputs, and the difference between the drug-like small molecule samples and the generated small molecules is used as training loss to iteratively optimize the molecular generation model.

5. The method for generating small molecules by cross-modal pre-training according to claim 4, characterized in that: The initial representation vector sample is subjected to collapse processing and destruction processing to obtain a representation vector sample, including: Determine the drug-like small molecule as the second modal data, and obtain an average value of the second modal data; The average value of the second modality data is subtracted from the initial representation vector sample, and alignment noise is added to obtain the representation vector sample.

6. The method for generating small molecules by cross-modal pre-training according to claim 5, characterized in that: The alignment noise is obtained according to the budget of a training sample set consisting of protein pocket samples and their corresponding ligand small molecule samples.

7. A small molecule generation device based on cross-modal pre-training, characterized in that: include: Protein pocket acquisition module, used to obtain a given protein pocket; A representation vector acquisition module, used to acquire a target representation vector of the protein pocket; A ligand small molecule generation module, used to generate a ligand small molecule corresponding to the protein pocket according to the target representation vector based on a pre-trained molecule generation model; The molecular generation model is obtained by training and optimizing a training sample set consisting of drug-like small molecule samples and their corresponding representation vector samples. The representation vector samples are encoded by a molecular encoder and the modal gap is reduced to simulate the representation vector of the protein pocket. The molecular encoder is a pre-trained contrastive learning model.

8. An electronic 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 method for generating small molecules based on cross-modal pre-training as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for generating small molecules based on cross-modal pre-training as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for generating small molecules based on cross-modal pre-training as claimed in any one of claims 1 to 6 is implemented.

Images