Drug virtual screening method and device based on deep learning

Abstract

The application provides a drug virtual screening method and device based on deep learning, wherein the method comprises the following steps: inputting all candidate compound small molecules in a candidate molecule database into a pre-trained molecular encoder respectively to obtain molecular vectorization representation of the candidate compound small molecules; constructing an index structure corresponding to the candidate compound small molecules based on the molecular vectorization representation of the candidate compound small molecules; inputting a protein target to be matched into a pre-trained protein target encoder to obtain protein target vectorization representation corresponding to the protein target; and matching the protein target vectorization representation based on the index structure to obtain a target compound small molecule corresponding to the protein target. The method establishes a full-amount mapping function of the target and the molecule through a vector calculation mode, realizes high-throughput virtual screening in a second-level calculation time, realizes rapid virtual screening of a full-amount candidate molecule library, improves the precision of the virtual screening, and increases the possibility of drug discovery.

Description

Technical Field

[0001] The present invention relates to the technical field of drug screening, and in particular to a method and device for drug virtual screening based on deep learning. Background Art

[0002] Currently, drug discovery is extremely costly and time-consuming due to the unavoidable and cumbersome nature of biological experiments. This is primarily due to the sheer number of drug candidates, which require individual validation. The discovery of new drugs is a lengthy process, beginning with initial biological screening of a vast pool of small molecule drugs, followed by validation across multiple functional domains and clinical trials.

[0003] Virtual screening methods in computer-aided drug design are an important technical means for lead compound discovery. Traditional virtual screening primarily relies on molecular docking technology, using a molecular force field-based scoring function to calculate binding energies, thereby obtaining energy rankings for compound screening. However, this method is computationally very time-consuming. According to a 2021 paper published in Nature, using molecular docking to complete virtual screening of a library of 10 billion molecules would require approximately 3,000 years using a single-core CPU and cost approximately $800,000 in computational costs.

[0004] Considering the speed limitation, virtual screening is usually performed on a small subset of the molecular library. Although this shortens the calculation time, this method greatly limits the accuracy of virtual screening and reduces the possibility of finding good lead compounds.

[0005] Therefore, how to solve the problem that existing virtual screening methods cannot quickly and effectively screen the entire candidate molecule library is an important issue that needs to be urgently addressed in the field of virtual screening technology. Summary of the Invention

[0006] The present invention provides a deep learning-based drug virtual screening method and device, which is used to overcome the defect that existing virtual screening methods cannot quickly and effectively perform full virtual screening of candidate molecule libraries, realize rapid virtual screening of the entire candidate molecule library, improve the accuracy of virtual screening, and increase the possibility of drug discovery.

[0007] On the one hand, the present invention provides a deep learning-based drug virtual screening method, comprising: inputting all candidate compound small molecules in a candidate molecule database into a pre-trained molecular encoder respectively to obtain molecular vector representations corresponding to the candidate compound small molecules; constructing an index structure corresponding to the candidate compound small molecules based on the molecular vector representations corresponding to the candidate compound small molecules; determining a protein target to be matched, and inputting the protein target into a pre-trained protein target encoder to obtain a protein target vector representation corresponding to the protein target; based on the index structure, matching the target compound small molecule corresponding to the protein target according to the protein target vector representation; wherein, the molecular encoder and the protein target encoder are obtained by comparative learning and training optimization using the matched protein targets and their ligands as training sample data sets.

[0008] Furthermore, the index structure corresponding to the candidate compound molecule is constructed based on the molecular vectorized representation corresponding to the candidate compound small molecule, including: vector division of the molecular vectorized representation corresponding to the candidate compound small molecule to obtain a preset number of first vectorized representations of preset dimensions; clustering the first vectorized representations to obtain the cluster center of each first vectorized representation; and organizing the index structure according to the first vectorized representations of the preset dimensions and their cluster centers.

[0009] Furthermore, based on the index structure, the target compound small molecule corresponding to the protein target is obtained according to the matching of the protein target vectorized representation, including: dividing the protein target vectorized representation into the preset number of second vectorized representations of the preset dimensions; determining the distance between each of the second vectorized representations and each of the cluster centers and summing them to obtain an actual distance value; according to the actual distance value, prioritizing all candidate compound small molecules in the candidate molecule database, and outputting the target compound small molecule corresponding to the protein target; wherein, the target compound small molecule is a candidate compound small molecule whose priority ranking is before the preset ranking.

[0010] Furthermore, the molecular encoder and the protein target encoder are subjected to comparative learning training optimization, specifically including: obtaining a training sample data set, wherein the training sample data set includes matched protein targets and their corresponding small compound molecules; for each matched protein target, the matched protein target and the small compound molecule corresponding to the matched protein target in the training sample data set are used as positive samples, and the matched protein target and other small compound molecules that do not correspond to the matched protein target in the training sample data set are used as negative samples, and the molecular encoder and the protein target encoder are trained and optimized respectively by a preset loss function.

[0011] Furthermore, the comparative learning and training optimization of the molecular encoder and the protein target encoder may include: pre-training the molecular encoder and the protein target encoder;

[0012] Furthermore, the pre-training of the protein target encoder and the molecular encoder includes: constructing a pre-training sample data set, the pre-training sample data set including a compound molecule pre-training data set and a protein target pre-training data set; pre-training the molecular encoder according to the compound molecule pre-training data set, and pre-training the protein target encoder according to the protein target pre-training data set; wherein each small compound molecule in the compound molecule pre-training data set includes a corresponding atomic category and topological structure, and each protein target in the protein target pre-training data set includes a corresponding three-dimensional atomic coordinate and atomic category.

[0013] Furthermore, the molecular encoder is pre-trained according to the compound molecule pre-training dataset, including: for each one-dimensional serialized representation of the compound small molecule in the compound molecule pre-training dataset, randomly masking an atom constituting the compound small molecule, and using other atoms constituting the compound small molecule to predict the atomic category of the masked atom; and / or, for each two-dimensional topological representation of the compound small molecule in the compound molecule pre-training dataset, randomly masking an atom constituting the compound small molecule, and using other atoms constituting the compound small molecule to predict the atomic category of the masked atom; and / or, determining that each compound small molecule in the compound molecule pre-training dataset includes corresponding three-dimensional atomic coordinates, and performing the three-dimensional topological representation of the atom constituting the compound small molecule. The method comprises the steps of randomly adding noise to the target and predicting the coordinates of the atoms to which the noise is added to eliminate the noise added to the coordinates of the atoms constituting the small molecule of the compound, thereby completing the pre-training of the molecular encoder; accordingly, pre-training the protein target encoder according to the protein target pre-training dataset comprises: for each protein target in the protein target pre-training dataset, randomly masking an atom constituting the protein target, and predicting the atomic category of the masked atom using other atoms constituting the protein target; and / or randomly adding noise to the atomic coordinates constituting the protein target, and predicting the coordinates of the atoms to which the noise is added to eliminate the noise added to the atomic coordinates of the protein target, thereby completing the pre-training of the protein target encoder.

[0014] Furthermore, the molecular encoder is one or more of a one-dimensional encoder, a two-dimensional encoder, and a three-dimensional encoder, and the protein target encoder is one or more of a one-dimensional encoder, a two-dimensional encoder, and a three-dimensional encoder.

[0015] In the second aspect, the present invention also provides a drug virtual screening device based on deep learning, including: a molecular vector representation generation module, which is used to input all candidate compound small molecules in the candidate molecule database into a pre-trained molecular encoder respectively to obtain the molecular vector representation corresponding to the candidate compound small molecule; an index structure construction module, which is used to construct an index structure corresponding to the candidate compound small molecule based on the molecular vector representation corresponding to the candidate compound small molecule; a protein target vector representation generation module, which is used to determine the protein target to be matched, and input the protein target into a pre-trained protein target encoder to obtain the protein target vector representation corresponding to the protein target; a target compound small molecule matching module, which is used to match the target compound small molecule corresponding to the protein target based on the index structure according to the protein target vector representation; wherein, the molecular encoder and the protein target encoder are obtained by comparative learning and training optimization using the matched protein target and its ligand as training sample data sets.

[0016] 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 program, the deep learning-based drug virtual screening method as described above is implemented.

[0017] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the deep learning-based drug virtual screening methods described above.

[0018] The deep learning-based drug virtual screening method provided by the present invention obtains molecular vector representations corresponding to the candidate compound small molecules by inputting all candidate compound small molecules in the candidate molecule database into a pre-trained molecular encoder respectively, and constructs an index structure corresponding to the candidate compound small molecules based on the molecular vector representations corresponding to the candidate compound small molecules, determines the protein target to be matched, and inputs the protein target into a pre-trained protein target encoder to obtain a protein target vector representation corresponding to the protein target, thereby obtaining the target compound small molecule corresponding to the protein target based on the index structure and matching according to the protein target vector representation. This method obtains the molecular vector representation of the candidate compound small molecule through a pre-trained molecular encoder, and constructs an index structure corresponding to the candidate compound small molecule. The protein target vector representation of the protein target to be matched is obtained through a pre-trained protein target encoder, and the target compound small molecule corresponding to the protein target is matched from the index structure based on the protein target vector representation. This process establishes a full mapping function between the target and the molecule in a vector calculation mode, realizes high-throughput virtual screening within a computing time of seconds, realizes rapid virtual screening of the full candidate molecule library, improves the accuracy of virtual screening, and increases the possibility of drug discovery. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 any creative work.

[0020] Figure 1 A schematic diagram of the process of a traditional virtual screening method provided by the prior art;

[0021] Figure 2 A schematic diagram of the process of the deep learning-based drug virtual screening method provided by the present invention;

[0022] Figure 3 A schematic diagram of the overall process of the deep learning-based drug virtual screening method provided by the present invention;

[0023] Figure 4 A schematic diagram of the structure of a deep learning-based drug virtual screening device provided by the present invention;

[0024] Figure 5 This is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0026] It should be noted that the traditional virtual screening method relies on molecular docking technology. For the same protein target, it is necessary to calculate the binding energy between it and each small molecule in the candidate molecule library to complete the preliminary virtual screening. Specifically, Figure 1 A schematic flow chart of a traditional virtual screening method provided by the prior art is shown.

[0027] like Figure 1 As shown, the full candidate molecule library has approximately 10 20 to 10 60 For a known protein pocket target, it is necessary to combine it with the above 10 20 to 10 60 A virtual screening of drug-like compounds based on protein-ligand docking simulation is performed to screen out a preset number (e.g., 100) of candidate small molecules with the highest scores from the candidate molecule library. Then, a wet experiment is performed on the protein pocket target and the small molecules obtained by virtual screening to determine whether the protein pocket target and the small molecules obtained by virtual screening can be combined.

[0028] However, due to the huge number of candidate molecule libraries, if we want to complete the virtual screening of the entire candidate molecule library, the traditional virtual screening method is slow and the whole process takes a lot of time.

[0029] In order to shorten the time spent on virtual screening, existing technologies usually choose to perform virtual screening on a small subset of molecular libraries, but this method greatly limits the accuracy of virtual screening and reduces the possibility of finding good lead compounds.

[0030] In view of this, the present invention proposes a drug virtual screening method based on deep learning, specifically, Figure 2 A schematic diagram of the process of the deep learning-based drug virtual screening method provided by the present invention is shown.

[0031] like Figure 2 As shown, the method includes:

[0032] S210, inputting all candidate compound small molecules in the candidate molecule database into a pre-trained molecular encoder to obtain molecular vectorized representations corresponding to the candidate compound small molecules.

[0033] It can be understood that all candidate compound small molecules in the known candidate molecule database are input into the pre-trained molecular encoder one by one. After encoding processing by the molecular encoder, the molecular vector representation corresponding to each candidate compound small molecule can be obtained and stored.

[0034] Among them, the molecular encoder can be any one of a one-dimensional encoder, a two-dimensional encoder and a three-dimensional encoder, or a combination of multiple ones of a one-dimensional encoder, a two-dimensional encoder and a three-dimensional encoder. It can be set according to actual conditions and is not specifically limited here.

[0035] The one-dimensional encoder can be a one-dimensional convolutional neural network, a one-dimensional transformer, or a long short-term memory neural network; the two-dimensional encoder can be a graph neural network, a two-dimensional transformer, or a two-dimensional graph structure; the three-dimensional encoder can be a three-dimensional convolutional neural network based on a three-dimensional point cloud image, a three-dimensional equivariant graph neural network, or a three-dimensional transformer structure.

[0036] In a specific embodiment, the molecular encoder is a three-dimensional encoder, specifically a "Transformer + graph neural network structure".

[0037] It should be noted that the molecular encoder is pre-trained, and the molecular vector representation output by the molecular encoder contains information related to the structure and properties of the small molecule itself.

[0038] The training method of the molecular encoder is not specifically described here. For example, in a specific embodiment, a pre-training method is used to train and optimize the molecular encoder.

[0039] S220, constructing an index structure corresponding to the candidate compound small molecule based on the molecular vector representation corresponding to the candidate compound small molecule.

[0040] It can be understood that in step S210, the candidate compound small molecule is input into the pre-trained molecular encoder to obtain the molecular vector representation corresponding to the candidate compound small molecule. Further, based on all the obtained molecular vector representations, an index structure corresponding to the candidate compound small molecule is constructed and stored for subsequent use.

[0041] Construct an index structure corresponding to the candidate compound small molecule. Specifically, the FAISS (Facebook AI Similarity Search) of meta can be used to construct an index structure corresponding to the candidate compound small molecule.

[0042] Specifically, each molecular vector representation can be divided into a preset number of small molecule vector representations of the same dimension, and the preset number of small molecule vector representations of the same dimension can be clustered separately to obtain the cluster centers corresponding to the preset number of small molecule vector representations of the same dimension, thereby compressing the molecular vector representation of the original dimension into a dimension consistent with the preset number, which can greatly speed up the subsequent matching search.

[0043] Then, based on the vectorized representation of small molecules and their corresponding cluster centers, the index structure corresponding to the candidate compound small molecules can be organized.

[0044] The preset number and the specific dimensions of the divided small molecule vector representations can also be set according to actual conditions. For example, in a specific embodiment, each molecule vector representation is divided into four 32-dimensional small molecule vector representations.

[0045] FAISS is an open-source library for clustering and similarity search developed by the Facebook AI team. It provides efficient similarity search and clustering for dense vectors, supporting searches on billions of vectors. It is a mature approximate nearest neighbor search library. It includes multiple algorithms for searching vector sets of any size, as well as supporting code for algorithm evaluation and parameter tuning.

[0046] FAISS includes many ways to construct index structures, including but not limited to Euclidean distance calculation, dot product calculation, scalar quantization, and index exact test plus clustering.

[0047] It should also be noted that the index structure may be a tree structure, a spherical structure, or other structures, which are not specifically limited here.

[0048] S230, determining a protein target to be matched, and inputting the protein target into a pre-trained protein target encoder to obtain a protein target vectorized representation corresponding to the protein target.

[0049] It can be understood that, on the basis of determining the protein pocket target, i.e., the protein target, of the small molecule compound to be matched, the protein target to be matched is input into a pre-trained protein target encoder. After encoding processing by the protein target encoder, a protein target vector representation corresponding to the protein target can be obtained.

[0050] Among them, like the above-mentioned molecular encoder, the protein target encoder can be a one-dimensional encoder, a two-dimensional encoder, or a three-dimensional encoder. The specific setting can be made according to actual conditions and is not specifically limited here.

[0051] In a specific embodiment, the protein target encoder is a three-dimensional encoder, specifically a "Transformer + graph neural network structure".

[0052] It should be noted that the protein target encoder is pre-trained, and the protein target vector representation output by the protein target encoder contains information related to the structure and properties of the protein target itself.

[0053] The training method of the protein target encoder is not specifically described here. For example, in a specific embodiment, a pre-training method is used to train and optimize the protein target encoder.

[0054] It should also be noted that the protein target encoder in this step is independent of the above-mentioned molecular encoder, and the structures of the two can be the same or different, and the words are not specifically limited.

[0055] It should be noted that there is no strict order of execution between this step and the above-mentioned steps S210 and S220. That is, step S230 can be executed at the same time as steps S210 and S220, or before or after steps S210 and S220, and no specific limitation is made here.

[0056] S240, based on the index structure, matches the target compound small molecule corresponding to the protein target according to the vectorized representation of the protein target.

[0057] It can be understood that, based on the construction of the index structure corresponding to the candidate compound small molecule in step S220 and the obtaining of the protein target vectorized representation corresponding to the protein target to be matched in step S230, further, based on the constructed index structure, the index structure is quickly retrieved according to the protein target vectorized representation to match the target compound small molecule corresponding to the protein target.

[0058] It should be noted that the target compound small molecule is a compound small molecule obtained through preliminary virtual screening, and then a wet experiment can be performed on the protein target and the target compound small molecule to determine whether the protein target and the target compound small molecule can be combined.

[0059] The target compound small molecules corresponding to the protein target are obtained according to the matching of the protein target vector representation. Specifically, the obtained protein target vector representation can be divided into a preset number of vector representations corresponding to the above small molecule vector representation, and then the distance between the preset number of vector representations and the cluster center corresponding to the above small molecule vector representation is calculated, so that the target compound small molecules corresponding to the protein target can be screened out according to the distance.

[0060] There are multiple small molecules of the target compound, which can be set according to actual needs.

[0061] For example, in a specific embodiment, there are 100 target compound small molecules matched to the protein target.

[0062] In this embodiment, all candidate compound small molecules in the candidate molecule database are input into a pre-trained molecular encoder to obtain a molecular vector representation corresponding to the candidate compound small molecule, and based on the molecular vector representation corresponding to the candidate compound small molecule, an index structure corresponding to the candidate compound small molecule is constructed, the protein target to be matched is determined, and the protein target is input into the pre-trained protein target encoder to obtain a protein target vector representation corresponding to the protein target, so that based on the index structure, the target compound small molecule corresponding to the protein target is matched according to the protein target vector representation. This method obtains the molecular vector representation of the candidate compound small molecule through a pre-trained molecular encoder, and constructs an index structure corresponding to the candidate compound small molecule, obtains the protein target vector representation of the protein target to be matched through a pre-trained protein target encoder, and matches the target compound small molecule corresponding to the protein target from the index structure according to the protein target vector representation. This process establishes a full mapping function between the target and the molecule in a vector calculation mode, realizes high-throughput virtual screening in seconds of computing time, realizes rapid virtual screening of the full candidate molecule library, improves the accuracy of virtual screening, and increases the possibility of drug discovery.

[0063] On the basis of the above embodiments, further, based on the molecular vectorized representation corresponding to the candidate compound small molecule, an index structure corresponding to the candidate compound molecule is constructed, including: vector division of the molecular vectorized representation corresponding to the candidate compound small molecule to obtain a preset number of first vectorized representations of preset dimensions; clustering the first vectorized representations to obtain the cluster center of each first vectorized representation; and organizing the index structure according to the first vectorized representations of preset dimensions and their cluster centers.

[0064] It can be understood that after obtaining the molecular vector representations corresponding to all candidate compound small molecules through the molecular encoder, the molecular vector representation corresponding to each candidate compound small molecule is vectorized and divided. For better description, the vector representation after vector division is recorded as the first vector representation, that is, after vector division, a preset number of preset dimensions of the first vector representation can be obtained.

[0065] Among them, the preset number and preset dimension can be set according to specific circumstances and are not specifically limited here.

[0066] For example, in a specific embodiment, all molecular vectorized representations constitute a 128-dimensional floating-point number vector, which is divided into four 32-dimensional first vectorized representations.

[0067] Next, each first vectorized representation is clustered, that is, the cluster center of each first vectorized representation can be obtained, thereby compressing all molecular vectorized representations into data dimensions consistent with a preset number, which can greatly speed up the matching search.

[0068] Then, an index structure can be organized according to the first vectorized representation of the preset dimension and its cluster center.

[0069] Furthermore, based on obtaining a protein target vectorized representation of the protein target to be matched through the protein target, the target compound small molecule corresponding to the protein target is matched in the index structure constructed according to the protein target vectorized representation.

[0070] Based on the index structure, the target compound small molecule corresponding to the protein target is obtained according to the protein target vectorized representation matching, including: dividing the protein target vectorized representation into a preset number of second vectorized representations of preset dimensions; determining the distance between each second vectorized representation and each cluster center and summing them to obtain an actual distance value; according to the actual distance value, prioritizing all candidate compound small molecules in the candidate molecule database, and outputting the target compound small molecule corresponding to the protein target; wherein the target compound small molecule is a candidate compound small molecule whose priority ranking is before the preset ranking.

[0071] Specifically, the same vector division method as the molecular vector representation is used to divide the protein target vector representation into a preset number of second vector representations of preset dimensions, and the distance value between each second vector representation and 256 cluster centers is pre-calculated and stored.

[0072] When matching target compound small molecules from the index structure based on the vectorized representation of the protein target, it is only necessary to determine the distance between each second vectorized representation and the corresponding cluster center of the first vectorized representation and sum them up to obtain the distance between the protein target to be matched and a candidate compound small molecule, that is, the actual distance value.

[0073] It should be noted that since the distance between each second vectorized representation and the 256 cluster centers has been pre-calculated, the process of determining the distance between each second vectorized representation and the corresponding cluster center of the first vectorized representation and summing them up can be achieved by directly looking up the table and adding them up to obtain the actual distance value, which requires little calculation, is fast in calculation speed, and takes little time.

[0074] On the basis of calculating the actual distance values ​​mentioned above, priority sorting is performed according to the size of the actual distance values, and the priority sorting of the vectorized representation of molecules in the database can be obtained, that is, the priority sorting of the candidate compound small molecules in the candidate molecule database. The smaller the actual distance value, the higher the priority sorting. Conversely, the larger the actual distance value, the lower the priority sorting.

[0075] The preset ranking is determined according to actual needs, and then the candidate compound small molecules whose priority ranking is within the preset ranking are extracted from the priority ranking of the candidate compound small molecules, that is, the target compound small molecules.

[0076] In a specific embodiment, the molecular vector representation corresponding to a candidate compound small molecule is vector-partitioned to obtain four 32-dimensional first vector representations, namely (m1, m2, m3, m4), where m1, m2, m3, and m4 respectively represent the cluster ID of each first vector representation. These four first vector representations are clustered to obtain corresponding cluster centers, and the vector formed by the cluster centers of the four first vector representations is represented as (c1, c2, c3, c4).

[0077] The protein target vectorized representation corresponding to the protein target to be matched is vectorized to obtain four 32-dimensional second vectorized representations, namely (p1, p2, p3, p4), and a distance value d is calculated between the second vectorized representation and the 256 cluster centers and stored.

[0078] Then, for each candidate compound small molecule corresponding to the molecular vector representation, the distance between it and the protein target to be matched can be directly obtained by looking up the table and adding it. Specifically, the actual distance value D = (p1, c1) + (p2, c2) + (p3, c3) + (p4, c4), where p i ∈R 32 , c i ∈R 32 , d(p i , c i )=||p i -c i ||2.

[0079] In this embodiment, based on the molecular vector representation corresponding to the candidate compound small molecule, an index structure corresponding to the candidate compound small molecule is constructed, and based on the index structure, the target compound small molecule corresponding to the protein target is matched according to the protein target vector representation. In this process, the target compound small molecule corresponding to the protein target is matched from the index structure, and a full mapping function of the target and the molecule is established in a vector calculation mode, and high-throughput virtual screening is realized within a computing time of seconds, thereby realizing rapid virtual screening of the entire candidate molecule library, improving the accuracy of virtual screening, and increasing the possibility of drug discovery.

[0080] On the basis of the above embodiment, the molecular encoder and the protein target encoder are further subjected to comparative learning training optimization, specifically including: obtaining a training sample data set, the training sample data set including matched protein targets and their corresponding small compound molecules; for each matched protein target, the matched protein target and the small compound molecule corresponding to the matched protein target in the training sample data set are used as positive samples, and the matched protein target and other small compound molecules that do not correspond to the matched protein target in the training sample data set are used as negative samples, and the molecular encoder and protein target encoder are trained and optimized respectively using a preset loss function.

[0081] It is understood that before using the molecular encoder to obtain the molecular vector representation corresponding to the candidate small molecule and using the protein target encoder to obtain the protein target vector representation corresponding to the protein target to be matched, it is necessary to perform comparative learning and training on the molecular encoder and the protein target encoder to obtain the optimized molecular encoder and protein target encoder so that the meaning of the molecular vector representation output by the molecular encoder is closer to the meaning of the protein target vector representation output by the protein target encoder.

[0082] First, compounds with matched protein target-ligand compositions are collected based on the existing large database, and then these compounds are split to obtain matched protein targets and their matching small molecules, thereby obtaining a training sample data set.

[0083] Among them, existing large databases include but are not limited to Pdbbind, Biolip and chembl. Each training sample in the training sample dataset includes a matched protein target and its corresponding small molecule compound.

[0084] It should be noted that the matched protein target and its corresponding small compound molecule can be a real protein target and its corresponding small compound molecule, or a protein target and its corresponding small compound molecule obtained through computer simulation, and no specific limitation is made here.

[0085] The molecular encoder is subjected to comparative learning and training optimization. In each round of training, a batch data set is sampled from the full training sample data set. The batch data set includes a certain number of training samples. The number of training samples in each batch data set can be limited according to actual conditions and is not specifically limited here.

[0086] For the training samples in each batch of data sets, these training samples are regarded as positive samples, that is, the matched protein targets and their corresponding small compound molecules.

[0087] For each matched protein target in the training sample, the matched protein target and other small compound molecules in the batch data set that do not match the matched protein target are used as negative samples.

[0088] Similarly, for each small molecule compound in a training sample, the small molecule compound and other protein targets in the batch data set that do not match the small molecule compound are also negative samples.

[0089] It should be noted that in addition to being constructed by splitting the protein target-ligand composition, the above-mentioned negative samples can also be constructed by obtaining small molecules from external large-scale molecular databases, such as zinc.

[0090] Specifically, for each protein target p in the above training sample data set, assuming that its corresponding small molecule compound as a positive example is I, a small molecule compound m different from I can be randomly selected from an external large-scale molecular database. The protein target p and the small molecule compound m can be used as global random negative examples, thereby constructing a large number of negative samples that meet the requirements.

[0091] In each round of training of the molecular encoder, the small molecule compound is used as input and the corresponding molecular vector representation is used as output. The training goal is to shorten the distance between the small molecule compound and the matching protein target (i.e., positive sample) as much as possible, and to increase the distance between the small molecule compound and the unmatched protein target (i.e., negative sample). Specifically, the above-mentioned training goal can be achieved through a preset loss function.

[0092] In one embodiment, the preset loss function is a softmax loss function, and the training process of the molecular encoder can be regarded as a multi-classification task. The expression of the preset loss function is as follows:

[0093]

[0094] in,

[0095]

[0096] In the above preset loss function expression, i and j are training sample numbers, n is the total number of training samples, k is the current training sample number, and y is the training sample number. k is a multi-classification vector used to represent the number of positive and negative samples, p i is the vectorized representation of the protein target of the training sample number i, m j is the vectorized representation of the molecule with training sample number j.

[0097] The protein target encoder is trained and optimized through comparative learning. For each round of training, the training process of the molecular encoder is the same as above and will not be elaborated here.

[0098] It should be noted that the training processes of the molecular encoder and the protein target encoder are independent of each other.

[0099] In this embodiment, a training sample data set including matched protein targets and their corresponding small compound molecules is obtained. For each matched protein target, the matched protein target and the small compound molecules corresponding to the matched protein target in the training sample data set are used as positive samples, and the matched protein target and other small compound molecules that do not correspond to the matched protein target in the training sample training set are used as negative samples. The molecular encoder and the protein target encoder are trained and optimized respectively through a preset loss function. During the training and optimization process, the distance between the positive samples is shortened as much as possible, and the distance between the negative samples is widened, so that the meaning of the molecule vectorized representation output by the molecular encoder is as close as possible to the meaning of the protein target vectorized representation output by the protein target encoder, thereby assisting in the rapid virtual screening of the full candidate molecule library.

[0100] On the basis of the above embodiment, the molecular encoder and the protein target encoder are further subjected to comparative learning training optimization, which includes: pre-training the molecular encoder and the protein target encoder, specifically, constructing a pre-training sample data set, the pre-training sample data set including a compound molecule pre-training data set and a protein target pre-training data set; pre-training the molecular encoder according to the compound molecule pre-training data set, and pre-training the protein target encoder according to the protein target pre-training data set; wherein each compound small molecule in the compound molecule pre-training data set includes a corresponding atomic category and topological structure, and each protein target in the protein target pre-training data set includes a corresponding three-dimensional atomic coordinate and atomic category.

[0101] It is understandable that in order to ensure that the molecular vector representation output by the molecular encoder can better characterize the structure and properties of the candidate compound small molecule itself, and the protein target vector representation output by the protein target encoder can better characterize the structure and properties of the protein target to be matched itself, the molecular encoder and the protein target encoder can be pre-trained before comparative learning and training optimization of the molecular encoder and the protein target encoder.

[0102] Specifically, protein targets and small molecules are first collected from existing large databases to generate corresponding protein target pre-training datasets and compound molecule pre-training datasets. The protein target encoder is then pre-trained based on the protein target pre-training data, while the molecule encoder is pre-trained based on the compound molecule pre-training dataset.

[0103] It should be noted that each small molecule compound is composed of multiple atoms. Therefore, each small molecule compound includes two properties: atomic type and topological structure.

[0104] Each protein target is also composed of multiple atoms. Therefore, each protein target includes two properties: three-dimensional atomic coordinates and atomic category.

[0105] Pre-train the molecular encoder based on the compound molecule pre-training dataset. The pre-training tasks may include any one or more of the following:

[0106] (1) For the one-dimensional serialized representation of each small molecule in the compound molecule pre-training dataset, one atom constituting the small molecule is randomly masked, and the atomic category of the masked atom is predicted using the other atoms constituting the small molecule.

[0107] (2) For each compound molecule in the compound molecule pre-training dataset, a two-dimensional topological representation is obtained by randomly masking an atom that constitutes the compound molecule, and using the other atoms that constitute the compound molecule to predict the atomic category of the masked atom.

[0108] (3) Determine that each small compound molecule in the compound molecule pre-training data set includes corresponding three-dimensional atomic coordinates, randomly add noise to the atomic coordinates constituting the small compound molecule, and predict the atomic coordinates with added noise to eliminate the noise added to the atomic coordinates constituting the small compound molecule, thereby completing the pre-training of the molecular encoder.

[0109] Similarly, the protein target encoder is pre-trained based on the protein target pre-training dataset. The pre-training tasks can also include any one or both of the following:

[0110] (1) For each protein target in the protein target pre-training dataset, a two-dimensional topological representation is used to randomly mask an atom that constitutes the protein target, and the atomic category of the masked atom is predicted using the other atoms that constitute the protein target;

[0111] (2) Randomly add noise to the atomic coordinates that constitute the protein target, and predict the atomic coordinates to which the noise is added to eliminate the noise added to the atomic coordinates of the protein target, thereby completing the pre-training of the protein target encoder.

[0112] In a specific embodiment, an atom constituting a small molecule of a compound is randomly masked, and the atomic category of the masked atom is predicted using the other atoms constituting the small molecule of the compound. Each pre-training is performed by inputting the small molecule of the compound with one atom masked into the molecular encoder, and the predicted probability q of the atomic category of the masked atom being r is obtained. r , r∈1,2,3,...,C, r is the atomic category, and the loss function corresponding to pre-training is as follows:

[0113]

[0114] in,

[0115]

[0116] In the above loss function l1, C represents the total number of atomic categories, k represents the atomic category of the atom, and y k is the true label of atom category k, q k is the probability that the atomic class of the masked atom is k.

[0117] In another specific embodiment, when it is determined that a small molecule compound includes three-dimensional atomic coordinates, noise is randomly added to the atomic coordinates constituting the small molecule compound, and the small molecule compound after adding noise is input into the molecular encoder to obtain the atomic coordinates of the atoms to which the noise is added predicted by the molecular encoder. Combined with the real noise coordinates, that is, the real atomic coordinates y∈R of the noise-added atoms 3 , to eliminate the noise added to the atoms, thereby completing the pre-training of the molecular encoder.

[0118] The loss function involved in this process can be the standard mean square error (MSE) loss, as follows:

[0119]

[0120] in, are the true atomic coordinates with added noise, and y are the atomic coordinates with added noise predicted by the molecular encoder.

[0121] It should be noted that these two specific embodiments are also applicable to the pre-training process of the protein target encoder, which will not be described in detail here.

[0122] It should also be noted that the pre-training processes of the molecular encoder and the protein target encoder are independent of each other.

[0123] In this embodiment, by pre-training the molecular encoder and the protein target encoder before performing comparative learning training optimization on the molecular encoder and the protein target encoder, the molecular encoder and the protein target encoder can learn the interaction relationship between atoms, as well as the structure and properties of the small molecule compound and the protein target itself, thereby making the comparative learning training process of the molecular encoder and the protein target encoder smoother and more optimized.

[0124] Additionally, in one embodiment, Figure 3 The figure shows the overall process diagram of the deep learning-based drug virtual screening method provided by the present invention.

[0125] like Figure 3 As shown, first, compounds composed of real protein targets and ligands are obtained from the existing data set and split to obtain protein targets and corresponding small molecules to construct a training sample data set.

[0126] Among them, protein targets correspond to protein target encoders, and small molecule compounds correspond to molecular encoders.

[0127] It should be noted that the molecular encoder and protein target encoder in the figure are pre-trained.

[0128] Then, in Figure 3 Left: The constructed training sample dataset is used to perform comparative learning and training optimization on the molecular encoder and protein target encoder to obtain the trained molecular encoder and protein target encoder.

[0129] After the training is completed, billions of candidate molecules are input into the molecular encoder one by one in an offline environment to obtain the molecular vector representations M1, M2, M3…M corresponding to all candidate compound small molecules. n , and based on the vectorized representation of these molecules, the corresponding index structure is organized and stored.

[0130] In specific applications, after determining the protein target to be matched, the protein target is input into the protein target encoder to obtain a protein target vector representation P1 corresponding to the protein target.

[0131] Finally, according to the vectorized representation P1 of the protein target, a rapid search is performed from the constructed index structure, and the target compound small molecule M2 corresponding to the protein target to be matched is obtained by virtual screening.

[0132] In another embodiment, the recall rates of the deep learning-based drug virtual screening method provided by the present invention and the existing virtual screening methods were compared on a set data set, as shown in the following table:

[0133]

[0134] As can be seen from the table above, the drug screening method (DrugClip) provided by the present invention has a recall rate of 0.299. In contrast, the corresponding recall rates of existing machine learning-based virtual screening methods are only about 52% of DrugClip at best, the recall rates of existing commercial simulation software are only about 74% of DrugClip at best, and the recall rates of existing free simulation software are only about 23% of DrugClip. Clearly, the recall rate of the drug virtual screening method provided by the present invention is much higher than that of existing virtual screening methods.

[0135] That is to say, compared with the existing virtual screening methods, the present invention establishes a full mapping function of targets and molecules through a vector calculation model, realizes high-throughput virtual screening within a computing time of seconds, realizes rapid virtual screening of the entire candidate molecule library, improves the accuracy of virtual screening, and increases the possibility of drug discovery.

[0136] Figure 4 The schematic diagram of the structure of the drug virtual screening device based on deep learning provided by the present invention is shown. Figure 4 As shown, the device includes: a molecule vector representation generation module 410, which is used to input all candidate compound small molecules in the candidate molecule database into a pre-trained molecule encoder respectively to obtain the molecule vector representation corresponding to the candidate compound small molecule; an index structure construction module 420, which is used to construct the index structure corresponding to the candidate compound small molecule based on the molecule vector representation corresponding to the candidate compound small molecule; a protein target vector representation generation module 430, which is used to determine the protein target to be matched, and input the protein target into a pre-trained protein target encoder to obtain the protein target vector representation corresponding to the protein target; a target compound small molecule matching module 440, which is used to match the target compound small molecule corresponding to the protein target based on the index structure according to the protein target vector representation; wherein, the molecule encoder and the protein target encoder are obtained by comparative learning and training optimization by using the real matched protein targets and their ligands as training sample data sets.

[0137] In this embodiment, all candidate compound small molecules in the candidate molecule database are input into a pre-trained molecular encoder through the molecular vector representation generation module 410 to obtain the molecular vector representation corresponding to the candidate compound small molecules. The index structure construction module 420 constructs the index structure corresponding to the candidate compound small molecules based on the molecular vector representation corresponding to the candidate compound small molecules. The protein target vector representation generation module 430 determines the protein target to be matched, and inputs the protein target into the pre-trained protein target encoder to obtain the protein target vector representation corresponding to the protein target. Thus, the target compound small molecule matching module 440 matches the target compound small molecule corresponding to the protein target based on the index structure and the protein target vector representation. The device obtains the molecular vector representation of the candidate compound small molecule through a pre-trained molecular encoder, and constructs an index structure corresponding to the candidate compound small molecule. It obtains the protein target vector representation of the protein target to be matched through a pre-trained protein target encoder, and matches the target compound small molecule corresponding to the protein target from the index structure based on the protein target vector representation. This process establishes a full mapping function between the target and the molecule in a vector calculation mode, realizes high-throughput virtual screening within a computing time of seconds, realizes rapid virtual screening of the full candidate molecule library, improves the accuracy of virtual screening, and increases the possibility of drug discovery.

[0138] Figure 5 An example of a physical structure diagram of an electronic device is shown below. Figure 5As shown, the electronic device may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call the logic instructions in the memory 530 to execute a deep learning-based drug virtual screening method, which includes: inputting all candidate compound small molecules in the candidate molecule database into a pre-trained molecular encoder to obtain a molecular vectorized representation corresponding to the candidate compound small molecule; constructing an index structure corresponding to the candidate compound small molecule based on the molecular vectorized representation corresponding to the candidate compound small molecule; determining a protein target to be matched, and inputting the protein target into a pre-trained protein target encoder to obtain a protein target vectorized representation corresponding to the protein target; based on the index structure, matching the target compound small molecule corresponding to the protein target according to the protein target vectorized representation; wherein the molecular encoder and the protein target encoder are obtained by comparative learning and training optimization using real matched protein targets and their ligands as training sample data sets.

[0139] In addition, the logic instructions in the above-mentioned memory 530 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, 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. The computer software product is stored in a storage medium and includes several instructions for enabling 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: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0140] 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 deep learning-based drug virtual screening method provided by the above-mentioned methods, the method comprising: inputting all candidate compound small molecules in the candidate molecule database into a pre-trained molecular encoder respectively to obtain the molecular vectorized representation corresponding to the candidate compound small molecules; constructing an index structure corresponding to the candidate compound small molecules based on the molecular vectorized representation corresponding to the candidate compound small molecules; determining the protein target to be matched, and inputting the protein target into a pre-trained protein target encoder to obtain the protein target vectorized representation corresponding to the protein target; based on the index structure, matching the target compound small molecule corresponding to the protein target according to the protein target vectorized representation; wherein, the molecular encoder and the protein target encoder are obtained by comparative learning and training optimization using real matched protein targets and their ligands as training sample data sets.

[0141] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0142] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling 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 certain parts of the embodiments.

[0143] 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 various embodiments of the present invention.

Claims

1. A drug virtual screening method based on deep learning, characterized in that: include: Input all candidate small molecules in the candidate molecule database into a pre-trained molecular encoder to obtain a molecular vector representation corresponding to the candidate small molecules; Based on the molecular vector representation corresponding to the candidate small molecule, construct an index structure corresponding to the candidate small molecule; Determining a protein target to be matched, and inputting the protein target into a pre-trained protein target encoder to obtain a protein target vectorized representation corresponding to the protein target; Based on the index structure, obtaining a target compound small molecule corresponding to the protein target according to the vectorized representation of the protein target; The molecular encoder and the protein target encoder are obtained by performing comparative learning and training optimization using matched protein targets and their ligands as training sample data sets; The step of constructing an index structure corresponding to the candidate compound molecule based on the molecular vectorized representation corresponding to the candidate compound small molecule includes: Performing vector division on the molecular vector representation corresponding to the candidate small molecule compound to obtain a first vector representation of a preset number of preset dimensions; performing clustering processing on the first vectorized representations to obtain a cluster center of each first vectorized representation; Organizing the index structure according to the first vectorized representation of the preset dimension and its cluster center; The method of obtaining a target compound small molecule corresponding to the protein target based on the index structure and according to the vectorized representation of the protein target includes: Dividing the protein target vectorized representation into second vectorized representations of the preset number of preset dimensions; Determine the distance between each of the second vectorized representations and each of the cluster centers and sum them to obtain an actual distance value; Prioritizing all candidate small molecules in the candidate molecule database according to the actual distance value, and outputting the target small molecule corresponding to the protein target; The target small molecule compound is a candidate small molecule compound having a priority ranking higher than a preset ranking.

2. The drug virtual screening method based on deep learning according to claim 1, characterized in that The molecular encoder and the protein target encoder are subjected to comparative learning and training optimization, specifically including: Acquire a training sample data set, wherein the training sample data set includes matched protein targets and their corresponding small molecules; For each matched protein target, the matched protein target and the small compound molecule corresponding to the matched protein target in the training sample data set are used as positive samples, and the matched protein target and other small compound molecules that do not correspond to the matched protein target in the training sample data set are used as negative samples. The molecular encoder and the protein target encoder are trained and optimized respectively using a preset loss function.

3. The drug virtual screening method based on deep learning according to claim 2, characterized in that The comparative learning and training optimization of the molecular encoder and the protein target encoder previously includes: Pre-training the molecular encoder and the protein target encoder; 4. The drug virtual screening method based on deep learning according to claim 3, characterized in that The pre-training of the protein target encoder and the molecule encoder comprises: Constructing a pre-training sample data set, wherein the pre-training sample data set includes a compound molecule pre-training data set and a protein target pre-training data set; Pre-training the molecular encoder according to the compound molecule pre-training dataset, and pre-training the protein target encoder according to the protein target pre-training dataset; Each small compound molecule in the compound molecule pre-training dataset includes a corresponding atomic category and topological structure, and each protein target in the protein target pre-training dataset includes a corresponding three-dimensional atomic coordinate and atomic category.

5. The drug virtual screening method based on deep learning according to claim 4, characterized in that The pre-training of the molecular encoder according to the compound molecule pre-training dataset includes: For each one-dimensional serialized representation of a small compound molecule in the compound molecule pre-training dataset, randomly masking an atom constituting the small compound molecule, and using other atoms constituting the small compound molecule to predict the atomic category of the masked atom; and / or, For each compound molecule in the compound molecule pre-training dataset, a two-dimensional topological representation is obtained by randomly masking an atom constituting the compound molecule, and predicting the atomic category of the masked atom using other atoms constituting the compound molecule; and / or, Determining that each small compound molecule in the compound molecule pre-training data set includes corresponding three-dimensional atomic coordinates, randomly adding noise to the atomic coordinates constituting the small compound molecule, and predicting the atomic coordinates to which the noise is added to eliminate the noise added to the atomic coordinates constituting the small compound molecule, thereby completing pre-training of the molecular encoder; Accordingly, the pre-training of the protein target encoder according to the protein target pre-training dataset comprises: For each protein target in the protein target pre-training dataset, a two-dimensional topological representation is used to randomly mask an atom constituting the protein target, and the atomic category of the masked atom is predicted using other atoms constituting the protein target; And / or, randomly adding noise to the atomic coordinates constituting the protein target, and predicting the atomic coordinates to which the noise is added, so as to eliminate the noise added to the atomic coordinates of the protein target, thereby completing the pre-training of the protein target encoder.

6. The drug virtual screening method based on deep learning according to claim 1, characterized in that The molecular encoder is one or more of a one-dimensional encoder, a two-dimensional encoder, and a three-dimensional encoder, and the protein target encoder is one or more of a one-dimensional encoder, a two-dimensional encoder, and a three-dimensional encoder.

7. A drug virtual screening device based on deep learning, characterized in that: include: A molecular vectorized representation generation module is used to input all candidate small molecules in the candidate molecule database into a pre-trained molecular encoder to obtain the molecular vectorized representation corresponding to the candidate small molecules; An index structure building module, configured to build an index structure corresponding to the candidate compound small molecule based on the molecular vector representation corresponding to the candidate compound small molecule; A protein target vectorized representation generation module is used to determine a protein target to be matched, and input the protein target into a pre-trained protein target encoder to obtain a protein target vectorized representation corresponding to the protein target; A target compound small molecule matching module is used to obtain a target compound small molecule corresponding to the protein target based on the index structure and the vectorized representation of the protein target; The molecular encoder and the protein target encoder are obtained by performing comparative learning and training optimization using matched protein targets and their ligands as training sample data sets; The step of constructing an index structure corresponding to the candidate compound molecule based on the molecular vectorized representation corresponding to the candidate compound small molecule includes: Performing vector division on the molecular vector representation corresponding to the candidate small molecule compound to obtain a first vector representation of a preset number of preset dimensions; performing clustering processing on the first vectorized representations to obtain a cluster center of each first vectorized representation; Organizing the index structure according to the first vectorized representation of the preset dimension and its cluster center; The method of obtaining a target compound small molecule corresponding to the protein target based on the index structure and according to the vectorized representation of the protein target includes: Dividing the protein target vectorized representation into second vectorized representations of the preset number of preset dimensions; Determine the distance between each of the second vectorized representations and each of the cluster centers and sum them to obtain an actual distance value; Prioritizing all candidate small molecules in the candidate molecule database according to the actual distance value, and outputting the target small molecule corresponding to the protein target; The target small molecule compound is a candidate small molecule compound having a priority ranking higher than a preset ranking.

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 program, the steps of the deep learning-based drug virtual screening method according to any one of claims 1 to 6 are 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 steps of the deep learning-based drug virtual screening method according to any one of claims 1 to 6 are implemented.

Images