New Target Compound Activity Prediction Method and System Based on Deep Learning

Abstract

The present invention belongs to the field of compound activity prediction, and specifically relates to a method and system for predicting the activity of new target compounds based on deep learning. The method aims to address the problem of low accuracy in compound activity prediction results due to the concentration, small amount, imbalance, and high similarity of new target active compound data. The present invention comprises the following steps: extracting data from public literature, patents, and databases to construct a data set; using a literature segmentation method to segment the data set into a training set and a test set; converting the compounds into SMILES formulas; data cleaning and slicing; constructing and training a molecular activity prediction model based on a deep neural network; and using the trained model to predict the activity of molecules in the screening set. The present invention effectively reduces the various deviations caused by the small, unbalanced, and similar sample size of the active compound data set disclosed for a completely new target, improves prediction performance, and achieves fast and accurate predictions.

Description

Technical Field

[0001] The present invention belongs to the field of compound activity prediction, and specifically relates to a method and system for predicting the activity of new target compounds based on deep learning. Background Art

[0002] In recent years, researchers have introduced deep learning (DL) methods into drug development, particularly in the field of virtual screening, with promising results. Compared to traditional virtual screening methods, DL offers rapid speed, high accuracy, and broad applicability. Furthermore, DL's end-to-end model requires minimal domain expertise, making it easier to apply and particularly well-suited for large-scale screening (millions or more).

[0003] However, data for new drug development is expensive and scarce. According to statistics, the average cost of new drug development for the discovery, generation, and optimization of lead compounds is approximately $800 million. These stages ultimately yield 10,000 to 20,000 lead compounds, with an average cost of $40,000 to $80,000 per lead compound. Meanwhile, publicly available literature and patents typically list only a few dozen to a few hundred independent active compounds for a new target, far short of the data volume required for deep learning.

[0004] There are two types of imbalances in the public data of active compounds for new targets. The first type is more common and can be well illustrated by the following data: a typical high-throughput screening (HTS) library generally has 10 6 -10 7 compounds, drug-like compounds are about 10 due to different constraints. 18 -10 200 The number of lead compounds is generally about 10. 3 -10 5 This order of magnitude results in a significant difference in the ratio of potential positive to negative sample sets, which is reflected in real-world data (such as HTS results). On the other hand, publicly available literature or patents typically do not disclose the vast majority of negative reactions in HTS or biological experiments. Instead, they primarily list active positive samples and a small number of negative samples with similar structures to the positive samples. Therefore, if only the results from literature or patents were collected, the number of negative samples obtained might be less than 20% of the number of positive samples, or even less.

[0005] Compared to small and unbalanced data, the issue of high data similarity is more easily overlooked. High similarity can manifest in two ways: The first is high similarity between samples of the same nature: a positive or negative sample set often consists of several groups of compounds with extremely similar properties. The second manifestation is when negative samples are extremely similar to certain positive samples, which is called activity cliffs.

[0006] Publicly available datasets for novel targets (including patents, documents, and databases) present numerous challenges for deep learning models due to their sparse sample size, imbalance, and similarity. Some methods can address or mitigate certain deficiencies in the dataset, but these solutions are often constrained by other issues. For example, upsampling methods designed to address data imbalance often exacerbate similarity issues, while the "maximum unbiased set" approach to similarity reduction reduces the number of available samples. Currently, no single deep learning model can comprehensively address these issues. Summary of the Invention

[0007] In order to solve the above-mentioned problems in the prior art, namely, the problem that the new target active compound data is concentrated, the data volume is small, the data is unbalanced, and the similarity is high, resulting in low accuracy of compound activity prediction results, the present invention provides a new target compound activity prediction method based on deep learning, the new target compound activity prediction method comprising:

[0008] Step P10, converting the molecules of the compound to be screened into a set number of SMILES formulas;

[0009] Step P20, preprocessing the SMILES data to obtain a one-hot vector set of the compounds to be screened;

[0010] Step P30, constructing and training a molecular activity prediction model, and predicting the activity of the compound to be screened based on the one-hot vector set of the compound to be screened using the trained molecular activity prediction model;

[0011] Wherein, the training method of the molecular activity prediction model is:

[0012] Step T10, obtaining similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set;

[0013] Step T20, dividing the sample set into a test sample set and a training sample set by using a document segmentation method;

[0014] Step T30, converting the molecules of each sample in the sample set into a set number of SMILES formulas;

[0015] Step T40: perform data preprocessing on each SMILES data to obtain a preprocessed sample set;

[0016] Step T50: constructing a molecular activity prediction model based on a deep neural network, and performing iterative training of the model based on the preprocessed sample set to obtain a trained molecular activity prediction model.

[0017] In some preferred embodiments, the SMILES formula is a simplified molecular linear input specification that represents the compound molecule, and the set number of SMILES formulas is one of 32, 64, 128, 256, and 512.

[0018] In some preferred embodiments, the data preprocessing includes:

[0019] The cleaned data are obtained by replacing the independent groups / ions, chiral atoms and double-character elements of the SMILES data with the set characters;

[0020] Slicing the cleaned data through a sliding window of a set length, cutting the cleaned data into a plurality of string segments of a set length;

[0021] retaining one fragment in the same group among the plurality of character string fragments of set lengths and deleting identical fragments that appear in different groups, to obtain a merged and deleted fragment set;

[0022] Each fragment in the merged and deleted fragment set is converted into a one-hot vector to obtain a one-hot vector set of the compound to be screened, thereby completing the data preprocessing operation.

[0023] In some preferred embodiments, the term "the same group" means that the compound molecules / fragments are in the same active sample set or in the same inactive sample set;

[0024] The heterogeneous group means that the compound molecules / fragments appear in both the active sample set and the inactive sample set.

[0025] In some preferred embodiments, the similar compounds refer to compounds with the same skeleton after the compound molecules are divided into a skeleton and fragments smaller than a set threshold.

[0026] In some preferred embodiments, the skeleton refers to a plurality of fragments obtained by breaking a single bond that is not in a ring in a compound molecule, wherein the fragments whose mass is greater than half of the molecular mass of the compound are called the skeleton.

[0027] In some preferred embodiments, the document segmentation method is:

[0028] The sample set is divided into documents. If a document is divided into a test sample set, all the data extracted from the document are divided into the test sample set. If a document is divided into a training sample set, all the data extracted from the document are divided into the training sample set.

[0029] In some preferred embodiments, when the activity category of the compound molecule is positive, all of its corresponding fragments are positive.

[0030] In some preferred embodiments, the deep neural network is a deep neural network for supervised learning.

[0031] In some preferred embodiments, the deep neural network used for supervised learning is one of a CNN neural network, an RNN neural network, an LSTM neural network, a biLSTM neural network, a GRU neural network, or a deep neural network based on an attention or self-attention mechanism.

[0032] In some preferred embodiments, the activity prediction of the compound to be screened is performed using the trained molecular activity prediction model in step P30, including:

[0033] Converting the one-hot vectors corresponding to the preprocessed fragments of the compound to be screened into word vectors to obtain a word vector set of the compound to be screened;

[0034] Obtaining the activity prediction value corresponding to each word vector in the word vector set of the compound to be screened through the trained molecular activity prediction model;

[0035] The activity prediction values are weighted and averaged to obtain the activity prediction value of the compound to be screened.

[0036] Another aspect of the present invention provides a new target compound activity prediction system based on deep learning, the new target compound activity prediction system comprising:

[0037] The data acquisition module is used to obtain the compound molecules to be screened, as well as similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set;

[0038] A sample segmentation module, configured to divide the sample set into a test sample set and a training sample set by using a document segmentation method;

[0039] A molecule conversion module is used to convert the molecules of the compound to be screened into a set number of SMILES formulas, and to convert the molecules of each sample in the sample set into a set number of SMILES formulas;

[0040] The data preprocessing module is used to preprocess the SMILES data to obtain a one-hot vector set of the compounds to be screened, and to preprocess each SMILES data corresponding to the samples in the sample set to obtain a preprocessed sample set;

[0041] A model building and training module is configured to build a molecular activity prediction model based on a deep neural network, and iteratively train the model based on the preprocessed sample set to obtain a trained molecular activity prediction model;

[0042] The activity prediction module is configured to predict the activity of the compound to be screened based on the one-hot vector set corresponding to the pre-processed fragments of the compound to be screened through the trained molecular activity prediction model.

[0043] Beneficial effects of the present invention:

[0044] (1) The new target compound activity prediction method based on deep learning in the present invention is more effective than existing methods. Compared with existing methods, it achieves the highest accuracy, F1score and MCC values.

[0045] (2) The present invention provides a novel deep learning-based method for predicting target compound activity, providing a unique two-step data amplification method, wherein a compound can be represented as a set number of SMILES formulas, each of which can be cut to produce a number of fragments. This method can easily amplify active compound data by 10,000-fold or more, fully meeting the data volume requirements of any deep learning method.

[0046] (3) The new target compound activity prediction method based on deep learning of the present invention does not require any method for processing unbalanced data sets. However, even in extremely unbalanced data (positive samples: negative samples are approximately 1:60), the method of the present invention can effectively avoid the impact caused by data imbalance.

[0047] (4) The present invention uses a method for predicting the activity of new target compounds based on deep learning. It uses unique fragments as the final data samples and greatly reduces the similarity in the data set by merging and deleting fragments, thereby highlighting the differences in the compounds. At the same time, the present invention also uses the "literature segmentation" method to segment the training set and the validation set. This method is extremely simple to implement, but while ensuring that the similarity is not increased, it also ensures the randomness of the segmentation of the training set and the validation set, effectively avoiding the problem of generalization. In summary, the method of the present invention can avoid the final test results from having an inflated accuracy rate, and compounds that are similar but have very different activities can also be well distinguished. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0049] Figure 1 Schematic diagram of the process of the new target compound activity prediction method based on deep learning of the present invention;

[0050] Figure 2 This is a schematic diagram of the molecular activity prediction model framework of an embodiment of the new target compound activity prediction method based on deep learning of the present invention. DETAILED DESCRIPTION

[0051] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are intended only to illustrate the relevant invention and are not intended to limit the invention. It should also be noted that, for ease of description, only portions relevant to the relevant invention are shown in the accompanying drawings.

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

[0053] The present invention provides a deep learning-based method for predicting the activity of novel target compounds. This deep learning-based virtual screening approach differs from traditional virtual screening methods in that it extracts data from publicly available literature, patents, and databases, enabling direct prediction of compound activity based on the SMILES structure of the compound molecule. This method facilitates rapid data processing, saving both time and money. Furthermore, this method effectively mitigates biases caused by the sparse, unbalanced, and similar sample size of publicly available datasets of active compounds targeting novel targets, thereby improving prediction performance.

[0054] In addition, since the classifier is implemented based on a deep neural network, this method does not require manual setting of molecular features. Features related to compound activity can be automatically extracted through the deep neural network. This method can further improve classification accuracy and achieve better performance.

[0055] The present invention provides a method for predicting the activity of a new target compound based on deep learning, the method comprising:

[0056] Step P10, converting the molecules of the compound to be screened into a set number of SMILES formulas;

[0057] Step P20, preprocessing the SMILES data to obtain a one-hot vector set of the compounds to be screened;

[0058] Step P30, constructing and training a molecular activity prediction model, and predicting the activity of the compound to be screened based on the one-hot vector set of the compound to be screened using the trained molecular activity prediction model;

[0059] Wherein, the training method of the molecular activity prediction model is:

[0060] Step T10, obtaining similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set;

[0061] Step T20, dividing the sample set into a test sample set and a training sample set by using a document segmentation method;

[0062] Step T30, converting the molecules of each sample in the sample set into a set number of SMILES formulas;

[0063] Step T40: perform data preprocessing on each SMILES data to obtain a preprocessed sample set;

[0064] Step T50: constructing a molecular activity prediction model based on a deep neural network, and performing iterative training of the model based on the preprocessed sample set to obtain a trained molecular activity prediction model.

[0065] In order to more clearly illustrate the new target compound activity prediction method based on deep learning of the present invention, the following Figure 1 Each step in the embodiment of the present invention is described in detail.

[0066] The first embodiment of the present invention is a method for predicting the activity of new target compounds based on deep learning. The steps are described in detail as follows:

[0067] Step P10: converting the molecules of the compound to be screened into a set number of SMILES formulas.

[0068] SMILES formula, that is, using the simplified molecular input lineentry system (SMILES) to represent compound molecules.

[0069] In practical applications, other character strings may be used to represent compound molecules, such as SMART, InChI, etc., which are not described in detail in the present invention.

[0070] Each compound molecule can be converted into multiple SMILES formulas, and the number can be set as needed, such as 32, 64, 128, 256 and 512.

[0071] In practical applications, the number of SMILES formulas converted from a compound molecule can be set as needed, and the present invention does not limit this.

[0072] Step P20: preprocess the SMILES data to obtain a one-hot vector set of the compounds to be screened.

[0073] Step P30: construct and train a molecular activity prediction model, and predict the activity of the compound to be screened based on the one-hot vector set of the compound to be screened by using the trained molecular activity prediction model.

[0074] The activity prediction of the compounds to be screened is performed using the trained molecular activity prediction model, including:

[0075] Converting the one-hot vectors corresponding to the preprocessed fragments of the compound to be screened into word vectors to obtain a word vector set of the compound to be screened;

[0076] Obtaining the activity prediction value corresponding to each word vector in the word vector set of the compound to be screened through the trained molecular activity prediction model;

[0077] The activity prediction values are weighted and averaged to obtain the activity prediction value of the compound to be screened.

[0078] The trained molecular activity prediction model is used to predict the activity of each compound in the set of compounds to be screened, that is, the set of compounds to be screened is classified to obtain molecules that are active against the target.

[0079] Wherein, the training method of the molecular activity prediction model is:

[0080] Step T10: obtaining similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set.

[0081] From the data extracted from literature, patents and databases, only no more than 60 active and similar compounds are taken as samples, while all inactive compound molecules similar to the active compounds are taken as samples.

[0082] Compounds that are not similar to any samples in the sample set but meet the inclusion criteria can all be used as samples.

[0083] In one embodiment of the present invention, no more than 15 active and similar compounds are sampled.

[0084] In the same group, it means that the compound molecules / fragments are in the same active sample set or in the same inactive sample set; in different groups, it means that the compound molecules / fragments appear in both the active sample set and the inactive sample set.

[0085] Similar compounds are defined as compounds with the same skeleton after the compound molecule is divided into a skeleton and fragments smaller than a set threshold. In one embodiment of the present invention, the threshold is set to 3, i.e., no more than 2 fragments.

[0086] The definition of the skeleton is: breaking certain single bonds in the compound that are not in the ring to obtain several fragments, among which the largest fragment (the fragment mass is greater than half the molecular mass of the compound) is called the skeleton, and the rest are fragments.

[0087] Step T20: dividing the sample set into a test sample set and a training sample set by using a document segmentation method.

[0088] The document segmentation method is:

[0089] The sample set is divided into documents. If a document is divided into a test sample set, all the data extracted from the document are divided into the test sample set. If a document is divided into a training sample set, all the data extracted from the document are divided into the training sample set.

[0090] Step T30: convert the molecule of each sample in the sample set into a set number of SMILES formulas.

[0091] One compound molecule can be converted into a certain number of SMILES. This number can be set as needed and can be the same as or different from that in step P10.

[0092] Step T40: perform data preprocessing on each SMILES data to obtain a preprocessed sample set.

[0093] The data preprocessing in step P20 and step T40 specifically includes:

[0094] The cleaned data is obtained by replacing the individual groups / ions, chiral atoms and double-character elements of the SMILES data with the set characters.

[0095] Data cleaning can be done by representing each independent group, chiral atom and double-character element with one character, that is, atoms or groups marked with multiple characters or "[]" in the SMILES formula are replaced with a single character.

[0096] The cleaned data is sliced through a sliding window of a set length, and the cleaned data is cut into a plurality of character string segments of a set length.

[0097] In one embodiment of the present invention, data slicing can be performed using a fixed-length sliding window to slice the compound molecules into a number of fixed-length string segments. If a compound molecule is a positive sample, then all the segments obtained after the segmentation are also positive samples, and so on.

[0098] Among the plurality of character string fragments of set length, one fragment in the same group is retained, and identical fragments appearing in different groups are deleted to obtain a merged and deleted fragment set.

[0099] In one embodiment of the present invention, data merging and deletion refers to retaining only one sample set fragment in the same group (both active samples or both inactive samples); and deleting the same sample in different groups (appearing in both active and inactive sample sets).

[0100] Each fragment in the merged and deleted fragment set is converted into a one-hot vector to obtain a one-hot vector set of the compound to be screened, thereby completing the data preprocessing operation.

[0101] Step T50: constructing a molecular activity prediction model based on a deep neural network, and performing iterative training of the model based on the preprocessed sample set to obtain a trained molecular activity prediction model.

[0102] The deep neural network here refers to the deep neural network that can be used for supervised learning, including but not limited to CNN, RNN, LSTM, biLSTM, GRU, deep neural networks based on attention or self-attention mechanisms, etc.

[0103] In one embodiment of the present invention, Figure 2 As shown, it is a schematic diagram of the molecular activity prediction model framework of an embodiment of the new target compound activity prediction method based on deep learning of the present invention. The deep neural network used at this time is a biLSTM neural network. The molecular activity prediction model constructed based on the biLSTM neural network includes: an input layer, an Ebedding layer, a biLSTM layer, a fully connected layer and an output layer.

[0104] In one embodiment of the present invention, the sample set is divided into a test sample set and a training sample set by a literature segmentation method. The training sample set is used for training the molecular activity prediction model, and the test sample set is used for verifying the molecular activity prediction model.

[0105] During model training, the dropout method is used to prevent overfitting. The activation function of the fully connected layer is sigmoid, the optimization method is adam, and the loss function is the cross entropy loss function.

[0106] Although the various steps in the above embodiment are described in the above-mentioned order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple changes are within the scope of protection of the present invention.

[0107] The second embodiment of the present invention provides a new target compound activity prediction system based on deep learning, comprising:

[0108] The data acquisition module is used to obtain the compound molecules to be screened, as well as similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set;

[0109] A sample segmentation module, configured to divide the sample set into a test sample set and a training sample set by using a document segmentation method;

[0110] A molecule conversion module is used to convert the molecules of the compound to be screened into a set number of SMILES formulas, and to convert the molecules of each sample in the sample set into a set number of SMILES formulas;

[0111] The data preprocessing module is used to preprocess the SMILES data to obtain a one-hot vector set of the compounds to be screened, and to preprocess each SMILES data corresponding to the samples in the sample set to obtain a preprocessed sample set;

[0112] A model building and training module is configured to build a molecular activity prediction model based on a deep neural network, and iteratively train the model based on the preprocessed sample set to obtain a trained molecular activity prediction model;

[0113] The activity prediction module is configured to predict the activity of the compound to be screened based on the one-hot vector set corresponding to the pre-processed fragments of the compound to be screened through the trained molecular activity prediction model.

[0114] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0115] It should be noted that the deep learning-based new target compound activity prediction system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be combined into one module or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are merely for the purpose of distinguishing the modules or steps and are not to be considered as improper limitations of the present invention.

[0116] An electronic device according to a third embodiment of the present invention includes:

[0117] at least one processor; and

[0118] a memory communicatively connected to at least one of the processors; wherein,

[0119] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned new target compound activity prediction method based on deep learning.

[0120] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, which are used to be executed by the computer to implement the above-mentioned new target compound activity prediction method based on deep learning.

[0121] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes and related instructions of the storage device and processing device described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0122] Those skilled in the art should be able to appreciate that, in conjunction with the modules and method steps of each example described in the embodiments disclosed herein, it is possible to implement them with electronic hardware, computer software, or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0123] The following describes in detail the implementation process and effects of the technical solution of the present invention in a specific application.

[0124] Example 1: Performance Testing (Taking the PD-1 / PD-L1 Pathway Inhibitor Dataset as an Example)

[0125] Step T10: A total of 70 articles related to small molecule inhibitors of the PD-1 / PD-L1 pathway and the articles / patents cited therein were collected from Web of Science and PubMed. 4-15 positive samples were randomly selected from each article / patent (try to select from different backbone molecules). In addition, all negative samples were included as much as possible.

[0126] Because the "standards" for active compounds in literature, patents, etc. vary, it is necessary to set the threshold for compound inclusion based on each different standard. For example, use Table 1 to determine the inclusion criteria.

[0127] Table 1

[0128] Positive Negative IC50 <300nM >500nM Splenocyte proliferation >80％ <20％ Kd <200nm >8uM

[0129] At the same time, inactive compounds (ID2316) of the PD-1 / PD-L1 pathway published on PubChem are also included.

[0130] Finally, the overall dataset is shown in Table 2:

[0131] Table 2

[0132] Positive samples Negative samples quantity 543 29046

[0133] Step T20: The data in the literature / patent are segmented according to the “literature segmentation” method, while the data on PubChem are segmented according to the random segmentation method, with the ratio of training set to validation set being 7:3.

[0134] Step T30: Convert the data in the dataset into a specified number of SMILES equations. Two methods are used here: the GL method, which converts one molecule into one SMILES equation, and the GL+ method, which converts one molecule into 64 SMILES equations.

[0135] Step T40: Replace each SMILES formula in the training set with multiple characters or atoms or groups marked with "[]" with a single character. Then, using a sliding window method, cut the cleaned string into equal lengths. For the cut segments, only one identical segment in the same group is retained; identical segments in different groups are deleted to form a set of positive and negative segments. Finally, convert the strings in the segment set into one-hot vectors.

[0136] Step T50: Input the training set into the neural network for training. The neural network consists of an embedding layer, a biLSTM layer, a fully connected layer, and an output layer.

[0137] Step P10-Step P30: Take a molecule from the test set, clean it, cut it into fragments, convert it into a one-hot vector, and then input it into the trained neural network to predict its activity and compare it with the actual value.

[0138] Comparative experiment: The training set and test set data assigned in step T20 were converted into PubChemFingerPrinter format (881-dimensional binary vectors) and trained and tested using methods such as support vector machine (SVM), random forest (RF), and gradient boosting iterative decision tree (GBDT).

[0139] The experimental results are shown in Table 3:

[0140] Table 3

[0141] FP FN Acc(%) f1-score MCC GL+ method 26.9 27.0 99.39 0.8440 0.8409 GL method 37.8 34.4 99.19 0.7931 0.7890 Support Vector Machine 26.1 50.2 99.14 0.7651 0.7632 RF 19 64 99.06 0.7239 0.7277 GBDT 30.6 52.1 99.06 0.7495 0.7469

[0142] Note: All test results are the average of 10 training cycles.

[0143] As shown in the table above, the method of the present invention outperforms other methods in metrics such as Acc, f1-score, and MCC for both SMILES amplification methods. On the other hand, the above results show that for the SVM, RF, and GBDT methods, the number of samples falsely classified as negative (FN) is much greater than the number of samples falsely classified as positive (FP), which is caused by sample imbalance. However, the method of the present invention does not have this problem.

[0144] Example 2: Expansion of sample number

[0145] Step T10: A total of 70 articles related to small molecule inhibitors of the PD-1 / PD-L1 pathway and the articles / patents cited therein were collected from Web of Science and PubMed. 4-15 positive samples were randomly selected from each article / patent (try to select from different backbone molecules). In addition, all negative samples were included as much as possible.

[0146] Because the "standards" for active compounds in literature, patents, etc. vary, it is necessary to set a threshold for compound inclusion based on each different standard. For example, Table 1 above can be used to determine the inclusion criteria.

[0147] At the same time, inactive compounds (ID2316) of the PD-1 / PD-L1 pathway published on PubChem are also included.

[0148] Finally, the overall dataset is shown in Table 2 above.

[0149] Step T20: The data in the literature / patent are segmented according to the “literature segmentation” method, while the data on PubChem are segmented according to the random segmentation method, with the ratio of training set to test set being 7:3.

[0150] Step T30: Convert the data in the dataset into a specified number of SMILES equations. Two methods are used: the GL method, which converts one molecule into one SMILES equation, and the GL+ method, which converts one molecule into 512 SMILES equations.

[0151] Step T40: Replace each SMILES formula in the training set with multiple characters or atoms or groups marked with "[]" with a single character. Then, using a sliding window method, cut the cleaned string into equal lengths. For each cut, only one identical fragment within the same group is retained; identical fragments in different groups are deleted. Finally, a set of positive and negative fragments is formed.

[0152] The experimental results are shown in Table 4:

[0153] Table 4

[0154] Number of positive samples Number of negative samples Before amplification 543 29046 After amplification by GL method 23260 728508 After amplification by GL+ method 5267118 98812816

[0155] As shown in Table 4, using the method of the present invention, the number of positive samples (active compounds) increased by approximately 42 times when using only one amplification method (sliding window cutting), while using both amplification methods increased the number of positive samples by approximately 9,700 times. Taking into account the impact of changing parameters, this number can be even higher. At least a million-level data volume is already able to meet the data requirements of deep learning.

[0156] Example 3: Impact of Sample Imbalance (Taking the PD-1 / PD-L1 Pathway Inhibitor Dataset as an Example)

[0157] Step T10: A total of 70 articles related to small molecule inhibitors of the PD-1 / PD-L1 pathway and the articles / patents cited therein were collected from Web of Science and PubMed. 4-15 positive samples were randomly selected from each article / patent (try to select from different backbone molecules). In addition, all negative samples were included as much as possible.

[0158] Because the "standards" for active compounds in literature, patents, etc. vary, it is necessary to set a threshold for compound inclusion based on each different standard. For example, Table 1 above can be used to determine the inclusion criteria.

[0159] At the same time, inactive compounds (ID2316) of the PD-1 / PD-L1 pathway published on PubChem are also included.

[0160] Finally, the overall dataset is shown in Table 2 above.

[0161] Step T20': Split the literature / patent data using the "literature segmentation" method. Randomly select several samples from the PubChem data to achieve a near-balance between the number of negative and positive samples. These randomly selected samples are then allocated to the training and test sets in a 7:3 ratio.

[0162] At this point, the total number of samples in the data set is shown in Table 5:

[0163] Table 5

[0164] Positive samples Negative samples quantity 543 543

[0165] Thus, the present invention obtains a data set with a balanced number of positive and negative samples.

[0166] Step T30: Convert the data in the dataset into a specified number of SMILES equations. Two methods are used: the GL method, which converts one molecule into one SMILES equation, and the GL+ method, which converts one molecule into 64 SMILES equations in an unbalanced dataset and 512 SMILES equations in a balanced dataset.

[0167] Step T40: Replace each SMILES formula in the training set with multiple characters or atoms or groups marked with "[]" with a single character. Then, using a sliding window method, cut the cleaned string into equal lengths. For the cut segments, only one identical segment within the same group is retained; identical segments in different groups are deleted to form a set of positive and negative segments. Finally, convert the strings in the segment set into one-hot vectors.

[0168] Step T50: Input the training set into the neural network for training. The neural network consists of an embedding layer, a biLSTM layer, a fully connected layer, and an output layer.

[0169] Steps P10-P30: Take a molecule from the test set, clean it, cut it into fragments, convert it into a one-hot vector, and then input it into the trained neural network to predict its activity and compare it with the actual value.

[0170] Comparative experiment: The training set and test set data assigned in step T20' were converted into PubChemFingerPrinter format (881-dimensional binary vectors) and trained and tested using methods such as support vector machine (SVM), random forest (RF) and gradient boosted iterative decision tree (GBDT).

[0171] Experimental results: Table 6 lists the differences between balanced and unbalanced data:

[0172] Table 6

[0173]

[0174] Note: 1. All test results are the average of 10 training runs. 2. The results for the unbalanced dataset are the same as those in Example 1. 3. The results for the balanced dataset may be slightly better than those for the unbalanced dataset due to the different test set sizes.

[0175] As shown in the table above, even when the number of positive and negative samples in the data set is quite unbalanced (close to 60:1), the FN:FP of the SVM, RF and GBDT methods is about 2:1 or even 3:1, while when the data set is balanced, the FN:FP value is close to 1 or even less than 1. This shows that the three machine learning methods will be disturbed by the imbalanced data set, and thus tend to judge most of the samples in the test set as negative samples. It is worth noting that the method of the present invention does not use any method to balance the data, but whether in the balanced sample set or the unbalanced sample set, the FN:FP value of the two different examples of the method of the present invention is close to 1, indicating that the method of the present invention is not affected by data imbalance compared to the above three machine learning methods.

[0176] Example 4: Impact of Sample Similarity (Taking the PD-1 / PD-L1 Pathway Inhibitor Dataset as an Example)

[0177] Step T10: A total of 70 articles related to small molecule inhibitors of the PD-1 / PD-L1 pathway and the articles / patents cited therein were collected from Web of Science and PubMed. 4-15 positive samples were randomly selected from each article / patent (try to select from different backbone molecules). In addition, all negative samples were included as much as possible.

[0178] Because the "standards" for active compounds in literature, patents, etc. vary, it is necessary to set a threshold for compound inclusion based on each different standard. For example, Table 1 above can be used to determine the inclusion criteria.

[0179] At the same time, inactive compounds (ID2316) of the PD-1 / PD-L1 pathway published on PubChem are also included.

[0180] Finally, the overall dataset is shown in Table 2 above.

[0181] Step T20": Randomly select several samples from the PubChem data to balance the number of negative samples with the number of positive samples, as shown in the table below. Then randomly assign the positive and negative samples to the training and validation sets in a ratio of 7:3.

[0182] At this time, the total number of samples in the data set is as shown in the above Table 4. However, at this time, the division of the training set and the validation set of the present invention is random, and the number of positive and negative samples is balanced.

[0183] Step T30: Convert the data in the dataset into a specified number of SMILES equations. Two methods are used here: the GL method, which converts one molecule into one SMILES equation, and the GL+ method, which converts one molecule into 64 SMILES equations.

[0184] Step T40: Replace each SMILES formula in the training set with multiple characters or atoms or groups marked with "[]" with a single character. Then, using a sliding window method, cut the cleaned string into equal lengths. For the cut segments, only one identical segment in the same group is retained; identical segments in different groups are deleted to form a set of positive and negative segments. Finally, convert the strings in the segment set into one-hot vectors.

[0185] Step T50: Input the training set into the neural network for training. The neural network consists of an embedding layer, a biLSTM layer, a fully connected layer, and an output layer.

[0186] Steps P10-P30: Take a molecule from the test set, clean it, cut it into fragments, convert it into a one-hot vector, and then input it into the trained neural network to predict its activity and compare it with the actual value.

[0187] Comparative experiment: The training set and test set data allocated in step T20 were converted into PubChemFingerPrinter format (881-dimensional binary vectors) and trained and tested using methods such as support vector machine (SVM), random forest (RF), and gradient boosted iterative decision tree (GBDT).

[0188] Experimental results: Table 7 lists the differences between using literature segmentation and using random segmentation:

[0189] Table 7

[0190]

[0191] Note: 1. All test results are the average of 10 training runs. 2. The results of the literature segmentation method are the results of Example 3.

[0192] When two structurally similar compounds, A and B, belonging to the same group, are assigned to the training set and the test set, respectively (for example, A is assigned to the training set and B is assigned to the test set), A is correctly classified due to the training process. However, during the test, B's score will be similar to A's and will also be correctly classified. This type of correct classification is similar to putting the same sample into both the training set and the test set, which will lead to artificially high test results.

[0193] As shown in the table above, the results obtained using the random segmentation method are far superior to those using document segmentation, indicating that the random segmentation method does indeed inflate the results due to similarity issues. The reduced error rate reflects the extent of the influence of similarity. The results show that similarity has the greatest impact on the results of SVM, while the impact on the method of the present invention is the smallest. However, even so, the error caused by similarity is still over 25%. This also shows that the "document segmentation" method used in the present invention can effectively reduce the impact of similarity.

[0194] The terms "first", "second", etc. are used to distinguish similar objects, rather than to describe or indicate a particular order or sequence.

[0195] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0196] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for predicting the activity of new target compounds based on deep learning, characterized in that: The method for predicting the activity of new target compounds includes: Step P10, converting the molecules of the compound to be screened into a set number of SMILES formulas; Step P20, preprocessing the SMILES data to obtain a one-hot vector set of the compounds to be screened; Step P30, constructing and training a molecular activity prediction model, and predicting the activity of the compound to be screened based on the one-hot vector set of the compound to be screened using the trained molecular activity prediction model; Wherein, the training method of the molecular activity prediction model is: Step T10, obtaining similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set; Step T20, dividing the sample set into a test sample set and a training sample set by using a document segmentation method; Step T30, converting the molecules of each sample in the sample set into a set number of SMILES formulas; Step T40: performing data preprocessing on each SMILES data to obtain a preprocessed sample set; Step T50, constructing a molecular activity prediction model based on a deep neural network, and performing iterative training of the model based on the preprocessed sample set to obtain a trained molecular activity prediction model; The data preprocessing includes: The cleaned data are obtained by replacing the independent groups / ions, chiral atoms and double-character elements of the SMILES data with the set characters; Slicing the cleaned data through a sliding window of a set length, cutting the cleaned data into a plurality of string segments of a set length; retaining one fragment from the same group among the plurality of character string fragments of set lengths and deleting identical fragments that appear in different groups, thereby obtaining a merged and deleted fragment set; Each fragment in the merged and deleted fragment set is converted into a one-hot vector to obtain a one-hot vector set of the compound to be screened, thereby completing the data preprocessing operation.

2. The method for predicting the activity of new target compounds based on deep learning according to claim 1, characterized in that: The SMILES formula is a simplified molecular linear input specification for representing compound molecules, and the set number of SMILES formulas is one of 32, 64, 128, 256, and 512.

3. The method for predicting the activity of new target compounds based on deep learning according to claim 1, characterized in that: The term "same group" means that the compound molecules / fragments are in the same active sample set or in the same inactive sample set; The heterogeneous group means that the compound molecules / fragments appear in both the active sample set and the inactive sample set.

4. The method for predicting the activity of new target compounds based on deep learning according to claim 1, characterized in that: The similar compounds refer to compounds with the same skeleton after the compound molecules are divided into a skeleton and fragments smaller than a set threshold.

5. The method for predicting the activity of new target compounds based on deep learning according to claim 4, characterized in that: The skeleton refers to a number of fragments obtained by breaking single bonds that are not in a ring in a compound molecule, wherein the fragments whose mass is greater than half of the molecular mass of the compound are called the skeleton.

6. The method for predicting the activity of new target compounds based on deep learning according to claim 1, characterized in that: The document segmentation method is: The sample set is divided into documents. If a document is divided into a test sample set, all the data extracted from the document are divided into the test sample set. If a document is divided into a training sample set, all the data extracted from the document are divided into the training sample set.

7. The method for predicting the activity of new target compounds based on deep learning according to any one of claims 2 to 5, characterized in that: When the activity category of the compound molecule is positive, all of its corresponding fragments are positive.

8. The method for predicting the activity of new target compounds based on deep learning according to claim 1, characterized in that: In step P30, the activity of the compound to be screened is predicted using the trained molecular activity prediction model, including: Converting the one-hot vectors corresponding to the preprocessed fragments of the compound to be screened into word vectors to obtain a word vector set of the compound to be screened; Obtaining the activity prediction value corresponding to each word vector in the word vector set of the compound to be screened through the trained molecular activity prediction model; The activity prediction values are weighted and averaged to obtain the activity prediction value of the compound to be screened.

9. A new target compound activity prediction system based on deep learning, characterized in that: The new target compound activity prediction system includes: The data acquisition module is used to obtain the compound molecules to be screened, as well as similar active compound molecules in the literature that are not higher than a set threshold, all inactive compound molecules similar to the active compound, and a set number of non-similar compound molecules as a sample set; A sample segmentation module, configured to divide the sample set into a test sample set and a training sample set by using a document segmentation method; A molecule conversion module is used to convert the molecules of the compound to be screened into a set number of SMILES formulas, and to convert the molecules of each sample in the sample set into a set number of SMILES formulas; The data preprocessing module is used to preprocess the SMILES data to obtain a one-hot vector set of the compounds to be screened, and to preprocess each SMILES data corresponding to the samples in the sample set to obtain a preprocessed sample set; A model building and training module is configured to build a molecular activity prediction model based on a deep neural network, and iteratively train the model based on the preprocessed sample set to obtain a trained molecular activity prediction model; The activity prediction module is configured to predict the activity of the compounds to be screened based on the one-hot vector set corresponding to the pre-processed fragments of the compounds to be screened using the trained molecular activity prediction model; The data preprocessing includes: The cleaned data are obtained by replacing the independent groups / ions, chiral atoms and double-character elements of the SMILES data with the set characters; Slicing the cleaned data through a sliding window of a set length, cutting the cleaned data into a plurality of string segments of a set length; retaining one fragment from the same group among the plurality of character string fragments of set lengths and deleting identical fragments that appear in different groups, thereby obtaining a merged and deleted fragment set; Each fragment in the merged and deleted fragment set is converted into a one-hot vector to obtain a one-hot vector set of the compound to be screened, thereby completing the data preprocessing operation.

Images