Method and device for predicting odor molecule and olfactory receptor reaction based on multi-modal deep learning network

Abstract

The application relates to a smell molecule and olfactory receptor reaction prediction method and device based on a multi-modal deep learning network. The method first uses the interaction data between smell molecules and olfactory receptors, extracts the physical and chemical descriptors of the smell molecules as the characteristics of the smell molecules, uses a deep learning network model to determine whether unknown smell molecules can stimulate olfactory receptors. Secondly, sequence data of the smell molecules are used to determine the possibility of the reaction between the smell molecules and each olfactory receptor through a network calculation model. Then, smell molecule structure image data are further introduced to calculate the reaction probability between the smell molecules and each olfactory receptor. Finally, in order to improve the accuracy and reliability of the recognition result, the obtained calculation results, i.e. two probability distributions, are analyzed to determine the final reaction olfactory receptor of the smell molecule. The application significantly improves the accuracy of the recognition of the smell molecules and the olfactory receptors, reduces the dependence on experimental methods, and provides a new tool for chemical and biological research.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of predicting the reaction of odor molecules and olfactory receptors, and particularly relates to a method and device for predicting the reaction of odor molecules and olfactory receptors based on a multi-modal deep learning network. BACKGROUND

[0002] Analyzing the interaction between odor molecules and olfactory receptors is crucial for drug design and fragrance development. In the field of drug design, accurately predicting the interaction between odor molecules and olfactory receptors is essential for discovering new drugs and optimizing lead compounds. In the fragrance industry, understanding and predicting how odor molecules interact with olfactory receptors can accelerate the discovery and development of new fragrances. In addition, the field of environmental science is also more concerned about the impact of chemical substances on the biological olfactory receptor system, which is of great significance for assessing environmental risks and developing safety standards.

[0003] In chemical and biological research, analyzing the olfactory receptors that react to odor molecules is important for understanding olfactory perception and studying cognitive functions and related diseases. The relationship between the structure of odor molecules and olfactory receptors is complex. Structurally similar odor molecules can react with different olfactory receptors, and structurally different odor molecules can also react with the same olfactory receptors. This complexity of the reaction between odor molecules and olfactory receptors allows the brain to trigger corresponding odors through the activation pattern of olfactory receptor combinations. Traditionally, the study of the interaction between odor molecules and olfactory receptors relies on experimental methods such as biological assays and physical and chemical analysis, but such methods are often time-consuming, costly, and difficult to expand to large-scale sample analysis. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a method and device for predicting the reaction of odor molecules and olfactory receptors based on a multi-modal deep learning network.

[0005] The purpose of the present application is achieved by the following technical solutions:

[0006] In a first aspect, a method for predicting the reaction of odor molecules and olfactory receptors based on a multi-modal deep learning network is provided,

[0007] The method for predicting the reaction of odor molecules and olfactory receptors based on a multi-modal deep learning network includes the following steps:

[0008] Step S1, constructing a first data set, wherein the first data set involves odor molecules reacting with olfactory receptors and odor molecules not reacting with olfactory receptors, each odor molecule having a plurality of physicochemical descriptors, the first data set including a training set and a test set, the physicochemical descriptors of each odor molecule being used as training data, the training set in the first data set involving two types of odor molecules, the first type of molecules being odor molecules reacting with olfactory receptors, and the second type of molecules being odor molecules not reacting with olfactory receptors, a deep learning network model being trained using the training set in the first data set, and the trained deep learning network model being tested using the test set in the first data set to construct an odor molecule-olfactory receptor reaction prediction model;

[0009] Step S2, based on the odor molecule-olfactory receptor reaction prediction model in step S1, selecting a plurality of odor molecules reacting with olfactory receptors to construct a second data set and a third data set, wherein the odor molecules in the second data set and the third data set have molecular sequence data and molecular structure images respectively, the molecular sequence data and the molecular structure images being used as feature representations of the corresponding odor molecules, the second data set and the third data set also including a training set and a test set, the molecular sequence data and the molecular structure images of the odor molecules being used as training data, a network calculation model and a deep neural network being trained respectively using the training set in the second data set and the third data set, and the trained network calculation model and deep neural network being tested respectively using the test set in the second data set and the third data set to construct a corresponding first odor molecule-olfactory receptor reaction prediction model and a second odor molecule-olfactory receptor reaction prediction model, and determine the probability distribution of all olfactory receptors based on the first odor molecule-olfactory receptor reaction prediction model and the second odor molecule-olfactory receptor reaction prediction model;

[0010] Step S3, inputting a set of odor molecules to be tested into the first odor molecule-olfactory receptor reaction prediction model and the second odor molecule-olfactory receptor reaction prediction model respectively to obtain probability values corresponding to the two groups of probability distributions, adding the two probability values and then taking the average to obtain an average probability value, setting a threshold value, comparing the average probability value with the threshold value to determine the reacting olfactory receptors of the odor molecules to be tested.

[0011] In some embodiments, the physicochemical descriptors include molecular weight, three-dimensional structural features of the molecule, molecular polarity, and lipid solubility parameters;

[0012] Each odor molecule m has a corresponding set of physicochemical descriptors

[0013]

[0014] wherein D is a set of odor molecules and physicochemical descriptors, M is a set of odor molecules that can stimulate olfactory receptors and odor molecules that cannot stimulate olfactory receptors, d is the number of physicochemical descriptors, x mi is the i-th physicochemical descriptor of odor molecule m.

[0015] In some embodiments, the deep learning network model in step S1 is F(·), W is the network parameter, and L(·) is the loss function, and the specific formula is as follows:

[0016]

[0017] wherein, is the predicted output of the network for odor molecule m, |M| is the number of odor molecules in the set M, is the actual label of the odor molecule, and l(·) is the loss function of a single sample.

[0018] In some embodiments, testing the trained deep learning network model using the test set in the first data set comprises:

[0019] The odor molecule to be tested m' is predicted using the trained deep learning network model F(·):

[0020]

[0021] wherein, is the predicted output of the odor molecule m', and by setting a suitable threshold value, whether the odor molecule m' can react with the olfactory receptor, is the physicochemical descriptor of the odor molecule m';

[0022] The deep learning network model comprises multiple layers of neurons and can capture the complex relationship between molecular features and olfactory receptor reactivity through nonlinear transformation.

[0023] In some embodiments, a corresponding first odor molecule and olfactory receptor reaction prediction model is constructed, and the probability distribution of all olfactory receptors based on the first odor molecule and olfactory receptor reaction prediction model is determined, comprising:

[0024] Step S211, select an odor molecule sample that can react with an olfactory receptor, use the molecular sequence data as its unique feature representation, and identify the olfactory receptor that the odor molecule can react with through the molecular sequence data, as formula (4), I is a data set containing odor molecules and their sequence data, and each odor molecule m has its unique molecular sequence data wherein [x m1 ,x m2,...x md ] is a molecular descriptor, [x md+1 ,x md+2 ,...x mg ] is a molecular fingerprint, and E is a set of odor molecules that can react with olfactory receptors:

[0025]

[0026] In step S212, the molecular sequence data is taken as a feature representation of the odor molecule, and the network computing model learns the complex mapping relationship between the odor molecule features and the corresponding olfactory receptors, thereby accurately identifying the reaction olfactory receptors of the odor molecule; as formula (5), Z(·) is the network computing model, and L(·) is the loss function:

[0027]

[0028] wherein θ is a model parameter, is the molecular sequence data, is the predicted output of the network for the odor molecule m, is the actual label of the odor molecule, represents whether the odor molecule m can react with the i-th olfactory receptor, N is a set of all olfactory receptors, and |N| is the number of olfactory receptors;

[0029] In step S213, for the odor molecule to be tested, the molecular sequence data of the odor molecule is taken as input, and the specific olfactory receptor that reacts with the odor molecule is identified by using the network computing model, as formula (6), for the odor molecule to be tested m', the trained network model Z(·) is used for prediction:

[0030]

[0031] wherein, is the predicted output of the odor molecule m', a corresponding first odor molecule and olfactory receptor reaction prediction model is constructed, and a probability distribution of all olfactory receptors based on the first odor molecule and olfactory receptor reaction prediction model is determined.

[0032] In some embodiments, the molecular structure image shows the planar structure information of the molecule, and the planar structure information of the molecule includes the connection mode between atoms, the type of chemical bonds, and the existence of rings.

[0033] In a second aspect, a smell molecule and olfactory receptor reaction prediction device based on a multi-modal deep learning network is provided, and the device comprises a memory and one or more processors, the memory storing executable code, and the processor executes the executable code to implement the smell molecule and olfactory receptor reaction prediction method based on the multi-modal deep learning network as described in any one of the first aspect.

[0034] The present application has the following beneficial effects: the present application first uses the interaction data between smell molecules and olfactory receptors, extracts the physicochemical descriptors of smell molecules as their features, uses a deep learning network model to determine whether unknown smell molecules can stimulate olfactory receptors, secondly, uses the molecular sequence data of smell molecules to determine the possibility of reaction between smell molecules and each olfactory receptor through a network calculation model, then further introduces the molecular structure image data of smell molecules to calculate the reaction probability between smell molecules and each olfactory receptor, and finally, to improve the accuracy and reliability of the recognition result, the final reaction olfactory receptor of the smell molecule is determined by analyzing the obtained calculation results, i.e. two probability distributions; the present application can accurately determine whether the molecule can react with the olfactory receptor by using the multi-modal deep learning network, and further identify the olfactory receptor with which the smell molecule reacts, which not only can process known molecular structures, but also can be generalized to unknown or complex molecular structures, thereby improving the applicability of the model. The present application uses a calculation method instead of experimental operation, reduces the consumption of laboratory resources, and reduces the cost of studying the reaction between smell molecules and olfactory receptors. In combination with advanced data processing technology and machine learning algorithm, the present application provides an efficient and accurate smell molecule and olfactory receptor reaction recognition method. In addition, the method of the present application can be widely applied in the fields of drug design, perfume development, environmental monitoring, etc., and provides a new technical means for the research and application in related fields, and has wide practical value. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0036] Figure 1 The flowchart of the smell molecule and olfactory receptor reaction prediction method based on the multi-modal deep learning network provided by the embodiments of the present application.

[0037] Figure 2 Part of the physicochemical descriptor graph provided by the embodiments of the present application.

[0038] Figure 3 Part of the molecular fingerprint provided for the embodiments of the present application.

[0039] Figure 4 Part of the molecular structure image provided for the embodiments of the present application.

[0040] Figure 5 The olfactory receptor type provided for the embodiments of the present application.

[0041] Figure 6 The ROC curve obtained by testing the network computing model provided for the embodiments of the present application based on molecular sequence data and molecular structure image respectively, and the ROC curve diagram obtained by analyzing the calculation results of the two ways.

[0042] Figure 7 The true positive rate value obtained by identifying the olfactory receptor based on the network computing model provided for the embodiments of the present application based on molecular sequence data and molecular structure image respectively, and the true positive rate value diagram obtained by analyzing the calculation results of the two ways.

[0043] Figure 8 The olfactory receptor category identification diagram provided for the embodiments of the present application.

[0044] Figure 9 The structure diagram of the odor molecule and olfactory receptor reaction prediction device based on the multi-modal deep learning network provided for the embodiments of the present application. DETAILED DESCRIPTION

[0045] In order to better understand the technical solutions of the present application, the embodiments of the present application will be described in detail below in combination with the drawings.

[0046] First, define the method of evaluating the performance of the model, respectively using accuracy, true positive rate and AUROC value for evaluation; need to evaluate from two aspects:

[0047] (1) The performance of the model to determine whether the odor molecule can react with the olfactory receptor; (2) The performance of the model to identify the specific olfactory receptor that the odor molecule reacts with.

[0048] The definitions of accuracy, true positive rate and AUROC value are as follows:

[0049] For evaluating the performance of the model to determine whether the odor molecule can react with the olfactory receptor, the accuracy is defined as follows:

[0050]

[0051] Wherein, k is the number of odor molecules correctly predicted whether it can react with the olfactory receptor, and n is the total number of test molecules.

[0052] For evaluating the performance of the model in identifying the specific olfactory receptors that the odor molecules react with, accuracy, true rate and AUROC value are used for evaluation.

[0053] 1) Accuracy is defined as follows:

[0054]

[0055] where n represents the number of test odor molecules, |N represents the number of olfactory receptor categories, c i represents the prediction of whether the i-th odor molecule can react with the identified olfactory receptor, where the number of correctly identified olfactory receptors, i.e. the number of olfactory receptors that can react with the odor molecule and the number of olfactory receptors that cannot react with the odor molecule are correctly identified, is higher, indicating that the overall identification performance of the model is better.

[0056] 2) True rate is defined as follows:

[0057]

[0058] where TP represents the number of true cases, i.e. the number of correctly identified reactive olfactory receptors, and FN represents the number of false negatives, i.e. the number of reactive olfactory receptors incorrectly identified as non-reactive olfactory receptors by the model. The value of TPR is between 0 and 1, and the higher the value, the better the model's ability to identify reactive olfactory receptors, and the better the model's performance.

[0059] 3) AUROC value is defined as follows:

[0060] AUROC (Area Under the Receiver Operating Characteristic Curve) is a statistical measure of the performance of a binary classification model, representing the area under the ROC curve, with a value between 0 and 1, and the higher the value, the better the model's classification performance, while the ROC (Receiver Operating Characteristic curve) curve is a graphical representation of the model's performance at different threshold settings by plotting the true rate against the false positive rate.

[0061] AUROC can be divided into two types: Micro-AUROC, the area under the curve (AUROC), and Macro-AUROC, the area under the curve (AUROC). Micro-AUROC first calculates the total number of true positives, false positives, and false negatives for all categories, then plots the ROC curve and calculates the AUROC value based on these totals. Micro-AUROC focuses on overall performance and is suitable for processing data sets with imbalanced categories. Macro-AUROC calculates the AUROC value for each category's ROC curve and then takes the average of these AUROC values. It gives equal weight to all categories, focusing on the performance of each category, and is suitable for cases with balanced categories. Finally, the larger the values ​​of Micro-AUROC and Macro-AUROC, the better the model performance.

[0062] We used specific examples of odor molecules and olfactory receptors, and selected a specific deep learning neural network for actual testing to verify the performance and feasibility of this solution. The specific experimental results are as follows.

[0063] In some embodiments, as Figure 1 — Figure 8 As shown, a method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network is provided. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network includes the following steps:

[0064] Step S1: constructing a first data set, wherein the first data set includes odor molecules that react with olfactory receptors and odor molecules that do not react with olfactory receptors, and each odor molecule has multiple physical and chemical descriptors.

[0065] In some embodiments, the physicochemical descriptors are a series of quantitative values ​​used to characterize the physical, chemical, geometric and electronic properties of molecules, including molecular weight, three-dimensional structural characteristics of molecules, molecular polarity, lipid solubility parameters, etc. Thousands of physicochemical descriptor values ​​are usually used to uniquely characterize molecules.

[0066] Each molecule m has a corresponding set of physical and chemical descriptors

[0067]

[0068] Where D is the set of odor molecules and physicochemical descriptors, which includes odor molecules and their corresponding physicochemical descriptors, M is the set of odor molecules that can stimulate olfactory receptors and those that cannot stimulate olfactory receptors, d is the number of physicochemical descriptors, and x mi is the i-th physicochemical descriptor of odor molecule m.

[0069] Figure 2Some of the physicochemical descriptors used in this example. The physicochemical descriptors used in this example were obtained from the alvaDesc tool based on the SMILES (Simplified Molecular-Input Line-Entry System) format string of the molecules.

[0070] The first dataset includes a training set and a test set, and the physicochemical descriptors of each odor molecule are used as training data.

[0071] In some embodiments, 4550 odor molecules are used as training data, including 2000 odor molecules that can react with olfactory receptors, 1550 odor molecules that cannot react with olfactory receptors, and 1247 physicochemical descriptors are used to represent the characteristics of each odor molecule. The label of each odor molecule is 1 or 0, 1 indicating that the odor molecule can react with the olfactory receptor, and 0 indicating that the odor molecule cannot react with the olfactory receptor.

[0072] The training set in the first dataset involves two types of odor molecules, the first type of odor molecule is an odor molecule that reacts with the olfactory receptor, and the second type of odor molecule is an odor molecule that does not react with the olfactory receptor. The training set in the first dataset is used to train the deep learning network model to learn the pattern that can distinguish whether the odor molecule can react with the olfactory receptor. During the training process, the deep learning network model simultaneously learns the characteristics of the odor molecules that can react with the olfactory receptor and the odor molecules that cannot react with the olfactory receptor, enhancing the generalization ability of the model.

[0073] In some embodiments, the ViT (Vision Transformer) network is used as a deep learning network model for model training, the ratio of training set to validation set is 8:2, the Adam optimizer is used, the learning rate is 0.001, the batch size is 64, the number of iterations is 2732, the training duration is 10 hours, and the test accuracy on the validation set is 93%.

[0074] The ViT network is a Transformer model applied to computer vision tasks, which converts images into a sequence of vectors by dividing the image into a series of small blocks and linearly mapping these small blocks into a fixed-dimensional vector sequence, thereby converting the image into a sequence of data that can be processed by the Transformer. The vectors are then processed by the encoder, and a multi-layer perceptron head is used for classification and other downstream tasks. The example converts the numerical sequence into a format suitable for processing by the ViT network, maps each number to a high-dimensional space, and converts the numerical sequence into a series of embedding vectors to achieve data feature extraction and training recognition tasks.

[0075] In some embodiments, the deep learning network model in step S1 is F(·), W is the network parameter, and L(·) is the loss function, and the specific formula is as follows:

[0076]

[0077] wherein, is the predicted output of the network for the odor molecule m, |M| is the number of odor molecules in the odor molecule set M, is the actual label of the odor molecule (1 or 0, i.e., can or cannot react with the olfactory receptor), and l(·) is the loss function of a single sample, such as mean square error loss.

[0078] The trained deep learning network model is tested using the test set in the first data set to construct an odor molecule and olfactory receptor reaction prediction model; according to the physicochemical descriptor data of the odor molecule to be tested, the inference ability of the trained deep learning network model is used to determine whether the odor molecule has the possibility of reacting with the olfactory receptor.

[0079] As formula (3), for the odor molecule m' to be tested, the trained deep learning network model F(·) is used for prediction:

[0080]

[0081] wherein, is the predicted output of the odor molecule m', and by setting a suitable threshold value, to obtain whether the odor molecule m' can react with the olfactory receptor, is the physicochemical descriptor of the odor molecule m'.

[0082] The deep learning network model includes multiple layers of neurons, which can capture the complex relationship between odor molecule features and olfactory receptor reactivity through nonlinear transformation. In addition, the deep learning network model is optimized through appropriate activation functions and loss functions to ensure the accuracy and reliability of the prediction results. The deep learning network model F(·) extracts odor molecule features and establishes a prediction model by learning a large amount of odor molecule and olfactory receptor interaction data D, and the training goal of the model is to minimize the loss function L(·) to improve the accuracy of the judgment of the odor molecule reactivity.

[0083] Model testing: The test set uses 1362 odor molecules that did not participate in the training to test, of which 700 odor molecules can react with the olfactory receptor and 662 odor molecules cannot react with the olfactory receptor. The accuracy on the test set is 87%, i.e., 87% of the odor molecules can be correctly predicted whether they can react with the olfactory receptor.

[0084] Step S2, based on the odor molecule and olfactory receptor reaction prediction model in step S1, select a plurality of odor molecules reacting with olfactory receptors to construct a second data set and a third data set, wherein the odor molecules in the second data set and the third data set have molecular sequence data and molecular structure image respectively, the molecular sequence data and the molecular structure image are used as the feature representation of the corresponding odor molecule, the second data set and the third data set also include training set and test set, the molecular sequence data and the molecular structure image of the odor molecule are used as the training data, the training set in the second data set and the third data set is used to train the network calculation model and the deep neural network respectively, and the test set in the second data set and the third data set is used to test the trained network calculation model and the deep neural network respectively, to construct the corresponding first odor molecule and olfactory receptor reaction prediction model and second odor molecule and olfactory receptor reaction prediction model, and determine the probability distribution of all olfactory receptors based on the first odor molecule and olfactory receptor reaction prediction model and the second odor molecule and olfactory receptor reaction prediction model.

[0085] In some embodiments, using molecular sequence data, the possibility of reaction between odor molecules and each olfactory receptor is judged by a network calculation model, including the following steps:

[0086] Step S211, select odor molecule samples capable of reacting with olfactory receptors;

[0087] In some embodiments, for each odor molecule, 1247 physicochemical descriptors and 166 molecular fingerprint data are used as molecular sequence data to further supplement the feature information of the odor molecule, and the molecular sequence data of the odor molecule is extracted from the odor molecule by using PaDEL tool, Figure 3 is part of the molecular sequence data used in this example.

[0088] Using molecular sequence data as its unique feature representation, the olfactory receptors that the odor molecule can react with are identified by molecular sequence data, as formula (4), I is a data set containing odor molecules and their sequence data, and each odor molecule m has its unique molecular sequence data where [x m1 ,x m2 ,...x md ] is the molecular physicochemical descriptor, [x md+1 ,x md+2 ,...x mg ] is the molecular fingerprint, and E is the set of odor molecules that can react with olfactory receptors:

[0089]

[0090] Step S212, the molecular sequence data is represented as the feature of the odor molecule, and the network calculation model learns the complex mapping relationship between the molecular feature and the corresponding olfactory receptor, so as to accurately identify the reaction olfactory receptor of the odor molecule; as formula (5), Z(·) is the network calculation model, and L(·) is the loss function:

[0091]

[0092] Wherein, θ is the model parameter, is the molecular sequence data, is the predicted output of the network for the odor molecule m, is the actual label of the odor molecule, indicates whether the odor molecule m can react with the i-th olfactory receptor, N is the set of all olfactory receptors, and |N| is the number of olfactory receptors;

[0093] In some embodiments, the sequence data of each odor molecule is represented as the feature of the molecule and as the training data of the network calculation model. In an example, 3240 odor molecules are used, and the part of the molecules can react with the olfactory receptors, corresponding to 52 olfactory receptors in total, and each odor molecule corresponds to one or more olfactory receptors. The extracted molecular sequence data is used as the input feature, and the DNN network is trained, the ratio of the training set to the verification set is 8:2, the Adam optimizer is used, the learning rate is 0.001, the batch size is 64, the iteration number is 3742, the training time is 6 hours, and the test accuracy on the verification set is 91%, indicating that the model can effectively learn the correlation between the molecular sequence feature and the reaction olfactory receptor.

[0094] Model test: test using odor molecules not participating in training to verify the model's ability to distinguish the reaction olfactory receptor of the molecule. 1356 odor molecules not participating in training are used for testing, and the accuracy is 86%, verifying the model's generalization ability on unknown data. And the micro-average area under the curve and the macro-average area under the curve are used to evaluate the model performance, Figure 6 (a) is the ROC curve of the test odor molecule of the example network, wherein the micro-average area under the curve is 0.873, and the macro-average area under the curve is 0.842, indicating that the trained network has good discrimination ability for the reaction olfactory receptor of the odor molecule. Figure 7 (a) is the true positive rate value obtained by identifying 52 olfactory receptors by the network of the present example, and has good prediction results for most olfactory receptors.

[0095] The corresponding first odor molecule and olfactory receptor reaction prediction model is completed.

[0096] Step S213, for the odor molecule to be tested, the molecular sequence data of the odor molecule is taken as input, and the specific olfactory receptor reacting with the odor molecule is identified by using the network calculation model, such as formula (6), for the odor molecule m', the trained network calculation model Z(·) is used for prediction:

[0097]

[0098] Wherein, is the predicted output of the odor molecule m';

[0099] The probability distribution of all olfactory receptors based on the first odor molecule and olfactory receptor reaction prediction model is determined.

[0100] In some embodiments, the possibility of reaction between the odor molecule and each olfactory receptor is determined by a deep neural network using molecular structure image data, including the following steps:

[0101] Step S221, select the odor molecule that can react with the olfactory receptor as the data set, and use the molecular structure image for the feature representation of the odor molecule. The molecular structure image can capture the planar structure information of the molecule, such as the connection mode between molecules and the type of chemical bond, as well as the characteristics of atomic type, chemical bond type and ring, etc., which are used as the features of each molecule and as the training data of the deep neural network. The instance uses the RDKit tool to obtain the molecular structure image from the SMILES format string of the molecule;

[0102] Step S222, use the structure image of each odor molecule as the training data of the deep neural network, use 3240 odor molecules as the training data, which correspond to 52 olfactory receptors in total, and each odor molecule corresponds to one or more olfactory receptors. The instance uses the ConvNeXt network for model training, the ratio of training set to validation set is 8:2, the Adam optimizer is used, the learning rate is 0.001, the batch size is 64, the iteration number is 3651, and the training time is 7.5 hours. The test accuracy on the validation set is 93%, indicating that the model has good discrimination ability for specific reaction olfactory receptors of odor molecules.

[0103] ConvNext network combines the advantages of convolutional neural network and Transformer to improve the performance of image classification task, adjusts the calculation ratio of different stages to match the pattern in Transformer, and uses deep separable convolution technology and large convolution kernel size to reduce the number of parameters and calculation while maintaining efficiency.

[0104] Model testing: 1356 odor molecules not involved in training are used for testing, and the test accuracy is 88%. Figure 6(b) is the ROC curve of the example network for the test molecules, where the micro-averaged area under the curve is 0.902 and the macro-averaged area under the curve is 0.863, indicating that the network has good recognition and discrimination ability for odorant molecules olfactory receptor reactivity. Figure 7 (b) is the true rate value obtained by the network of the present example for identifying 52 kinds of olfactory receptors, and has good prediction results for most olfactory receptors.

[0105] The corresponding second odorant molecule and olfactory receptor reaction prediction model is completed.

[0106] Step S223, for the odorant molecule to be tested, the deep neural network analyzes the molecular structure image features to obtain the corresponding reaction olfactory receptor. For the odorant molecule m' to be tested, the trained deep neural network G(·) is used for prediction and identification:

[0107]

[0108] Where ω is the model parameter, s m′ is the structure image of the odorant molecule m', is the predicted output result of the odorant molecule m'.

[0109] The probability distribution of all olfactory receptors based on the second odorant molecule and olfactory receptor reaction prediction model is determined.

[0110] Step S3, the set of molecules to be tested is respectively input into the first odorant molecule and olfactory receptor reaction prediction model and the second odorant molecule and olfactory receptor reaction prediction model, to obtain the probability values corresponding to the two groups of probability distributions, add the two probability values and then take the average to obtain the average probability value, set a threshold value, compare the average probability value with the threshold value, to determine the reaction olfactory receptor of the odorant molecule to be tested.

[0111] In some embodiments, the specific steps of determining the reaction olfactory receptor of the odorant molecule to be tested are as follows:

[0112] Step S31, for the odorant molecule m' to be tested, the network calculation model Z(·) is used to predict the probability distribution of all olfactory receptors based on the molecular sequence data The deep neural network G(·) is used to predict another group of probability distributions of all olfactory receptors based on the molecular structure image

[0113] Step S32, for the two groups of probability distributions of olfactory receptors obtained, the corresponding probability values are added and then averaged to obtain another group of probability distributions p m′ of all olfactory receptors, and a suitable threshold value T is set to select the final reaction olfactory receptor set R therefrom. As formula (8):

[0114]

[0115] wherein, and respectively represent the probability value corresponding to the i-th olfactory receptor predicted by Z(·) and G(·).

[0116] In some embodiments, based on the molecular sequence data, a DNN network is used to identify a group of probabilities of the odor molecule reacting with all olfactory receptors; then, based on the molecular structure image, a ConvNeXt network is used to identify another group of probabilities of the odor molecule reacting with all olfactory receptors. For the two groups of reaction probabilities, the corresponding probability values are added and averaged, and a threshold of 0.5 is set to determine the reaction olfactory receptor of the odor molecule.

[0117] The test is performed using 1356 odor molecules, and the accuracy is improved from 86% and 88% to 93%. Figure 6 (c) is the ROC curve of the method for olfactory receptors, wherein the micro-average area under the curve is 0.962, and the macro-average area under the curve is 0.924. Figure 7 (c) is the true rate value of each olfactory receptor tested by the method, which is significantly improved compared with the true rate value of judging the reaction olfactory receptor of the odor molecule by using only a single method.

[0118] The present application first uses the interaction data between odor molecules and olfactory receptors, extracts the physicochemical descriptors of odor molecules as their characteristics, uses a deep learning network model to judge whether unknown odor molecules can stimulate olfactory receptors, secondly, uses the molecular sequence data of odor molecules, calculates the possibility of the reaction between odor molecules and each olfactory receptor through a network calculation model, then further introduces the molecular structure image data of odor molecules, calculates the reaction probability between odor molecules and each olfactory receptor, and finally, to improve the accuracy and reliability of the identification result, the calculated results, i.e. two probability distributions, are analyzed to determine the final reaction olfactory receptor of the odor molecule; the present application can accurately judge whether the odor molecule can react with the olfactory receptor by using a multi-modal deep learning network, and further identify the olfactory receptor with which the odor molecule reacts, which not only can process known molecular structures, but also can be generalized to unknown or complex molecular structures, improving the applicability of the model. The present application uses a calculation method instead of experimental operation, reduces the consumption of laboratory resources, and reduces the cost of studying the reaction between odor molecules and olfactory receptors. Combined with advanced data processing technology and machine learning algorithm, the present application provides an efficient and accurate identification method for the reaction between odor molecules and olfactory receptors. In addition, the method of the present application can be widely applied in the fields of drug design, perfume development, environmental monitoring, etc., and provides a new technical means for the research and application in related fields, which has wide practical value.

[0119] Corresponding to the foregoing embodiment of the method for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network, the present application also provides an embodiment of an apparatus for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network.

[0120] Referring to Figure 9 The apparatus for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network provided by the embodiment of the present application comprises a memory and one or more processors, the memory stores executable code, and the processor, when executing the executable code, is configured to implement the method for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network in the foregoing embodiment.

[0121] The apparatus for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The apparatus embodiment can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability where the device is located. From the hardware level, as shown in Figure 9 As shown in the figure, it is a hardware structure diagram of the device with data processing capability where the apparatus for predicting the reaction of an odor molecule and an olfactory receptor based on a multi-modal deep learning network is located. In addition to the processor, the memory, the network interface, and the non-volatile memory shown in the figure, the device with data processing capability where the apparatus is located in the embodiment usually includes other hardware according to the actual functions of the device with data processing capability, and details are not described here. Figure 9

[0122] The implementation process of the functions and roles of each unit in the above apparatus is specifically described in the implementation process of the corresponding steps in the above method, and is not described here.

[0123] For the apparatus embodiment, since it basically corresponds to the method embodiment, the related parts can be referred to the part of the method embodiment. The apparatus embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e. they can be located in one place, or distributed on multiple network units. According to the actual needs, part or all of the modules can be selected to achieve the purpose of the present application scheme. Those skilled in the art can understand and implement it without creative labor.

[0124] ​The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the method for predicting the reaction between an odor molecule and an olfactory receptor based on a multi-modal deep learning network.

[0125] The computer readable storage medium can be an internal storage unit of any data processing capable device of any of the preceding embodiments, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, and the like. Further, the computer readable storage medium can include both an internal storage unit and an external storage device of any data processing capable device. The computer readable storage medium is used to store computer programs and other programs and data required by any data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0126] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising a" does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0127] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different than the order in the embodiments and still achieve the desired result. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be utilized or can be advantageous.

[0128] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different than the order in the embodiments and still achieve the desired result. In addition, the processes depicted in the figures do not necessarily require the particular order shown or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be utilized or can be advantageous.

[0129] The above merely provides the preferred embodiment of one or more embodiments of the present specification, and is not intended to limit one or more embodiments of the present specification. Any modification made within the spirit and principle of one or more embodiments of the present specification shall fall within the scope of the present specification.

Claims

1. A method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network, characterized in that: The following steps are involved: Step S1: constructing a first data set, wherein the first data set includes odor molecules that react with olfactory receptors and odor molecules that do not react with olfactory receptors, each odor molecule has multiple physicochemical descriptors, and the first data set includes a training set and a test set. The physicochemical descriptors of each odor molecule are used as training data. The training set in the first data set includes two types of odor molecules: a first type of molecule is an odor molecule that reacts with olfactory receptors, and a second type of molecule is an odor molecule that does not react with olfactory receptors. A deep learning network model is trained using the training set in the first data set, and the trained deep learning network model is tested using the test set in the first data set to construct a prediction model for the reaction of odor molecules to olfactory receptors. Step S2: Based on the odor molecule and olfactory receptor reaction prediction model in step S1, multiple odor molecules that react with olfactory receptors are selected to construct a second data set and a third data set, wherein the odor molecules in the second data set and the third data set respectively have molecular sequence data and molecular structure images, and the molecular sequence data and the molecular structure images are both used as feature representations of the corresponding odor molecules. The second data set and the third data set also include a training set and a test set. The molecular sequence data and the molecular structure images of the odor molecules are used as training data. The training sets in the second data set and the third data set are used to train a network computing model and a deep neural network, respectively. The trained network computing model and the deep neural network are tested using the test sets in the second data set and the third data set, respectively, to construct the corresponding first odor molecule and olfactory receptor reaction prediction model and the second odor molecule and olfactory receptor reaction prediction model, and determine the probability distribution of all olfactory receptors corresponding to the first odor molecule and olfactory receptor reaction prediction model and the second odor molecule and olfactory receptor reaction prediction model; In step S3, the set of odor molecules to be tested is input into the first odor molecule and olfactory receptor reaction prediction model and the second odor molecule and olfactory receptor reaction prediction model respectively to obtain the probability values ​​corresponding to the two sets of probability distributions, the two probability values ​​are added and then averaged to obtain the average probability value, a threshold is set, and the average probability value is compared with the threshold to determine the olfactory receptor that reacts to the odor molecules to be tested.

2. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network according to claim 1, characterized in that: The physicochemical descriptors include molecular weight, three-dimensional structural characteristics of the molecule, molecular polarity and lipid solubility parameters; Each odor molecule m has a corresponding set of physical and chemical descriptors Where D is the set of odor molecules and physicochemical descriptors, which includes odor molecules and their corresponding physicochemical descriptors, M is the set of odor molecules that can stimulate olfactory receptors and those that cannot stimulate olfactory receptors, d is the number of physicochemical descriptors, and x mi is the i-th physicochemical descriptor of odor molecule m.

3. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network according to claim 2, characterized in that: In step S1, the deep learning network model is F(·), W is the network parameter, and L(·) is the loss function. The specific formula is as follows: in, is the network's predicted output for odor molecule m, |M| is the number of odor molecules contained in the odor molecule set M, is the actual label of the odor molecule, and l(·) is the loss function for a single sample.

4. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network according to claim 3, characterized in that: Testing the trained deep learning network model using the test set in the first dataset includes: The odor molecule m′ to be tested is predicted using the trained deep learning network model F(·): in, is the predicted output of odor molecule m′, which is distinguished by setting an appropriate threshold Get whether the odor molecule m' can react with the olfactory receptor, is the physicochemical descriptor of the odor molecule m′; The deep learning network model includes multiple layers of neurons and can capture the complex relationship between odor molecule characteristics and olfactory receptor reactivity through nonlinear transformation.

5. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network according to claim 2, characterized in that: Constructing a corresponding prediction model for the reaction between the first odor molecule and the olfactory receptor, and determining the probability distribution of all olfactory receptors based on the prediction model for the reaction between the first odor molecule and the olfactory receptor includes: Step S211, select an odor molecule sample that can react with the olfactory receptor, use the odor molecule sequence data as its unique feature representation, and identify the olfactory receptor with which the odor molecule can react through the odor molecule sequence data, as shown in formula (4), I is a data set containing odor molecules and their sequence data, and each odor molecule m has its unique molecular sequence data where [x m1 ,x m2 ,...x md ] is the molecular physicochemical descriptor, [x md+1 ,x md+2 ,...x mg ] is the molecular fingerprint, and E is the set of odor molecules that can react with the olfactory receptors: In step S212, the molecular sequence data is used as the feature representation of the odor molecule, and the network computing model learns the complex mapping relationship between the odor molecule features and their corresponding olfactory receptors, thereby accurately identifying the olfactory receptors that respond to the odor molecule. As shown in formula (5), Z(·) is the network computing model and L(·) is the loss function: Among them, θ is the model parameter, is the molecular sequence data, is the network’s predicted output for the odor molecule m, is the actual label of the odor molecule, Indicates whether the odor molecule m can react with the i-th olfactory receptor, N is the set of all olfactory receptors, and |N| is the number of olfactory receptors; In step S213, for the odor molecule to be tested, the molecular sequence data of the odor molecule is used as input, and the network computing model is used to identify the specific olfactory receptor that reacts with the odor molecule, as shown in formula (6). For the odor molecule to be tested m′, the trained network model Z(·) is used to make a prediction: in, is the predicted output of the odor molecule m′, a corresponding first odor molecule and olfactory receptor reaction prediction model is constructed, and the probability distribution of all olfactory receptors based on the first odor molecule and olfactory receptor reaction prediction model is determined.

6. The method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network according to claim 1, characterized in that: The molecular structure image shows the planar structural information of the molecule, which includes the connection mode between atoms, the type of chemical bonds, and the existence of rings.

7. A device for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, it is used to implement the method for predicting the reaction between odor molecules and olfactory receptors based on a multimodal deep learning network as described in any one of claims 1 to 6.

Images