Olfactory receptor function prediction method based on transfer learning and molecular docking

Abstract

The invention discloses an olfactory receptor function prediction method based on transfer learning and molecular docking. The olfactory receptor function prediction method comprises the following steps: acquiring a source domain data set and a target domain data set; a feature extraction module is used for merging feature vectors and inputting the merged feature vectors into a full-connection neural network to build a function prediction model; pre-training the function prediction model by using the source domain data set, and finely adjusting the function prediction model by using the target domain data set to obtain a first-step transfer learning model; judging whether a target species of a to-be-predicted olfactory receptor function is consistent with a species corresponding to the target domain data set or not, and if yes, predicting the olfactory receptor function by using the first-step transfer learning model; if not, one or more steps of fine tuning are carried out on the basis of the first step of transfer learning model until a second step of transfer learning model for olfactory receptor function prediction is obtained. According to the method, experimental data, molecular simulation and deep learning technologies are effectively combined, and rapid and accurate prediction of the functions of the large-scale olfactory receptors is achieved.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics and artificial intelligence technology, and in particular to a method for predicting olfactory receptor function based on transfer learning and molecular docking. Background Art

[0002] Olfactory receptors (ORs) are a class of key proteins responsible for olfactory perception in organisms. They can specifically recognize and bind to various volatile organic compounds (VOCs). The functional characterization of ORs is of great significance for understanding the olfactory mechanism, environmental adaptability, and social behavior of organisms. However, due to the high sequence variability of ORs and the complexity and time-consuming nature of experimental methods, only a few ORs have been clearly characterized, and the functions of most ORs remain unknown.

[0003] Traditional OR function prediction methods mainly rely on experimental methods, mainly including constructing in vivo or in vitro protein expression systems for physiological and biochemical testing to determine the binding strength of OR and VOC, that is, the function of OR. Although these experimental methods can provide accurate functional information, they are limited by experimental conditions, cost and time, and are difficult to be widely used in large-scale ORs function screening.

[0004] With the development of computational biology, structure-based protein function prediction methods have gradually become a hot topic of research. In particular, molecular docking technology predicts the binding mode between proteins and ligands by simulating the spatial structural complementarity and energy minimization between them. Molecular docking technology plays an important role in the virtual screening process in drug discovery and is also used for the functional prediction of ORs to improve the efficiency and success rate of experiments. However, molecular docking technology also has certain limitations, such as the possibility of false positives in the calculation results, and for ORs with high sequence variability, a single molecular docking method is difficult to meet the requirements of prediction accuracy. In addition, due to the limitations of experimental data, it is challenging to build a comprehensive and accurate OR function prediction model.

[0005] As an important branch of artificial intelligence, deep learning (DL) has shown great potential in dealing with complex bioinformatics problems. DL technology can improve the accuracy of protein function prediction by learning data representation through multiple levels of abstraction. However, the training of DL models requires a large amount of labeled data, while the experimental data currently available for training is relatively limited, which limits the application of DL technology in OR function prediction.

[0006] In order to overcome the problem of insufficient data, transfer learning technology was proposed and applied to OR function prediction. Transfer learning allows a pre-trained model to be fine-tuned on a new task to solve the problem of data scarcity. By transferring the knowledge of the pre-trained model to a new dataset, the prediction performance of the model can be significantly improved, and better prediction results can be obtained even with less data.

[0007] Although some methods have been used to predict OR function, how to effectively combine experimental data, molecular simulation and deep learning technology to achieve rapid and accurate prediction of large-scale ORs functions remains an urgent problem to be solved. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a method for predicting olfactory receptor function based on transfer learning and molecular docking, which combines the advantages of a large amount of molecular docking data and limited experimental results, including a two-step transfer learning process, firstly adjusting the pre-trained model based on molecular docking data to a small-scale experimental data set through training, and then migrating the optimized model of one species to another species with partial experimental data. Compared with the known methods, this method shows superior prediction accuracy and generalization ability.

[0009] The present invention provides a method for predicting olfactory receptor function based on transfer learning and molecular docking, which is characterized by comprising:

[0010] Step 1: Obtain source domain dataset and target domain dataset;

[0011] Step 2: Use the feature extraction module to build a function prediction model by merging the feature vectors and inputting them into a fully connected neural network;

[0012] Step 3: pre-training the function prediction model using the source domain dataset, and fine-tuning the function prediction model using the target domain dataset so that the function prediction model can be predicted for the experimental results of function identification, thereby obtaining the first step transfer learning model;

[0013] Step 4: Determine whether the target species for which the olfactory receptor function is to be predicted is consistent with the species corresponding to the target domain data set. If they are consistent, use the first-step transfer learning model to predict the olfactory receptor function; if they are inconsistent, use the olfactory receptor function identification experimental data of the target species on the basis of the first-step transfer learning model to perform one or more steps of fine-tuning until a second-step transfer learning model that can be used to predict the olfactory receptor function of the target species is obtained, and then use the second-step transfer learning model to predict the olfactory receptor function.

[0014] Furthermore, large-scale molecular docking data are used as the source domain dataset.

[0015] Furthermore, the target domain dataset comes from existing batch function identification data.

[0016] Furthermore, a protein and small molecule language model based on a natural language processing method is used as the feature extraction module.

[0017] Furthermore, in step 2, the amino acid sequence of the olfactory receptor and the simplified molecular linear input specification of the small molecule are used as model inputs to obtain feature vector representations of the protein and the small molecule. The two feature vectors are then merged and input into a fully connected neural network, and the binding score between the protein and the small molecule is used as a label to build a functional prediction model.

[0018] Furthermore, a 5-layer deep fully connected neural network model is used to learn and fit the feature vector.

[0019] Furthermore, in the step three, the affinity between the protein and the small molecule is used as a label for the pre-training.

[0020] Furthermore, in the step 4, the species corresponding to the experimental data has a close relationship with the target species.

[0021] Preferably, for target species for which there is no experimental data for identification of olfactory receptor function, the experimental data sets of multiple different species are merged into a mixed data set and then the first step of transfer learning is performed, and the model obtained by the first step of transfer learning is used to predict the olfactory receptor function of the target species.

[0022] The olfactory receptor function prediction method based on transfer learning and molecular docking of the present invention effectively combines experimental data, molecular simulation and deep learning technology, and can achieve rapid and accurate prediction of large-scale olfactory receptor functions. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 Demonstrate the basic principle of multi-step transfer learning.

[0024] Figure 2 Shows the basic process of multi-step transfer learning.

[0025] Figure 3 The deep learning model architecture used by the method of the present invention is shown.

[0026] Figure 4 A flow chart showing the prediction of olfactory receptor function using the method of the present invention. DETAILED DESCRIPTION

[0027] In order to make those skilled in the art better understand the technical solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and the present invention is not limited thereto.

[0028] Figure 1 Show the basic principle of transfer learning, Figure 2 The basic process of multi-step transfer learning is shown. The function prediction model is pre-trained on a large batch of molecular docking datasets to obtain the overall distribution trend of the interaction between OR protein and VOC molecules; then further training is performed on a certain scale of experimental datasets to obtain a more accurate function prediction model; for species with insufficient experimental data, we can further optimize the existing function prediction model using the small amount of experimental data available on the species, thereby realizing OR function prediction for species with insufficient available data.

[0029] The method of the present invention can be used for rapid function prediction of olfactory receptors. Under the premise of having certain functional identification experimental data, its prediction accuracy is higher than that of traditional molecular docking methods. Taking the prediction of insect olfactory receptor function as an example, the prediction of olfactory receptor function using this method is described, which includes the following steps:

[0030] (1) Obtain source domain datasets and target domain datasets. Prepare large-scale molecular docking data as the source domain dataset. The Insect Olfactory Receptor Structure and Function Prediction Database (iORbase, https: / / www.iorbase.com / ) provides large-scale molecular docking data between insect olfactory receptors and gas molecules, including molecular docking results between nearly 6,000 insect olfactory receptors of 59 insect species and nearly 2,000 gas molecules. It contains most types of insect olfactory receptors and compound types, which can better reflect the general laws of the interaction between insect olfactory receptors and gas molecules. Therefore, we choose to use the data in this database as the pre-training dataset. The target domain dataset comes from existing batch functional identification data, which is mainly collected from published literature.

[0031] For species without existing molecular docking data sets, various existing molecular docking software can be used for molecular docking. Here, VINA is used as an example to briefly explain the steps for molecular docking. First, the structure of the olfactory receptor needs to be predicted based on its amino acid sequence. Open source protein structure prediction methods, such as alphafold2, can be used; secondly, the three-dimensional structure of the protein and small molecule of the compound is used to calculate its binding free energy, that is, the result of molecular docking, using molecular docking software.

[0032] (2) Build a function prediction model. The function prediction model architecture used in this method is as follows: Figure 3 As shown, first, a protein and small molecule language model based on natural language processing methods is used as a feature extraction module, and the amino acid sequence of the olfactory receptor and the simplified molecular linear input specification (SMILES) of the small molecule are used as the input of the model to obtain the feature vector representation of the protein and the small molecule. Then, the two parts of the feature vector are merged and input into a fully connected neural network, and the binding score between the protein and the small molecule is used as a label to train the model.

[0033] For the feature extraction part, we used the currently mature protein language model and small molecule language model. These models have been pre-trained on ultra-large-scale protein and small molecule datasets in advance. They can extract features from protein sequences and small molecule SMILES format sequences, and finally convert them into abstract feature vector representations. The language model used for feature extraction comes from the open source website hugging face (https: / / huggingface.co / ). Using the open source language model, it is relatively easy to convert the input sequence into an array-shaped feature vector for subsequent fully connected model training. The protein feature extraction module converts amino acid sequences of length 300-400 into feature vectors of length 1024, while the small molecule feature extraction module converts small molecule SMILES sequences of varying lengths into feature vectors of length 768. The two vectors are spliced ​​head to tail to obtain the feature representation vectors of both proteins and small molecules. After that, the spliced ​​feature vectors can be input into the fully connected network model for subsequent binding score prediction.

[0034] We use a 5-layer deep fully connected neural network model to learn and fit the concatenated feature vectors. The input layer of the fully connected neural network is consistent with the length of the concatenated feature vector, which is 1024+768=1792. After the input layer, 3 layers of fully connected layers with 512 nodes are stacked, and the final output layer length is 1, which is used to output a certain value, that is, the binding score given by the model. In the fully connected model, we add a batch normalization layer after each fully connected layer to scale the data to the same range and eliminate the dimensional differences between the features, thereby increasing the convergence speed of the model. The feature vector passes through the neural network model and finally obtains a one-dimensional output score, which is the direct interaction score between the corresponding protein and the small molecule. The mean square error (MSE loss) is used as the loss function of the regression task:

[0035]

[0036] Where n is the number of samples, is the prediction result, y iThe difference between the label and the score output by the model is calculated, and the neural network weights are updated using backpropagation.

[0037] (3) Function prediction model pre-training and the first step of transfer learning. The first step of transfer learning is to migrate from the molecular docking dataset to the functional identification experimental dataset. First, a large-scale molecular docking data (source domain dataset) is used to pre-train the function prediction model. During the training, the affinity (affinity score) between the protein and the small molecule is used as the label for model training, so as to train a model for predicting molecular docking. Afterwards, the model trained with the molecular docking data is trained again using the functional identification experimental data, that is, the model is further fine-tuned so that it can predict the experimental results of functional identification. The source domain dataset of the first step of transfer learning is the molecular docking dataset, and its source has been explained in step (1). Then, a certain amount of functional identification experimental data (target domain dataset) is used to fine-tune the model. The functional identification datasets involved are shown in Table 1.

[0038] Table 1 Functional identification experimental dataset information

[0039]

[0040]

[0041] If the target species for predicting the olfactory receptor function is consistent with the species corresponding to the experimental data set, the function prediction can be performed directly. Since the olfactory receptor sequences and functions between different species may be quite different, the model of the first step of transfer learning is not suitable for directly predicting the OR function of different species. It is necessary to perform the second step of transfer learning, that is, transfer between different species.

[0042] (4) Second-step transfer learning. If the target species is inconsistent with the species corresponding to the experimental dataset, the second-step transfer learning can be used to further fine-tune the model. Based on the first-step transfer learning model, a small amount of OR function identification experimental data of the target species is used for retraining, that is, the first-step transfer learning model is fine-tuned until the olfactory receptor function prediction model for the target species is obtained. The first-step transfer learning is the migration from the molecular docking results (source domain) to the experimental results (target domain), while the second-step transfer learning is the migration from the large-scale experimental results (source domain) of a certain species to the individual experimental results (target domain) of the target species. Generally speaking, the species corresponding to the experimental dataset should have a close relationship with the target species. For insects, transfer learning can be effectively carried out between two insect species belonging to the same order. If the relationship between the two species is distant and the difference is large, resulting in a large difference between the source domain and the target domain of the second transfer learning, the reliability of the prediction results may be reduced.

[0043] Olfactory receptor function prediction in other situations. For target species with no experimental data available, the model obtained using the first-step transfer learning can be directly used to predict the target olfactory receptor function. After comparison, it was found that the first-step transfer learning after merging the experimental data sets of multiple different species into a mixed data set has a higher accuracy in predicting the olfactory receptor function of target species with no experimental data available than the first-step transfer learning model obtained using a single data set.

[0044] In summary, the applicable scenarios and corresponding operations of the olfactory receptor function prediction method proposed in the present invention are as follows: (1) If there are a large number of functional identification results for the target species, the first step of transfer learning is performed using the existing experimental data, and the first step of transfer learning model is used to predict the olfactory receptor function of the target species; (2) If there are a small number of functional identification results for the target species, the second step of transfer learning can be performed on the basis of the first step of transfer learning model, and the second step of transfer learning model is used to predict the olfactory receptor function of the target species. (3) For target species with no available experimental data, the first step of transfer learning is performed using a mixed data set obtained by merging multiple experimental data sets, and the corresponding model is used to predict the olfactory receptor function of the target species. The process of predicting olfactory receptor function using the method of the present invention is as follows: Figure 4 shown.

[0045] The above are only preferred embodiments of the present invention. It should be noted that the above preferred embodiments should not be regarded as limiting the present invention, and the protection scope of the present invention should be based on the scope defined by the claims. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the spirit and scope of the present invention, which should also be regarded as falling within the protection scope of the present invention.

Claims

1. A method for predicting olfactory receptor function based on transfer learning and molecular docking, characterized in that: include: Step 1: Obtain source domain dataset and target domain dataset; Step 2: Use the feature extraction module to build a function prediction model by merging the feature vectors and inputting them into a fully connected neural network; Step 3: pre-training the function prediction model using the source domain dataset, and fine-tuning the function prediction model using the target domain dataset so that the function prediction model can be predicted for the experimental results of function identification, thereby obtaining the first step transfer learning model; Step 4: Determine whether the target species for which the olfactory receptor function is to be predicted is consistent with the species corresponding to the target domain dataset. If they are consistent, use the first step transfer learning model to predict the olfactory receptor function. If there is inconsistency, that is, based on the first-step transfer learning model, the olfactory receptor function identification experimental data of the target species is used to perform one or more steps of fine-tuning until a second-step transfer learning model that can be used to predict the olfactory receptor function of the target species is obtained, and then the second-step transfer learning model is used to predict the olfactory receptor function.

2. The method for predicting olfactory receptor function according to claim 1, characterized in that: Large-scale molecular docking data were used as the source domain dataset.

3. The method for predicting olfactory receptor function according to claim 1, characterized in that: The target domain dataset comes from existing batch function identification data.

4. The method for predicting olfactory receptor function according to claim 1, characterized in that: A protein and small molecule language model based on a natural language processing method is used as the feature extraction module.

5. The method for predicting olfactory receptor function according to claim 1, characterized in that: In the step 2, the amino acid sequence of the olfactory receptor and the simplified molecular linear input specification of the small molecule are used as model inputs to obtain feature vector representations of the protein and the small molecule. The two feature vectors are then merged and input into a fully connected neural network, and the binding score between the protein and the small molecule is used as a label to build a functional prediction model.

6. The method for predicting olfactory receptor function according to claim 1, characterized in that: A 5-layer deep fully connected neural network model is used to learn and fit the feature vector.

7. The method for predicting olfactory receptor function according to claim 1, characterized in that: In the step three, the affinity between the protein and the small molecule is used as a label for the pre-training.

8. The method for predicting olfactory receptor function according to claim 1, characterized in that: In the step 4, the species corresponding to the experimental data has a close relationship with the target species.

9. The method for predicting olfactory receptor function according to claim 1, characterized in that: For target species that do not have experimental data for olfactory receptor function identification, the experimental data sets of multiple different species are merged into a mixed data set and then the first step of transfer learning is performed. The model obtained from the first step of transfer learning is used to predict the olfactory receptor function of the target species.

Images