Insect bionic odor detection method based on ligand recognition mechanism and aversion training

Abstract

The invention discloses an insect bionic odor detection method based on a ligand recognition mechanism and aversion training, and relates to the technical field of biological detection, and the method comprises the steps: matching an insect with a target VOC based on the ligand recognition mechanism; performing aversive olfaction training on the matched insects and target VOCs; for the insects and the target VOC which are subjected to the aversion olfaction training, identification and spatial distribution of various insects are distinguished on the basis of a CNN hybrid model, and the relation between the spatial distribution of the insects and an odor source is obtained; based on the relation between insect space distribution and odor sources, a VOC distribution map of an insect target area is described through multi-point recognition, and insect bionic odor detection based on a ligand recognition mechanism and aversion training is completed. The method solves the problem that an existing method is not excellent enough in detection sensitivity, repeatability and stability.

Description

Technical Field

[0001] The invention relates to the technical field of biological detection, and in particular to an insect bionic odor detection method based on ligand recognition mechanism and aversion training. Background Art

[0002] Compared with dogs, insects have incomparable advantages. However, the advantages of insects being small and numerous also become difficulties in actual use. People cannot restrain or track them like dogs. After they are released, they are too small, fly too fast, and even have mimicry ability in some environments, so it is difficult for users to identify and track them and their activity trajectories. The use of this biological sense of smell is mostly achieved through preference training, such as using bees' preference for nectar and police dogs' preference for meat to form associations with targets. There is no doubt that this association has the advantage of natural existence, but it also has many limitations, especially in the use of insect sense of smell. Generally speaking, we use insects' hunger and the need for reproduction and mating for training, but the preference based on hunger is easily affected by the state of the insect. The preference of bees for nectar is a special case, because bees are social animals, and the natural duty of flower bees is to find nectar, etc., so regardless of the state of the bee, it will tend to show the behavior of finding "nectar". However, our commonly used model organisms such as fruit flies do not have this characteristic. The reproduction and mating needs based on pheromones are more susceptible to interference. For example, this preference is often affected by factors such as insect age, mating status, gender, light, temperature, etc., and the same insect will behave differently at different stages. Therefore, preference training is obviously difficult to stably maintain insects in a sensitive state.

[0003] Compared with preference training, punishment-based aversion training is theoretically more suitable for our goals. But nowadays, most of the conditional training conducted by researchers belongs to irrational training. According to the principle mechanism of biological olfaction, different species have different sensitivities to the same odor. This is because their olfactory receptors have different abilities to bind to odor molecules, as well as the subsequent neural responses. Therefore, when we select subjects for training, we usually do not know whether the selected insects can smell the target odor molecules or whether they are sensitive to the target odor molecules. In other words, most of the olfactory training we conduct is blind, and these will affect the behavior that insects ultimately exhibit. Importantly, this irrational olfactory training means that the training results it brings are also unstable. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an insect bionic odor detection method based on ligand recognition mechanism and aversion training to solve the problem that the existing methods are not superior in detection sensitivity, repeatability and stability.

[0005] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: an insect bionic odor detection method based on ligand recognition mechanism and aversion training, comprising the following steps:

[0006] S1: Matching insects and target VOCs based on ligand recognition mechanism;

[0007] S2: Aversive olfactory training with matching insects and target VOCs;

[0008] S3: For the insects and target VOCs that have completed the disgust olfactory training, the CNN hybrid model is used to discriminate the identification and spatial distribution of various insects, and the relationship between the spatial distribution of insects and the odor source is obtained;

[0009] S4: Based on the relationship between the spatial distribution of insects and the odor source, multi-point recognition is used to depict the VOC distribution map of the insect target area, completing insect bionic odor detection based on ligand recognition mechanism and aversion training.

[0010] Furthermore, the S1 includes the following sub-steps:

[0011] S11: Search and compare the target VOC with the molecules that the insects can smell in the iORbase database;

[0012] S12: Based on the search and comparison results, determine whether the target VOC is already in the iORbase database. If yes, complete the matching of the insect and the target VOC. If not, replace the training insect using molecular simulation technology to complete the matching of the insect and the target VOC.

[0013] Furthermore, S2 includes the following sub-steps:

[0014] S21: A set number of insects were placed in a training tube lined with a copper mesh and smelled the target VOC and then the ambient gas;

[0015] S22: stimulation is performed when the insect smells the target VOC, constituting a training cycle;

[0016] S23: Use a gradient-increasing voltage and perform three training cycles for each voltage to obtain a suitable training voltage to complete the aversive olfactory training for the matched insects and target VOCs.

[0017] Furthermore, S3 includes the following sub-steps:

[0018] S31: placing the insects that have completed the aversion olfactory training in a transparent white-bottom octagonal cage, placing an odor source containing the target VOC outside the transparent white-bottom octagonal cage, and observing the spatial distribution data, density distribution data, and time data required to form a stable distribution of the insects in the transparent white-bottom octagonal cage;

[0019] S32: preprocessing the spatial distribution data, the density distribution data and the time data required to form a stable distribution;

[0020] S33: using the preprocessed spatial distribution data, density distribution data, and time data required to form a stable distribution to train and evaluate the CNN hybrid model, and obtain the final CNN hybrid model;

[0021] S34: The final CNN hybrid model is deployed to actual application scenarios to identify various insects and their spatial distribution, and to obtain the relationship between the spatial distribution of insects and the odor source.

[0022] Furthermore, the data preprocessing in S32 includes image data normalization and data enhancement, the image data normalization includes scaling the image pixel values ​​to a range of 0-1, and the data enhancement includes rotation, scaling and cropping.

[0023] Furthermore, the CNN hybrid model in S33 includes a CNN feature extraction layer, a feature processing and fusion layer, and other model layers connected in sequence;

[0024] The CNN feature extraction layer includes an input layer, a convolution layer, an activation layer and a pooling layer connected in sequence, wherein the input layer is used to receive image data, the convolution layer uses multiple convolution kernels to extract local features of the image, the activation layer applies a ReLU activation function, and the pooling layer is used to reduce feature dimensions while retaining important information;

[0025] The feature processing and fusion layer includes a fully connected layer, which flattens the features extracted by the CNN feature extraction layer into a one-dimensional vector;

[0026] The other model layers include a loop layer and a classification and regression layer connected in sequence. The loop layer uses a long short-term memory network LSTM to process sequence data. The classification and regression layer uses a fully connected layer to perform classification tasks, regression tasks, or accesses a support vector machine SVM for prediction tasks.

[0027] The beneficial effects of the present invention are:

[0028] It is proposed to use mathematical models based on algorithms to automatically identify insect individuals and distributions, realize insect olfactory bionics, and solve the bottleneck that existing technologies are unable to track and monitor changes in insect behavior.

[0029] The insect olfactory structure and function database iORbase and molecular simulation technology are used to achieve rational selection of insect olfactory training molecules and avoid interference of insensitive odor molecules in the experiment.

[0030] The proposed insect bionic olfactory recognition system will have the characteristics of strong portability and more comprehensive odor source identification information. For example, it has the adaptability to air, space, sea and ground platforms, the device can be used with different mobile machines and equipment, and can simultaneously determine the direction and concentration of the odor source. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of an insect bionic odor detection method based on ligand recognition mechanism and aversion training.

[0032] Figure 2 Schematic diagram of a T-shaped maze.

[0033] Figure 3 Schematic diagram of the reverse tracking algorithm principle.

[0034] Figure 4 This is an overall schematic diagram of the odor source detection and prediction system.

[0035] Figure 5 This is a schematic diagram of the test results with 9 repetitions.

[0036] Figure 6 This is the distribution diagram of fruit flies' responses to acetaldehyde and air as a control when they are not trained.

[0037] Figure 7 This is the distribution diagram of fruit flies' responses to acetaldehyde during training and air as the control.

[0038] Figure 8 This is a schematic diagram of the octagonal cage test results. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0040] like Figure 1 As shown, an insect bionic odor detection method based on ligand recognition mechanism and aversion training includes the following steps:

[0041] S1: Matching insects and target VOCs based on ligand recognition mechanism;

[0042] S2: Aversive olfactory training with matching insects and target VOCs;

[0043] S3: For the insects and target VOCs that have completed the disgust olfactory training, the CNN hybrid model is used to discriminate the identification and spatial distribution of various insects, and the relationship between the spatial distribution of insects and the odor source is obtained;

[0044] S4: Based on the relationship between the spatial distribution of insects and the odor source, multi-point recognition is used to depict the VOC distribution map of the insect target area, completing insect bionic odor detection based on ligand recognition mechanism and aversion training.

[0045] The S1 includes the following sub-steps:

[0046] S11: Search and compare the target VOC with the molecules that the insects can smell in the iORbase database;

[0047] S12: Based on the search and comparison results, determine whether the target VOC is already in the iORbase database. If so, complete the matching of the insect and the target VOC. If not, use molecular simulation technology to replace the training insect (or use MD to analyze its structure and compare the molecules with the closest structure) to complete the matching of the insect and the target VOC.

[0048] The S2 includes the following sub-steps:

[0049] S21: A set number of insects were placed in a training tube lined with a copper mesh and smelled the target VOC and then the ambient gas;

[0050] S22: stimulation is performed when the insect smells the target VOC, constituting a training cycle;

[0051] S23: Use a gradient-increasing voltage and perform three training cycles for each voltage to obtain a suitable training voltage to complete the aversive olfactory training for the matched insects and target VOCs.

[0052] In one embodiment of the present invention, about 100 fruit flies are placed in a training tube lined with a copper mesh, and smell two odors (target VOC (CS+) and environmental gas (CS-)) successively, each lasting for 1 minute, and fresh air is passed through for 45 seconds after the end. When and only when the fruit flies smell CS+, they will be stimulated by the electric shock stimulator. This electric shock stimulation lasts for 1 minute together with CS+, and consists of 12 1.5s AC pulses, so that one process constitutes a training cycle. Using a gradient increasing voltage, each voltage is trained for three cycles to obtain the most suitable training voltage.

[0053] During the test, the trained fruit flies were placed in a T-maze, such as Figure 2As shown, A is a training tube with an electric shock stimulator, B is a maze corridor, C and D are sorting bins, the blue arrow indicates the direction of insect movement, the red arrow is the direction of odor molecules entering, and the two arms of the maze are respectively connected with two odors that fruit flies smell during training, CS+ with electric shock and CS- without electric shock. Fruit flies have 2 minutes to choose between the two sorting bins, and after the choice, they are separated and counted. Then, they choose CS+ and compare it with the new odor molecule to observe its avoidance effect on CS+.

[0054] The S3 includes the following sub-steps:

[0055] S31: placing the insects that have completed the aversion olfactory training in a transparent white-bottom octagonal cage, placing an odor source containing the target VOC outside the transparent white-bottom octagonal cage, and observing the spatial distribution data, density distribution data, and time data required to form a stable distribution of the insects in the transparent white-bottom octagonal cage;

[0056] In theory, insects should be concentrated roughly in the direction of the straight line connecting the cage and the odor source, and the closer to the odor source, the greater the distribution density. Points with higher concentrations require less time to establish a steady state.

[0057] S32: preprocessing the spatial distribution data, the density distribution data and the time data required to form a stable distribution;

[0058] S33: using the preprocessed spatial distribution data, density distribution data, and time data required to form a stable distribution to train and evaluate the CNN hybrid model, and obtain the final CNN hybrid model;

[0059] S34: The final CNN hybrid model is deployed to actual application scenarios to identify various insects and their spatial distribution, and to obtain the relationship between the spatial distribution of insects and the odor source.

[0060] The data preprocessing in S32 includes image data normalization and data enhancement. The image data normalization includes scaling the image pixel values ​​to a range of 0-1, and the data enhancement includes rotation, scaling and cropping.

[0061] The CNN hybrid model in S33 includes a CNN feature extraction layer, a feature processing and fusion layer, and other model layers connected in sequence;

[0062] The CNN feature extraction layer includes an input layer, a convolution layer, an activation layer and a pooling layer connected in sequence, wherein the input layer is used to receive image data, the convolution layer uses multiple convolution kernels to extract local features of the image, the activation layer applies a ReLU activation function, and the pooling layer is used to reduce feature dimensions while retaining important information;

[0063] The feature processing and fusion layer includes a fully connected layer, which flattens the features extracted by the CNN feature extraction layer into a one-dimensional vector;

[0064] The other model layers include a loop layer and a classification and regression layer connected in sequence. The loop layer uses a long short-term memory network LSTM to process sequence data. The classification and regression layer uses a fully connected layer to perform classification tasks, regression tasks, or accesses a support vector machine SVM for prediction tasks.

[0065] In this embodiment, the CNN hybrid model is mainly composed of a recurrent neural network (RNN), a long short-term memory network (LSTM), a support vector machine (SVM), a decision tree, etc., which is used to process sequence data, enhance the non-spatial characteristics of features, or perform classification and regression tasks. According to the algorithm principle, in the actual detection process, each search point corresponds to a two-dimensional distribution state of an insect. However, it may also be a situation that the distribution in the device should be indifferent when it is far away from the target and when it is on the target. Therefore, the entire search process can be recorded by video, and the distribution of each point is obtained by extracting the video frame. Although it is generally sufficient to obtain the plane data relative to the lens, in this case, the LSTM-based RNN analysis search process must be used to assist in determining whether it is close to the target and predict the final landing point of the insect on the z-axis. After judging that it is closer to the target, the data after dimensionality increase can be classified and judged by SVM, and the whole process can be connected in series using a decision tree model.

[0066] The model training process includes:

[0067] Initialization: Randomly initialize weights for CNN and other model layers;

[0068] Forward propagation: extract features through CNN and pass them to other model layers;

[0069] Loss function calculation: Select the loss function according to the task type, such as cross entropy loss for classification tasks;

[0070] Back propagation: calculate gradients and update network weights;

[0071] Optimization: Use optimization algorithms such as Adam and SGD to adjust weights.

[0072] The model evaluation process includes:

[0073] Use the validation set to evaluate model performance;

[0074] Adjust hyperparameters to optimize the model.

[0075] Finally, the trained model is deployed to actual applications, and necessary model compression and acceleration can be performed to run efficiently on mobile devices or servers.

[0076] In this embodiment, based on the fruit fly optimization algorithm (FOA), this solution sets the following algorithm, which is called the reverse tracking algorithm:

[0077] Different from FOA, this scheme fixes the origin of the coordinate system at the center of the device, that is, the coordinate system is established with itself as the reference, and the rectangular coordinate system is replaced by the polar coordinate system, such as Figure 3 As shown; for any fruit fly individual X i , and their spatial positions in the device are all expressed as (ρ i ,θ i ) to describe; when the fruit fly population in the device becomes stable, a certain distribution is formed. At this time, we i A threshold value is set for ρ, and the coordinate system θ([0,2π]) is divided into eight equal intervals. The qualified individuals are classified and merged according to the θ value, and the interval with the largest number of individuals is selected; because aversive learning is adopted, the iteration direction is the opposite direction of the median of the interval selected in the above step, and the above steps are repeated.

[0078] like Figure 4 As shown, A is the object to be tested (smell source), B is the insect behavior chamber, C is the video collector (camera), D is the neural network processor (computer or single-chip microcomputer), E is the result output device (display), and F is the insect used. After the insect sniffs the odor source in the behavior chamber, its distribution will eventually stabilize. At this time, the distribution can be photographed and recorded using the video collector located directly above the behavior chamber, and transmitted to the neural network processor for processing and analysis, and finally presented on the result output device.

[0079] In this embodiment, the fruit flies selected through training can sniff the target odor molecules, and then the insects are placed in a device, and the device is divided into eight compartments by partitions, so theoretically different compartments have different concentrations of odor molecules. After the odor source is placed and a location is selected, the trained insects will sense the odor molecules at that point, forming a certain distribution within the device. The camera located on the device takes a picture, and the image is analyzed by the trained mathematical model based on the convolutional neural network to obtain the spatial relationship between the point and the odor source, achieving an effect similar to that of a compass. Repeating this process and identifying and analyzing multiple points, you can get a distribution map of odor molecules in the area, and realize the search and detection of the odor source, such as Figure 5 shown.

[0080] Next, acetaldehyde was used as the target VOC, and the trained insects were fruit flies. By consulting the data, it was found that fruit flies did not have a clear preference for acetaldehyde, so without training, there was no difference in the distribution of fruit flies in the case of acetaldehyde and the control air, such as Figure 6 Then, electrical stimulation was used for punishment training. The training conditions were AC pulse wave, 30% duty cycle, 20V voltage, 1min electrical stimulation, 1min control, 45s air purification, 8 cycles, 25.6℃ temperature, 52% humidity, and Y-shaped maze to test the training effect. The test duration was 2min. Figure 7 As shown in the figure, A is the beginning of the test, acetaldehyde is introduced on the left side and air is introduced on the right side. The chi-square test shows that there is no difference in distribution; B is 2 minutes after the test, the fruit flies have formed obvious avoidance behavior towards acetaldehyde. The chi-square test shows that the distribution of the results is significantly different, indicating that the training is successful. Finally, the octagonal cage test is carried out. Figure 8 As shown in the figure, A is the actual octagonal cage test system, the square box is in a closed state except for the air inlet and outlet, the left side is the air inlet, and the right side is the air outlet; B is the test result of acetaldehyde molecules in the octagonal cage, the red arrow is the vector synthesis direction of the system airflow field, so the distinguishing point for the chi-square test is the red line segment on the right side of B. After acetaldehyde gas was introduced and tested for 2 minutes, more fruit flies chose the right side of the distinguishing point, that is, the fruit flies formed an avoidance behavior towards acetaldehyde. The chi-square test showed that the distribution of the results was significantly different, and the training was successful.

[0081] The present invention realizes rational aversion olfactory training based on ligand recognition mechanism, that is, using molecular simulation, molecular docking and other means to predict OR-VOC matching relationship as a prerequisite for aversion training, realizes rational experimental guidance, and solves the obstacles and limitations of preference training that are easily disturbed. At the same time, it realizes "insect olfactory bionics" based on theories such as deep learning. Through the means of computational biology, it reflects the group behavior of insects towards odor sources, imitates the decision-making process of the brain when insects sniff, and improves the speed and reliability of judgment and decision-making on targets.

[0082] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the invention.

Claims

1. An insect bionic odor detection method based on ligand recognition mechanism and aversion training, characterized in that: The following steps are involved: S1: Matching insects and target VOCs based on ligand recognition mechanism; S2: aversive olfactory training with matching insects and target VOCs; S3: For the insects and target VOCs that have completed the disgust olfactory training, the CNN hybrid model is used to discriminate the identification and spatial distribution of various insects, and the relationship between the spatial distribution of insects and the odor source is obtained; S4: Based on the relationship between the spatial distribution of insects and the odor source, multi-point recognition is used to depict the VOC distribution map of the insect target area, completing insect bionic odor detection based on ligand recognition mechanism and aversion training.

2. The insect bionic odor detection method based on ligand recognition mechanism and aversion training according to claim 1 is characterized in that: The S1 includes the following sub-steps: S11: Search and compare the target VOC with the molecules that the insects can smell in the iORbase database; S12: Based on the search and comparison results, determine whether the target VOC is already in the iORbase database. If yes, complete the matching of the insect and the target VOC. If not, replace the training insect using molecular simulation technology to complete the matching of the insect and the target VOC.

3. The insect bionic odor detection method based on ligand recognition mechanism and aversion training according to claim 1 is characterized in that: The S2 includes the following sub-steps: S21: A set number of insects were placed in a training tube lined with a copper mesh and smelled the target VOC and then the ambient gas; S22: stimulation is performed when the insect smells the target VOC, constituting a training cycle; S23: Use a gradient-increasing voltage and perform three training cycles for each voltage to obtain a suitable training voltage and complete the aversive olfactory training for the matched insects and target VOCs.

4. The insect bionic odor detection method based on ligand recognition mechanism and aversion training according to claim 1, characterized in that: The S3 includes the following sub-steps: S31: placing the insects that have completed the aversion olfactory training in a transparent white-bottom octagonal cage, placing an odor source containing the target VOC outside the transparent white-bottom octagonal cage, and observing the spatial distribution data, density distribution data, and time data required to form a stable distribution of the insects in the transparent white-bottom octagonal cage; S32: preprocessing the spatial distribution data, the density distribution data and the time data required to form a stable distribution; S33: training and evaluating the CNN hybrid model using the preprocessed spatial distribution data, density distribution data, and time data required to form a stable distribution, to obtain a final CNN hybrid model; S34: The final CNN hybrid model is deployed to actual application scenarios to identify various insects and their spatial distribution, and to obtain the relationship between the spatial distribution of insects and the odor source.

5. The insect bionic odor detection method based on ligand recognition mechanism and aversion training according to claim 4, characterized in that: The data preprocessing in S32 includes image data normalization and data enhancement. The image data normalization includes scaling the image pixel values ​​to a range of 0-1, and the data enhancement includes rotation, scaling and cropping.

6. The insect bionic odor detection method based on ligand recognition mechanism and aversion training according to claim 4, characterized in that: The CNN hybrid model in S33 includes a CNN feature extraction layer, a feature processing and fusion layer, and other model layers connected in sequence; The CNN feature extraction layer includes an input layer, a convolution layer, an activation layer and a pooling layer connected in sequence, wherein the input layer is used to receive image data, the convolution layer uses multiple convolution kernels to extract local features of the image, the activation layer applies a ReLU activation function, and the pooling layer is used to reduce feature dimensions while retaining important information; The feature processing and fusion layer includes a fully connected layer, which flattens the features extracted by the CNN feature extraction layer into a one-dimensional vector; The other model layers include a loop layer and a classification and regression layer connected in sequence. The loop layer uses a long short-term memory network LSTM to process sequence data. The classification and regression layer uses a fully connected layer to perform classification tasks, regression tasks, or accesses a support vector machine SVM for prediction tasks.

Images