Method for predicting topological structure of organism brain based on spatial-temporal characteristics of neural activity

Abstract

The invention provides a neural activity spatio-temporal characteristic-based organism brain topological structure prediction method, which combines a deep learning technology with a traditional imaging technology, and comprises the following steps of: firstly, obtaining a connection relationship of partial neurons by using a traditional method; constructing a graph neural network based on the known connection relationships, introducing a long short-term memory (LSTM) network into the graph neural network, and constructing a brain topological structure prediction model; training the prediction model through calcium activity data of known neurons; after model training is completed, calcium activity data of unknown neurons are input into the model, and the connection relation between the unknown neurons and the known neurons can be predicted. Along with the increase of the number of known neurons, the neural connection map is continuously expanded, and the performance of the brain topological structure prediction model constructed based on the neural connection map is continuously improved through continuous iteration, so that the prediction accuracy of the model is effectively improved, and the mining speed of the neural connection relationship is increased.

Description

Technical Field

[0001] The present invention relates to the technical field of brain neuron network analysis, and in particular to a method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity. Background Art

[0002] The brain is composed of a vast number of neurons, interconnected to form a complex network that enables advanced thinking and creativity. To understand how this network of neurons generates advanced cognitive behaviors, we need to delve deeper into the connections between neurons, nuclei, and brain regions.

[0003] When faced with a vast brain network, reconstructing the brain's topological structure on a whole-brain scale, that is, a neural connectivity map, can reveal the relationship between brain structure and function, which is the basis for understanding how the brain works. The mapping of neural connectivity maps mainly relies on traditional imaging technologies such as nuclear magnetic resonance, optical microscopy, and electron microscopy. Specifically, the structural relationship between neurons is often discovered through changes in the concentration of labeled molecules or the movement trajectory of tracer viruses under a microscope. Nuclear magnetic resonance, optical microscopy, and electron microscopy techniques mainly include two steps: the first is sample imaging, which involves the research of labeled molecules, tracer viruses, and imaging equipment; the second is data reconstruction, which focuses on the development of reconstruction algorithms such as data parsing and relationship mining.

[0004] In recent years, continuous advances in multidimensional and multimodal imaging technologies and image data processing have greatly promoted the development of neural connectivity mapping. However, two challenges remain: 1) With the continuous improvement of imaging resolution, electron microscopy technology can obtain more detailed images of neural structures, resulting in an exponential increase in the amount of imaging data. This surge in data has severely limited the data parsing and relationship mining capabilities of traditional electron microscopy technology when faced with large-scale neural connectivity mapping. Existing methods struggle to efficiently process and extract key information from complex neural networks, thus limiting our comprehensive understanding of the nervous system. 2) Current neural connectivity maps mainly provide basic information such as the number of neurons and neural connectivity relationships, but they are unable to reveal the strength of neural connections and their age-related changes. Neural connection strength is a key structural property of neural circuits and is closely related to neural activity and motor behavior. However, because neural connectivity maps lack detailed information on neural connection strength and age-related changes, many studies based on connectivity maps have failed to fully consider the impact of neural connection strength.

[0005] In addition, recent advances in whole-brain calcium imaging have led to a continuous increase in the amount of biological calcium activity data from model organisms such as nematodes, zebrafish, and mice. Simultaneously, the rapid development of artificial intelligence and big data technologies has facilitated the realization of data-driven connectivity mining and connection strength measurement.

[0006] Therefore, proposing a biologically plausible method to measure neural connection strength is of great significance for improving the neural connection map. Summary of the Invention

[0007] The purpose of the present invention is to provide a prediction method that can effectively predict neural connection relationships and improve the efficiency of reconstructing the topological structure of the brain of an organism.

[0008] To achieve the above objectives, the present invention proposes a method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity, including: 1) obtaining calcium activity data of the organism through calcium imaging processing; 2) introducing a long short-term memory (LSTM) network into a graph neural network to construct a topological structure prediction model of the brain of an organism; 3) inputting the calcium activity data into the topological structure prediction model, and identifying neuronal features through the long short-term memory (LSTM) network and the graph neural network in the model, thereby predicting the neuronal connection relationship.

[0009] Furthermore, the topology prediction model is combined with imaging technology to continuously expand the neural connection map, so that the model is continuously iterated, which specifically includes the following steps:

[0010] S1: Identify the model organism to be studied, obtain the partial neuronal connectivity of the organism through traditional imaging technology, and construct a local neural connectivity map;

[0011] S2: Perform whole-brain calcium imaging on the organism to be studied to obtain calcium activity time series data of known neurons and unknown neurons;

[0012] S3: Based on the local neural connection map, a prediction model of the biological brain topology structure is constructed through a graph neural network, and a long short-term memory (LSTM) network is introduced into the graph neural network;

[0013] S4: Inputting calcium activity data of known neurons into a graph neural network that incorporates a long short-term memory (LSTM) network to train and optimize the prediction model of the biological brain topology;

[0014] S5: Input the calcium activity data of unknown neurons into the trained biological brain topology prediction model to predict the neural connection relationship between unknown neurons and known neurons, and verify the neural connection relationship with high prediction probability;

[0015] S6: Add the verified neural connection relationship to the local neural connection map to obtain the extended neural connection map, record the unverified neural connection relationship and wait for further optimization of the model before re-verification;

[0016] S7: Based on the expanded neural connection map, execute step S3, so that the neural connection map is continuously expanded, and the brain topology structure prediction model constructed based on the neural connection map is continuously iterated to continuously optimize the prediction results.

[0017] As the above process proceeds, the neural connectivity map continues to expand. The brain topology prediction model built based on the neural connectivity map continues to improve through continuous iteration, further accelerating the expansion of the neural connectivity map and increasing the efficiency of neural connectivity map reconstruction. It can be seen that the prediction model effectively indicates to electron microscopy which neural connection relationships are more likely to exist and which are less likely to exist, thus preventing the electron microscopy imaging technology from indiscriminately searching and thus reducing efficiency.

[0018] Furthermore, in step S2, the whole brain calcium imaging processing method for the organism to be studied is:

[0019] 1) Introducing fluorescent calcium indicator proteins into the nervous system of an organism;

[0020] 2) The organism is anesthetized and fixed on a microfluidic chip, and in vivo imaging is performed using a high-resolution laser scanning confocal microscope;

[0021] 3) Extract the fluorescence intensity of each neuron in the imaging video frame by frame to generate the corresponding neural calcium activity time series data.

[0022] Furthermore, the method for constructing the biological brain topology prediction model is:

[0023] S3.1. Graph Structure Construction: Based on the non-Euclidean properties of biological neural systems, the biological nervous system is modeled as a graph, and the interactions between neurons are used to predict their connectivity information. The biological nervous system is denoted as G = (V, E), where neurons and neural connections are represented by the node set V and the edge set E, respectively. A graph neural network is used to aggregate the signals of adjacent neurons to obtain neuronal features. Long-short-term memory (LSTM) networks are introduced into the graph neural network framework to effectively model the historical activity of neurons and obtain a representation of neural activity.

[0024] S3.2: Dataset division: The number of neurons and connections in an organism are represented by |V| and |E|, respectively. The neurons used for model training, verification, and testing and their actual connections are represented by G tra_pos , G val_pos and G tes_pos The neurons used for model training, validation and testing and their actual non-existent connections are represented as G tra_neg , G val_neg and G tes_neg ; In order to avoid the impact of data imbalance on training, G tra_pos , Gval_pos and G tes_pos The number of connections and G tra_neg , G val_neg and G tes_neg The number of connections contained in is equal. The graph neural network in the model is based on G tra_pos The neurons contained in the network and their actual connections are constructed.

[0025] S3.3: Model design optimization: Use a graph neural network-based optimization strategy to train the model framework to predict neuronal connections and use the cross-entropy loss function to optimize the model;

[0026] S3.4: Prediction evaluation promotion: Use three indicators, prediction loss Loss, prediction accuracy ACC, and area under the curve AUC, to evaluate the prediction performance of the model. By combining these evaluation indicators, the model's performance in the neural connection prediction task is comprehensively measured.

[0027] Furthermore, the process of predicting neuronal connectivity using the aforementioned brain topology prediction model is as follows: first, preliminary features are extracted from the organism's calcium activity data using an LSTM network. These preliminary features are then used as input to a graph neural network to obtain final neuronal features. Finally, the final neuronal features are integrated and input into a fully connected network to predict whether there are actual connections between neurons.

[0028] Furthermore, in step M1, the SAGEConv module in the graph neural network is used to aggregate calcium signals and obtain neuronal features. In normal physiological activities of an organism, neural nodes integrate signals from other neural nodes. From the perspective of deep learning, neurons in a graph neural network aggregate signals from adjacent neurons and perform nonlinear transformations on the aggregated signals to obtain features that can characterize neurons.

[0029] The manifestation is:

[0030]

[0031] h i =σ(W·concat(h i ,h N(i) )

[0032] h i =norm(h i )

[0033] Among them, h i and h j represent the signals of the i-th and j-th neurons respectively, h N(i) represents the aggregate signal of all neighboring neurons N(i) of the i-th neuron; eji is the scalar weight of the connection from the jth neuron to the ith neuron, W is a learnable weight tensor; aggregate(·), σ(·), and norm(·) represent the aggregation function, nonlinear mapping function, and normalization function of the state features, respectively.

[0034] Furthermore, neural systems exhibit not only complex spatial structural characteristics but also significant temporal dependencies. These temporal characteristics of neural activity are crucial for understanding neural network function. In this context, traditional graph neural networks focus solely on modeling spatial adjacency relationships, making it difficult to capture the temporal dynamics of neural activity. In step S3.1, to fully reflect the temporal dependencies of neural activity, a long short-term memory (LSTM) network is introduced into the graph neural network framework. This allows the extracted neuronal features to better capture the temporal variations in neural activity.

[0035] Assume that after preprocessing the calcium signal activity data of each neuron, the obtained time series is represented as X t ={x t |t=1,2,…,L}, where x t represents the calcium signal intensity at time t, and L is the length of the data sequence. To effectively integrate this temporal dependency information into the graph neural network, we input it into an LSTM network. Once the calcium activity time series data is fed into the LSTM network, the network captures and retains long-term dependencies through memory cells (cell states) and gating mechanisms (such as input gates, forget gates, and output gates), obtaining a representation of neural activity and effectively modeling the neuron's historical activity. The features extracted by the LSTM network serve as the neuronal features of the time series and are used as the input for each neuron in the graph neural network. The introduction of the LSTM network ensures that the dynamic characteristics of each neuron not only rely on its spatial adjacency but also consider the historical activity between neurons in the temporal dimension, thereby obtaining a more accurate representation of neural activity.

[0036] The mathematical expression of LSTM network can be written as:

[0037]

[0038] h t =o t ⊙tanh(c t )

[0039] Among them, f t 、i t 、c t 、o t 、 and h tRepresent the output variables of each component in the LSTM network, σ and tanh are activation functions, W and b are weight tensors and bias respectively. The symbol ⊙ represents the element-by-element multiplication operation. Finally, the feature H extracted by the LSTM network is t ={h t |t=1,2,…,L} is used as the input of each neuron in the graph neural network.

[0040] Furthermore, the operation process of the LSTM network is as follows: First, the input gate i t Determines the impact of the current input on the neuron state; forget gate f t Determine the previous memory state c t-1 The degree of forgetting at the current moment; then, the candidate memory unit c t Generate new candidate memory information at the current moment, and finally obtain the updated memory state c by combining the output of the input gate and the forget gate t。 Output gate o t Determine the neuron output h at the current moment t , and enhances its expressive power through the nonlinear transformation of the activation function tanh. The advantage of LSTM network in extracting neuron time features is its ability to dynamically select and update information through the gating mechanism, which enables it to effectively capture the long-term dependencies in neuron activities. t After being calculated and extracted, they will be input into the subsequent graph network as time series neuron features, further improving the spatiotemporal modeling capabilities of the neural activity map.

[0041] Furthermore, the training process of the biological brain topology prediction model is as follows: 1) extracting preliminary neuron features from calcium activity data through an LSTM network; 2) inputting the extracted preliminary neuron features into a graph neural network to extract the final neuron features; 3) integrating the final neuron features and inputting them into a fully connected neural network to obtain predicted values of neural connections; 4) applying a gradient descent algorithm to iteratively update weight variables in the framework based on the predicted values and true values of the neural connections.

[0042] Furthermore, the cross entropy loss function is used to optimize the model. The specific form is as follows:

[0043] L={l1,l2,...l n ,...,l N} T

[0044] l n =-[y n ·logx n +(1-y n )·log(1-x n)]

[0045] Loss = mean(L)

[0046] Among them, Loss is the value of the binary cross entropy function, N is the number of samples in each batch during training; x n and y n are the prediction scores of the binary labels, and their values range from (0,1).

[0047] Furthermore, during the calcium imaging data processing experiment, in order to ensure the accuracy and reliability of the data and eliminate errors, individual differences and random interference, the following data preprocessing measures were adopted:

[0048] 1) Large-Scale Experimental Design: Under consistent experimental conditions such as temperature, humidity, and light, each organism will be cultured on a large scale to ensure that each data set is derived from multiple individuals. Organisms of each age group will be imaged under the same experimental conditions to ensure representative and comparable data.

[0049] 2) Automated Imaging and Manual Interpretation: Automated calcium imaging video analysis tools are used to rapidly track and label neurons in the video. Neurons with stable movements and strong fluorescence signals are tracked in real time using an automated tracking algorithm based on image features, and their fluorescence intensity is extracted. Neurons with violent movements or weak fluorescence signals are manually labeled to ensure accurate signal extraction. The calcium activity signal for each neuron is calculated from the mean fluorescence intensity of that neuron.

[0050] 3) Data Denoising and Smoothing: First, the neuronal signals extracted from the same batch of calcium imaging videos are denoised. High-frequency noise is removed using a Gaussian filter or low-pass filter to minimize interference from the external environment or instrumentation. Furthermore, a signal smoothing algorithm (such as a moving average) is used to further smooth the calcium signals to ensure they accurately reflect neuronal activity.

[0051] 4) Time Series Normalization: To eliminate the effects of fluorescence intensity differences and instrument variations that may occur during the experiment, the calcium activity time series of each neuron was normalized using the Z-score normalization method. This ensures comparability of data across individuals and experimental conditions and improves data robustness.

[0052] 5) Signal Segmentation and Time Alignment: The continuous imaging video is segmented into several time windows for better analysis of the short-term dynamics of neuronal activity. The data within each time window is time-aligned to ensure that the signals and time series of different neurons can be synchronously compared during analysis.

[0053] Compared with the prior art, the advantages of the present invention are:

[0054] 1. The present invention takes into account that the activity data of the organism's nervous system not only has spatial characteristics but also has significant time dependence. By introducing the long short-term memory network (LSTM) on the basis of the graph neural network, the graph neural network (GNN) and the long short-term memory network (LSTM) are used as two modules of the organism's brain topology structure prediction model. End-to-end training technology is used to fully explore the spatiotemporal characteristics of the organism's neural activity. Based on these spatiotemporal characteristics, the model can effectively predict the neural connection relationship, and the prediction performance is also significantly improved.

[0055] 2. Compared with traditional imaging techniques, this method combines deep learning with traditional imaging techniques. It first uses traditional methods to obtain the connectivity relationships of some neurons. It then constructs a graph neural network based on these known connectivity relationships. The graph neural network is then trained using calcium activity data from known neurons. Once the graph neural network model is trained, the calcium activity data from unknown neurons is input into the model to predict the connectivity relationships between the unknown neurons and the known neurons. As the number of known neurons increases, the neural connectivity map continues to expand. The performance of the brain topology prediction model constructed based on the neural connectivity map continues to improve through continuous iteration, effectively increasing the model's prediction accuracy and accelerating the discovery of neural connectivity relationships.

[0056] 3. By continuously expanding the neural connectivity map through the present method, the brain topology prediction model constructed based on the neural connectivity map continues to improve through continuous iteration, further accelerating the expansion of the neural connectivity map, thereby improving the efficiency of neural connectivity map reconstruction. In addition, this prediction model effectively indicates which neural connection relationships are more likely to exist and which are less likely to exist for traditional electron microscopy imaging, thereby avoiding the indiscriminate search of electron microscopy imaging, which reduces efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a diagram showing the architecture of the method of the present invention;

[0058] Figure 2 Schematic diagram of the process of the present invention;

[0059] Figure 3 Schematic diagram of the brain topology prediction model in the method of the present invention;

[0060] Figure 4 Schematic diagram of node connections within the neural network module in the method of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0062] This example takes nematodes as an example and proposes a method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity. Figure 1 As shown, this method mainly includes two modules: calcium imaging data acquisition and processing and topological structure prediction model design. The calcium activity data obtained through whole-brain calcium imaging processing is input into the topological structure prediction model, and the graph neural network with long short-term memory (LSTM) network is introduced into the model to identify neuronal features, thereby predicting the neuronal connection relationship.

[0063] Among them, such as Figure 1 As shown, the whole-brain calcium imaging method for C. elegans includes fluorescent tracing, live imaging, data recording, and data processing. The specific steps are as follows: First, the fluorescent calcium indicator protein GCaMP is introduced into the nervous system of adult C. elegans. Then, the worms are anesthetized with 1mM Levamisol and immobilized on a custom-made microfluidic chip for live imaging using a high-resolution laser scanning confocal microscope. Next, the fluorescence intensity of each neuron in the imaging video is extracted frame by frame to generate corresponding neural calcium activity time series data. To ensure data accuracy and reliability and eliminate errors, individual differences, and random interference, the following data preprocessing methods are used:

[0064] 1) Large-Scale Experimental Design: Adult (Day 1) C. elegans were cultured on a large scale, maintaining consistent experimental conditions such as temperature, humidity, and light intensity, ensuring that each data set was derived from multiple individuals. E. elegans of each age group were imaged under identical experimental conditions to ensure representative and comparable data.

[0065] 2) Automated Imaging and Manual Interpretation: Automated calcium imaging video analysis tools are used to rapidly track and label neurons in the video. Neurons with stable movements and strong fluorescence signals are tracked in real time using an automated tracking algorithm based on image features, and their fluorescence intensity is extracted. Neurons with violent movements or weak fluorescence signals are manually labeled to ensure accurate signal extraction. The calcium activity signal for each neuron is calculated from the mean fluorescence intensity of that neuron.

[0066] 3) Data Denoising and Smoothing: First, the neuronal signals extracted from the same batch of calcium imaging videos are denoised. High-frequency noise is removed using a Gaussian filter or low-pass filter to minimize interference from the external environment or instrumentation. Furthermore, a signal smoothing algorithm (such as a moving average) is used to further smooth the calcium signals to ensure they accurately reflect neuronal activity.

[0067] 4) Time Series Normalization: To eliminate the effects of fluorescence intensity differences and instrument variations that may occur during the experiment, the calcium activity time series of each neuron was normalized using the Z-score normalization method. This ensures comparability of data across individuals and experimental conditions and improves data robustness.

[0068] 5) Signal Segmentation and Time Alignment: The continuous imaging video is segmented into several time windows for better analysis of the short-term dynamics of neuronal activity. The data within each time window is time-aligned to ensure that the signals and time series of different neurons can be synchronously compared during analysis.

[0069] On the other hand, the design method of the biological brain topology prediction model includes graph structure construction, data set partitioning, model design optimization and prediction evaluation promotion, specifically:

[0070] 1. Graph Structure Construction: Considering the non-Euclidean nature of biological nervous systems, the neural system of the model organism Caenorhabditis elegans is modeled as a graph, and its connectivity information is predicted using the interactions between neurons. The C. elegans nervous system is denoted as G = (V, E), where neurons and neural connections are represented by the node set V and the edge set E, respectively. In the normal physiological activities of C. elegans, neural nodes integrate signals from other neural nodes, which is expressed as follows:

[0071]

[0072] h i =σ(W·concat(h i ,h N(i) )

[0073] h i =norm(h i )

[0074] Among them, h i and h j represent the signals of the i-th and j-th neurons respectively, h N(i) represents the aggregate signal of all neighboring neurons N(i) of the i-th neuron; e ji is the scalar weight of the connection from the jth neuron to the ith neuron, W is a learnable weight tensor; aggregate(·), σ(·) and norm(·) represent the aggregation function, nonlinear mapping function and normalization function of the state features respectively. Figure 4As shown in the figure, the connection diagram of each node in the graph network is as follows. From the perspective of deep learning, the neurons in the graph neural network will aggregate signals from adjacent neurons and perform nonlinear transformation on the aggregated signals to obtain features that can characterize the neurons. In this embodiment, the SAGEConv module in the graph neural network is used to aggregate calcium signals and obtain neuronal features.

[0075] In addition, considering that the nervous system of Caenorhabditis elegans not only exhibits complex spatial structural characteristics but also exhibits significant temporal dependence, the temporal characteristics of these neural activities are crucial for understanding the function of neural networks. In this context, traditional graph neural networks only focus on modeling spatial adjacency relationships and have difficulty capturing the temporal dynamics of neuronal activity. In order to fully reflect the temporal dependence of neural activity, such as Figure 3 As shown in Figure 1, the long short-term memory (LSTM) network is introduced into the graph neural network framework, so that the extracted neuron features can better capture the temporal changes of neural activity. The specific method is: the calcium signal activity data obtained by calcium imaging is preprocessed to obtain the calcium activity time series X t ={x t |t=1,2,…,L}, where x t represents the calcium signal intensity at time t, and L is the length of the data sequence. To effectively integrate this time-dependent information into the graph neural network, it is input into an LSTM network. When calcium activity time series data is fed into the LSTM network, the network captures and retains long-term dependencies through memory cells (cell states) and gating mechanisms (such as input gates, forget gates, and output gates), obtaining a representation of neural activity and effectively modeling the neuron's historical activity. The features extracted by the LSTM network serve as the neuronal features of the time series and are used as the input for each neuron in the graph neural network. The introduction of the LSTM network ensures that the dynamic characteristics of each neuron not only rely on its spatial adjacency but also consider the historical activity between neurons in the temporal dimension, thereby obtaining a more accurate representation of neural activity.

[0076] In this embodiment, the mathematical expression of the LSTM network can be written as:

[0077]

[0078] h t =o t ⊙tanh(c t )

[0079] Among them, f t 、i t 、c t 、o t 、 and ht Represent the output variables of each component in the LSTM network, σ and tanh are activation functions, W and b are weight tensors and bias respectively. The symbol ⊙ represents the element-by-element multiplication operation. Finally, the feature H extracted by the LSTM network is t ={h t |t=1,2,…,L} is used as the input of each neuron in the graph neural network.

[0080] The operation process of the LSTM network is as follows: First, the input gate i t Determines the impact of the current input on the neuron state; forget gate f t Determine the previous memory state c t-1 The degree of forgetting at the current moment; then, the candidate memory unit c t Generate new candidate memory information at the current moment, and finally obtain the updated memory state c by combining the output of the input gate and the forget gate t。 Output gate o t Determine the neuron output h at the current moment t , and enhances its expressive power through the nonlinear transformation of the activation function tanh. The advantage of LSTM network in extracting neuron time features is its ability to dynamically select and update information through the gating mechanism, which enables it to effectively capture the long-term dependencies in neuron activities. t After being calculated and extracted, they will be input into the subsequent graph network as time series neuron features, further improving the spatiotemporal modeling capabilities of the neural activity map.

[0081] In this embodiment, by combining the graph neural network with the LSTM network, the spatiotemporal characteristics of neurons can be characterized by mining the calcium activity data of Caenorhabditis elegans.

[0082] 2. Dataset division: The number of neurons and the number of connections of C. elegans are represented by |V| and |E| respectively, and the neurons and their actual connections used for model training, verification and testing are represented by G tra_pos , G val_pos and G tes_pos The neurons used for model training, validation and testing and their actual non-existent connections are represented as G tra_neg , G val_neg and G tes_neg ; In order to avoid the impact of data imbalance on training, G tra_pos , G val_pos and G tes_pos The number of connections and G tra_neg , G val_neg and G tes_neg The number of connections contained in the model is equal. The graph neural network in the model is based on Gtra_pos The model is constructed based on the neurons contained in the graph and their actual connections. The workflow is as follows: First, preliminary features are extracted from the calcium activity data of C. elegans using an LSTM network. These features are then used as input to a graph neural network to obtain the final neuron features. Finally, the final neuron features are integrated and input into a fully connected network to predict whether there are actual connections between neurons.

[0083] 3. Model design optimization: A graph neural network-based optimization strategy is used to train the model framework to predict neuronal connections, and the cross-entropy loss function is used to optimize the model. The training process mainly includes the repeated iteration of the following three steps: (1) Calcium activity data is input into the LSTM network to extract preliminary neuronal features, which are then input into the graph neural network to further extract the final neuronal features; (2) The final neuronal features are integrated and input into the fully connected neural network to obtain the predicted values of the neural connections; (3) Based on the predicted values and the true values of the neural connections, the gradient descent algorithm is applied to iteratively update the weight variables in the framework.

[0084] During the training process, the cross entropy loss function is used to optimize the model. The specific form is as follows:

[0085] L={l1,l2,...l n ,...,l N} T

[0086] l n =-[y n ·logx n +(1-y n )·log(1-x n )]

[0087] Loss = mean(L)

[0088] Among them, Loss is the value of the binary cross entropy function, N is the number of samples in each batch during training; x n and y n are the prediction scores of the binary labels, and their values range from (0,1).

[0089] 4. Prediction evaluation and promotion: In order to evaluate the prediction performance of the model, three indicators are used to evaluate the prediction performance of the model: prediction loss, prediction accuracy, and area under the curve (AUC). By combining these evaluation indicators, the model's performance in the neural connection prediction task is comprehensively measured.

[0090] This invention combines traditional imaging technology with a brain topology prediction model. The specific workflow is as follows:

[0091] S1: Identify the model organism to be studied, obtain the partial neuronal connectivity of the organism through traditional imaging technology, and construct a local neural connectivity map;

[0092] S2: Perform whole-brain calcium imaging on the organism to be studied to obtain calcium activity time series data of known neurons and unknown neurons;

[0093] S3: Based on the local neural connection map, a prediction model of the biological brain topology structure is constructed through a graph neural network, and a long short-term memory (LSTM) network is introduced into the graph neural network.

[0094] S4: Inputting calcium activity data of known neurons into a graph neural network that incorporates a long short-term memory (LSTM) network to train and optimize the prediction model of the biological brain topology;

[0095] S5: Input the calcium activity data of unknown neurons into the trained biological brain topology prediction model to predict the neural connection relationship between unknown neurons and known neurons, and verify the neural connection relationship with high prediction probability;

[0096] S6: Add the verified neural connection relationship to the local neural connection map to obtain the extended neural connection map, record the unverified neural connection relationship and wait for further optimization of the model before re-verification;

[0097] S7: Based on the expanded neural connection map, execute step S3, so that the neural connection map is continuously expanded, and the brain topology structure prediction model constructed based on the neural connection map is continuously iterated to continuously optimize the prediction results.

[0098] As the above process proceeds, the neural connectivity map continues to expand. The brain topology prediction model built based on the neural connectivity map continues to improve through continuous iteration, further accelerating the expansion of the neural connectivity map and increasing the efficiency of neural connectivity map reconstruction. It can be seen that the prediction model effectively indicates to electron microscopy which neural connection relationships are more likely to exist and which are less likely to exist, thus preventing the electron microscopy imaging technology from indiscriminately searching and thus reducing efficiency.

[0099] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity, characterized in that: include: 1) Obtain calcium activity data of organisms through calcium imaging; 2) Introducing a long short-term memory (LSTM) network into a graph neural network to construct a topological structure prediction model for the brain of an organism; 3) Inputting calcium activity data into the topological structure prediction model, identifying neuronal features through the long short-term memory (LSTM) network and graph neural network in the model, and thus predicting neuronal connectivity.

2. The method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The topology prediction model is combined with imaging technology to continuously expand the neural connection map and iterate the model, which specifically includes the following steps: S1: Obtain the partial neuronal connectivity of the organism to be studied and construct a local neural connectivity map; S2: Perform whole-brain calcium imaging on the organism to be studied to obtain calcium activity time series data of known neurons and unknown neurons; S3: Based on the local neural connection map, a prediction model of the biological brain topology structure is constructed through a graph neural network, and a long short-term memory (LSTM) network is introduced into the graph neural network; S4: Inputting calcium activity data of known neurons into a graph neural network that incorporates a long short-term memory (LSTM) network to train and optimize the prediction model of the biological brain topology; S5: Input the calcium activity data of unknown neurons into the trained biological brain topology prediction model to predict the neural connection relationship between unknown neurons and known neurons, and verify the neural connection relationship with high prediction probability; S6: Add the verified neural connection relationship to the local neural connection map to obtain the extended neural connection map, record the unverified neural connection relationship and wait for further optimization of the model before re-verification; S7: Based on the expanded neural connection map, execute step S3, so that the neural connection map is continuously expanded, and the brain topology structure prediction model constructed based on the neural connection map is continuously iterated to continuously optimize the prediction results.

3. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The calcium imaging processing method is: 1) Introducing fluorescent calcium indicator proteins into the nervous system of an organism; 2) The organism is anesthetized and fixed on a microfluidic chip, and in vivo imaging is performed using a high-resolution laser scanning confocal microscope; 3) Extract the fluorescence intensity of each neuron in the imaging video frame by frame to generate the corresponding neural calcium activity time series data.

4. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The method for constructing the biological brain topology prediction model is: M1. Graph Structure Construction: Based on the non-Euclidean properties of biological neural systems, the biological nervous system is modeled as a graph, denoted as G = (V, E), where neurons and neural connections are represented by the node set V and the edge set E, respectively. A graph neural network is used to aggregate the signals of adjacent neurons to obtain neuron features. Introducing the long short-term memory (LSTM) network into the graph neural network framework effectively models the historical activity of neurons and obtains neural activity representation; M2: Dataset division: The number of neurons and connections of an organism are represented by |V| and |E| respectively, and the neurons used for model training, verification and testing and their actual connections are represented by G tra_pos , G val_pos and G tes_pos The neurons used for model training, validation and testing and their actual non-existent connections are represented as G tra_neg , G val_neg and G tes_neg Among them, G tra_pos , G val_pos and G tes_pos The number of connections and G tra_neg , G val_neg and G tes_neg The number of connections contained in . M3: Model Design Optimization: Use a graph neural network-based optimization strategy to train the model framework to predict neuron connections and use the cross-entropy loss function to optimize the model; M4: Prediction evaluation promotion: Use three indicators: prediction loss Loss, prediction accuracy ACC and area under the curve AUC to evaluate the prediction performance of the model.

5. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 4, characterized in that: In step M1, the SAGEConv module in the graph neural network is used to aggregate calcium signals and obtain neuronal features. In normal physiological activities of an organism, neural nodes integrate signals from other neural nodes. Neurons in the graph neural network aggregate signals from adjacent neurons and perform nonlinear transformations on the aggregated signals to obtain features that can characterize neurons. The manifestation is: h i =σ(W·concat(h i ,h N(i) ) h i =norm(h i ) Among them, h i and h j represent the signals of the i-th and j-th neurons respectively, h N(i) represents the aggregate signal of all neighboring neurons N(i) of the i-th neuron; e ji is the scalar weight of the connection from the jth neuron to the ith neuron, W is a learnable weight tensor; aggregate(·), σ(·), and norm(·) represent the aggregation function, nonlinear mapping function, and normalization function of the state features, respectively.

6. The method for predicting the topological structure of the brain of an organism based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The long short-term memory (LSTM) network is introduced into the graph neural network framework. When the calcium activity time series data is input into the LSTM network, the LSTM network captures and retains long-term dependencies through memory units and gating mechanisms to obtain neural activity representations. The features extracted by the LSTM network are used as neuronal features of the time series and are used as input for each neuron in the graph neural network.

7. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The operation process of the LSTM network is as follows: First, the input gate i t Determines the impact of the current input on the neuron state; forget gate f t Determine the previous memory state c t-1 The degree of forgetting at the current moment; then, the candidate memory unit c t Generate new candidate memory information at the current moment, and finally obtain the updated memory state c by combining the output of the input gate and the forget gate t。 Output gate o t Determine the neuron output h at the current moment t , and enhance its expressive power through the nonlinear transformation of the activation function tanh.

8. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 4, characterized in that: The training process of the biological brain topology prediction model is as follows: 1) extracting preliminary neuronal features from calcium activity data through an LSTM network; 2) Input the extracted preliminary neuron features into the graph neural network to extract the final neuron features; 3) The final neuron features are integrated and input into a fully connected neural network to obtain the predicted value of the neural connection; 4) Based on the predicted value and the true value of the neural connection, the gradient descent algorithm is applied to iteratively update the weight variables in the framework.

9. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 4, characterized in that: The cross entropy loss function is used to optimize the model, which is as follows: L＝{l1,l2,...l n ,...,l N} T l n =-[y n ·logx n +(1-y n )·log(1-x n )] Loss = mean(L) Among them, Loss is the value of the binary cross entropy function, N is the number of samples in each batch during training; x n and y n are the prediction scores of the binary labels, and their values range from (0,1).

10. The method for predicting the topological structure of the biological brain based on the spatiotemporal characteristics of neural activity according to claim 1, characterized in that: The process of predicting neuronal connectivity using the biological brain topology prediction model is as follows: 1) inputting the calcium activity time series data into the biological brain topology prediction model; 2) extracting preliminary features from the biological calcium activity data using the LSTM network in the model; 3) using the extracted preliminary features as input to the graph neural network to obtain final neuronal features; 4) The final neuron features obtained are integrated and input into the fully connected network to predict whether there is an actual connection between neurons.

Images