Multi-scale human brain simulation system integrating data, structure, mechanism and behavior

Abstract

The present invention belongs to the field of artificial neural networks and human brain simulation, and specifically relates to a multi-scale human brain simulation system that integrates data, structure, mechanism and behavior, which aims to solve the problem that the existing technology cannot take into account both the simulation of large-scale whole-brain neural networks and the fine simulation of neuronal discharge activities in brain regions. The system includes: a database construction module, configured to collect brain image data and build a storage database; a cortical column model construction module, configured to build a multi-layer cortical column model of the corresponding brain region; a full-scale model construction module, configured to combine the human brain connection map in the storage database and the multi-layer cortical column model of each brain region, connect the brain regions, and obtain a full-scale model; a model training module, configured to train the full-scale model; a simulation module, configured to simulate neural activities through a trained full-scale model. The present invention achieves the balance between the simulation of large-scale whole-brain neural networks and the fine simulation of neuronal discharge activities in brain regions.

Description

Technical Field

[0001] The present invention belongs to the field of artificial neural network and human brain simulation, and specifically relates to a multi-scale human brain simulation system integrating data, structure, mechanism and behavior. Background Art

[0002] The existing treatment options for brain function diseases mainly include drug therapy, non-invasive neural stimulation (such as light, electricity, magnetism, ultrasound, etc.), invasive neural stimulation, surgical treatment, etc. In order to speed up the drug development cycle and improve the effect of treatment, it is necessary to simulate the effects of drugs or neural stimulation on brain neural activity. The existing brain simulation technology mainly includes the following two methods:

[0003] 1. Simulation of mesoscopic structure or functional connectivity for diseases: Inferring functional connectivity and constructing mesoscopic connectivity group models through clinical observation data (such as EEG, fMRI, DTI). This modeling method is often used in the identification of brain functional diseases (such as epilepsy, depression, Alzheimer's disease, etc.) to reveal pathological mechanisms. The main processes of this method are as follows: 1. Data acquisition; 2. Data preprocessing; 3. Functional connectivity analysis; 4. Connectome model construction; 5. Disease model construction; 6. Treatment strategy development.

[0004] 2. Microscopic local neural network simulation: Process the microscopic image data to extract the neuron type and connection structure, so as to simulate the neural activity at the microscopic level, and establish the neuron and synapse model to simulate the activity of fine neural circuits. This method mainly includes the following steps: 1. Data acquisition; 2. Image segmentation, identification of neuron type and synaptic connection; 3. Establishment of neuron and synapse model; 4. Establishment of microscopic local neural network; 5. Introducing stimulation to simulate network discharge; 6. Analyzing the functional principle of neural circuits.

[0005] Microscopic local neural network simulation technology requires relatively detailed microscopic data and complex data processing to extract neuron types and connection structures, and the detailed modeling of neurons and synapses is difficult to simulate large-scale whole-brain neural networks. The complex discharge of neurons also increases the difficulty of analyzing the function of neural circuits.

[0006] Mesoscopic structure or functional connectivity simulation methods cannot accurately simulate the neuronal firing activity within brain regions. Summary of the invention

[0007] In order to solve the above problems in the prior art, that is, to solve the problem that the prior art cannot take into account both the simulation of large-scale whole-brain neural networks and the fine simulation of neuronal discharge activities in brain regions, which leads to poor accuracy and robustness of neural activity simulation. The present invention proposes a multi-scale human brain simulation system that integrates data, structure, mechanism and behavior, which includes:

[0008] A database construction module is configured to collect brain image data of patients with set pathological types and preprocess them to obtain preprocessed data; based on the preprocessed data, a human brain connection map is drawn, and a storage database is constructed in combination with a knowledge set corresponding to the pathological type and clinical data of each patient;

[0009] A cortical column model building module, configured to build a multi-layer cortical column model of a corresponding brain region according to the selected neuron model in combination with the knowledge set in the storage database;

[0010] A full-scale model building module is configured to connect brain regions by combining the human brain connection map and the multi-layer cortical column model of each brain region in the storage database to obtain a large-scale neural network in which brain regions influence each other, i.e., a full-scale model;

[0011] A model training module, configured to train the full-scale model based on the clinical data in the storage database; during training, a tensor parallel distributed training acceleration method of a sparse matrix is ​​used to accelerate the training;

[0012] A simulation module, configured to simulate neural activity using a trained full-scale model.

[0013] In some preferred embodiments, the multi-scale human brain simulation system integrating data-structure-mechanism-behavior further includes a virtual screening module and an optimization module;

[0014] The virtual screening module is configured to evaluate and screen different neural activity simulations through the full-scale model;

[0015] The optimization module is configured to adjust the full-scale model parameters and optimize the neural activity simulation scheme according to the neural activity simulation effect.

[0016] In some preferred embodiments, the neuron model includes a LIF model, a GLIF model, a HH model, and a multi-compartmental neuron model;

[0017] The LIF model is:

[0018]

[0019] Where V represents the neuron membrane potential, t is time, C is capacitance, G is L represents leakage conductance, Isyn represents input current from synapse, and Iinj represents external stimulation current.

[0020] In some preferred embodiments, a multi-layer cortical column model of a corresponding brain region is constructed by:

[0021] Classifying neurons into different types, placing them at corresponding levels in the cortical column according to the brain region knowledge in the knowledge set, and allocating the proportion of each type of neurons in each layer of the cortical column;

[0022] Neurons of different types and levels form synaptic connections using connection probability matrices;

[0023] Randomly assign weights that follow the Gamma distribution to each synaptic connection;

[0024] After allocation, the positive and negative values ​​of the weights of each synaptic connection are determined according to the type of neuron, and then a multi-layer cortical column model of the corresponding brain area is obtained.

[0025] In some preferred embodiments, the synaptic connections and synaptic connection weights of neurons in the constructed multi-layer cortical column model of each brain region are stored using a sparse matrix.

[0026] In some preferred embodiments, the brain regions are connected by combining the human brain connection map and the multi-layer cortical column model of each brain region in the storage database to obtain a large-scale neural network in which the brain regions influence each other, and the method is as follows:

[0027] For each connected brain area, the brain area before the connection is taken as the source brain area, and the brain area after the connection is taken as the target brain area; the pulse firing rates of the excitatory neurons and inhibitory neurons in the source brain area are calculated, and multiplied by the connection weights between the brain areas respectively to obtain the external stimulation input into the neurons of the target brain area, and then obtain a large-scale neural network in which the brain areas influence each other.

[0028] In some preferred embodiments, the full-scale model is trained based on the clinical data in the storage database by:

[0029] Inputting random background current into the full-scale model to generate discharge activity and obtain a simulated signal;

[0030] Performing time-frequency analysis on the simulated signal and the real signal of each brain region respectively to obtain a simulated signal time-frequency diagram and a real signal time-frequency diagram; the real signal is the clinical data;

[0031] Calculating the loss between the simulated signal time-frequency diagram and the real signal time-frequency diagram, and calculating the gradient of the parameters through the back propagation algorithm, performing gradient descent, and optimizing the model gradient;

[0032] The full-scale model is trained in a loop until a trained full-scale model is obtained.

[0033] In some preferred embodiments, when performing back-propagation algorithm calculations, only the gradients of non-zero elements are calculated in sparse matrix multiplication.

[0034] In some preferred embodiments, a sparse matrix tensor parallel distributed training acceleration method is used to accelerate training, and the method is:

[0035] During the training process, the GPU of the multi-scale human brain simulation system of data-structure-mechanism-behavior fusion uses the cuSparse acceleration library to perform large-scale sparse tensor calculations; the dynamic equations of complex neurons and the refined dendritic nonlinear operations are fused by operators; the complex neurons include HH neurons and multi-compartmental neurons; the refined dendrites include synapses of synaptic currents simulated by the nonlinear dynamic equations of NMDA receptors and synapses of synaptic currents simulated by the nonlinear dynamic equations of GABA receptors;

[0036] Interaction of neuron information is carried out through the distributed multi-card communication framework NCCL;

[0037] Statistical methods are used to evaluate the distribution characteristics of communication volume of unstructured sparse pulse data and high fan-in and fan-out neuron connection topology, identify nodes and areas with high communication demand, and use graph segmentation algorithm to divide the neural network. Under the condition of node load balancing, neurons with high communication volume are allocated to the same computing node; the high fan-in and fan-out neurons are neurons with in-degree and out-degree exceeding the previously set percentage.

[0038] In some preferred embodiments, when performing back-propagation algorithm calculations, when the pulses emitted by neurons are not differentiable, the gradient of the pulses is approximated by a proxy gradient.

[0039] Beneficial effects of the present invention:

[0040] The present invention achieves a balance between simulation of large-scale whole-brain neural networks and fine simulation of neuronal discharge activities within brain regions, thereby improving the accuracy and robustness of neural activity simulation.

[0041] 1) The present invention adopts a multi-scale combined brain simulation method. The model not only simulates the functional or structural connections between brain regions, but also finely models the cortical columns, neurons and synapses in the brain regions. It can more finely analyze the effects of different types of neurons and synaptic parameters on simulated neural activity, which helps to simulate and analyze drug stimulation targets more accurately;

[0042] 2) The parameters in the present invention can be trained by the back-propagation method, which can make the simulated neural activity closer to the observed data and improve the credibility of the simulation;

[0043] 3) The modeling methods and parameters of neurons and synapses in the present invention can be flexibly adjusted, with a high degree of operational freedom. Full coverage from microscopic (ion channels, synapses, neurons)-mesoscopic (neural circuits)-macroscopic (whole brain connections). It has the ability to observe macroscopic neural activity phenomena corresponding to modifications to microscopic structures; in order to be able to flexibly adapt to differentiated human brain connection groups, brain area connection ports are reserved to achieve high compatibility from computing components to connection methods, laying the foundation for individualized brain simulation. Compared with other brain simulation technologies and platforms, this allows users to process and change model settings in a high degree of freedom, thereby performing brain simulation more accurately;

[0044] 4) The brain simulation engine of the present invention has the ability to achieve diversification and heterogeneity of computing components. Different from common brain simulation technologies, multiple components of different types or with different parameter settings can coexist in the same model of the platform. Different neuron dynamic models are used for specific functions and diseases (for example, the HH neuron model focusing on ion channel changes, the LIF neuron model focusing on membrane parameters, the multi-compartmental neuron model focusing on dendrite calculations, etc.), which are easy to operate and more adaptable;

[0045] 5) The present invention uses the collected neural and clinical medical data to establish a model and then simulate neural activity, providing a high-degree-of-freedom simulation platform for processing brain activity and working methods that can be explained for medical and clinical research. This platform allows users to conduct low-cost, accurate and reliable neural dynamics experiments, and helps users observe, study and predict neural activity by simulating drug targets or transcranial stimulation, greatly accelerating the revelation of the working mechanism of the nervous system and the causes of neural diseases, as well as the research and formulation of treatment plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings.

[0047] Figure 1 It is a schematic diagram of the framework of a multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to an embodiment of the present invention;

[0048] Figure 2 is a schematic diagram of a construction process of a full-scale model of an embodiment of the present invention;

[0049] Figure 3 is a schematic diagram of a full-scale model training process of an embodiment of the present invention;

[0050] Figure 4 It is a flowchart of a method for simulating a human brain driven by neuroscience data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0053] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0054] A multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to a first embodiment of the present invention comprises:

[0055] A database construction module is configured to collect brain image data of patients with set pathological types and preprocess them to obtain preprocessed data; based on the preprocessed data, a human brain connection map is drawn, and a storage database is constructed in combination with a knowledge set corresponding to the pathological type and clinical data of each patient;

[0056] A cortical column model building module, configured to build a multi-layer cortical column model of a corresponding brain region according to the selected neuron model in combination with the knowledge set in the storage database;

[0057] A full-scale model building module is configured to connect brain regions by combining the human brain connection map and the multi-layer cortical column model of each brain region in the storage database to obtain a large-scale neural network in which brain regions influence each other, i.e., a full-scale model;

[0058] A model training module, configured to train the full-scale model based on the clinical data in the storage database; during training, a tensor parallel distributed training acceleration method of a sparse matrix is ​​used to accelerate the training;

[0059] The simulation module is configured to simulate neural activity through a trained full-scale model.

[0060] In order to more clearly explain the multi-scale human brain simulation system of data-structure-mechanism-behavior fusion of the present invention, the following is combined with the attached Figure 1 Each module in an embodiment of the system of the present invention is described in detail.

[0061] The present invention uses actual neural and clinical medical data to establish a model and then simulate neural activity data, providing a high-degree-of-freedom platform for simulating brain activity and working methods for medical and clinical research, allowing researchers to conduct low-cost, precise and controllable neural dynamics experimental platforms, and helping researchers study and predict neural activity through simulations such as drug targets or transcranial stimulation, assisting the development of neural stimulation therapy and drug development, accelerating drug development efficiency, improving the accuracy and therapeutic effect of neural stimulation therapy, or reducing surgical risks. The details are as follows:

[0062] The present invention provides a multi-scale human brain simulation system integrating data, structure, mechanism and behavior, including a database construction module, a cortical column model construction module, a full-scale model construction module and a model training module;

[0063] A database construction module is configured to collect brain image data of patients with set pathological types and preprocess them to obtain preprocessed data; based on the preprocessed data, a human brain connection map is drawn, and a storage database is constructed in combination with a knowledge set corresponding to the pathological type and clinical data of each patient;

[0064] In this embodiment, the patient's EEG (Electroencephalogram), fMRI (functional magnetic resonance imaging) or DTI (diffusion tensor image) data is collected through clinical equipment, and sufficient data preprocessing (including standardization, smoothing, etc.) is performed. Through coherence analysis, correlation analysis, cross-correlation analysis, causal analysis and other methods, the functional or structural connections between brain regions are calculated for modeling individual brain models (i.e., building a human brain connection map). Combined with expert knowledge and brain science knowledge engine brainknow (see: http: / / www.brain-knowledge-engine.org / ), knowledge of brain regions and neural circuits related to the disease is collected to form a knowledge set. The knowledge set related to the disease is combined with the clinical data set of each individual patient to build a corresponding storage database.

[0065] A cortical column model building module, configured to build a multi-layer cortical column model of a corresponding brain region according to the selected neuron model in combination with the knowledge set in the storage database;

[0066] According to their complexity, neuron models are divided into LIF model (Leaky Integrate-and-Fire model), GLIF model (Generalized leaky integrate-and-fire model), HH model (Hodgkin-Huxley model, a neuron model that accurately simulates neuron membrane potential and ion channel dynamics), and multi-compartmental neuron model.

[0067] In this embodiment, drawing on the working mechanism of pulse discharge of neurons in the brain, different neuron model calculation components can be selected according to the precision requirements of the simulation. Then, based on the knowledge of brain areas, neurons are divided into different types, placed in the corresponding layers of the cortical column, and the proportion of each type of neurons in each layer of the cortical column is allocated. Synaptic connections are formed between neurons of different types and levels with corresponding probabilities (i.e., connection probability matrices). Weights that obey the Gamma distribution are randomly assigned to each synaptic connection; after assignment, the positive and negative weights of each synaptic connection are determined according to the type of neuron (i.e., when the presynaptic neuron is an excitatory neuron, the synapse is an excitatory synapse, and the synaptic weight is positive. When the presynaptic neuron is an inhibitory neuron, the synapse is an inhibitory synapse, and the synaptic weight is negative), thereby obtaining a multi-layer cortical column model of the corresponding brain area.

[0068] In addition, since the density of neuronal connections within brain regions is often sparse, in order to save memory, the present invention proposes using sparse matrices to store the connections and connection weights between neurons, greatly reducing the memory requirements of subsequent large-scale training steps of the whole-brain model.

[0069] The full-scale model building module is configured to connect brain regions by combining the human brain connection map and the multi-layer cortical column model of each brain region in the storage database to obtain a large-scale neural network of mutual influence between brain regions, i.e., a full-scale model; Figure 2 shown.

[0070] In this embodiment, by referring to the resting state and functional state connection groups in the brain (such as functional connection, structural connection and other mechanisms), the brain regions are connected through long-range synapses, specifically:

[0071] According to the connection matrix between brain regions, weights are assigned to the connections between brain regions;

[0072] Combine macro brain connection with micro brain simulation. Let the number of neurons in brain region i be Ni. If the synapses are refined to simulate the long-range connections between brain regions, then a matrix of size (sum_i(Ni)*sum_i(Ni)) is required to store the synapses and weights of the whole brain. The memory consumption of this method is unacceptable. Therefore, the present invention proposes a large-scale brain simulation with multi-scale combination, in the following way: for each connection, the brain region before the connection is the source brain region, and the brain region after the connection is the target brain region. The pulse firing rate of the excitatory neurons and inhibitory neurons in the source brain region is calculated, and the connection weights between brain regions independently obtained by the database construction module are multiplied to obtain the external stimulation input to the neurons in the target brain region. In combination with the cortical column model construction module, a sparse matrix is ​​used to store the fine connections inside each brain region. The method of the present invention greatly reduces the memory requirement of the whole brain simulation by multiplying the connection weights of the mesoscopic brain regions, and the connection weights are independently obtained, making it possible to further train the model.

[0073] The model training module is configured to train the full-scale model based on the clinical data in the storage database; during training, a tensor parallel distributed training acceleration method of a sparse matrix is ​​used to accelerate the training.

[0074] In this embodiment, the model is adjusted based on clinical data to make the neural activity of the simulated brain area correspond to the actual observation results. Specifically:

[0075] The model is driven by random background current to generate discharge activity, and the correlation of activities between calculated brain regions and the EEG of simulated brain regions are obtained, i.e., simulated signals.

[0076] Calculate the error between the simulated brain area activity (i.e., the simulated signal) and the actual data (real signal) (for example: perform time-frequency analysis on the simulated signal and the real signal of each brain area respectively, obtain the simulated signal time-frequency diagram and the real signal time-frequency diagram, and calculate the loss between the simulated signal time-frequency diagram and the real signal time-frequency diagram (including the mean square error, etc.)).

[0077] Among them, a parallel distributed training acceleration method is used for sparse matrix tensors, specifically:

[0078] Based on the sparse tensor cuSparse acceleration library, the underlying training and inference framework for large-scale spiking neural networks is designed to give full play to the sparse characteristics of spiking neural networks and the parallel computing capabilities of GPUs; operator fusion is performed on the dynamic equations of complex neurons and refined dendritic nonlinear operations to improve the GPU computing and bandwidth utilization; complex neurons include HH neurons and multi-compartmental neurons; refined dendrites include synapses whose synaptic currents are simulated by the nonlinear dynamic equations of NMDA receptors and synapses whose synaptic currents are simulated by the nonlinear dynamic equations of GABA receptors.

[0079] Operator fusion is an existing technology, refer to https: / / chenzomi12.github.io / 03Compiler03Frontend / 03OPFusion.html.

[0080] Based on the high-performance distributed multi-card communication framework NCCL, the neuron information interaction of brain simulation is realized, so as to achieve large-scale expansion of brain simulation tasks so that it is not limited to the video memory capacity of a single-card GPU; for the connection topology of unstructured sparse pulse data and neurons with high fan-in and fan-out (neurons with high fan-in and fan-out are neurons with in-degree and out-degree exceeding a previously set percentage (preferably 25% in the present invention)), communication mode analysis and graph segmentation and algorithm optimization are performed, statistical methods are used to evaluate the distribution characteristics of communication volume, identify nodes and areas with high communication requirements, and use graph segmentation algorithms to divide the neuron network, distribute neurons with high communication volume to the same computing node as much as possible, and consider load balancing of different computing nodes to ensure effective utilization of computing resources, thereby reducing communication overhead and communication delay between cards.

[0081] The gradient of the parameters in the network is calculated by the back-propagation method. In the existing network training framework, the training of sparse matrix weights has not been configured. The optimizer of the existing model also supports only dense parameters by default. The present invention proposes to calculate only the gradient of non-zero elements in sparse matrix multiplication, and customize the optimizer so that the elements that were originally 0 in the sparse matrix do not participate in the network training. In this way, a training method that does not need to convert sparse matrices into dense matrices during training is realized, and combined with a multi-card parallel computing method, a large-scale whole-brain model can be trained within an acceptable memory requirement.

[0082] When encountering a non-differentiable situation where a neuron emits a pulse, the proxy gradient method is used to approximate the gradient of the pulse: the proxy gradient formula is preferably as follows:

[0083]

[0084] Wherein, H is the pulse function, α represents the coefficient of the control function width, and x is the independent variable of the function. When x>0, H(x)=1, otherwise H(x)=0.

[0085] The model parameters are iteratively optimized using stochastic gradient descent.

[0086] A simulation module, configured to simulate neural activity using a trained full-scale model.

[0087] In this embodiment, neural activity simulation includes helping users observe, study and predict neural activity by simulating drug targets or transcranial stimulation; that is, simulating the metabolic process of drugs in a digital brain model (i.e., a full-scale model), and how drugs interact with specific receptors in the brain, or simulating the effects of stimulation therapy on targets. Use the model to predict the effects of corresponding treatments on EEG activity, neurotransmitter balance or other related biomarkers. The specific simulation process is as follows:

[0088] Determine the set of parameters of brain regions affected by the treatment and the targets of stimulation produced by the treatment;

[0089] Adjust these parameters accordingly, apply corresponding stimulation to the target, observe the changes in the brain simulation activities of interest, and simulate the effects of treatment;

[0090] The effects of treatment on the model are compared with clinical data, and drug dosage or stimulation parameters are matched to the adjustment amplitude of parameters or the intensity of stimulation in the simulation model.

[0091] The multi-scale human brain simulation system integrating data-structure-mechanism-behavior also includes virtual screening module and optimization module;

[0092] A virtual screening module, configured to evaluate and screen different neural activity simulations through the full-scale model, that is, to evaluate and screen the potential efficacy and safety of different drugs through the full-scale model;

[0093] In this embodiment, virtual screening is performed in a digital brain model, and the dynamic changes of the neural network are observed through quantitative indicators such as brain region pulse firing rate, simulated EEG, brain region activity correlation, and coherence to evaluate the potential efficacy and safety of different candidate treatment options. The effects of different treatment options are compared through the model to optimize the treatment option.

[0094] An optimization module, configured to adjust the full-scale model parameters and optimize the neural activity simulation scheme according to the neural activity simulation effect; that is, to adjust the full-scale model parameters and optimize the treatment scheme according to the treatment effect;

[0095] In this embodiment, after treatment, the treatment effect is evaluated and the actual results are fed back to the digital brain. The model parameters are dynamically adjusted according to the treatment effect to continuously optimize the treatment plan.

[0096] In order to more clearly illustrate the technical solution of the present invention, the following describes the implementation method by taking the construction of a human brain simulation system to assist in the development of drugs for depression as a specific example.

[0097] 1) Collect EEG, MRI, DTI and other data and calculate the functional connection matrix. Combined with the brain science knowledge engine brainknow (http: / / www.brain-knowledge-engine.org / ), collect knowledge about brain regions and neural circuits related to depression to form a knowledge set. The disease-related knowledge set is combined with the clinical data set of each individual patient to build a corresponding storage database. The specific method includes:

[0098] Use MRI scanners to perform high-resolution structural imaging to obtain detailed information about the brain's anatomical structure. Use diffusion tensor imaging (DTI) or more advanced diffusion imaging techniques (such as high-angle resolution diffusion imaging, HARDI) to depict the diffusion path of water molecules in the brain's white matter, thereby inferring the direction of nerve fibers and analyzing the functional connections between brain regions at rest through computers. Subsequently, a standardized processing process (motion artifact correction, spatial normalization, smoothing, etc.) is used to draw a human brain connection map;

[0099] Using the brain science knowledge engine brainknow to collect knowledge about brain regions related to depression;

[0100] Integrate raw data, human brain connection maps and knowledge collection to build a storage database;

[0101] 2) Based on the knowledge of disease-related brain regions in 1), a microscopic neuron and synaptic connection model of the corresponding brain region is established. Neurons and synaptic connections form a hierarchical cortical column structure.

[0102] The neuron is modeled as a LIF model, and the model dynamics are as follows:

[0103]

[0104] Where V represents the neuron membrane potential, t is time, C is capacitance, G is L represents leakage conductance, Isyn represents input current from synapse, and Iinj represents external stimulation current. When the membrane potential exceeds the threshold Vt, the neuron releases a pulse at this moment, that is, the output value is 1.

[0105] Neurons are divided into excitatory neurons and inhibitory neurons. The cortical column has 6 layers, and the excitatory and inhibitory neurons in each layer make up a total of 12 types of neurons. The proportion of each type of neuron in each layer is allocated according to the knowledge of brain regions, as well as a 12*12 connection probability matrix between the 12 types of neurons. Synaptic connections are formed between neurons of different types and levels using the connection probability matrix. Weights that obey the Gamma distribution are randomly assigned to each synaptic connection. After the assignment, the positive and negative weights of each synaptic connection are determined according to the type of neuron, and a neuron and synaptic connection model in the cortical column (or a multi-layer cortical column model of the corresponding brain region) is constructed.

[0106] 3) After modeling the cortical columns of multiple brain regions, based on the structural or functional connections obtained in 1), connections are established between brain regions through long-range synapses to form a large-scale neural network in which brain regions influence each other. Specifically:

[0107] According to the connection matrix between brain regions, weights are assigned to the connections between brain regions; for each connection, the brain region before the connection is the source brain region, and the brain region after the connection is the target brain region. The pulse firing rates of the excitatory neurons and inhibitory neurons in the source brain region are calculated and multiplied by the connection weights between brain regions independently obtained in 1) to obtain the external stimulation input to the neurons in the target brain region.

[0108] 4) Based on clinical data, train the parameters in the model so that the neural activity of the simulated brain area corresponds to the actual observation results. In this example, EEG is used as clinical observation data (i.e., clinical data); Figure 3 As shown;

[0109] The model is driven by random background current (or background stimulation current) to generate discharge activity, and the average synaptic current of the excitatory synapses in the brain region is used to approximate the EEG signal; the time-frequency analysis of the EEG signals in each brain region is performed to obtain the time-frequency diagram of the simulated and real EEG signals (i.e. Figure 3 The real time-frequency diagram and simulated time-frequency diagram in the figure are used to calculate the mean square error between the time-frequency diagrams.

[0110] The back-propagation algorithm is used to calculate the gradient of the parameters, perform gradient descent, optimize the model gradient, and combine with the multi-card parallel computing method to enable large-scale whole-brain models to be trained within acceptable memory requirements;

[0111] The model parameters are iteratively optimized using the stochastic gradient descent method, and the training is repeated until a trained full-scale model (i.e. Figure 3 Multiscale digital brain model in .

[0112] 5) Therapeutic stimulation simulation. The metabolic process of drugs used to treat depression is simulated in the digital brain model. The model is used to predict the effects of drugs on brain electrical activity, neurotransmitter balance, or other relevant biomarkers. The specific steps are as follows:

[0113] Determine the set of parameters of brain regions affected by the drug and how the drug affects these parameters; adjust these parameters accordingly and observe whether the pulse firing rate of brain regions that are lower than normal increases after being affected by the drug;

[0114] Compare the increase in firing rate and changes in brain functional connectivity caused by the drug with clinical data to evaluate the effect of the drug:

[0115] Obtain EEG data after drug treatment, perform the same processing from 1) to 4) on the data to obtain a new brain model and corresponding parameters; compare the parameters related to drug metabolism in the new model with those in the previous model, determine the corresponding relationship between the parameter changes in the previous model and the drug dosage, and evaluate the drug effect;

[0116] 6) Virtual screening and optimization: Conduct virtual screening in digital brain models to compare the extent to which different drugs increase the pulse firing rate in brain regions and evaluate the potential efficacy and safety of different drugs;

[0117] 7) Effect evaluation and long-term optimization. After using the selected drug for treatment, evaluate the treatment effect and feed the actual results back to the digital brain. Dynamically adjust the model parameters according to the treatment effect and continuously optimize the treatment plan.

[0118] In summary, the present invention uses actual brain neural structure data to construct a simulated nervous system. Through multi-scale collaborative modeling of brain function diseases, the system will produce accurate and reliable simulated neural activity that matches the brain. The simulation model of the present invention is based on clinical data sets, and the microstructure is targeted and simplified to achieve more sophisticated simulation within acceptable computing resource requirements. For the settings of the running model or the targeted adjustments and modifications of the targets, the model produces corresponding meaningful neural dynamic changes, thereby predicting the changes and effects of the neural activity corresponding to the manipulation of the neural target.

[0119] It should be noted that the multi-scale human brain simulation system of data-structure-mechanism-behavior fusion provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps and are not regarded as improper limitations of the present invention.

[0120] A human brain simulation method based on neuroscience data driven according to a second embodiment of the present invention, such as Figure 4 As shown, the method comprises the following steps:

[0121] Data and knowledge collection: collecting brain image data of patients with set pathological types and preprocessing them to obtain preprocessed data; drawing a human brain connection map based on the preprocessed data, and building a storage database in combination with the knowledge set corresponding to the pathological type and the clinical data of each patient;

[0122] Establishing a cortical column model; combining the knowledge set in the storage database and constructing a multi-layer cortical column model of the corresponding brain region according to the selected neuron model;

[0123] Establishing a full-scale model; combining the human brain connection map stored in the database and the multi-layer cortical column model of each brain region, connecting brain regions to obtain a large-scale neural network that influences each other between brain regions, i.e., a full-scale model;

[0124] Model adjustment based on clinical data; training the full-scale model based on the clinical data in the storage database; during training, using a tensor parallel distributed training acceleration method of a sparse matrix to accelerate training;

[0125] Therapeutic stimulation simulation; neural activity simulation through trained full-scale models;

[0126] Virtual screening and optimization: evaluating and screening different neural activity simulations through the full-scale model;

[0127] Effect evaluation and long-term optimization: According to the effect of neural activity simulation, adjust the full-scale model parameters and optimize the neural activity simulation scheme.

[0128] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the method described above can refer to the corresponding process in the aforementioned system embodiment, and will not be repeated here.

[0129] A third embodiment of the present invention is a human brain simulation device driven by neuroscience data, comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned human brain simulation method driven by neuroscience data.

[0130] A fourth embodiment of the present invention is a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by a computer to implement the above-mentioned human brain simulation method driven by neuroscience data.

[0131] Technicians in the technical field can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the above-described human brain simulation device driven by neuroscience data and computer-readable storage medium can refer to the corresponding process in the aforementioned system example and will not be repeated here.

[0132] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the technical field. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0133] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.

[0134] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not expressly listed, or also includes elements inherent to such process, method, article, or apparatus / device.

[0135] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A multi-scale human brain simulation system integrating data, structure, mechanism and behavior, characterized in that: The system includes: A database construction module is configured to collect brain image data of patients with set pathological types and preprocess them to obtain preprocessed data; based on the preprocessed data, a human brain connection map is drawn, and a storage database is constructed in combination with a knowledge set corresponding to the pathological type and clinical data of each patient; A cortical column model building module, configured to build a multi-layer cortical column model of a corresponding brain region according to the selected neuron model in combination with the knowledge set in the storage database; The neuron models include LIF model, GLIF model, HH model, and multi-compartment neuron model; The LIF model is: Where V represents the neuron membrane potential, t is time, C is capacitance, G is L represents leakage conductance, Isyn represents input current from synapse, and Iinj represents external stimulation current; Construct a multi-layer cortical column model corresponding to the brain region by: Classifying neurons into different types, placing them at corresponding levels in the cortical column according to the brain region knowledge in the knowledge set, and allocating the proportion of each type of neurons in each layer of the cortical column; Neurons of different types and levels form synaptic connections using connection probability matrices; Randomly assign weights that follow the Gamma distribution to each synaptic connection; After the allocation, the positive and negative weights of each synaptic connection are determined according to the type of neuron, and then a multi-layer cortical column model of the corresponding brain region is obtained; The synaptic connections and synaptic connection weights of neurons in the constructed multi-layer cortical column model of each brain region are stored using sparse matrices; The full-scale model building module is configured to connect brain regions by combining the human brain connection map and the multi-layer cortical column model of each brain region in the storage database to obtain a large-scale neural network that influences each other between brain regions, i.e., a full-scale model: For each connected brain region, the brain region before the connection is taken as the source brain region, and the brain region after the connection is taken as the target brain region; the pulse firing rates of the excitatory neurons and inhibitory neurons in the source brain region are calculated, and the pulse firing rates are multiplied by the connection weights between the brain regions, respectively, to obtain the external stimuli input into the neurons of the target brain region, and then to obtain a large-scale neural network of mutual influence between brain regions; A model training module, configured to train the full-scale model based on the clinical data in the storage database; during training, a tensor parallel distributed training acceleration method of a sparse matrix is ​​used to accelerate the training; A simulation module, configured to simulate neural activity using a trained full-scale model.

2. The multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to claim 1 is characterized in that: The multi-scale human brain simulation system integrating data, structure, mechanism and behavior also includes a virtual screening module and an optimization module; The virtual screening module is configured to evaluate and screen different neural activity simulations through the full-scale model; The optimization module is configured to adjust the full-scale model parameters and optimize the neural activity simulation scheme according to the neural activity simulation effect.

3. The multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to claim 2 is characterized in that: Based on the clinical data in the storage database, the full-scale model is trained by: Inputting random background current into the full-scale model to generate discharge activity and obtain a simulated signal; Performing time-frequency analysis on the simulated signal and the real signal of each brain region respectively to obtain a simulated signal time-frequency diagram and a real signal time-frequency diagram; the real signal is the clinical data; Calculating the loss between the simulated signal time-frequency diagram and the real signal time-frequency diagram, and calculating the gradient of the parameters through the back propagation algorithm, performing gradient descent, and optimizing the model gradient; The full-scale model is trained in a loop until a trained full-scale model is obtained.

4. The multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to claim 3 is characterized in that: When performing back-propagation algorithm calculations, only the gradients of non-zero elements are calculated in the sparse matrix multiplication.

5. The multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to claim 4 is characterized in that: The sparse matrix tensor parallel distributed training acceleration method is used to accelerate training. The method is as follows: During the training process, the GPU of the multi-scale human brain simulation system of data-structure-mechanism-behavior fusion uses the cuSparse acceleration library to perform large-scale sparse tensor calculations; the dynamic equations of complex neurons and the refined dendritic nonlinear operations are fused by operators; the complex neurons include HH neurons and multi-compartmental neurons; the refined dendrites include synapses of synaptic currents simulated by the nonlinear dynamic equations of NMDA receptors and synapses of synaptic currents simulated by the nonlinear dynamic equations of GABA receptors; Interaction of neuron information is carried out through the distributed multi-card communication framework NCCL; Statistical methods are used to evaluate the distribution characteristics of communication volume of unstructured sparse pulse data and high fan-in and fan-out neuron connection topology, identify nodes and areas with high communication demand, and use graph segmentation algorithm to divide the neural network. Under the condition of node load balancing, neurons with high communication volume are allocated to the same computing node; the high fan-in and fan-out neurons are neurons with in-degree and out-degree exceeding the previously set percentage.

6. The multi-scale human brain simulation system integrating data, structure, mechanism and behavior according to claim 4, characterized in that: When performing back-propagation algorithm calculations, when the pulse emitted by a neuron is not differentiable, the gradient of the pulse is approximated by a proxy gradient.

Images