Evolutionary spiking neural network cancer driver gene screening method, system and apparatus

Abstract

The application provides an evolutionary pulse neural network cancer pathogenic gene screening method, system and device. The method comprises the following steps: obtaining gene mutation data and pathway correlation data in a sample; screening candidate genes to obtain a significant mutation gene set; constructing a complex complex network according to the hierarchical relationship between pathways and the gene relationship contained in the pathways; constructing an evolutionary pulse neural network on the basis of the complex complex network; converting input data into a pulse sequence in a pulse coding mode, and learning and training the network; and completing subtype classification and pathogenic gene screening according to an output layer dynamic evolution mechanism. The scheme improves the biomedical interpretability and classification accuracy of cancer pathogenic gene screening by embedding biological pathway information into an evolutionary pulse neural network model, and is suitable for subtype classification, pathogenic gene screening and auxiliary medical identification scenarios.

Description

Technical Field

[0001] This invention belongs to the fields of biomedical engineering, bioinformatics, artificial intelligence, and gene data processing. Specifically, it relates to an evolutionary spiking neural network method, system, and device for screening cancer-causing genes by embedding gene association pathway information, so as to use the evolutionary spiking neural network to screen cancer-causing genes and classify them into corresponding subtypes. Background Technology

[0002] This invention is the research result of the major project "Brain Science and Brain-like Research" under the Science and Technology Innovation 2030 Program, with the project number 2021ZD0201302 and the title "Cognitive-Learning Mechanism and Method of Spiral Neural Networks".

[0003] With the development of high-throughput sequencing technology, artificial intelligence technology, and mathematical modeling methods, using gene data to screen cancer-causing genes and classify gene subtypes has become an important direction for interdisciplinary research in medicine and mathematics.

[0004] Currently, for cancer genes, gene-based subtype classification methods can be broadly categorized into unsupervised machine learning methods, supervised machine learning methods, and deep learning methods. Traditional unsupervised methods often cluster based on sample similarity or low-rank representations, supervised methods frequently employ classifiers such as support vector machines, and deep learning methods learn multi-omics features through autoencoders, graph neural networks, or generative models. While most current neural network-based cancer gene subtype classification models can improve recognition capabilities to some extent, they often fail to adequately consider intrinsic biological information and lack biomedical interpretability, resulting in poor classification performance and inaccurate identification.

[0005] On the other hand, cancer-related gene data is usually extremely high in dimensionality, and the number of genes involved in case samples is enormous. If all gene features are directly input into a neural network model, it can easily lead to a decrease in learning efficiency, insufficient model generalization ability, and a lack of stability in training results.

[0006] Furthermore, existing models typically focus primarily on statistical features at the gene level, while rarely embedding structured biological information at the pathway level directly into neural networks.

[0007] Therefore, there is an urgent need for a new cancer-causing gene screening technology that can effectively reduce dimensionality and screen high-dimensional gene data, directly embed gene pathway biological information into spiking neural networks, and utilize the dynamic structural learning ability of evolutionary spiking neural networks to improve the interpretability and accuracy of gene subtype classification and pathogenic gene identification. Summary of the Invention

[0008] To address the shortcomings of existing cancer-causing gene identification models, such as insufficient biomedical interpretability, poor learning performance on high-dimensional gene data, and difficulty in fully utilizing gene pathway information, this invention provides an evolutionary spiking neural network method for cancer-causing gene identification that incorporates gene association pathway information. Specifically, this invention provides the following technical solution: On one hand, this invention provides a method for screening cancer-causing genes using evolutionary spiking neural networks, the method comprising: S1. Obtain gene mutation data, gene copy number variation data, and gene pathway relationship data from cancer case samples, and construct the original sample set; S2. Based on the gene mutation frequency, candidate genes in the original sample set are screened to obtain a set of key genes; S3. Based on the pathway hierarchy and pathway-included gene relationships in the pathway database, construct a complex network with embedded gene-associated pathway information. S4. Based on the aforementioned complex network, construct an evolutionary spiking neural network, using single nucleotide variation index and copy number variation index as input layer features, and establish a spiking neuron model for the network nodes. S5. Perform pulse coding on the input layer features to obtain the input pulse sequence, and train the evolutionary spiking neural network based on the pulse temporal dependence plasticity rule; S6. Generate category representation neurons based on the dynamic evolution mechanism of the output layer during training, and complete the matching between the test sample and the output layer neurons; output subtype classification results and / or pathogenic gene screening results based on the matching results.

[0009] Preferably, in step S2, a candidate gene set is determined based on gene mutation data, and the candidate gene set is screened for significance. Genes that meet the significance threshold are identified as a set of significant mutant genes, which serve as the key gene set. The gene copy number variation data is used as a feature of the subsequent input layer to participate in the evolutionary spiking neural network modeling, or as a supplementary source for the candidate gene set.

[0010] Preferably, the complex network constructed in S3 includes an input layer, a gene layer, a pathway layer, and an evolutionary output layer connected in sequence; The pathway layer is constructed as follows: based on the directed hierarchical relationship between pathways in the Reactome database, a multi-layered pathway node from layer 1 to layer 5 is constructed, where each layer of pathway node represents a different level of human biological pathway. Then, edges are connected according to the hierarchical relationship between pathways to form a pathway layer. The gene layer is determined by the intersection of the genes contained in the first layer of the pathway layer and the set of key genes; The number of neurons in the input layer is three times the number of neurons in the gene layer, representing whether the gene has undergone somatic mutation, copy number increase, or copy number decrease, respectively.

[0011] Preferably, the input layer and the gene layer are connected by a fully connected method; the pathway layer is connected by a sparse connection method based on pathway relationships; and the gene layer and the pathway layer are connected by a sparse connection method based on gene-pathway attribution relationships.

[0012] Preferably, in S4, the spiking neuron model adopts the LIF neuron model, and membrane potential states are established for the input layer nodes, gene layer nodes, pathway layer nodes, and evolutionary output layer nodes in the complex network; for the first The membrane potential of a neuron is updated as follows: in, Indicates the membrane time constant. Indicates the first At time 1 neuron membrane potential, Indicates the resting membrane potential. Indicates membrane impedance, Indicates the first At time 1 neuron The received input current; the input current It is obtained by weighted summation of the impulse outputs of the precursor neurons via synaptic weights; when Reaching or exceeding the preset discharge threshold At that time, the first Each neuron fires a pulse, denoted as _____. And reset the membrane potential to the reset membrane potential. ;when The preset discharge threshold was not reached. At that time, the first A neuron that does not fire a pulse is denoted as _____. .

[0013] Preferably, in step S5, a frequency-based pulse coding method is used to encode the input layer features, which are single nucleotide variation indicators and copy number variation indicators.

[0014] Preferably, the frequency-based pulse coding method is as follows: First, calculate the modulation factor. Its value is determined by the gene. Significance of mutations Value and degree of the first layer node in the pathway layer Decide: in, The significance q-value output by the MutSigCV algorithm. For genes The number of path nodes connected in the first layer of the path layer. The average node degree of all key genes in the first layer of the pathway layer. These are adjustable parameters; Then, a discrete pulse sequence is generated using Poisson pulse coding or Bernoulli pulse coding. When using Poisson pulse coding, the reference firing strength Modulated and corrected to When using Bernoulli pulse coding, the firing probability is... Revised to .

[0015] Preferably, in S5, the pulse timing-dependent plasticity rule is: the order in which presynaptic neurons and postsynaptic neurons fire pulses determines the direction and magnitude of synaptic weight enhancement or weakening; The change in synaptic weights satisfies: in, Represents neurons With neurons The change in synaptic weight between them Represents the synaptic enhancement constant. This represents the synaptic attenuation constant. and These represent the times when neurons i and j fire their impulses, respectively. and These represent the time-scale window adjustment constants for enhancement and reduction, respectively.

[0016] Preferably, in step S6, the dynamic evolution mechanism specifically includes: when the similarity between a new input sample and an existing output layer neuron is lower than a preset threshold, adding a new output layer neuron; when the similarity between a new input sample and an existing output layer neuron is higher than a preset threshold, merging and updating the connection weights of the corresponding output layer neurons; the merging and updating method is as follows: in, This represents the weights of the output layer neurons to be updated. This represents the weight of the neuron corresponding to the new sample. This indicates the number of similar sample neurons that have been fused with the current output layer neurons.

[0017] Preferably, the weights of the new sample on the neuron The value of is a vector composed of the average firing rates of the new input sample on each neuron in the gene layer after forward propagation, i.e.: in, For the neurons in the gene layer at time The pulse firing state vector, This is the length of the pulse coding time window; This indicates the number of similar sample neurons that have been fused with the current output layer neurons.

[0018] Preferably, step S6 further includes: inputting the sample to be tested into the trained evolutionary spiking neural network, comparing its similarity with the output layer neurons, and determining the subtype classification result; based on this, statistically analyzing the spiking frequency of each neuron in the gene layer and pathway layer of the sample to be tested, and determining nodes whose spiking frequency exceeds a preset activation threshold as activated; tracing the gene nodes and pathway nodes associated with the activated states of the matched classification results; for multiple samples under the same classification result, statistically analyzing the activation frequency, average firing rate, and activation frequency of the pathway nodes connected to each gene node; when the activation frequency of a gene node exceeds a preset proportion threshold, and its corresponding pathway node is in an activated state, or the gene node is connected to multiple activated pathway nodes, the gene is determined as a candidate pathogenic gene and output as a pathogenic gene screening result.

[0019] In a second aspect, the present invention also provides an evolutionary spiking neural network cancer-causing gene screening system, which is used to perform the method described above, the system comprising: The data acquisition module is used to acquire gene mutation data, copy number variation data, and pathway relationship data from cancer case samples; and to obtain candidate genes based on the gene mutation data and copy number variation data. The gene screening module is used to perform significant screening on candidate genes to obtain a set of key genes; A network construction module is used to construct complex networks and evolutionary spiking neural networks based on the Reactome pathway hierarchy and the set of key genes. The neuron modeling module is used to build LIF neuron models for network nodes; The pulse coding module is used to perform pulse coding on the input layer features; The network training module is used to learn and train the evolutionary spiking neural network based on the pulse time-dependent plasticity rule; The dynamic output module is used to generate or update category representation neurons according to the dynamic evolution mechanism of the output layer; The classification and identification module is used to output subtype classification results and / or pathogenic gene screening results.

[0020] Thirdly, the present invention also provides an evolutionary spiking neural network cancer-causing gene screening device, which includes at least a memory and a processor, wherein the processor invokes computer instructions in the memory to execute the method described above.

[0021] Compared with the prior art, the technical solution of the present invention has at least the following advantages: 1. By directly embedding gene association pathway information into the neural network, the biomedical interpretability of target gene identification is improved.

[0022] 2. By introducing the MutSigCV gene screening step before modeling, the impact of the original high-dimensional gene space on the model's learning efficiency and classification stability is reduced.

[0023] 3. By constructing a network structure of input layer-gene layer-pathway layer-evolutionary output layer, the coupled expression of gene information and pathway information was achieved.

[0024] 4. By employing LIF neurons, frequency pulse coding, STDP learning, and dynamic evolution mechanism of the output layer, the model possesses both pulse timing modeling capability and network structure adaptive capability.

[0025] 5. Experiments show that, under both balanced and imbalanced sample conditions, this model achieves certain improvements in accuracy, recall, and F1 score compared to traditional neural network models and various comparative models. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a schematic diagram illustrating the overall principle of an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the gene screening and P-ESNN network construction process according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] Those skilled in the art should understand that the following specific embodiments or implementation methods are a series of optimized configurations listed to further explain the specific content of the invention. These configuration methods can be combined or used in conjunction with each other, unless the invention explicitly states that some or a specific embodiment or implementation method cannot be associated with or used in conjunction with other embodiments or implementation methods. Furthermore, the following specific embodiments or implementation methods are merely optimized configurations and are not intended to limit the scope of protection of the invention.

[0030] This invention is a research result of the major project "Brain Science and Brain-like Research" (No. 2021ZD0201302, titled "Cognitive-Learning Mechanism and Method of Spiral Neural Networks") under the Science and Technology Innovation 2030 Program.

[0031] Example 1

[0032] The principle of this invention can be summarized as follows: First, the gene mutation data and copy number variation data in the case samples are structured, and key genes with high mutation significance are screened using the MutSigCV algorithm. Based on this, a complex network coupling gene interactions and pathway associations is constructed according to the pathway hierarchy and gene relationships in the Reactome database. Further, an evolutionary spiking neural network embedding gene association pathway information is constructed, using single nucleotide variant indicators and copy number variation indicators as input layers. A LIF neuron model is established for the network nodes, trained using pulse coding and STDP learning, and the gene subtype classification and pathogenic gene screening and identification are completed through a dynamic evolution mechanism in the output layer. Figure 1 , 2 The key implementation steps of this solution will be elaborated here.

[0033] Step 1: Data Acquisition and Gene Screening In a preferred embodiment of this example, gene mutation data and copy number variation data from cancer case samples are first acquired, and a candidate gene set is constructed based on these two types of data. First, genes exhibiting single nucleotide variants, insertion / deletion variants, or somatic mutation events are extracted from the gene mutation data to form a first candidate gene subset. Genes exhibiting copy number increase or decrease events are extracted from the copy number variation data to form a second candidate gene subset. The first and second candidate gene subsets are then merged and deduplicated to obtain the original candidate gene set. Preferably, the gene mutation data comes from a standardized mutation annotation file, and includes the mutated gene name, sample identifier, variant type, variant classification, and corresponding allele information. The copy number variation data comes from a copy number fragment file, a gene-level copy number matrix, or a copy number increase / decrease event matrix after thresholding. Since the original sample contains a large number of genes, directly using all genes as network nodes would result in poor model learning performance. Therefore, this embodiment preferably uses the MutSigCV algorithm to perform significant screening of candidate genes. The MutSigCV algorithm calculates the significance p-value and q-value of each gene by comparing the difference between the current gene and the background mutation situation, combined with the mutation type and the degree of influence, and selects driver genes based on the significance threshold. In this invention, the input gene range of the MutSigCV algorithm is limited to the candidate gene set obtained by merging and deduplicating gene mutation data and copy number variation data. After significance screening, the selected significant mutated genes are intersected with the pathway-containing genes in the Reactome pathway database. Only genes that simultaneously meet the mutation significance condition and the pathway association condition are retained as key gene nodes in the subsequent complex network. Specifically, firstly, based on the single nucleotide variation data in the case samples, the mutation frequency, mutation type and background mutation rate of each gene in different samples are statistically analyzed; then, combined with gene length, sequencing coverage and other covariate information, the mutation significance of each gene is corrected, and the significance p-value and q-value corresponding to each gene are calculated. Finally, significant mutated genes are selected from the original candidate gene set according to the preset significance threshold as the key gene set in the subsequent network construction.

[0034] Preferably, the following conditions will be met: Genes with significant mutations were identified. This screening process yielded a set of significant mutant genes (i.e., a set of key genes), and listed representative genes with relatively small q values, such as TP53, SPOP, FOXA1, APC, and PTEN.

[0035] This step can reduce the spatial dimension of the original gene while preserving the biological significance of the potential pathogenic gene, thereby reducing the interference of redundant noise on the training process of the spiking neural network and improving the learning efficiency and classification stability of the subsequent model.

[0036] It should be noted that in step 1 of this embodiment, the candidate gene set is preferably obtained based on gene mutation data. That is, genes that have undergone single nucleotide variants, insertion / deletion variants, or somatic mutation events are extracted from the gene mutation data to form the candidate gene set, which is then further screened for significance using the MutSigCV algorithm. Gene copy number variation data is not used as the core input of the MutSigCV algorithm, but is used as a feature of the subsequent input layer in the evolutionary spiking neural network modeling. As an optional implementation, genes that have undergone significant copy number increases or decreases can also be used as a supplementary source of the candidate gene set, and merged with the candidate gene set obtained based on gene mutation data to remove duplicates.

[0037] Step 2: Construction of complex composite networks: After completing the screening of key genes, a set of key genes is obtained, which is then combined with... Figure 3 As shown, a complex network embedding gene-related pathway information is constructed based on pathway relationships in the Reactome database. This complex network comprises an input layer, a gene layer, a pathway layer, and an evolutionary output layer. The pathway layer is constructed as a multi-layered directed acyclic structure based on the pathway hierarchy in the Reactome database. The gene layer is determined by performing an intersection operation between the genes involved in the Reactome pathways and the aforementioned set of significantly mutated genes. The input layer carries the variation features corresponding to each gene node and establishes connections with the gene layer. The evolutionary output layer dynamically generates category representation neurons during training. Furthermore, nodes at different levels in the pathway layer are directed according to the upstream-downstream relationship of the pathway, and gene layer nodes are mapped to corresponding pathway nodes according to their respective pathways, thus forming a complex network structure that simultaneously couples "inter-gene interactions" and "inter-pathway associations." By directly embedding biological pathway information into the network, the model can have a clearer biomedical explanatory basis when classifying cancer subtypes and screening pathogenic genes.

[0038] Specifically, the pathway layer is constructed first. Preferably, the hierarchical relationship between human pathways can be extracted from the Reactome database. The top-level pathway without a lower-level pathway is taken as the 5th-level node. Then, the pathway nodes of the 4th, 3rd, 2nd and 1st levels are constructed by backtracking layer by layer, and the edges are connected according to the relationship between the pathways to form the pathway layer.

[0039] The final gene layer is constructed before the pathway layer. Preferably, a temporary gene layer is first constructed based on the genes contained in the first-layer pathway in the Reactome database. Then, the intersection of this temporary gene layer and the key gene set obtained by the aforementioned MutSigCV algorithm is taken as the final gene layer. This ensures that the gene nodes entering the network are both mutationally significant and directly related to the biological pathway structure.

[0040] An input layer is constructed before the gene layer. Preferably, the number of neurons in the input layer is set to three times the number of neurons in the final gene layer, corresponding to whether the gene has undergone somatic mutation, whether the gene has undergone copy number increase, and whether the gene has undergone copy number decrease, respectively. The input layer and the gene layer are connected using a fully connected method; the gene layer and the pathway layer are connected using a sparse connection method based on gene-pathway attribution; and the pathway layer itself is connected using a sparse connection method based on pathway relationships.

[0041] Step 3: Establishing the Evolutionary Spike Neural Network: After constructing the complex composite network, an evolutionary spiking neural network (ESN) is built upon it. This ESN is a dynamic network structure, where neurons and synaptic connections can be continuously added, removed, and adjusted during the learning process. In this embodiment, a LIF neuron model is built for each node in the complex composite network. To match the different biological habits of the gene layer and the pathway layer, heterogeneous and differentiated settings are applied to membrane potential, firing threshold, and reset mechanism. Differential membrane potential settings: membrane time constants of neurons in the gene layer Smaller size allows it to focus more on transient variable signal inputs; the neurons in the pathway layer The larger size allows it to accumulate presynaptic activity for a longer period of time, which is consistent with the cascade amplification and delayed transmission characteristics of signals in biological pathways.

[0042] Adaptive firing threshold setting: An adaptive threshold mechanism is used instead of a fixed threshold. The neuron firing threshold is dynamically adjusted according to its historical firing frequency. That is, when a certain pathway node is frequently activated, the threshold automatically increases to suppress redundant firing; conversely, the threshold of nodes that have not been activated for a long time decreases, enhancing sensitivity to weak and variable signals.

[0043] Reset mechanism optimization: Instead of a direct zero-based hard reset, a subtraction reset strategy is adopted. When the membrane potential... Reaching the threshold After the pulse is emitted, instead of returning to zero, the threshold potential is subtracted, and the excess is retained. This method can better preserve the input signal strength information and improve the network's ability to distinguish the intensity of gene mutations.

[0044] Resting potential baseline difference: The resting potential baseline of neurons in the gene layer is lower than that in the pathway layer, which suppresses single-point noise and ensures that only real and significant variant signals can effectively activate downstream pathway nodes.

[0045] By employing the aforementioned heterogeneous configuration, the evolutionary spiking neural network can more accurately capture the hierarchical activation patterns of cancer-causing genes and abnormal pathways during training, thereby further improving classification accuracy and biomedical interpretability.

[0046] Furthermore, the spiking neuron model adopts the LIF neuron model, establishing membrane potential states for the input layer nodes, gene layer nodes, pathway layer nodes, and evolutionary output layer nodes in the complex network; for the first... The membrane potential of a neuron is updated as follows: in, Indicates the membrane time constant. Indicates the first At time 1 neuron membrane potential, Indicates the resting membrane potential. Indicates membrane impedance, Indicates the first At time 1 neuron The received input current; the input current It is obtained by weighted summation of the impulse outputs of the precursor neurons through synaptic weights.

[0047] when Reaching or exceeding the preset discharge threshold At that time, the first Each neuron fires a pulse, denoted as _____. And reset the membrane potential to the reset membrane potential. ;when The preset discharge threshold was not reached. At that time, the first A neuron that does not fire a pulse is denoted as _____. .

[0048] Because the LIF neuron model has the advantages of concise mathematical expression, high implementation efficiency, and preservation of basic biological neurodynamic characteristics, it can effectively model the spiking process in gene-pathway complex networks while ensuring network computability.

[0049] Step 4: Input pulse code: The input layer features are derived from the gene mutation data and gene copy number variation data obtained in step S1, specifically including three types of indicators: somatic mutation indicators, which indicate whether there are non-synonymous somatic single nucleotide variants or short insertion / deletion variants in the gene, preferably represented by a binary value, where 1 indicates the presence of somatic mutations affecting the protein sequence and 0 indicates the absence of such mutations; copy number increase indicators, which indicate whether the gene copy number is significantly higher than the normal diploid baseline, preferably determined based on the GISTIC threshold score, with significant amplification recorded as 1 and otherwise as 0; and copy number decrease indicators, which indicate whether the copy number is significantly lower than the normal baseline, with significant deletion recorded as 1 and otherwise as 0.

[0050] The above indicators are preprocessed as follows before entering pulse coding: the original values ​​of each indicator are normalized to... For somatic mutation indicators, the binarization results can be further corrected by combining the frequency of mutated alleles, so that different mutation abundances correspond to different normalized values, preserving information on mutation heterogeneity. For copy number increase / decrease indicators, discrete markers can be adjusted to continuous values ​​by combining the magnitude of fragment copy number changes, reflecting the degree of copy number change. After preprocessing, each key gene is represented by three input layer neurons, corresponding to somatic mutation, copy number increase, and copy number decrease features, respectively.

[0051] This embodiment introduces an improved coding method based on frequency-based pulse coding. Considering the significance of gene mutations and the influence of the gene's topological position in the pathway network on signal propagation, a modulation factor is introduced when generating the pulse sequence to weightedly adjust the pulse firing intensity or probability. Modulation factor From genes Significance of mutations Value and degree of the first layer node in the pathway layer The decision is made, and the calculation method is as follows: in, The significance q-value output by the MutSigCV algorithm. For genes The number of path nodes connected in the first layer of the path layer. The average node degree of all key genes in the first layer of the pathway layer. This is an adjustable parameter. The modulation factor assigns greater weight to genes with high mutation significance and high pathway connectivity, enabling their variant signals to achieve higher firing intensity during pulse coding, thus more effectively activating downstream pathway nodes.

[0052] When using improved Poisson pulse coding, the reference firing intensity Modulated and corrected to The pulse sequence satisfies: in, This sets the number of pulse firings within a preset time window. If Bernoulli pulse coding is used, the firing probability will be... Revised to This process then generates discrete pulse sequences. In this way, the biological prior information of the gene is directly embedded in the input encoding stage, forming a synergy with the subsequent gene-pathway network structure, thereby improving the sensitivity to pathogenic gene signals and subtype classification.

[0053] Furthermore, Bernoulli pulse coding is introduced as follows: it maps normalized input feature values ​​to pulse firing probabilities at each time step, generating discrete pulse sequences that follow a Bernoulli distribution. It satisfies: in, This represents the probability of firing a pulse at a single moment. This indicates the number of moments within the time window. The Poisson pulse coding method will not be elaborated upon here.

[0054] By further converting continuous or discrete variational features into impulse signals, input information can enter the network in an event-driven manner in the time dimension, thereby enhancing the model's ability to express temporal propagation patterns.

[0055] It should be noted that pulse coding methods can also directly use Poisson pulse coding or Bernoulli pulse coding, but these existing coding methods cannot effectively utilize the potential prior information in the data, and their performance is inferior to the preferred scheme of this embodiment in terms of actual signal response and classification sensitivity.

[0056] Step 5: Network Training The pulse-encoded input pulse sequence is fed into the evolutionary spiking neural network, and sample learning training is performed based on the following pulse timing-dependent plasticity rule: the order in which presynaptic and postsynaptic neurons fire pulses determines the direction and magnitude of synaptic weight enhancement or reduction. Specifically, during sample learning training, the strength of corresponding synaptic connections is updated by enhancing or suppressing them according to the relative order of their firing times; when the presynaptic neuron fires its pulse before the postsynaptic neuron, the corresponding connection weight increases according to the enhancement rule; when the postsynaptic neuron fires its pulse before the presynaptic neuron, the corresponding connection weight decreases according to the reduction rule.

[0057] Preferably, the weight update formula is: in, Represents neurons With neurons The change in synaptic weight between them Represents the synaptic enhancement constant. This represents the synaptic attenuation constant. and These represent the times when neurons i and j fire their respective pulses. and These represent the timescale window adjustment constants for enhancement and attenuation, respectively. Through this training process, the network can progressively learn the cancer gene mutation patterns, pathway association patterns, and category differences contained in the input samples.

[0058] Through this unsupervised learning mechanism, the network can gradually learn the pulse timing differences of different subtype samples at the gene and pathway layers during the training process, enabling the network connection structure to adaptively match the biological laws contained in the samples.

[0059] Step 6: Output layer dynamic evolution mechanism: Unlike static classifiers, the output layer in this embodiment employs a dynamic evolution mechanism, generating or merging output neurons for samples of the same type. Specifically, as samples are progressively input into the network for training, if the similarity between the current input sample and existing output layer class representation neurons is lower than a preset threshold, a new neuron is generated in the output layer to represent that type of sample; if the similarity between the current input sample and existing class representation neurons is higher than a preset threshold, existing output layer neurons are merged and updated to compress redundant class representations and enhance the aggregation effect of samples of the same type.

[0060] The similarity is calculated as follows: after the current input sample is propagated forward, the peak value or average firing rate of its membrane potential response is recorded at each neuron in the output layer, which constitutes the response vector of the sample in the output layer. For neurons with existing category representations Take its weight vector ;calculate and The cosine similarity between them is used as a measure of the degree of matching, that is: in, Indicates the first Each category represents a neuron's weight vector, whose value is the average firing rate vector of the same type of samples fused by that neuron across all neurons in the gene layer. When the cosine similarity between the test sample and a neuron representing a certain category is higher than a preset similarity threshold, the similarity is determined to meet the merging condition; otherwise, it is determined not to meet the condition, triggering the addition of a new neuron in the output layer.

[0061] Through the above similarity calculation method, the output layer structure can be adaptively updated according to the actual response patterns between samples, improving its adaptability to changes in class boundaries and sample distribution.

[0062] Preferably, the weight merging and updating method satisfies: in, This represents the weights of the output layer neurons to be updated. This represents the weight of the neuron corresponding to the new sample. Its value is a vector composed of the average firing rate of each neuron in the gene layer after the new input sample has undergone forward propagation. in, For the neurons in the gene layer at time The pulse firing state vector, This is the length of the pulse coding time window; This indicates the number of neurons from the same type of sample that have been fused with the current output layer neurons. Through the aforementioned dynamic evolution mechanism, the output layer structure can adaptively update as training samples are continuously input, thereby improving its adaptability to changes in class boundaries and sample distribution.

[0063] Through the dynamic evolution mechanism of the output layer, the evolutionary spiking neural network can adaptively adjust the structure of the output layer according to the characteristics of sample distribution, thereby improving the model's adaptability and classification performance under balanced samples, unbalanced samples, and different sample sizes.

[0064] Step 7: Subtype Classification and Pathogenic Gene Identification Output: After training, the test sample is input into the evolutionary spiking neural network. The similarity between the response vector of the test sample on each neuron in the output layer and the weight vector of each existing category representation neuron is calculated. The category with the highest similarity exceeding a preset matching threshold is taken as the subtype classification result. Based on this, the spiking frequency of each neuron in the gene layer and pathway layer of the test sample is statistically analyzed. Nodes with firing frequencies exceeding a preset activation threshold are determined to be in an activated state. Gene nodes and pathway nodes with activated states associated with the matched category (i.e., the matched classification result) are traced. Genes that are repeatedly activated in samples of the same category and are located at key convergence positions of activation pathways are output as pathogenic gene screening results.

[0065] Taking two types of prostate cancer as examples, this embodiment experimentally validates the P-ESNN model proposed in this invention. P-ESNN is an abbreviation for the "evolutionary spiking neural network with embedded gene association pathway information" as defined in this invention. It represents the overall network model that combines Reactome pathway information, key gene nodes, LIF spiking neurons, STDP learning rules, and the dynamic evolution mechanism of the output layer, and is not a commonly used model name.

[0066] In the balanced sample experiment, 660 samples were selected, with 600 samples used as the training set and 60 samples as the test set, equally representing primary prostate cancer cases and metastatic prostate cancer cases. In the imbalanced sample experiment, 990 samples were selected, with 900 samples used as the training set and 90 samples as the test set; the training set contained 600 primary prostate cancer cases and 300 metastatic prostate cancer cases, while the test set contained 60 primary prostate cancer cases and 30 metastatic prostate cancer cases. Decision trees, K-nearest neighbors, SGD classifiers, ANN, random forests, and Naive Bayes were selected as comparative models in the experiments, and accuracy, recall, and F1 score were used as classification performance evaluation indicators. Table 1 shows the classification results of each model under the balanced sample condition, and Table 2 shows the classification results of each model under the imbalanced sample condition.

[0067] Table 1 Table 2 The experimental results above show that, under balanced sample conditions, the P-ESNN model achieves an accuracy of 85%, a recall of 90%, and an F1 score of 0.8572, which is higher than that of the comparative models such as decision trees, K-nearest neighbors, SGD classifiers, ANN, random forests, and Naive Bayes. Under imbalanced sample conditions, the P-ESNN model achieves an accuracy of 82.22%, a recall of 80%, and an F1 score of 0.7501, again higher than the aforementioned comparative models. These results demonstrate that the P-ESNN model proposed in this invention can improve the accuracy and stability of cancer subtype classification and pathogenic gene screening while maintaining the biomedical interpretability of gene pathways.

[0068] Through the above steps, this embodiment constructs an evolutionary spiking neural network model with both biological interpretability and classification recognition capabilities based on high-dimensional cancer gene data, realizing joint modeling of cancer pathogenic gene identification and cancer subtype classification.

[0069] Example 2

[0070] In a second embodiment, the present invention also provides an evolutionary spiking neural network cancer pathogenic gene screening system embedded with gene association pathway information, which is used to perform the method described in Embodiment 1 above. Combined with Figure 4 As shown, the system includes: The data acquisition module is used to acquire gene mutation data, copy number variation data, and pathway relationship data from cancer case samples. The gene screening module is used to screen key gene sets based on the MutSigCV algorithm; The network construction module is used to construct complex composite networks and evolutionary spiking neural networks consisting of an input layer, a gene layer, a pathway layer, and an evolutionary output layer. The neuron modeling module is used to build LIF neuron models for network nodes; A pulse coding module for performing Poisson pulse coding and / or Bernoulli pulse coding; The network training module is used to learn and train the network based on STDP rules; The dynamic output module is used to execute the dynamic evolution mechanism of the output layer; The classification and identification module is used to output cancer subtype classification results and cancer pathogenic gene identification results.

[0071] Example 3

[0072] In yet another embodiment, the present invention also provides an evolutionary spiking neural network cancer pathogenic gene screening device with embedded gene association pathway information. The device includes at least a memory and a processor. The processor calls a computer program in the memory to execute the method steps described in Embodiment 1 above, or to execute the functions of each module in the system described in Embodiment 2 above.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for screening cancer-causing genes using evolutionary spiking neural networks, characterized in that, The method includes: S1. Obtain gene mutation data, gene copy number variation data, and gene pathway relationship data from cancer case samples, and construct the original sample set; S2. Based on the gene mutation frequency, candidate genes in the original sample set are screened to obtain a set of key genes; S3. Based on the pathway hierarchy and pathway-included gene relationships in the pathway database, construct a complex network with embedded gene-associated pathway information. S4. Based on the aforementioned complex network, construct an evolutionary spiking neural network, using single nucleotide variation index and copy number variation index as input layer features, and establish a spiking neuron model for the network nodes. S5. Perform pulse coding on the input layer features to obtain the input pulse sequence, and train the evolutionary spiking neural network based on the pulse temporal dependence plasticity rule; S6. Generate category representation neurons based on the dynamic evolution mechanism of the output layer during training, and complete the matching between the test sample and the output layer neurons; output subtype classification results and / or pathogenic gene screening results based on the matching results.

2. The method according to claim 1, characterized in that, In step S2, a candidate gene set is determined based on gene mutation data, and the candidate gene set is screened for significance. Genes that meet the significance threshold are identified as a significant mutation gene set, which serves as the key gene set. The gene copy number variation data is used as a feature of the subsequent input layer to participate in the evolutionary spiking neural network modeling, or as a supplementary source of the candidate gene set.

3. The method according to claim 1, characterized in that, The complex network constructed in S3 includes an input layer, a gene layer, a pathway layer, and an evolutionary output layer connected in sequence. The pathway layer is constructed as follows: based on the directed hierarchical relationship between pathways in the Reactome database, a multi-layered pathway node from layer 1 to layer 5 is constructed, where each layer of pathway node represents a different level of human biological pathway. Then, edges are connected according to the hierarchical relationship between pathways to form a pathway layer. The gene layer is determined by the intersection of the genes contained in the first layer of the pathway layer and the set of key genes; The number of neurons in the input layer is three times the number of neurons in the gene layer.

4. The method according to claim 1, characterized in that, In S4, the spiking neuron model adopts the LIF neuron model, and membrane potential states are established for the input layer nodes, gene layer nodes, pathway layer nodes, and evolutionary output layer nodes in the complex network; for the first The membrane potential of a neuron is updated as follows: in, Indicates the membrane time constant. Indicates the first At time 1 neuron membrane potential, Indicates the resting membrane potential. Indicates membrane impedance, Indicates the first At time 1 neuron The received input current; when Reaching or exceeding the preset discharge threshold At that time, the first Each neuron fires a pulse, denoted as _____. And reset the membrane potential to the reset membrane potential. ;when The preset discharge threshold was not reached. At that time, the first A neuron that does not fire a pulse is denoted as _____. .

5. The method according to claim 1, characterized in that, In step S5, a frequency-based pulse coding method is used to encode the input layer features, which are single nucleotide variation indicators and copy number variation indicators.

6. The method according to claim 1, characterized in that, In S5, the pulse timing-dependent plasticity rule is: the order in which presynaptic neurons and postsynaptic neurons fire pulses determines the direction and magnitude of synaptic weight enhancement or weakening; The change in synaptic weights satisfies: in, Represents neurons With neurons The change in synaptic weight between them Represents the synaptic enhancement constant. This represents the synaptic attenuation constant. and These represent the times when neurons i and j fire their impulses, respectively. and These represent the time-scale window adjustment constants for enhancement and reduction, respectively.

7. The method according to claim 1, characterized in that, In S6, the dynamic evolution mechanism specifically includes: when the similarity between a new input sample and an existing output layer neuron is lower than a preset threshold, a new output layer neuron is added; when the similarity between a new input sample and an existing output layer neuron is higher than a preset threshold, the connection weights of the corresponding output layer neurons are merged and updated; the merging and updating method is as follows: in, This represents the weights of the output layer neurons to be updated. This represents the weight of the neuron corresponding to the new sample. This indicates the number of similar sample neurons that have been fused with the current output layer neurons.

8. The method according to claim 1, characterized in that, S6 further includes: inputting the test sample into the trained evolutionary spiking neural network, comparing its similarity with the output layer neurons, and determining the subtype classification result; based on this, statistically analyzing the spiking frequency of each neuron in the gene layer and pathway layer of the test sample, and determining the nodes whose spiking frequency exceeds a preset activation threshold as activated; tracing the gene nodes and pathway nodes associated with the activated states of the matched classification results; for multiple samples under the same classification result, statistically analyzing the activation frequency, average firing rate, and activation frequency of the pathway nodes connected to each gene node; when the activation frequency of a gene node exceeds a preset proportion threshold, and its corresponding pathway node is in an activated state, or the gene node is connected to multiple activated pathway nodes, the gene is identified as a candidate pathogenic gene and output as a pathogenic gene screening result.

9. An evolutionary spiking neural network-based cancer-causing gene screening system, characterized in that, The system is configured to perform the method as described in any one of claims 1 to 8, the system comprising: The data acquisition module is used to acquire gene mutation data, copy number variation data, and pathway relationship data from cancer case samples; and to obtain candidate genes based on the gene mutation data and copy number variation data. The gene screening module is used to perform significant screening on candidate genes to obtain a set of key genes; A network construction module is used to construct complex networks and evolutionary spiking neural networks based on the Reactome pathway hierarchy and the set of key genes. The neuron modeling module is used to build LIF neuron models for network nodes; The pulse coding module is used to perform pulse coding on the input layer features; The network training module is used to learn and train the evolutionary spiking neural network based on the pulse time-dependent plasticity rule; The dynamic output module is used to generate or update category representation neurons according to the dynamic evolution mechanism of the output layer; The classification and identification module is used to output subtype classification results and / or pathogenic gene screening results.

10. An evolutionary spiking neural network cancer pathogenic gene screening device, characterized in that, The device includes at least a memory and a processor, the processor calling computer instructions in the memory to perform the method as described in any one of claims 1 to 8.

Images