A neural network computing method for gene expression regulation analysis and device thereof

Abstract

The application discloses a neural network calculation method and device for gene expression regulation analysis, and relates to the technical field of bioinformatics.The method comprises the following steps: obtaining first feature data and second feature data; constructing input features containing multiple regulation levels; inputting the input features of the multiple regulation levels and the second feature data into a target neural network model, and outputting a prediction result of a gene expression state.The neural network calculation method provided by the application can deeply integrate chromatin accessibility and three-dimensional spatial interaction data through a dynamic routing module, can solve the "black box" problem that a traditional deep learning model is inaccurate and difficult to explain, and can accurately identify key gene regulation pathways in a manner of explicitly simulating real biological regulation mechanisms.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, and in particular to a neural network computation method and apparatus for gene expression regulation analysis. Background Technology

[0002] Precise regulation of gene expression is a fundamental biological process determining cell function and fate, and its dysregulation is closely related to many major diseases, including cancer. The transcriptional activity of a gene is not determined solely by its coding sequence, but is controlled by a complex network of cis-regulatory elements such as promoters and enhancers. Distal enhancers, in particular, can interact with the promoters of target genes across long distances through chromatin folding in three-dimensional space, thereby precisely regulating gene expression. Therefore, a deep understanding of the synergistic regulatory mechanisms between the one-dimensional sequence information of the genome, chromatin accessibility, and three-dimensional spatial structure is a core issue in modern life science research.

[0003] With advancements in high-throughput sequencing technology, researchers are able to systematically acquire diverse regulatory data across the entire genome. For example, ATAC-seq technology can reveal open regions of chromatin to identify potential active regulatory elements; while chromatin conformation capture technologies such as Hi-C can depict three-dimensional spatial interaction maps of the genome, providing crucial data for resolving long-distance interactions between enhancers and promoters. The massive and complex diversity of omics data has greatly propelled the development of computational biology methods, particularly machine learning and deep learning models, which are widely used to predict gene expression states based on various genomic features. These models, such as convolutional neural networks, attempt to construct a mapping from genomic information to gene expression activity by learning patterns in DNA sequences and epigenetic data.

[0004] However, existing computational methods still have significant limitations in accurately analyzing gene regulatory networks. First, many models perform poorly in effectively integrating multimodal heterogeneous data. They may focus on one-dimensional sequence information but fail to explicitly and effectively integrate crucial three-dimensional chromatin spatial structure information into the model, limiting the accuracy of model predictions. Second, most current deep learning models are "black box" characteristics, with opaque internal decision-making logic, resulting in a severe lack of biological interpretability. Even if the model can make accurate predictions, researchers find it difficult to understand which specific regulatory elements or pathways play a decisive role, hindering the discovery of new biological knowledge. Furthermore, the structural design of existing models often fails to fully simulate the intrinsic biological logic of gene regulation. For example, the lack of explicit modeling of the hierarchical information flow from distal elements to promoters limits their ability to handle complex and dynamic biological processes.

[0005] In summary, current computational methods in the field of gene expression regulation face several common challenges. Existing technologies are insufficient in effectively integrating multi-omics data such as chromatin three-dimensional structure; the "black box" problem of prediction models hinders in-depth exploration of regulatory mechanisms; and the design of the model architecture itself fails to fully reflect the hierarchical and spatial characteristics of gene regulation. Therefore, there is an urgent need in this field for novel computational methods that can overcome these shortcomings to achieve more accurate and interpretable modeling and analysis of gene expression regulatory networks. Summary of the Invention

[0006] In a first aspect, the present invention provides a neural network computation method for analyzing gene expression regulation, comprising: acquiring first feature data representing chromatin accessibility and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements; constructing input features comprising multiple regulatory levels based on the first feature data; wherein each regulatory level corresponds to a type of regulatory element; inputting the input features of the multiple regulatory levels and the second feature data into a target neural network model, and outputting a prediction result of gene expression state through the target neural network model; wherein the target neural network model comprises at least one dynamic routing module for determining and adjusting the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

[0007] In an optional implementation, the step of constructing an input feature comprising multiple control levels based on the first feature data; wherein each control level corresponds to a type of control element, including:

[0008] The genomic region centered on the transcription start site of the target gene is divided into multiple contiguous units, which serve as the first regulatory level.

[0009] Enhancer regions of the target gene are identified and screened as the second regulatory level;

[0010] The promoter region near the transcription start site of the gene is defined as the third regulatory level.

[0011] In an optional implementation, the first regulatory hierarchy is constructed as follows: a genomic region totaling 5 Mb centered on the transcription start site is selected, and the region is divided into 1000 contiguous units of 5 kb each; and / or,

[0012] The second regulatory hierarchy includes the 100 nearest enhancer regions selected based on their linear distance from the transcription start site; and / or,

[0013] The third regulatory hierarchy comprises the 30 nearest promoter regions sampled within a 1kb range upstream and downstream of the transcription start site.

[0014] In an optional implementation, the step of inputting the input features of the multiple regulatory levels and the second feature data into the target neural network model, and outputting a prediction of gene expression status through the target neural network model, includes:

[0015] The first dynamic routing module maps the input features of the first control level to the first intermediate features based on the three-dimensional spatial interaction between the first control level and the second control level.

[0016] The first intermediate feature is combined with the input feature of the second control level, and the combined feature is mapped to the second intermediate feature through the second dynamic routing module based on the three-dimensional spatial interaction relationship between the second control level and the third control level.

[0017] The second intermediate feature is combined with the input feature of the third regulatory level, and the combined feature is mapped to the prediction result of gene expression state by the third dynamic routing module based on the three-dimensional spatial interaction relationship between the first regulatory level and the third regulatory level.

[0018] In an optional implementation, the information transmission weight adjusted by the first dynamic routing module is: The calculation method is as follows:

[0019] ;

[0020] in, The basic weight matrix representing the connection between the first and second layers; This represents the dynamic weight mask generated by the first dynamic routing module, and , Representing the first layer of features, Represents the second layer of features. Represents the spatial interaction matrix between the first and second layers; and / or,

[0021] The information transmission weight adjusted by the second dynamic routing module is The calculation method is as follows:

[0022] ;

[0023] in, The basic weight matrix representing the connection between the second and third layers; This represents the dynamic weight mask generated by the second dynamic routing module, and , Represents the second layer of features. Represents the third layer of features. Represents the spatial interaction matrix between the second and third layers; and / or,

[0024] The information transmission weight adjusted by the third dynamic routing module is: The calculation method is as follows:

[0025] ;

[0026] in, Represents the basic weight matrix; This represents the dynamic weight mask generated by the third dynamic routing module, and , Representing the first layer of features, Represents the third layer of features. This represents the spatial interaction matrix between the first and third layers.

[0027] In an optional implementation, each of the information transmission weights corresponds to a learnable sparse threshold.

[0028] When the absolute value of the information transmission weight is less than its corresponding sparse threshold, the information transmission weight is reset to zero.

[0029] In an optional implementation, the method for constructing the target neural network model includes:

[0030] Set the input nodes corresponding to the first, second, and third control levels;

[0031] Construct multiple linear mapping modules to perform feature mapping between different control levels;

[0032] Multiple dynamic routing modules are constructed and configured to control the information transmission weights of the linear mapping module; wherein, the configuration includes enabling the dynamic routing module to receive features of the source layer, features of the target layer, and data representing the three-dimensional spatial interaction relationship between the source layer and the target layer, and generating dynamic weights based on these.

[0033] In an optional implementation, the neural network computation method for gene expression regulation analysis further includes training the target neural network model, including:

[0034] The parameters of the model are optimized using a binary cross-entropy loss function, the formula for which is:

[0035] ;

[0036] Where N represents the total number of samples; The true label represents the i-th sample; This represents the model's predicted probability for the i-th sample.

[0037] In an optional implementation, the true label of the sample is determined in the following way:

[0038] Obtain RNA sequencing expression data of the target gene;

[0039] When the expression level of the target gene is greater than a first preset threshold, its true label is determined as a positive sample;

[0040] When the expression level of the gene is less than a second preset threshold, its true label is determined to be a negative sample.

[0041] In an optional implementation, the neural network computation method for gene expression regulation analysis further includes interpretability analysis of the prediction results of the target neural network model, including:

[0042] Select at least one of the plurality of control levels as the target level to be explained, and set a reference input baseline for the target level;

[0043] An attribution algorithm based on contribution backtracking is used to calculate the marginal contribution value of each input feature in the target level to the prediction result;

[0044] The input features in the target level are sorted according to the marginal contribution value, and a predetermined number of high contribution features are selected as the identified key functional control elements.

[0045] In a second aspect, the present invention provides a neural network computing device for analyzing gene expression regulation, comprising:

[0046] The acquisition module is used to acquire first feature data representing chromatin accessibility and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements.

[0047] The construction module is used to construct an input feature containing multiple control levels based on the first feature data; wherein each control level corresponds to a type of control element.

[0048] The prediction module is used to input the input features of the multiple regulatory levels and the second feature data into the target neural network model, and output the prediction result of gene expression state through the target neural network model; wherein, the target neural network model includes at least one dynamic routing module, used to determine and adjust the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

[0049] Thirdly, the present invention provides a computer system comprising a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the neural network computation method for gene expression regulation analysis as described in any of the foregoing embodiments.

[0050] Fourthly, the present invention provides a computer storage medium storing a computer program, which, when executed on a processor, implements a neural network computation method for gene expression regulation analysis according to any one of the foregoing embodiments.

[0051] This application provides a neural network computation method and apparatus for analyzing gene expression regulation. The neural network computation method for analyzing gene expression regulation brings significant benefits through its unique structure and data processing method, mainly in improving the accuracy and biological interpretability of gene expression state prediction.

[0052] First, this method enables the deep fusion of multi-dimensional, heterogeneous genomic data. In the real biological process of gene expression regulation, chromatin accessibility and three-dimensional spatial structure are two inseparable key factors. By simultaneously acquiring and processing feature data representing chromatin accessibility (such as ATAC-seq data) and feature data representing three-dimensional spatial interactions (such as Hi-C data), this computational method can construct a more comprehensive and three-dimensional regulatory picture. This integration of multimodal data allows the model to comprehensively determine whether a regulatory element is not only in an "activatable" state but also has the physical basis for "effective interaction" with other elements. Therefore, it is better able to capture complex regulatory logic than computational methods that rely on a single data source, thereby improving the accuracy and reliability of prediction results.

[0053] Secondly, the core advantage of this method lies in its dynamically designed routing module, which explicitly integrates prior biological knowledge into the structure of the neural network. Traditional neural networks, when dealing with such problems, typically learn the connection weights between different features implicitly, making it difficult to reflect the true biological regulatory mechanisms. The unique feature of this method's dynamic routing module is that it uses the "three-dimensional spatial interaction relationship" between the source and target regulatory elements as the direct basis for adjusting the information transmission weights. This means that the closer the physical contact between two regulatory elements in three-dimensional space, the more dynamically the module will enhance the information flow between them, and vice versa. This design directly simulates the core biological principle of enhancers contacting and activating promoters through chromatin loops, making the model's decision-making process no longer a simple mathematical fit, but a simulation of the real regulatory path, thus making the predictions more accurate and robust.

[0054] Finally, this explicit modeling based on biological mechanisms greatly enhances the biological interpretability of the results. Thanks to the dynamic routing module, the model is no longer a complete "black box." After the model completes its predictions, by analyzing the connection paths assigned higher information transmission weights, researchers can infer which distal regulatory elements play a crucial role in activating the expression of the target gene. This provides researchers with a powerful tool not only to know whether a gene is expressed, but also to explore "why" it is expressed, clearly identifying potential key regulatory elements and pathways. This capability elevates computational prediction from a simple classification task to a discovery tool that can guide subsequent biological experiments and generate verifiable scientific hypotheses, addressing the significant deficiency of poor interpretability prevalent in existing computational methods. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and therefore should not be considered as a limitation on the scope of protection of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of the hardware operating environment involved in an embodiment of the neural network computation method for gene expression regulation analysis of the present invention;

[0057] Figure 2 This is a flowchart illustrating Example 1 of the neural network computation method for gene expression regulation analysis of the present invention;

[0058] Figure 3 This is a detailed flowchart of step S200 in Example 2 of the neural network computation method for gene expression regulation analysis of the present invention.

[0059] Figure 4 This is a detailed flowchart of step S300 in Example 3 of the neural network computation method for gene expression regulation analysis of the present invention.

[0060] Figure 5 This is a flowchart illustrating steps S400 to S600 in Example 4 of the neural network computation method for gene expression regulation analysis of the present invention.

[0061] Figure 6 This is a flowchart illustrating steps S700 to S900 in Example 5 of the neural network computation method for gene expression regulation analysis of the present invention.

[0062] Figure 7 This is a schematic diagram of the deep neural network model in Example 6 of the neural network computation method for gene expression regulation analysis of the present invention;

[0063] Figure 8 This is a flowchart illustrating the method used in Example 7 of the neural network computation method for gene expression regulation analysis of the present invention.

[0064] Figure 9 This is a schematic diagram of the module connections of the neural network computing device for gene expression regulation analysis of the present invention. Detailed Implementation

[0065] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0066] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0067] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0068] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0069] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0070] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0071] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0072] like Figure 1 The diagram shown is a structural schematic of the hardware operating environment of the terminal involved in an embodiment of the present invention.

[0073] The neural network computing system for gene expression regulation analysis according to embodiments of the present invention can be a PC, or a mobile terminal device such as a smartphone, tablet computer, or portable computer. This neural network computing system for gene expression regulation analysis may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, or a remote control; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001. Optionally, the neural network computing system for gene expression regulation analysis may also include RF (Radio Frequency) circuitry, audio circuitry, a Wi-Fi module, etc.

[0074] Those skilled in the art will understand that Figure 1 The neural network computational system for gene expression regulation analysis shown is not intended to limit it and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a data interface control program, a network connection program, and a neural network calculation program for gene expression regulation analysis.

[0075] Example 1

[0076] Reference Figure 2 This embodiment provides a neural network computation method for analyzing gene expression regulation, including:

[0077] Step S100: Obtain first feature data representing chromatin accessibility and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements.

[0078] This step is the data preparation phase of the entire computational method. It requires the collection of two core types of biological data.

[0079] The first characteristic data mentioned above refers to which regions of the DNA strand within the cell nucleus are "open" and easily accessible to proteins. This "openness" or "accessibility" is a key prerequisite for determining whether a DNA fragment has regulatory activity.

[0080] The second characteristic data mentioned above refers to the folding of the DNA strand in three-dimensional space. It quantifies the physical proximity of any two regions on the genome (even those that are far apart in a linear sequence).

[0081] After this step is completed, two independent sets of raw datasets are obtained. One set is a genome-wide chromatin accessibility map (e.g., accessibility score for each genomic location), and the other set is a genome three-dimensional interaction matrix (e.g., a large matrix where each element represents the contact frequency between a pair of genomic regions).

[0082] The advantage of this step is that it establishes a multimodal and multidimensional data foundation. It does not view the genome in isolation, but collects information from two levels simultaneously: "regional status" (accessibility) and "inter-regional relationships" (three-dimensional interactions), providing the necessary data support for building a more comprehensive model that is closer to biological reality.

[0083] This step can typically be achieved using high-throughput biological experimental techniques.

[0084] The first characteristic data can be obtained through chromatin open sequencing technologies (such as ATAC-seq). This technology can identify open chromatin regions across the entire genome. The second characteristic data can be obtained through chromatin conformation capture technologies (such as Hi-C). This technology can capture and quantify the three-dimensional spatial contact frequency between all regions across the entire genome.

[0085] Step S200: Based on the first feature data, construct an input feature containing multiple control levels; wherein each control level corresponds to a type of control element.

[0086] The above steps involve structuring and preprocessing the "chromatin accessibility" data obtained in the first step. Instead of using all accessibility data as a whole, this step organizes them into different "hierarchies" or categories based on their biological functions.

[0087] After this step, you will obtain a set of structured, categorized input feature vectors. For example, you can obtain features at three levels:

[0088] (1) A level represents a broad genomic region (bin);

[0089] (2) A level represents an enhanced sub-region;

[0090] (3) Another level represents the starter subregion.

[0091] The features at each level are derived from the original chromatin accessibility data.

[0092] The advantage of this step is that it introduces prior biological knowledge into the model. By stratifying the input features, the model is informed from the outset of the different biological roles that different regions may play (such as enhancers and promoters). This structured input helps the model learn more efficiently the specific patterns of different categories of regulatory elements and their interrelationships, rather than blindly exploring from a bunch of undifferentiated raw data.

[0093] This step is a data processing step, which can be implemented using an algorithm. For example, a program can be written:

[0094] The first level (bin) takes the transcription start site (TSS) of each target gene as the center, divides a certain range of genomic regions (e.g., 5Mb) upstream and downstream of it into continuous units (bins) of equal length, and then extracts the average chromatin accessibility signal value within each bin.

[0095] The second / third level (enhancers / promoters) can identify the enhancer and promoter positions in the above regions based on known genome annotation databases, and then extract chromatin accessibility signal values ​​at these specific positions.

[0096] Step S300: Input the input features of the multiple regulatory levels and the second feature data into the target neural network model, and output the prediction result of gene expression state through the target neural network model; wherein, the target neural network model includes at least one dynamic routing module, used to determine and adjust the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

[0097] This step performs the core computation. It feeds all the data prepared in the previous two steps—the structured multi-level features and three-dimensional interaction data—into a neural network model for computation. The final result of this step is a prediction of the target gene's expression state by the model. This result is typically a numerical value, such as a probability between 0 and 1, indicating how likely the gene is to be "on" or "off".

[0098] The advantage of this step lies in utilizing the powerful computational tool of neural networks. Neural networks excel at learning nonlinear patterns and relationships from complex, high-dimensional data, which is precisely what genomic data analysis requires. It can automatically discover complex patterns related to gene expression from input features.

[0099] Specifically, a neural network model can be built, trained, and executed using programming languages ​​(such as Python) and deep learning frameworks (such as PyTorch and TensorFlow).

[0100] The "target neural network model" contains a special component called a "dynamic routing module." This component intelligently controls the flow of information within the model. Specifically, when information needs to be transferred from one level (source) to another level (target), this module makes a "decision."

[0101] The decision is not random, but can be based on three key pieces of information:

[0102] (1) Characteristics of the source hierarchy: For example, the openness of an enhanced subregion.

[0103] (2) Characteristics of the target level: For example, the openness of a starter subregion.

[0104] (3) The three-dimensional spatial interaction between the two: that is, the physical distance between the enhancer and the promoter in the cell nucleus (the contact frequency).

[0105] After processing the three inputs mentioned above, this module outputs a dynamic "information transfer weight". This weight is not a fixed parameter, but is dynamically calculated for each specific gene and each pair of specific regulatory elements.

[0106] In terms of the high degree of simulation of biological mechanisms in the target neural network model, the core biological principle that "the closer the physical contact between regulatory elements in three-dimensional space, the stronger the regulatory effect" is directly transformed into the model's internal computational logic. This makes the model's predictions not only based on the correlation of data, but also on the simulation of biological mechanisms, thus resulting in more accurate and reliable results.

[0107] Since the weights of information transmission are explicitly determined by biological characteristics (especially three-dimensional interactions), the model is no longer a complete "black box." By analyzing which connections are assigned high weights, it is possible to infer which regulatory pathways the model considers key, thus providing verifiable hypotheses for biological research.

[0108] Suppose the model is analyzing the regulatory effect of enhancer A on promoter P. The dynamic routing module receives features of A, features of P, and the three-dimensional interaction values ​​between A and P. If Hi-C data shows that A and P are in close contact (high interaction value), the module calculates a high weight, allowing a large amount of information from A to flow to P; conversely, if their interaction values ​​are low, the module calculates a low weight, effectively "cutting off" or weakening the information flow between them.

[0109] Example 2

[0110] Reference Figure 3 This embodiment provides a neural network computation method for gene expression regulation analysis. In step S200, based on the first feature data, an input feature structure is constructed containing multiple regulatory levels; wherein each regulatory level corresponds to a type of regulatory element, including:

[0111] Step S210: Divide the genomic region centered on the transcription start site of the target gene into multiple continuous units as the first regulatory level.

[0112] This step first identifies a target gene on the genome and locates its transcription start site (TSS) as the central anchor point for localization. Then, starting from this central point, a broad genomic region is delineated by extending upstream and downstream along the DNA sequence. Finally, this broad region is divided, like slicing a cake, into many equal-length, interconnected "continuous units" (i.e., bins). These divided units collectively constitute the first layer of the model's input features—the most basic and broadest layer.

[0113] The direct result of this step is a set of ordered genomic coordinate segments. For example, this yields a list of 1000 consecutive genomic segments, each 5kb in length. This list defines the "first regulatory level."

[0114] The advantage of this step is that it provides a comprehensive and unbiased panoramic view. Instead of pre-judging which region is more important, it includes the entire vast area surrounding the gene in the analysis. This ensures that the model can capture the influence of unknown regulatory elements that may exist far from the gene, avoiding the loss of important information due to limited field of view.

[0115] This step can be achieved using an algorithm that involves simple coordinate calculations:

[0116] (1) Input: Transcription start site coordinates of the target gene (TSS_pos), total region width (Total_Width), and number of units (Num_Units).

[0117] (2) Calculate the start and end coordinates of the region: Start_Pos = TSS_pos - (Total_Width / 2); End_Pos = TSS_pos + (Total_Width / 2).

[0118] (3) Calculate the width of each unit: Unit_Width = Total_Width / Num_Units.

[0119] By looping, starting from Start_Pos, Num_Units consecutive coordinate segments with a width of Unit_Width are generated sequentially.

[0120] Assuming a gene's TSS is located at position 10,000,000, a region covering 5 Mb (5,000,000 base pairs) needs to be constructed and divided into 1,000 units. This step first determines the range from 7,500,000 to 12,500,000. Then, it divides this range into 1,000 units of length 5,000. The coordinates of the first unit are [7500000, 7504999], the second is [7505000, 7509999], and so on, until the last unit.

[0121] Step S220: Identify and screen the enhancer regions of the target gene as the second regulatory level.

[0122] This step involves searching for and selecting specific functional elements as "enhancers" within the broad genomic region defined in the first step. Enhancers are key regulatory sequences that can significantly activate gene expression. This process includes "identification" (finding all possible candidates) and "screening" (selecting the part most relevant to the target gene).

[0123] The result of this step is a filtered list of genomic coordinates representing enhancer regions. This list is the "second regulatory level," which focuses more on a specific class of functionally important regulatory elements than the first regulatory level.

[0124] The advantage of this step is that it improves the focus and efficiency of the analysis. It directs the model's attention from the broad genomic background to a class of known elements with powerful regulatory functions. In this way, the model can more easily learn the key roles of enhancers in gene regulation, rather than engaging in inefficient exploration of a vast amount of background regions.

[0125] Specifically, this can be achieved by combining biological experimental data and computational algorithms:

[0126] (1) Identification: Using the “first feature data” (chromatin accessibility data) mentioned in the aforementioned implementation method, “open regions” (usually called “peaks”) with significant signals are searched on the genome. These regions are potential enhancer candidate regions.

[0127] (2) Screening: Sort and screen all candidate regions. A common screening algorithm is based on linear distance, that is, calculate the distance between each candidate enhancer and the target gene TSS on the one-dimensional DNA sequence, and then select the N closest regions as the final enhancer regions.

[0128] For example, within the aforementioned 5Mb region, 500 potential enhancers may have been identified by analyzing chromatin accessibility data. The screening algorithm in this step calculates the distance of each of these 500 regions to the gene's TSS, and then selects the 100 closest regions to form the features of the second regulatory level.

[0129] Step S230: Define the promoter region near the transcription start site of the gene as the third regulatory level.

[0130] This step involves defining the core region directly related to the initiation of gene expression—the promoter. The promoter is a short DNA sequence adjacent to the gene's TSS (transmission sequence segment), serving as the "platform" for transcription machinery (such as RNA polymerase) to bind and begin functioning. This step is very straightforward: it involves defining a clearly defined, relatively small window around the TSS.

[0131] The result of this step is a well-defined list of genomic coordinates representing promoter regions. This list constitutes the "third regulatory level," which is the level most closely related to gene location and most directly related to function among the three levels.

[0132] The advantage of this step is that it provides the model with the most critical and direct regulatory information input. Whether a gene can ultimately be expressed largely depends on the state of its promoter region. Treating the promoter as an independent, high-priority level allows the model to clearly learn how all distant regulatory signals (from enhancers, etc.) converge at this final "decision point" and produce an effect, perfectly simulating the final integration process of regulatory information.

[0133] Specifically, a coordinate-based definition method can be used:

[0134] (1) Input: Transcription start site coordinates of the target gene (TSS_pos), promoter window size (Promoter_Window_Size).

[0135] (2) Calculate the start and end coordinates of the start sub-region: Start_Pos = TSS_pos - Promoter_Window_Size; End_Pos = TSS_pos + Promoter_Window_Size.

[0136] For example, if a gene’s TSS is located at 10,000,000 and the promoter region is defined as the region within ±1kb of the TSS, then this step will generate one or more coordinate segments covering the range [9999000, 10001000] as the third regulatory level.

[0137] In some implementations, the first regulatory hierarchy is constructed by selecting a genomic region totaling 5 Mb centered on the transcription start site and dividing the region into 1000 contiguous units of 5 kb each.

[0138] The above steps specify the numerical parameters for constructing the first regulatory level (i.e., continuous units or bin layers). The processing may include:

[0139] (1) Selection range: The transcription start site (TSS) of the target gene is the center, and it extends 2.5Mb (megabase pairs) upstream and downstream to form a broad genomic region with a total length of 5Mb.

[0140] (2) Divide the region into 1,000 equal segments, each segment being 5 kb (kilobase pairs).

[0141] The output of this step is a list of 1,000 consecutive genomic coordinate regions, each 5kb in length, which together constitute the first regulatory level.

[0142] The advantage of this approach lies in providing a well-defined and validated preferred implementation. The 5Mb range is sufficient to cover the vast majority of known distal regulatory elements, while the 5kb resolution strikes a good balance between computational efficiency and biological accuracy. It concretizes a vague concept (“genomic region”) into a reproducible, standardized procedure.

[0143] For example, this can be achieved using a precise coordinate calculation algorithm:

[0144] Given the TSS coordinates TSS_pos.

[0145] Starting coordinates Start_Pos = TSS_pos - 2,500,000.

[0146] End coordinates End_Pos = TSS_pos + 2,500,000.

[0147] By looping, starting from Start_Pos, 1000 continuous coordinate segments with a length of 5kb are generated.

[0148] In some embodiments, the second regulatory hierarchy includes the 100 nearest enhancer regions selected based on linear distance from the transcription start site.

[0149] This step provides specific criteria for selecting the final candidates that constitute the second regulatory level from numerous candidate enhancer regions. The process can be as follows: first, calculate the distance between all identified candidate enhancers and the target gene TSS on the one-dimensional genome sequence; then, sort them in order of distance from closest to furthest; and finally, select the top 100.

[0150] The output of this step is a list of coordinates for 100 enhancer regions, which are the enhancers that are linearly closest to the target gene.

[0151] Its advantage lies in providing an objective, quantifiable, and computationally simple screening criterion. Although distance in three-dimensional space is more important, proximity in one-dimensional linear distance remains an important prior feature for judging the likelihood of a regulatory relationship. Fixing the number to 100 ensures a sufficient number of candidate enhancers to participate in model calculations while controlling computational complexity and dimensionality, making it an efficient and practical strategy.

[0152] Specifically, this can be achieved in the following way:

[0153] (1) Input: TSS coordinates TSS_pos and a list of N candidate enhancer coordinates Enhancer_List.

[0154] (2) For each enhancer in Enhancer_List, calculate the absolute value of the distance between its center point and TSS_pos.

[0155] (3) Sort Enhancer_List in ascending order based on the calculated distance.

[0156] (4) Select the first 100 enhancers after sorting.

[0157] In some implementations, the third regulatory hierarchy includes 30 of the nearest promoter regions sampled within a 1 kb range upstream and downstream of the transcription start site.

[0158] This step provides the precise range and structure of the third regulatory level (initiator sublayer). The processing may include:

[0159] (1) Define the scope: take TSS as the center and delineate a 1kb area upstream and downstream (a total of 2kb) as the core start-up sub-region.

[0160] (2) Sampling: Identify the 30 nearest promoter subregions within this 2kb core region.

[0161] The output of this step is a list of coordinates of 30 promoter-related regions, all located within a 2kb range of the TSS.

[0162] The advantage of this step lies in its high-resolution, finely detailed characterization of the "decision core region" of gene transcription. The ±1kb range precisely covers the core promoter and proximal regulatory regions, which are key locations for the binding of most transcription factors and the transcriptional machine. Subdividing this region into 30 features allows the model to learn and understand the complex molecular events occurring within this critical region with greater precision, thereby more accurately determining the final integration effect of regulatory signals.

[0163] Specifically, this can be achieved in the following way:

[0164] Input: TSS coordinates TSS_pos.

[0165] Define the coordinate range of the core bootstrap subregion: [TSS_pos - 1000, TSS_pos + 1000].

[0166] Within this 2kb region, 30 sub-regions or feature points are defined and sampled according to preset rules (e.g., straight-line distance from the transcription start site of the target gene).

[0167] Example 3

[0168] Reference Figure 4 This embodiment provides a neural network computation method for gene expression regulation analysis. Step S300 involves inputting the input features of the multiple regulatory levels and the second feature data into a target neural network model, and outputting a prediction of the gene expression state through the target neural network model, including:

[0169] Step S310: Through the first dynamic routing module, based on the three-dimensional spatial interaction relationship between the first control level and the second control level, the input features of the first control level are mapped to the first intermediate features.

[0170] The above steps constitute the first core computational step within the model. They process input features from the "first regulatory level" (i.e., the broad contiguous units / bins of the genome). The processing method is a "mapping," whereby the original, relatively basic first-level features are transformed into a new, higher-level feature expression through a mathematical transformation. This mapping process is not static but is subject to real-time regulation by the "first dynamic routing module," whose core regulatory basis is the "three-dimensional spatial interaction relationship" data between the first and second levels.

[0171] The output of this step is the "first intermediate feature". This set of features can be understood as information after "relative condensation" of the first-level original information, which has absorbed information about the three-dimensional spatial structure.

[0172] The advantage of the above steps lies in the explicit introduction of three-dimensional spatial regulation logic at the initial stage of information processing. Instead of blindly processing all first-level features, it selectively and strategically transforms these features based on three-dimensional interaction data. This allows the model to simulate the biological principle that "spatial proximity of elements has a stronger regulatory effect," making feature transformations more biologically meaningful from the outset.

[0173] Specifically, in a neural network, this "mapping" can be implemented by a linear mapping module (such as a fully connected layer). The role of the "dynamic routing module" is to dynamically calculate and adjust the connection weights of this linear module.

[0174] The algorithm works as follows: the first dynamic routing module receives first-level features, second-level features, and the three-dimensional interaction data between them as input. Based on these inputs, a set of dynamic weights is calculated. This set of dynamic weights is applied to the linear mapping module to weight and transform the first-level input features, ultimately obtaining the "first intermediate feature".

[0175] For example, suppose the first level has features in 1000 bins. This step will convert these 1000 features into, for example, 100 "first intermediate features". During the conversion, if a bin (e.g., bin #5) is in close contact with an enhancer in the second level (e.g., enhancer #10) in three-dimensional space, the dynamic routing module will increase the weight of the information path from the input of bin #5 to the intermediate feature associated with enhancer #10.

[0176] Step S320: Combine the first intermediate feature with the input feature of the second control level, and through the second dynamic routing module, map the combined feature to the second intermediate feature based on the three-dimensional spatial interaction relationship between the second control level and the third control level.

[0177] This step is the second core step in information processing. It first performs feature fusion, merging the "first intermediate feature" obtained in the previous step with the original "second regulatory level input feature" (i.e., the feature of the enhancer itself). Then, this fused new feature set is mapped to the next stage through a process similar to the first step, controlled by a "second dynamic routing module" based on three-dimensional interaction relationships.

[0178] The output of this step is another set of higher-level features, known as "second intermediate features". This set of features not only contains the enhancer's own state information, but also incorporates information from its broader genomic environment (represented by the first intermediate features).

[0179] The advantage of this step lies in the effective integration of information across different levels. The model here learns not isolated enhancer features, but enhancer features "in a specific genomic environment." This design allows the model to capture more complex regulatory logics. For example, an enhancer may have strong openness, but if its surrounding genomic environment is unfavorable for regulation, its final effect may be compromised. This integration of information makes the model's decision-making more comprehensive.

[0180] Specifically, standard algorithms such as vector concatenation or element-wise addition can be used to merge the "first intermediate feature" and the "input feature of the second regulatory level" into a longer feature vector. Then, similar to the first step, a linear mapping module controlled by the "second dynamic routing module" is used to transform this combined feature vector to obtain the "second intermediate feature." The decision basis of this module is the three-dimensional interaction data between the second level (enhancer) and the third level (promoter).

[0181] For example, suppose the first intermediate feature has 100 dimensions, and the original features of the second regulatory level (enhancer) also have 100 dimensions. This step first merges them into a 100-dimensional feature vector. Then, when mapping it to (for example) a 30-dimensional "second intermediate feature", the second dynamic routing module will focus on strengthening the weights of the information pathways corresponding to those enhancers that have close three-dimensional contact with the third level (promoter).

[0182] Step S330: The second intermediate feature is combined with the input feature of the third regulatory level, and the combined feature is mapped to the prediction result of gene expression state by the third dynamic routing module based on the three-dimensional spatial interaction relationship between the first regulatory level and the third regulatory level.

[0183] In this final step, the highly condensed and integrated "second intermediate feature" obtained in the previous step is fused with the most crucial "third regulatory level input feature" (i.e., the feature of the promoter itself). Then, an output module will make a judgment on the gene expression status based on this final and most comprehensive feature representation.

[0184] The final output of this step is a single prediction, namely, "the prediction of gene expression status".

[0185] The advantage of this step is that it simulates the process by which biological regulatory signals ultimately converge at the promoter. Regulatory information from the entire genome, selected and integrated layer by layer from distant to near (represented by the second intermediate feature), is finally combined with the state of the promoter, the direct "switch" of gene expression, to make a final decision that most closely resembles biological reality.

[0186] Specifically, similar to the second step, the "second intermediate feature" and the "input feature of the third regulatory layer" can be merged. This final fused feature is then input into one or more output layers (usually fully connected layers), ultimately compressing the high-dimensional features into a single output value. This output value typically passes through an activation function (such as the sigmoid function), converting it into a probability value between 0 and 1.

[0187] For example, suppose the second intermediate feature has 30 dimensions, and the original features of the third regulatory level (promoter) also have 30 dimensions. They are combined into a final 30-dimensional feature vector. This vector is fed into the output layer, and a value, such as 0.85, is calculated, indicating that the model predicts that the gene has an 85% probability of being in a high-expression state.

[0188] In some implementations, the information transmission weight adjusted by the first dynamic routing module is: The calculation method (Formula 1) is as follows:

[0189] ;

[0190] in, The basic weight matrix representing the connection between the first and second layers; This represents the dynamic weight mask generated by the first dynamic routing module, and , Representing the first layer of features, Represents the second layer of features. This represents the spatial interaction matrix between the first and second layers.

[0191] This specifies in detail the information transfer strength (i.e., weights) from the first regulatory bin to the second regulatory enhancer in the model. How is it calculated? It can be divided into two steps:

[0192] (1) Dynamic mask generation: First, the function of f_router1 (i.e., the first dynamic routing module) will receive three inputs: features of the first level Second-level features and their three-dimensional spatial interactions Based on these three inputs, it calculates and generates a dynamic "weight mask" for the current sample. .

[0193] (2) Final weight calculation: Then, this dynamically generated weight will be used. This is in contrast to a relatively fixed "base weight matrix" learned during model training. Multiply the results to obtain the final effective information transmission weight.

[0194] This M × W design combines stability and flexibility. Matrix can be trained to learn universal regulatory patterns that are independent of specific samples. And W 12 Masks can be based on the unique features of each sample ( , ) and spatial structure ( This allows for dynamic and personalized fine-tuning of this universal law. This combination makes the model both robust and accurate.

[0195] Specifically, the f_router1 function itself can be a small neural network, such as a two-layer perceptron. It takes x1, x2, and edge_l12 as input and outputs a value similar to M. 12 A mask matrix W that matches the shape 12 .

[0196] In some implementations, the information transmission weight adjusted by the second dynamic routing module is: The calculation method (Formula 2) is as follows:

[0197] ;

[0198] in, The basic weight matrix representing the connection between the second and third layers; This represents the dynamic weight mask generated by the second dynamic routing module, and , Represents the second layer of features. Represents the third layer of features. This represents the spatial interaction matrix between the second and third layers.

[0199] This specifies the information transmission weights from the second level of regulators (enhancers) to the third level of regulators (promoters). The calculation method is exactly the same as the first item, except that it operates on the next stage of the model's information flow.

[0200] (1) Dynamic mask generation: generated by the second dynamic routing module f_router2 based on the second-level features. Third-level features And the three-dimensional interaction between them, edge_l23, generates a dynamic weight mask. .

[0201] (2) Final weight calculation: With the basic weight matrix Multiply them to get the final weights. .

[0202] The core of this method lies in applying the same dynamic regulatory mechanism to enhancer-promoter connections. This is because it directly simulates one of the most important regulatory events in biology—enhancer activation of promoters through three-dimensional contact. This allows the model to determine the influence of enhancers on gene expression based on the most crucial biological principles.

[0203] For a given enhancer-promoter pair, if its features (x2, x3) indicate that they are both active and their 3D interaction value edge_l23 is high (indicating physical close contact), then f_router2 will generate a W value close to 1. 23 Mask. This will make the base weights M 23 If the interaction value is fully preserved, information can be smoothly passed from the enhancer to the promoter. Conversely, if the interaction value is low, W... 23 The value will then tend to 0, effectively "closing" this regulatory pathway.

[0204] In some implementations, the target neural network model further includes a third dynamic routing module, whose information transmission weights The calculation method (Formula 3) is as follows:

[0205] ;

[0206] in, Represents the basic weight matrix; This represents the dynamic weight mask generated by the third dynamic routing module, and , Representing the first layer of features, Represents the third layer of features. This represents the spatial interaction matrix between the first and third layers.

[0207] In some implementations, each of the information transmission weights corresponds to a learnable sparse threshold; when the absolute value of the information transmission weight is less than its corresponding sparse threshold, the information transmission weight is reset to zero.

[0208] This embodiment provides a dynamic connection filtering mechanism. Its specific processing procedure is as follows:

[0209] The calculated weight of each piece of information ( , , Each model assigns and learns a corresponding, independent "sparse threshold" (τ). This threshold is not a fixed hyperparameter, but rather, like other model parameters, it is automatically optimized and learned during training.

[0210] After each calculation, the model takes the absolute value of the information transmission weight (i.e. its strength, regardless of whether it is activation or inhibition) and compares it with the corresponding sparsity threshold.

[0211] If the absolute value of a weight is less than this learnable threshold, it means the connection is too weak and is considered redundant or noisy, and the model will force the weight to be reset to zero. If the absolute value of a weight is greater than or equal to the threshold, its original calculated value remains unchanged.

[0212] After performing this step, you will get a sparsed information transfer weight matrix.

[0213] In this matrix, many weak connections that originally had small non-zero values ​​were "removed," leaving only those strong connections that were strong enough to pass the threshold test. This makes the internal connections of the entire neural network more sparse.

[0214] This mechanism automatically prunes a large number of weak connections, forcing the model to rely on a few of the most critical regulatory pathways for prediction. When analyzing the model, the remaining non-zero connections represent the most important and highest-confidence biological regulatory pathways identified by the model, making the model's decision-making logic clear; a sparser model generally means lower complexity. By removing redundant connections, the risk of "overfitting" on the training data can be effectively reduced, making the learned patterns more universal and performing better when faced with new and unseen data; a weight matrix with a large number of zero values ​​is more space-efficient in storage and may also be more computationally efficient, which has practical application value in processing large-scale genomic data.

[0215] The above mechanism can be implemented as an independent computational layer or operation during the training and inference process of a neural network.

[0216] The sparse threshold (τ) is defined as a trainable parameter of the model.

[0217] During the forward propagation of the model, after each weight calculation step, an operation can be added: final = if abs(ω)>= τ else 0.

[0218] During the backpropagation (training) process of the model, an appropriate gradient estimation method (such as a straight-through estimator) can be used to ensure that the gradient can flow through this discontinuous decision point, so that the threshold (τ) can also be updated and learned by the optimizer (such as Adam) like other parameters.

[0219] For example, suppose the model obtains the information transfer weights between enhancer A and promoter P in a single calculation. _{23} is '0.05'. Meanwhile, through training, the model has learned an optimal sparse threshold tau$ of 0.1 for this connection. The model compares the absolute value of the weights abs(0.05) with the threshold 0.1. Because 0.05 < 0.1, the condition is met, so the final weight of this connection is forced to 0. In this round of calculation, the regulatory pathway of enhancer A on promoter P is temporarily "shut down".

[0220] Example 4

[0221] Reference Figure 5 This embodiment provides a neural network computation method for gene expression regulation analysis. The method for constructing the target neural network model includes:

[0222] Step S400: Set the input nodes corresponding to the first control level, the second control level and the third control level.

[0223] This step is the first step in building the neural network model architecture, namely defining the model's input interface. This step requires designing three independent, dedicated input channels for the model to receive feature data from the first, second, and third control layers, respectively.

[0224] The output of this step is a neural network structure with three different input ports. These three ports are structurally separate, and the dimension (i.e., the number of nodes) of each port is matched to the number of features of the control layer it is to receive.

[0225] The advantage of this step is that it preserves the inherent hierarchy of biological data from the very bottom of the model's design. It avoids inputting feature data from all different sources and with different meanings together, instead structurally distinguishing them from the outset. This design makes the model's subsequent learning and information processing more orderly and closer to biological logic.

[0226] When programming with deep learning frameworks such as PyTorch or TensorFlow, this step corresponds to defining three independent input tensors or placeholders in the initialization or construction function of the model class, and specifying the expected shape for each of them.

[0227] For example, assuming that the first, second, and third regulation levels have 1000, 100, and 30 features respectively, then when building the model, the code will explicitly define three inputs: input_1 with the shape of [batch_size, 1000], input_2 with the shape of [batch_size, 100], and input_3 with the shape of [batch_size, 30].

[0228] Step S500: Construct multiple linear mapping modules for feature mapping between different control levels.

[0229] After defining the input interface, this step involves building the core information processing pathway within the model. This requires constructing multiple "linear mapping modules," which are computational units that enable the transfer and transformation of information from one control level to the next.

[0230] The output of this step is a neural network skeleton with basic information flow. For example, a module is built that can convert first-level input information into second-level representation, and a module can convert second-level information into third-level representation.

[0231] The advantage of this step lies in establishing a hierarchical information processing architecture. It constructs a directional and hierarchical computational path, which corresponds to the biological process in gene regulation where signals gradually converge from distal elements (such as enhancers) to proximal elements (such as promoters), laying a structural foundation for subsequent realization of more refined regulatory simulations.

[0232] In deep learning frameworks, "linear mapping modules" typically refer to fully connected layers (or dense layers). Building these modules means instantiating these layers in the model's code and defining their respective input and output dimensions.

[0233] For example, based on the previous example, this step will construct a Linear_Module_1 with an input dimension of 1000 and an output dimension of 100. Simultaneously, a Linear_Module_2 will be constructed with an input dimension of 100 (or the dimension after combining with the second-level features) and an output dimension of 30.

[0234] Step S600: Construct multiple dynamic routing modules and configure the dynamic routing modules to control the information transmission weights of the linear mapping module; wherein, the configuration includes: enabling the dynamic routing module to receive features of the source layer, features of the target layer, and data representing the three-dimensional spatial interaction relationship between the source layer and the target layer, and generating dynamic weights based on these.

[0235] This step is the most crucial and innovative part of the entire model-building methodology: installing "smart valves" in the information processing pathway. This step consists of two parts:

[0236] (1) Construction: Add a special component called "dynamic routing module" to the architecture of the model.

[0237] (2) Configuration: These modules are configured to correctly receive and process three specific sets of input data: features at the source level, features at the target level, and the three-dimensional spatial interaction data between them. Ultimately, the modules are configured to use these three sets of inputs to generate a "dynamic weight".

[0238] The output of this step is a complete and unique neural network architecture. It is no longer a standard network that passively transmits information, but a network in which the information flow is actively and intelligently controlled. Its control logic is pre-designed (i.e., "configured") to rely on externally introduced three-dimensional spatial interaction data.

[0239] The advantage of this step is that it hard-codes explicit biological prior knowledge into the model's structural design. Through this "configuration" step, the model is forced to consider the physical proximity in three-dimensional space when determining the strength of information transmission. This design creates a model that is inherently more biologically interpretable and more realistic, fundamentally different from general models that rely solely on data fitting.

[0240] Specifically, the "dynamic routing module" can first be implemented as an independent sub-network (e.g., a two-layer perceptron). Then, in the main computation flow of the model (i.e., the forward method), code logic is written to ensure that the correct inputs (e.g., input_1, input_2, edge_l12) are passed to the corresponding dynamic routing module (e.g., Router_1). Then, the output of this module (i.e., the "dynamic weights") is calculated (e.g., element-wise multiplication) with the weights of the corresponding linear mapping module (e.g., Linear_Module_1), thereby controlling the information flow.

[0241] In some embodiments, the neural network computation method for gene expression regulation analysis further includes training the target neural network model, including:

[0242] The parameters of the model are optimized using a binary cross-entropy loss function, and the calculation formula for the binary cross-entropy loss function (Formula 4) is as follows:

[0243] ;

[0244] Where N represents the total number of samples; The true label represents the i-th sample; This represents the model's predicted probability for the i-th sample.

[0245] This embodiment provides the core mechanism for model training. The so-called "training" or "optimization" refers to a process of repeatedly adjusting all learnable parameters within the model (such as the basic weight matrix M mentioned in the previous embodiment), with the ultimate goal of making the model's prediction results as consistent as possible with the actual biological results.

[0246] This process is achieved through a "loss function." The loss function is used to quantify the difference or "error level" between the model's current prediction and the "standard answer" (i.e., the true label). This implementation explicitly specifies the use of the "binary cross-entropy loss function," which is a very effective and commonly used loss function specifically designed for handling binary classification problems (i.e., where there are only two possible outcomes, such as "gene expression" and "no expression").

[0247] At each step of training, a single value is calculated: the "loss". This value represents how poorly the model performs on the current batch of data. The goal of training is to minimize this loss value by continuously adjusting the model parameters.

[0248] Using binary cross-entropy as the loss function has significant advantages. It is a smooth and differentiable function that can calculate a definite gradient (i.e., the "direction to be adjusted") for each parameter of the model. This allows gradient-based modern optimization algorithms (such as Adam and SGD) to work efficiently, thus ensuring the stability and convergence of the model training process and effectively finding a set of optimal parameters that make the model's predictions most accurate.

[0249] For example, this can be achieved through the following machine learning training process:

[0250] (1) The model receives a batch of input features, performs forward propagation, and obtains a batch of predicted probabilities σ(p). i ).

[0251] (2) Compare these predicted probabilities with the corresponding true labels y i They are input together into a predefined binary cross-entropy loss function to calculate the loss value.

[0252] (3) Calculate the gradient of the loss value with respect to all parameters of the model using the backpropagation algorithm.

[0253] (4) An optimizer makes small updates to the model parameters based on these gradients.

[0254] (5) Repeat the above steps until the loss value converges to a sufficiently small value or reaches the preset number of training rounds.

[0255] In some implementations, the true label of the sample is determined by: acquiring RNA sequencing expression data of the target gene; when the expression level of the target gene is greater than a first preset threshold, its true label is determined as a positive sample; when the expression level of the gene is less than a second preset threshold, its true label is determined as a negative sample.

[0256] This embodiment provides a complete algorithm for converting continuous gene expression abundance values ​​into discrete binary classification labels. The first step is to query the expression level of each target gene from the results of an RNA sequencing (RNA-seq) experiment. This is typically a continuous value representing gene transcriptional activity (e.g., FPKM or TPM value). Next, instead of directly using this continuous value, the method employs a dual-threshold filtering approach. It sets a higher "first preset threshold" and a lower "second preset threshold."

[0257] If the expression level of a gene exceeds a high threshold, it means that it is a clearly expressed, highly expressed gene, and therefore its "true label" is determined as a positive sample (usually denoted as 1 in the calculation).

[0258] If the expression level of a gene is below a low threshold, it means that it is a clearly defined gene that is hardly expressed, and therefore its "true label" is determined as a negative sample (usually recorded as 0 in calculations).

[0259] The final output of this step is high-quality binary "true labels" (y) used for model training. i For each gene sample, it is either assigned a positive sample label or a negative sample label.

[0260] The greatest advantage of this method lies in creating a high-confidence training dataset. By setting two thresholds, this method cleverly ignores genes with intermediate states—those whose expression levels are neither high nor low, and whose states are ambiguous. Training the model only with genes whose expression states are clearly defined as "extremely high expression" and "extremely low expression" prevents the model from being misled by samples with ambiguous boundaries, thus learning clearer and more robust regulatory patterns. Allowing the model to focus on learning the most fundamental difference between the "on" and "off" states helps the model converge faster and achieve higher prediction accuracy.

[0261] For example, suppose the first preset threshold (high threshold) is set to 10 and the second preset threshold (low threshold) is set to 1.

[0262] The RNA-seq expression value of gene A is 50. Because 50 > 10, its true label y A It was identified as a positive sample (1).

[0263] The RNA-seq expression value of gene B is 0.2. Since 0.2 < 1, its true label y is... B It was identified as a negative sample (0).

[0264] The RNA-seq expression value of gene C is 5. Because it falls between 1 and 10, according to this rule, it will not be assigned a positive or negative label and therefore will not be included in this batch of high-confidence training samples.

[0265] Example 5

[0266] Reference Figure 6 This embodiment provides a neural network computation method for analyzing gene expression regulation. The method further includes interpretability analysis of the prediction results of the target neural network model, including:

[0267] Step S700: Select at least one of the plurality of control levels as the target level to be interpreted, and set a reference input baseline for the target level.

[0268] This step involves preparation and configuration before conducting interpretability analysis. It includes two key actions:

[0269] (1) Selecting the target level: The analyst can select one or more levels of interest from the multiple input levels of the model (such as bins of the first level, enhancers of the second level, promoters of the third level) as the focus of this interpretation.

[0270] (2) Set a reference baseline: Create a "neutral" or "no information" reference input for the selected target level, i.e., a baseline. The role of this baseline is to serve as a "zero point" or "control group", and the contribution values ​​of all features calculated subsequently are relative to this baseline.

[0271] The output of this step is to set a clear scope and a starting point for comparison for the upcoming interpretability analysis.

[0272] The advantage of this step is that it enables focused and meaningful analysis. Analysts do not have to face the contribution values ​​of all input features of the model (which may be thousands), but can ask targeted questions, such as "Which of these 100 enhancers are the most important?" Setting a reference baseline ensures the scientific rigor and consistency of the contribution value calculation, guaranteeing the reliability of the results.

[0273] In the program, this is achieved by specifying the input feature tensor to be analyzed (e.g., choosing a tensor representing the second-level enhancer features). The most common implementation is to create a tensor with the exact same shape as the target level feature tensor, but with all elements being zero (i.e., an "all-zero vector" or "zero matrix"), which represents a "zero state" without any biological signal input.

[0274] For example, if it is necessary to investigate which enhancers activate a certain gene, in this step, he would select the "second regulatory level (enhancers)" as the target level to be interpreted, and set a 100-dimensional vector with all values ​​of 0 as the reference input baseline.

[0275] Step S800: Using an attribution algorithm based on contribution backtracking, calculate the marginal contribution value of each input feature in the target level to the prediction result.

[0276] This is the core computational step in interpretability analysis. It employs a special type of algorithm, "contribution-based attribution algorithm," to perform the computation. This type of algorithm starts from the model's final output (i.e., the "prediction"), traces back along the computational path of the neural network, and "attributes" the final prediction to each input feature. This process calculates a "marginal contribution value" for each feature in the target layer (e.g., each enhancer).

[0277] The output of this step is a list or vector of contribution values, with a length equal to the number of features in the target layer. Each value in the list quantifies the degree to which the corresponding input feature contributes to the final prediction result. Positive values ​​typically indicate that the feature promotes positive prediction (such as gene expression), while negative values ​​indicate an inhibitory effect. The absolute magnitude of the value represents the strength of the influence.

[0278] The advantage of this step is that it provides an objective and quantitative measure of feature importance. It transforms a vague concept of "importance" into a precisely comparable numerical value, allowing researchers to clearly see the extent of each regulatory element's "voice" in the model's decision-making process.

[0279] Specifically, this can be achieved by calling a dedicated interpretable algorithm library (such as Captum). The "contribution-based attribution algorithm" mentioned in the claims is a general term for this type of algorithm; specific examples include DeepLIFT and Integrated Gradients. The algorithm's inputs are the trained model, real sample inputs, and the reference baseline set in the first step; the output is the contribution value of each feature.

[0280] For example, taking the 100 enhancers selected in the aforementioned embodiment as an example, this step will output a list containing 100 values. For example, [the contribution value of enhancer 1 = 0.5, the contribution value of enhancer 2 = -0.1, the contribution value of enhancer 3 = 0.02, ...].

[0281] Step S900: Sort the input features in the target level according to the marginal contribution value, and select a predetermined number of high contribution features as the identified key functional control elements.

[0282] This step is the final step in the analysis, aiming to transform the numerical results calculated in the previous step into conclusions that have direct guiding significance for biologists. The process includes: sorting all input features at the target level according to their corresponding marginal contribution values ​​from high to low; and selecting the top-ranked features (e.g., Top 10) from the sorted list.

[0283] The final output of this step is a sorted and filtered short list containing the "key functional regulatory elements" that the model believes contribute the most to the expression of the gene.

[0284] The advantage of this step is that it provides clear, actionable biological hypotheses. It distills complex analytical results into a well-defined "list of suspects." Researchers can then directly use this list to design subsequent biological experiments (e.g., using gene editing technology to knock out the enhancer with the highest contribution) to verify the model's predictions, thereby efficiently advancing scientific discovery.

[0285] This is a standard sorting and selection algorithm: pair the feature list with the contribution value list; call a sorting function to sort the pairing list in descending order according to the contribution value; and extract the first N elements (N is a "predetermined number") from the sorted list.

[0286] For example, after sorting the 100 enhancers and their contribution values, this step will select the 10 enhancers with the highest contribution values ​​and output a list: "Key enhancers are: enhancer #77, enhancer #12, enhancer #54, ...".

[0287] Example 6

[0288] To better illustrate the neural network computation method for gene expression regulation analysis provided in the foregoing embodiments, specific experimental examples are given below. However, it should be understood that these experimental examples are only for more detailed explanation and should not be construed as limiting the present invention in any way.

[0289] Experimental methods:

[0290] 1. Data collection:

[0291] To construct an interpretable neural network computational model that integrates prior biological knowledge, this method systematically integrates high-quality omics datasets from multiple public databases, covering omics information from typical human cell lines such as GM12878, K562, NHEK, and HMEC.

[0292] This included acquiring high-resolution Hi-C data to construct a three-dimensional spatial interaction matrix of chromatin, thereby identifying spatial interactions between different gene regulatory regions; and collecting ATAC-seq data to assess chromatin accessibility, in order to determine whether DNA regions are in an open state, and thus to determine their potential regulatory activity.

[0293] Simultaneously, by combining genome annotation information provided by databases such as GENCODE, the transcription start site (TSS) and its upstream and downstream regulatory regions of each gene were precisely located. RNA-seq expression profile data were also introduced as supervised learning labels to distinguish between highly expressed and low-expressed genes. Furthermore, to further verify the biological rationality of the model's predictions, transcription factor binding site prediction data from databases such as JASPAR were used to help explain the functional mechanisms of the cis-regulatory elements identified by the model.

[0294] Taking the K562 cell line as an example, a total of 20,260 genes were collected from ATAC-seq, Hi-C and RNA-seq data. The multi-source integration of data provided a comprehensive and reliable input foundation for the model, ensuring that the neural network training process takes into account both the advantages of data-driven and knowledge-guided approaches.

[0295] 2. Dataset Construction:

[0296] Taking the human leukemia cell line K562 as an example, it covers approximately 20,260 coding genes. In this embodiment, a structured data file is generated independently for each gene, and model input and supervision labels are constructed. The core data modeling process is as follows:

[0297] (1) Region delineation: The transcription start site (TSS) of the target gene is the center, extending 2.5Mb upstream and downstream, for a total range of 5Mb. This region is divided into 1000 equal-length 5kb bins, forming the Layer 1 characteristic region.

[0298] (2) Functional area definition:

[0299] 1) The promoter is defined as the TSS±1kb region, and a total of 30 nearest promoter regions are sampled (Layer 3).

[0300] 2) Enhancers are identified by chromatin accessibility signal (ATAC-se) and the 100 nearest enhancers are selected by combining the linear distance with TSS (Layer 2).

[0301] (3) Construction of the interaction matrix:

[0302] Based on Hi-C data, the following three sets of spatial interaction matrices were calculated: A, edge_l12: 1000 × 100, contact strength between bin and enhancer; B, edge_l13: 1000 × 30, contact strength between bin and promoter; C, edge_l23: 100 × 30, contact strength between enhancer and promoter.

[0303] (4) Quantification of openness:

[0304] The ATAC-seq intensities of the aforementioned regions were extracted from the chromatin accessibility map and used as input features for three layers: 1) Layer 1: 1000 dimensions; 2) Layer 2: 100 dimensions; 3) Layer 3: 30 dimensions.

[0305] (5) Expression level annotation and tag generation: Extract the FPK value of the target gene from the RNA-seq expression data; genes with an expression level >10 are labeled as positive samples (Pos); genes with an expression level <1 are labeled as negative samples (Neg); and samples in between are labeled as "None".

[0306] This process constructs high-quality multimodal input samples on a single-gene basis, ensuring that the model training data has a clear structure and well-defined hierarchical features.

[0307] 3. Model design and model construction:

[0308] Overall Model Design Concept: This embodiment proposes a deep neural network model that integrates chromatin spatial interaction structure and hierarchical representation of regulatory element functions for gene expression regulation modeling tasks. The framework diagram is shown below. Figure 7 As shown, this model extracts multi-level feature inputs based on ATAC-seq, Hi-C, and other data, and constructs a structurally explicit routing control mechanism to better simulate the flow path of regulatory information between different levels of CREs (such as bins, enhancers, and promoters), thereby improving prediction performance and biological interpretability.

[0309] Network structure composition: This embodiment provides a neural network model structure for predicting gene regulation relationships. Its overall architecture includes three levels of input features and three types of spatial interaction information, and the output is the gene expression state (binary classification result).

[0310] (1) Constructing multi-level control input features and spatial interaction data:

[0311] 1) The input to the data to be predicted includes three levels of regulatory feature vectors, namely:

[0312] A. First layer features (Layer 1): The original bin feature vector, with a dimension of 1000;

[0313] B. Second Layer Features: Enhanced sub-region features, with a dimension of 100;

[0314] C. Third layer features: These are the features of the promoter region, with a dimension of 30.

[0315] 2) Furthermore, to describe the spatial control relationships between different levels, three types of spatial interaction matrices are introduced, including:

[0316] edge_l12: Represents the spatial interaction matrix between Layer 1 and Layer 2, with a shape of [batch, 100000];

[0317] edge_l23: Represents the spatial interaction matrix between Layer 2 and Layer 3, with a shape of [batch, 3000].

[0318] edge_l13: Represents the shortcut interaction matrix between Layer 1 and Layer 3, with a shape of [batch, 30000].

[0319] (2) Construct a neural network backbone structure that includes a dynamic routing mechanism:

[0320] The constructed neural network model backbone consists of three linear mapping modules and multiple dynamic routing control modules, used to capture the nonlinear mapping relationships of inter-layer modulation features. It includes the following three-layer mapping structure:

[0321] 1) Linear 1 (bin → enhancer): The input is 1000-dimensional bin features, and the output is a 100-dimensional representation. The explicit control module is Router 1. The weight calculation method is: ω_12=M_12*W_12, where W_12=f_router1(x_1,x_2,edge_l12). Connections in the weight mask that are below the preset threshold will be set to zero.

[0322] 2) Linear 2 (enhancer → promoter): The input is a 100-dimensional enhancer feature + a 100-dimensional representation of the output of Linear 1, and the output is a 30-dimensional representation. The explicit control module is Router 2. The weight calculation method is: ω_23=M_23*W_23, where W_23=f_router2(x_2,x_3,edge_l23). Connections in the weight mask that are below the preset threshold will be set to zero.

[0323] 3) Linear 3 (Output Layer): Input is 30-dimensional promoter features + 30-dimensional representation of the output from Linear 2. The explicit control module is Router 3, and the weights are calculated as: ω_13=M_13*W_13, where W_13=f_router3(x_1,x_3,edge_l13). Connections in the weight mask below a preset threshold are set to zero. The output is the predicted probability of gene expression (processed by the Sigmoid activation function).

[0324] (3) Design a router module to achieve connection feasibility modeling: The router module is used to model the validity of inter-layer connections and control whether connections are activated or suppressed. Each pair of connections is controlled by an independent router. The router module is designed as a two-layer perceptron structure, and its inputs include:

[0325] A. Source feature vectors (such as bin or enhancer features);

[0326] B. Target feature vector (such as enhancer or promoter features);

[0327] C. Boundary values ​​(edge_ij) in the corresponding spatial interaction matrix.

[0328] The Router module outputs a scalar mask value, which is normalized using the Sigmoid activation function to adjust the weight of the corresponding connection.

[0329] (4) Sparsity control of connections based on mask gating mechanism: To improve the compressibility, generalization ability and subsequent interpretability of the model, a sparsity gating mechanism is introduced into the connection weights, as follows: Each connection is associated with a learnable sparse threshold τ; if the connection weight satisfies |τ×τ|<τ, then the connection is set to 0; otherwise, the connection is retained as an effective regulatory path. This mechanism enables the model to automatically select important regulatory pathways and eliminate redundant connections, which is beneficial to improving the credibility of biological explanations.

[0330] 4. Model training:

[0331] The model training process uses the binary cross-entropy loss function (BCEWithLogitsLoss), i.e., the aforementioned formula 4, with Adam as the optimizer, an initial learning rate of 0.005, 60 training epochs, and a batch size of 256. During each training epoch, the training set loss and the AUC of the training / validation / test sets are recorded, and the current model parameters are saved when the validation set AUC reaches its optimum. After model training is complete, the optimal model weights are saved according to the experimental cell lines (such as K562, GM12878, mouse_8cell, etc.) for downstream tasks and interpretability analysis. To enhance model interpretability, an explicit threshold parameter is introduced in the structural design to learnably suppress low-strength connections, thereby highlighting high-contribution pathways.

[0332] In addition, to verify the model performance, the AUC change trend of each round is recorded in real time during the training process, and the best validation set AUC and the corresponding test set performance are output at the end as the final evaluation metric.

[0333] 5. Introduction of interpretable methods:

[0334] To improve the interpretability of model predictions, this embodiment integrates the DeepLIFT algorithm module from the Captum framework outside the model architecture to provide a structured interpretation of the target model's prediction output. This method does not rely on modifications to the model's structure itself; instead, it uses the trained neural network model to backtrack the contribution of the feature inputs of each sample to be interpreted. The interpretation process is as follows:

[0335] (1) Loading the model and organizing the input data: First, load the parameters of the trained neural network model and obtain the sample data to be interpreted. Organize the sample data into an input format acceptable to the model, including the three-layer structure of regulatory features and their corresponding spatial interaction matrices, as follows: Layer 1: original bin features; Layer 2: enhanced sub-region features; Layer 3: activating sub-region features; edge_l12: represents the spatial interaction matrix between Layer 1 and Layer 2; edge_l23: represents the spatial interaction matrix between Layer 2 and Layer 3; edge_l13: represents the spatial interaction matrix between Layer 1 and Layer 3.

[0336] (2) Constructing the DeepLIFT interpreter object: Based on the DeepLift class in the Captum framework, instantiate the model interpreter object to perform model interpretability analysis. Taking the interpretation of enhancers as an example, the specific process includes: setting the enhancer feature (Layer 2, i.e., x_l2) as the main interpretation target; constructing a zero vector of the corresponding dimension as the reference input (baseline); initializing the DeepLift interpreter object: dl; and calling the attribute method to perform feature attribution analysis.

[0337] (3) Obtaining and processing contribution values: After executing the above call, the model returns the marginal contribution value (attributes vector) of each enhancer feature to the final output. The vector represents the relative influence of each enhancer feature on the model's prediction result. The contribution values ​​are sorted, and the Top-N high-contribution enhancer regions are selected and compared with known experimental annotation regions or functional annotation databases to verify the consistency between the model's interpretation results and the actual regulatory mechanism.

[0338] (4) Exporting and applying the interpretation results: To support subsequent analysis, the model supports exporting the ranking results of enhancer contribution values ​​as a .txt file. The output includes the index of each enhancer feature and its corresponding contribution score. This ranking list can be used to: draw visualizations of enhancer regulatory importance; perform significant enrichment analysis; and for functional annotation and target selection. This interpretation module can be applied to the output of any model based on a neural network structure without requiring any modification to the original network architecture, demonstrating good versatility and scalability.

[0339] 6. Model performance and interpretability verification:

[0340] Gene expression state prediction tasks were performed in multiple cell lines to test the model's performance. Predictions were compared with those of a traditional multilayer perceptron (MLP) model built using the same input information to validate the advantages of the model structure in expression modeling tasks. Furthermore, the model's parameter efficiency was verified by comparing the total number of parameters with a fully connected model. Regarding model interpretability, the interpretable results were compared with those of the classic ABC (Activity-By-Contact) model or the regulatory results of classic regulatory genes such as MYC to verify the model's interpretability performance.

[0341] 7. Downstream applications of the model in complex biological processes:

[0342] This embodiment can be applied to the mining of regulatory mechanisms in complex biological developmental processes. For example, a key event in early mammalian embryonic development is zygotic gene activation (ZGA), which marks the embryo's transition from maternally dependent transcripts to autonomous gene expression, accompanied by large-scale chromatin remodeling. Traditional models struggle to effectively capture its regulatory logic. The model can be applied to the analysis of multiple developmental time points to reveal the systemic functions of different gene regeneration processes (CREs) within the regulatory framework. It is worth emphasizing that the model's application value is not limited to embryonic development research. Due to its modeling architecture's universality based on the biological structures of regulatory pathways, the model can also be widely used in various biomedical research fields such as cancer development and cell differentiation, particularly in identifying key functional CREs in disease states, thereby providing theoretical support and target screening for precision intervention strategies such as gene editing and targeted regulation.

[0343] Example 7

[0344] To better illustrate the neural network computation method for gene expression regulation analysis provided in the foregoing embodiments, please refer to... Figure 8 The flowchart illustrates a process for modeling and interpretability analysis of regulatory pathways across multiple regulatory regions, based on the regulatory prediction method proposed in the preceding embodiments and targeting the MYC gene in the K562 cell line. Specifically, it includes:

[0345] 1. Obtain target data samples and known regulatory labels:

[0346] The human leukemia cell line K562 was selected as the research subject. Targeting the MYC gene, omics data were obtained from its 5MB binary region, spatial enhancers, and promoter regions to construct a three-layer input feature model. Each regulatory region contains the following information:

[0347] (1) Original features: ATAC-seq fragment density (Layer 1, Layer 2, Layer 3).

[0348] (2) Spatial interaction matrix: The interaction strength between layer1-layer2, layer2-layer3, and layer1-layer3 calculated by Hi-C.

[0349] (3) Tag data: Gene expression tags obtained from RNA-seq data.

[0350] 2. Data preprocessing process: Perform the following preprocessing steps:

[0351] Construct a three-layer input feature matrix:

[0352] (1) Layer 1: 1000-dimensional bin features (MYC gene ± 2.5Mb range);

[0353] (2) Layer 2: 100 spatial interaction enhancer features;

[0354] (3) Layer 3: Features of 30 promoter regions (within ±2kb range of TSS);

[0355] Interaction matrix processing: edge_l12 (1000 × 100), edge_l23 (100 × 30), and edge_l13 (1000 × 30) are all reshaped into one-dimensional flat vector input models.

[0356] 3. Model Input and Feature Flow

[0357] The above three layers of features and three sets of interaction matrices are input into the various modules of the model of this invention, as follows:

[0358] (1) bin → enhancer: mapped through Linear_1 and controlled by Router_1, fused with edge_l12;

[0359] (2) enhancer → promoter: through Linear_2 mapping, Router_2 control, and fusion with edge_l23;

[0360] (3) Output fusion: Through Linear_3 mapping and Router_3 control, edge_l13 is fused to perform classification prediction and output the MYC gene expression status (binary classification).

[0361] 4. Model interpretability analysis (taking enhancers as an example):

[0362] After the model is trained, the DeepLIFT algorithm is used to back-interpret the prediction results of the MYC samples. The operation steps are as follows:

[0363] (1) Select a reference input as a vector of all zeros;

[0364] (2) Calculate DeepLIFT attribution values ​​for the Layer 2 input nodes (enhancers) in the model;

[0365] (3) After sorting, the marginal contribution value of each enhancer to the MYC prediction result is obtained;

[0366] (4) The top 10 enhancers by contribution value were compared with known regulatory regions, and several high-scoring enhancers were found to be consistent with prior knowledge, thus verifying the accuracy of the model interpretation.

[0367] 5. Results Output and Functional Verification

[0368] The interpretation results will be output to two files:

[0369] (1) l2_zero.txt: List of enhanced sub-indices sorted by contribution value;

[0370] (2) l2_zero_importance.txt: Attribution score for each enhancer;

[0371] 6. Analysis and Explanation: This embodiment, using the MYC gene as an example, demonstrates the modeling, learning, and interpretation capabilities of multi-layered regulatory pathways in regulatory prediction tasks, supported by the actual three-dimensional interaction structure of chromatin. It showcases the robustness and transferability of this invention under complex multimodal input conditions. DeepLIFT-assisted interpretation provides causal clues behind the model predictions, making the results highly biologically readable.

[0372] refer to Figure 9 This application provides a neural network computing device for analyzing gene expression regulation, comprising:

[0373] The acquisition module 10 is used to acquire first feature data representing chromatin accessibility and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements.

[0374] The construction module 20 is used to construct an input feature containing multiple control levels based on the first feature data; wherein each control level corresponds to a type of control element.

[0375] The prediction module 30 is used to input the input features of the multiple regulatory levels and the second feature data into the target neural network model, and output the prediction result of gene expression state through the target neural network model; wherein, the target neural network model includes at least one dynamic routing module, used to determine and adjust the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

[0376] It is understood that the device in this embodiment corresponds to the neural network computation method for gene expression regulation analysis in the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0377] This application also provides a computer system, exemplary of which includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the aforementioned neural network computation method for gene expression regulation analysis.

[0378] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0379] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0380] This application also provides a computer storage medium for storing the computer program used in the aforementioned computer device. The computer storage medium can be a readable storage medium, a non-volatile storage medium, or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0381] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0382] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0383] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0384] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A neural network computation method for analyzing gene expression regulation, characterized in that, include: Acquire first feature data representing chromatin accessibility, and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements; Based on the first feature data, an input feature containing multiple control levels is constructed; wherein each control level corresponds to a type of control element. The input features of the multiple regulatory levels and the second feature data are input into the target neural network model, and the target neural network model outputs the prediction result of gene expression state; wherein, the target neural network model includes at least one dynamic routing module, which is used to determine and adjust the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

2. The neural network computation method for gene expression regulation analysis as described in claim 1, characterized in that, Based on the first feature data, an input feature structure is constructed containing multiple control levels; wherein each control level corresponds to a type of control element, including: The genomic region centered on the transcription start site of the target gene is divided into multiple contiguous units, which serve as the first regulatory level. Enhancer regions of the target gene are identified and screened as the second regulatory level; The promoter region near the transcription start site of the gene is defined as the third regulatory level.

3. The neural network computation method for gene expression regulation analysis as described in claim 2, characterized in that, The first regulatory hierarchy is constructed as follows: a genomic region totaling 5 Mb, centered on the transcription start site, is selected, and the region is divided into 1000 contiguous units, each 5 kb in length; and / or, The second regulatory hierarchy includes the 100 nearest enhancer regions selected based on their linear distance from the transcription start site; and / or, The third regulatory hierarchy comprises the 30 nearest promoter regions sampled within a 1kb range upstream and downstream of the transcription start site.

4. The neural network computation method for gene expression regulation analysis as described in claim 1, characterized in that, The step of inputting the input features of the multiple regulatory levels and the second feature data into the target neural network model, and outputting a prediction of gene expression status through the target neural network model, includes: The first dynamic routing module maps the input features of the first control level to the first intermediate features based on the three-dimensional spatial interaction between the first control level and the second control level. The first intermediate feature is combined with the input feature of the second control level, and the combined feature is mapped to the second intermediate feature through the second dynamic routing module based on the three-dimensional spatial interaction relationship between the second control level and the third control level. The second intermediate feature is combined with the input feature of the third regulatory level, and the combined feature is mapped to the prediction result of gene expression state by the third dynamic routing module based on the three-dimensional spatial interaction relationship between the first regulatory level and the third regulatory level.

5. The neural network computation method for gene expression regulation analysis as described in claim 4, characterized in that, The information transmission weight adjusted by the first dynamic routing module is The calculation method is as follows: ； in, The basic weight matrix representing the connection between the first and second layers; This represents the dynamic weight mask generated by the first dynamic routing module, and , Representing the first layer of features, Represents the second layer of features. Represents the spatial interaction matrix between the first and second layers; and / or, The information transmission weight adjusted by the second dynamic routing module is The calculation method is as follows: ； in, The basic weight matrix representing the connection between the second and third layers; This represents the dynamic weight mask generated by the second dynamic routing module, and , Represents the second layer of features. Represents the third layer of features. Represents the spatial interaction matrix between the second and third layers; and / or, The information transmission weight adjusted by the third dynamic routing module is: The calculation method is as follows: ； in, Represents the basic weight matrix; This represents the dynamic weight mask generated by the third dynamic routing module, and , Representing the first layer of features, Represents the third layer of features. This represents the spatial interaction matrix between the first and third layers.

6. The neural network computation method for gene expression regulation analysis as described in claim 5, characterized in that, Each of the aforementioned information transmission weights corresponds to a learnable sparse threshold; When the absolute value of the information transmission weight is less than its corresponding sparse threshold, the information transmission weight is reset to zero.

7. The neural network computation method for gene expression regulation analysis as described in claim 1, characterized in that, The method for constructing the target neural network model includes: Set the input nodes corresponding to the first, second, and third control levels; Construct multiple linear mapping modules to perform feature mapping between different control levels; Multiple dynamic routing modules are constructed and configured to control the information transmission weights of the linear mapping module; wherein, the configuration includes enabling the dynamic routing module to receive features of the source layer, features of the target layer, and data representing the three-dimensional spatial interaction relationship between the source layer and the target layer, and generating dynamic weights based on these.

8. The neural network computation method for gene expression regulation analysis as described in claim 1, characterized in that, The neural network computation method for analyzing gene expression regulation further includes training the target neural network model, including: The parameters of the model are optimized using a binary cross-entropy loss function, the formula for which is: ； Where N represents the total number of samples; The true label represents the i-th sample; This represents the model's predicted probability for the i-th sample.

9. The neural network computation method for gene expression regulation analysis as described in claim 8, characterized in that, The true label of the sample is determined in the following way: Obtain RNA sequencing expression data of the target gene; When the expression level of the target gene is greater than a first preset threshold, its true label is determined as a positive sample; When the expression level of the gene is less than a second preset threshold, its true label is determined to be a negative sample.

10. The neural network computation method for gene expression regulation analysis as described in claim 1, characterized in that, The neural network computation method for analyzing gene expression regulation further includes interpretability analysis of the prediction results of the target neural network model, including: Select at least one of the plurality of control levels as the target level to be explained, and set a reference input baseline for the target level; An attribution algorithm based on contribution backtracking is used to calculate the marginal contribution value of each input feature in the target level to the prediction result; The input features in the target level are sorted according to the marginal contribution value, and a predetermined number of high contribution features are selected as the identified key functional control elements.

11. A neural network computing device for analyzing gene expression regulation, characterized in that, include: The acquisition module is used to acquire first feature data representing chromatin accessibility and second feature data representing the three-dimensional spatial interaction relationship between different regulatory elements. The construction module is used to construct an input feature containing multiple control levels based on the first feature data; wherein each control level corresponds to a type of control element. The prediction module is used to input the input features of the multiple regulatory levels and the second feature data into the target neural network model, and output the prediction result of gene expression state through the target neural network model; wherein, the target neural network model includes at least one dynamic routing module, used to determine and adjust the information transmission weight from the source level to the target level based on the input features of the source level, the input features of the target level, and the three-dimensional spatial interaction relationship between the source level and the target level.

12. A computer system, characterized in that, The computer system includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the neural network computation method for gene expression regulation analysis according to any one of claims 1-10.

13. A computer storage medium, characterized in that, It stores a computer program, which, when executed on a processor, implements a neural network computation method for gene expression regulation analysis according to any one of claims 1-10.

Images