Muscle fiber specialization core target point mining method and system based on multi-modal learning

Abstract

The invention discloses a muscle fiber specialization core target point mining method and system based on multi-modal learning, particularly relates to the technical field of poultry muscle biology research, and is used for solving the problem of target point mining reliability caused by insufficient space-time alignment precision of multi-modal data in existing living body detection. The method comprises the following steps: acquiring living goose muscle tissue imaging and molecular omics data, performing spatial registration and time sequence alignment processing to generate time-space synchronous multi-modal data, calculating a time sequence correlation coefficient, judging whether the time sequence correlation coefficient exceeds a threshold value or not, generating a biological pathway activation intensity attention map based on the molecular omics data, and performing feature fusion. And finally, mining a core target spot through the target spot identification model and outputting a list, thereby realizing accurate identification of the muscle fiber specialization key target spot in a living body state.

Description

Technical Field

[0001] This invention relates to the field of poultry muscle biology research technology, and more specifically, to a method and system for mining muscle fiber specialization core targets based on multimodal learning. Background Technology

[0002] Research on poultry muscle biology has received increasing attention in recent years, particularly regarding the mechanisms of muscle fiber type transformation in economically important poultry such as geese. Compared to mice, a model animal, geese exhibit significant differences in muscle growth patterns, fiber type composition, and metabolic characteristics. Therefore, studying the regulatory mechanisms of muscle quality and meat production performance is crucial for livestock and poultry breeding. Currently, research in this field is beginning to employ multimodal data integration and analysis strategies. By combining in vivo imaging, tissue biopsy sampling, and molecular omics detection, these methods attempt to dynamically track the muscle fiber type transformation process in vivo and identify key regulatory targets. Existing techniques typically involve multiple in vivo examinations and data collections on the same batch of experimental animals to obtain temporal multimodal data for subsequent analysis.

[0003] However, in the process of integrating and analyzing multimodal data based on liveness detection, the inherent spatial positioning uncertainty of liveness sampling and the temporal asynchrony of data acquisition from different modalities make it difficult to achieve accurate alignment and fusion of heterogeneous data from the same experimental subject in the spatiotemporal dimensions. Existing data processing methods are unable to effectively eliminate the systematic errors introduced by liveness sampling site deviations and detection time point offsets, resulting in reduced consistency and correlation reliability among multimodal data, directly affecting the accuracy and reliability of subsequent myofiber specialization core target discovery. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a method and system for mining muscle fiber specialization core targets based on multimodal learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multimodal learning-based method for identifying core targets for muscle fiber specialization includes the following steps: S1. Obtain muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window; S2. Spatial registration processing is performed on the muscle tissue imaging data and muscle tissue molecular omics data to generate spatially aligned multimodal data. S3. Perform temporal alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; S4. Calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and determine whether the temporal correlation coefficient exceeds the preset correlation threshold. S5. If the preset relevant threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data, and the muscle tissue imaging data is fused according to the biological pathway activation intensity attention map to generate a fused feature vector. S6. Based on the fusion feature vector, the target recognition model is used to mine the core target of muscle fiber specialization and output a list of core targets.

[0006] Furthermore, acquire imaging and molecular omics data of muscle tissue from live geese collected within the same time window, including: Ultrasound scanning of live goose muscle tissue was used to obtain imaging data of the muscle tissue; The muscle tissue of the same live goose was sampled by puncture biopsy and molecular omics data of the muscle tissue was obtained by mass spectrometry analysis. Record the timestamps of ultrasound scans and biopsy sampling to ensure that the data are within the same time window.

[0007] Furthermore, spatial registration was performed on the muscle tissue imaging data and muscle tissue molecular omics data respectively to generate spatially aligned multimodal data, including: Anatomical landmarks were extracted from muscle tissue imaging data as spatial registration references. Determine the sampling location coordinates of muscle tissue molecular omics data in muscle tissue; Map the sampling location coordinates of muscle tissue molecular omics data to the anatomical landmark coordinate system of muscle tissue imaging data; Spatially aligned multimodal data is generated based on coordinate mapping relationships.

[0008] Furthermore, the spatially aligned multimodal data undergoes temporal alignment processing to generate spatiotemporally synchronized multimodal data, including: Obtain the timestamp information of muscle tissue imaging data and muscle tissue molecular omics data in spatially aligned multimodal data; Interpolation methods were used to align the time series data of muscle tissue imaging data and muscle tissue molecular omics data. Establish a correspondence between muscle tissue imaging data and muscle tissue molecular omics data on a unified time axis; Spatiotemporally synchronized multimodal data is generated based on the correspondence on a unified time axis.

[0009] Furthermore, the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data is calculated, and it is determined whether the temporal correlation coefficient exceeds a preset correlation threshold, including: Extracting time series of muscle tissue imaging data and time series of muscle tissue molecular omics data from spatiotemporally synchronized multimodal data; The Pearson correlation coefficient between the time series of muscle tissue imaging data and the time series of muscle tissue molecular omics data was calculated as the time series correlation coefficient. The calculated time-series correlation coefficient is compared with a preset correlation threshold. Based on the comparison results, determine whether the time-series correlation coefficient exceeds the preset correlation threshold.

[0010] Furthermore, if the threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data. Then, feature fusion is performed on the muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector, including: Gene set enrichment analysis was performed on muscle tissue molecular omics data to identify significantly activated biological pathways; Calculate the activation intensity value of each biological pathway based on gene expression levels; An attention map of biological pathway activation intensity is generated based on the activation intensity values ​​of each biological pathway. Weighted fusion processing of muscle tissue imaging data was performed using biological pathway activation intensity attention maps. The weighted and fused features are combined to generate a fused feature vector.

[0011] Furthermore, gene set enrichment analysis was performed on the molecular omics data of muscle tissue to identify significantly activated biological pathways, including: Molecular omics data of muscle tissue were analyzed using gene set enrichment analysis. Compare gene expression profiles with gene sets of known biological pathways; Calculate the enrichment score and statistical significance of each biological pathway; Biological pathways with enrichment scores exceeding the significance threshold were selected as significantly activated biological pathways.

[0012] Furthermore, the activation intensity values ​​of each biological pathway are calculated based on gene expression levels, including: Extract the expression levels of all genes contained in significantly activated biological pathways; Calculate the weighted average of the expression levels of all genes within this biological pathway; The activation intensity of each biological pathway was determined based on a weighted average. The activation intensity value is normalized to a preset range.

[0013] Furthermore, based on the fused feature vectors, a target identification model is used to mine core targets for myofibril specialization, outputting a list of core targets, including: The fused feature vectors are input into a pre-trained random forest classifier for feature importance evaluation; Candidate targets are ranked based on the feature importance assessment results; Candidate targets with importance scores higher than a set threshold are selected as core targets; The core targets are ranked from highest to lowest importance to generate a list of core targets.

[0014] On the other hand, the present invention provides a muscle fiber specialization core target mining system based on multimodal learning, comprising the following modules: The data acquisition module is used to acquire muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window. The registration processing module is used to perform spatial registration processing on muscle tissue imaging data and muscle tissue molecular omics data respectively, generating spatially aligned multimodal data; The timing processing module is used to perform timing alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; The correlation calculation module is used to calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and to determine whether the temporal correlation coefficient exceeds the preset correlation threshold. The feature fusion module is used to generate a biological pathway activation intensity attention map based on muscle tissue molecular omics data if the threshold is exceeded, and to perform feature fusion on muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector. The target mining module is used to mine core targets for muscle fiber specialization based on the fused feature vector and the target recognition model, and outputs a list of core targets.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By establishing a precise spatiotemporal alignment mechanism, the spatial registration and temporal synchronization problems of multimodal data in in vivo detection were effectively solved. Utilizing anatomical landmark coordinate mapping and a unified time axis construction, precise spatiotemporal matching of muscle tissue imaging data and molecular omics data was achieved, significantly improving the consistency and reliability of correlation among multimodal data. Thresholding based on temporal correlation coefficients ensured the quality of data used in subsequent analyses, providing a reliable data foundation for studying the dynamic changes of muscle fibers in vivo.

[0016] 2. An attention map of biological pathway activation intensity was introduced to guide multimodal feature fusion, fully utilizing the pathway functional information contained in molecular omics data. Significantly activated biological pathways were identified through gene set enrichment analysis, and attention weights were generated based on pathway activation intensity, achieving an organic fusion of imaging features and molecular functional features. This fusion approach effectively preserves the biological significance of in vivo detection data, improving the accuracy and interpretability of subsequent target discovery. Attached Figure Description

[0017] Figure 1 This is a flowchart of the muscle fiber specialization core target mining method based on multimodal learning according to the present invention; Figure 2 This is a schematic diagram of the structure of the muscle fiber specialization core target mining system based on multimodal learning of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 The present invention provides a method for mining core targets of muscle fiber specialization based on multimodal learning, which includes the following steps: S1. Obtain muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window; S2. Spatial registration processing is performed on the muscle tissue imaging data and muscle tissue molecular omics data to generate spatially aligned multimodal data. S3. Perform temporal alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; S4. Calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and determine whether the temporal correlation coefficient exceeds the preset correlation threshold. S5. If the preset relevant threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data, and the muscle tissue imaging data is fused according to the biological pathway activation intensity attention map to generate a fused feature vector. S6. Based on the fusion feature vector, the target recognition model is used to mine the core target of muscle fiber specialization and output a list of core targets.

[0020] S1. Obtain imaging and molecular omics data of muscle tissue from live geese collected within the same time window. Specific implementation includes: To acquire multimodal data of live goose muscle tissue, a specific muscle tissue region was first scanned using an ultrasound diagnostic instrument. Specifically, the area with the greatest thickness in the middle of the goose's pectoral muscle was selected as the scanning region. A linear probe was used to perform dual-plane scanning in both transverse and longitudinal sections at a center frequency of 15 MHz. The scanning depth was set to 3 cm, and the gain parameter was adjusted to 65 dB to obtain muscle tissue imaging data. This muscle tissue imaging data includes a sequence of grayscale images of muscle fiber arrangement, intermuscular septum distribution characteristics, and the contour of the epimysium. The image resolution is no less than 1920 × 1080 pixels, and the data is stored in DICOM medical image format.

[0021] Immediately after the ultrasound scan, a biopsy was performed on the same muscle tissue site from the same live goose. Using a 16G automated biopsy needle, precise positioning was achieved under real-time ultrasound guidance. The sampling point was selected at the geometric center of the ultrasound scan area, with a sampling depth controlled at 2.5 cm, yielding approximately 20 mg of muscle tissue. The obtained muscle tissue sample was immediately placed in a cryogenic environment of -80 degrees Celsius for preservation, and sample pretreatment was completed within 24 hours.

[0022] The pretreatment of muscle tissue samples included the following steps: After removing the samples from the low-temperature environment, they were washed in pre-cooled phosphate buffer solution. The muscle tissue was then homogenized using a tissue homogenizer at 4 degrees Celsius, and lysis buffer was added to extract total protein. After determining the protein concentration using the BCA method, 100 micrograms of the protein solution were subjected to enzymatic digestion using sequencing-grade trypsin at 37 degrees Celsius for 12 hours.

[0023] The enzymatically digested peptide samples were analyzed using liquid chromatography-mass spectrometry (LC-MS) with a Q Exactive series mass spectrometer. The primary mass spectrometry resolution was set to 70,000 MΩ, the secondary mass spectrometry resolution to 17,500 MΩ, and the collision energy to 27 eV. The mass spectrometry scan range was set to a mass-to-charge ratio of 350 to 1500, and the dynamic exclusion time was set to 30 seconds. Each sample was analyzed three times to obtain the final muscle tissue molecular omics data. This muscle tissue molecular omics data includes peptide sequence information, relative protein quantification data, and identification results of post-translational modification sites.

[0024] During ultrasound scanning and biopsy sampling, the timestamps of both procedures are recorded using a time synchronization system. This system consists of a central clock server and multiple distributed time recording terminals, with all terminals maintaining a time deviation of no more than 1 millisecond from the server. At the end of the ultrasound scan, the system automatically records the timestamp of the scan completion moment; during biopsy sampling, the timestamp of the sampling moment is recorded at the instant the biopsy needle is triggered.

[0025] The time window was determined as follows: the absolute time difference between the ultrasound scan completion timestamp and the biopsy sampling timestamp was calculated. When this time difference was less than or equal to 15 minutes, the two sets of data were considered to be within the same time window. This 15-minute time threshold was determined based on studies of the stability of metabolic activity in goose muscle tissue; within this time range, the physiological state and molecular composition of the muscle tissue did not change significantly. All timestamp data were stored in an international standard time format, including year, month, day, hour, minute, second, and millisecond information.

[0026] To ensure spatial consistency in data acquisition, the sampling area was marked on the goose's skin surface before ultrasound scanning and biopsy. A 3 cm × 3 cm square area was drawn on the surface of the goose's pectoral muscle using a medical marker, and this area was divided into nine equal 1 cm × 1 cm squares. Scanning and sampling were performed within this marked area. This spatial marking method ensures the accuracy of subsequent spatial registration processing and provides a reliable anatomical reference for multimodal data alignment.

[0027] Immediately after each data acquisition, the acquired muscle tissue imaging and molecular omics data undergo quality checks. Quality checks for muscle tissue imaging data include signal-to-noise ratio (SNR) assessment, contrast ratio testing, and resolution verification, requiring a SNR of at least 30 dB, a contrast ratio greater than 0.3, and a spatial resolution of 0.2 mm. Quality checks for muscle tissue molecular omics data include peptide identification quantity checks, protein coverage assessment, and quantitative repeatability verification, requiring at least 5000 identified peptides per sample, protein coverage of at least 40%, and a correlation coefficient between technical repeats of at least 0.95. Only data that passes these quality checks proceeds to subsequent processing steps.

[0028] S2. Spatial registration processing is performed on muscle tissue imaging data and muscle tissue molecular omics data respectively to generate spatially aligned multimodal data. Specific implementation includes: In the spatial registration process, anatomical landmarks are first extracted from the muscle tissue imaging data. A scale-invariant feature transform algorithm is used to identify stable anatomical features from the muscle tissue imaging data. These features include anatomical structures with significant geometric characteristics, such as myofiber bundle intersections, vascular bifurcation points, and intermuscular septal connections. Each feature point is described by a 128-dimensional feature vector, which contains location, orientation, and scale information. A contrast threshold of 0.03 and an edge threshold of 10 are set during feature point extraction to ensure sufficient discriminative power and stability of the extracted features. The extracted feature points are then screened, removing duplicate and low-quality features, retaining at least 50 evenly distributed anatomical landmarks as spatial registration benchmarks.

[0029] The sampling location coordinates for muscle tissue molecular omics data were determined using a three-dimensional reconstruction method. Based on the spatial positioning data from the biopsy, including puncture depth, puncture angle, and puncture point planar coordinates, a three-dimensional spatial model of the sampling location was established. The puncture depth was determined by the scale reading of the biopsy needle, accurate to 0.1 mm; the puncture angle was recorded by an angle sensor on the biopsy gun, including both elevation and azimuth parameters, with an accuracy of 0.5 degrees; the puncture point planar coordinates were taken from the coordinates of marker points on the ultrasound image. These parameters were used to calculate the three-dimensional coordinates of the sampling location within the muscle tissue. The origin was set as the center point of the ultrasound scanning area, with the X-axis along the muscle fiber direction, the Y-axis perpendicular to the muscle fiber direction, and the Z-axis along the depth direction.

[0030] The Prototype analysis method was used to map the sampling location coordinates of muscle tissue molecular omics data to the anatomical landmark coordinate system of muscle tissue imaging data. First, the correspondence between the two coordinate systems was established by calculating the rigid body transformation parameters between the two sets of coordinate points, including the rotation matrix and translation vector. The rotation matrix was solved using singular value decomposition, and the translation vector was calculated using the difference in centroid coordinates. During the mapping process, the iterative nearest-point algorithm was used to optimize the transformation parameters, with 100 iterations and a convergence threshold of 0.001 mm. For each sampling location coordinate, the calculated transformation matrix was applied to transform it to the anatomical landmark coordinate system, ensuring sub-millimeter accuracy in the coordinate mapping.

[0031] When generating spatially aligned multimodal data based on coordinate mapping, a unified spatial reference frame is first established. This frame is based on an anatomical landmark coordinate system, with the origin set as the geometric center of all anatomical landmarks, and the coordinate axes aligned with the anatomical direction of the muscle tissue. Muscle tissue imaging data is resampled into a unified spatial grid with a resolution of 0.2 mm × 0.2 mm × 0.2 mm, covering the entire region of interest. Muscle tissue molecular omics data is assigned to the corresponding spatial grid according to the mapped coordinate positions. Each grid cell contains both the imaging features and molecular omics measurements for that location.

[0032] Bilinear interpolation is used during spatial alignment to address data resampling issues caused by coordinate mapping. For muscle tissue imaging data, cubic spline interpolation is used to maintain image smoothness; for muscle tissue molecular omics data, nearest neighbor interpolation is used to maintain data specificity. Boundary handling strategies are implemented during interpolation, employing mirror filling for regions exceeding the original data range. The resulting spatially aligned multimodal data contains complete spatial location and value information, with each data point bearing precise spatial coordinate labels.

[0033] To verify the accuracy of spatial registration, marker distance error was used as the evaluation metric. The Euclidean distance between the mapped molecular omics data sampling locations and the nearest anatomical markers was calculated, requiring an average registration error of less than 1.5 mm and a maximum registration error of no more than 3 mm. Data points with registration errors exceeding the threshold underwent manual verification and re-registration to ensure that the spatial alignment quality of all data met the requirements for subsequent analysis. The registered multimodal data was stored in HDF5 format, including spatial coordinate information, data type identifiers, and data quality identifiers.

[0034] S3. Perform temporal alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data. Specific implementation includes: During the temporal alignment process, timestamp information is first extracted from the metadata of the spatially aligned multimodal data. The timestamp information includes three fields: acquisition start time, acquisition end time, and acquisition duration, with millisecond-level time precision. Timestamps for muscle tissue imaging data are read from the AcquisitionDateTime field of the DICOM file, while timestamps for muscle tissue molecular omics data are extracted from the RunStartTimeStamp field of the raw mass spectrometry file. All timestamps are uniformly converted to relative time values ​​calculated from a reference time point, typically chosen as the start time of the first data acquisition, with the time unit uniformly set to seconds.

[0035] Cubic spline interpolation was used to align the time series data of muscle tissue imaging and molecular omics data. First, a unified time axis was defined, ranging from the minimum to the maximum value of all timestamps, with a temporal resolution of 1 second. For the muscle tissue imaging data, the grayscale value at each spatial location was transformed into a time series curve over time. These curves were resampled, and the interpolated grayscale value was calculated at each unified time point. For the muscle tissue molecular omics data, the data points showing the expression level of each protein over time were fitted into a continuous curve, and then sampled at a unified time point. Natural boundary conditions were set during interpolation to ensure that the second derivative of the interpolated curve was zero at its endpoints.

[0036] When establishing the correspondence between muscle tissue imaging data and muscle tissue molecular omics data on a unified timeline, the principle of temporal nearest neighbor matching is adopted. For each time point on the unified timeline, the data point closest to that time point in the muscle tissue imaging data time series is found, and simultaneously, the data point closest to that time point in the muscle tissue molecular omics data time series is found. These two data points are then linked. When multiple data points are equidistant from the same unified time point, the data point with the smaller timestamp is selected as the corresponding point. After the correspondence is established, a mapping table is generated, recording the muscle tissue imaging data index and the muscle tissue molecular omics data index corresponding to each unified time point.

[0037] When generating spatiotemporally synchronized multimodal data based on correspondences on a unified time axis, a four-dimensional data structure is first created, consisting of three spatial dimensions and one temporal dimension. The spatial dimensions maintain the same resolution and coordinate system as the spatially aligned multimodal data, while the temporal dimension uses a unified time axis. Each four-dimensional grid point stores the grayscale values ​​of muscle tissue imaging data and the protein expression levels of muscle tissue molecular omics data. For grid points without corresponding original data, spatiotemporal interpolation is used for padding, employing bilinear interpolation spatially and linear interpolation temporally.

[0038] The spatiotemporally synchronized multimodal data employs a hierarchical storage structure. The bottom layer stores the raw acquired data, the middle layer stores the aligned data, and the top layer stores the complete spatiotemporal data after interpolation and padding. Each data layer contains a quality identifier matrix, recording the data's source and reliability. The quality identifier matrix values ​​range from 0 to 1, where 0 represents fully interpolated data, 1 represents raw acquired data, and intermediate values ​​represent partially interpolated data. Data storage utilizes the ZFP compression algorithm with a compression ratio of 4:1, reducing storage space requirements while maintaining data accuracy.

[0039] To verify the quality of temporal alignment, a temporal consistency index was calculated for the aligned data. For each spatial location, the cross-correlation coefficient between the time series of muscle tissue imaging data and the time series of muscle tissue molecular omics data was calculated, requiring an average cross-correlation coefficient greater than 0.85. Regions with low cross-correlation coefficients underwent manual review, and temporal alignment was re-performed if necessary. The final spatiotemporally synchronized multimodal data contains complete temporal evolution and spatial distribution information, which can be used for subsequent temporal correlation analysis and multimodal feature fusion.

[0040] S4. Calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and determine whether the temporal correlation coefficient exceeds a preset correlation threshold. Specific implementation includes: When calculating the temporal correlation coefficient, time series data were first extracted from spatiotemporally synchronized multimodal data. For each spatial location, grayscale sequences of muscle tissue imaging data and protein expression sequences of muscle tissue molecular omics data were extracted along the time dimension. The extracted time series were uniformly set to contain at least 30 time points with a time point interval of 1 second to ensure sufficient length for correlation analysis. During the extraction process, the data were standardized using the z-score standardization method to convert each time series into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of different data units.

[0041] To enhance robustness, a sliding window method was employed when calculating the Pearson correlation coefficient. The window size was set to 5 time points, with a sliding step of 1 time point. Within each window, the covariance and standard deviation of the time series data from muscle tissue imaging and molecular omics were calculated. The Pearson correlation coefficient was calculated using the standard formula: the covariance of the two time series divided by the product of their respective standard deviations. Outliers were handled during the calculation process; data points deviating from the mean by more than 3 standard deviations were removed using the 3σ principle, and missing values ​​were filled using linear interpolation. Each spatial location ultimately yielded a correlation coefficient value ranging from -1 to 1, representing the temporal correlation strength between the muscle tissue imaging and molecular omics data at that location.

[0042] The preset correlation threshold is determined based on the principle of statistical significance testing. By collecting a large amount of historical experimental data, the Pearson correlation coefficient distribution between the time series of muscle tissue imaging data and the time series of muscle tissue molecular omics data is calculated. A significance level of α=0.01 is used as the threshold determination standard; that is, when the correlation coefficient reaches this threshold, it indicates that the correlation between the two time series is statistically significant. Specifically, the correlation coefficient is converted into a t-statistic using a t-test, and then the corresponding correlation coefficient threshold is determined based on the degrees of freedom and the significance level. The degrees of freedom are calculated by subtracting 2 from the number of time points. For example, when there are 30 time points, the degrees of freedom are 28, and the corresponding correlation coefficient threshold is approximately 0.463. This threshold is dynamically adjusted according to the actual length of the time series to ensure the accuracy of the statistical test.

[0043] When comparing the calculated time-series correlation coefficient with a preset correlation threshold, a point-by-point comparison strategy is adopted. For each spatial location point, its Pearson correlation coefficient value is numerically compared with the preset correlation threshold. The comparison result is divided into three cases: when the correlation coefficient is greater than the preset correlation threshold, it indicates a positive correlation; when the correlation coefficient is less than the negative preset correlation threshold, it indicates a negative correlation; when the absolute value of the correlation coefficient is less than the preset correlation threshold, it indicates no significant correlation. A hysteresis interval is set during the comparison process to avoid frequent switching of judgment results near the threshold. The width of the hysteresis interval is set to 10% of the preset correlation threshold. For example, when the preset correlation threshold is 0.5, the hysteresis interval is 0.45 to 0.55.

[0044] Based on the comparison results, a binary mask matrix is ​​generated to determine whether the temporal correlation coefficient exceeds a preset correlation threshold. For each spatial location, if the absolute value of the temporal correlation coefficient exceeds the preset correlation threshold, the location is marked as 1, indicating that it has passed the quality check; otherwise, it is marked as 0, indicating that it has failed the quality check. Morphological processing is performed on the binary mask matrix, using 3×3 circular structuring elements for a dilation-erosion closing operation to fill small holes and smooth the boundaries. Finally, only connected regions with an area greater than 10 pixels are retained, and isolated noise points are removed to ensure that the regions that pass the check have sufficient spatial continuity and biological significance.

[0045] To verify the reliability of the time-series correlation calculation, a bootstrap method was used for stability assessment. Samples were randomly drawn with replacement from the original time series, and the Pearson correlation coefficient was repeatedly calculated 100 times. The mean and standard deviation of these correlation coefficients were then calculated. The standard deviation of the correlation coefficients was required to be less than 0.1, and the width of the 95% confidence interval was required to be less than 0.2, ensuring sufficient stability of the calculation results. For regions with insufficient stability, the time series length was increased or the sampling frequency was increased, and the calculation was repeated until the stability requirements were met. The final output time-series correlation coefficient matrix has the same spatial and temporal dimensions as the spatiotemporally synchronized multimodal data, providing quality assurance for subsequent feature fusion and target discovery.

[0046] S5. If the preset threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data, and feature fusion is performed on the muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector. Specific implementation includes: When generating attention maps of biological pathway activation intensity, gene set enrichment analysis was first performed on the muscle tissue molecular omics data. The GSEA algorithm was used to analyze the gene expression profiles in the muscle tissue molecular omics data, using standard pathway gene sets from the Molecular Signatures Database as a reference set, including pathway sets from databases such as KEGG, Reactome, and BioCarta. The minimum gene set size was set to 15 genes, and the maximum gene set size to 500 genes. The enrichment degree of each pathway was assessed by calculating cumulative enrichment scores and permutation tests. The number of permutations was set to 1000. Gene set enrichment analysis software was used to calculate the standardized enrichment score and false discovery rate for each pathway, with a false discovery rate threshold set to 0.05, to screen for significantly enriched biological pathways.

[0047] When calculating the activation intensity of each biological pathway based on gene expression levels, the expression levels of all genes in significantly activated biological pathways are first extracted. For each pathway, a weighted average method is used to calculate the activation intensity. The weights are determined based on the importance of the gene in the pathway, quantified by gene connectivity and centrality indicators. Genes with high connectivity are assigned larger weights, and genes with low connectivity are assigned smaller weights, with weight values ​​ranging from 0 to 1, and the sum of all weights being 1. The calculated weighted average is then logarithmically transformed and standardized to ensure comparability of activation intensity values ​​across pathways, ultimately normalizing the activation intensity values ​​to a range of 0 to 100.

[0048] When generating attention maps of biological pathway activation intensity based on the activation intensity values ​​of various biological pathways, a spatial mapping method is used to project pathway activation information onto the anatomical space of muscle tissue. First, a correspondence between gene expression and spatial location is established, and spatial weights are assigned according to the specificity of gene expression in muscle tissue. For each spatial location, the weighted sum of the influence of each pathway is calculated, with the weight determined by the spatial correlation between that location and the expression patterns of pathway-related genes. Gaussian smoothing is used when generating the attention map, with a smoothing kernel size of 3×3 pixels and a standard deviation of 0.5, ensuring the spatial continuity of the attention map. The final generated biological pathway activation intensity attention map has the same spatial dimension and resolution as the muscle tissue imaging data.

[0049] When using biological pathway activation intensity attention maps to perform weighted fusion processing on muscle tissue imaging data, a pixel-by-pixel weighted fusion strategy is employed. For each pixel in the muscle tissue imaging data, the fusion weight is determined based on its corresponding value in the biological pathway activation intensity attention map. The weight calculation uses the sigmoid function to map the attention value between 0 and 1; a higher attention value results in a greater weight. During the fusion process, multiple feature channels of the muscle tissue imaging data are weighted separately, maintaining the feature independence of each channel. The weighted feature maps are then summed across the channels to generate a preliminary fused feature representation.

[0050] When generating the fused feature vector by combining the weighted and fused features, a multi-scale feature aggregation method is employed. First, the fused features are extracted at different spatial scales, including the original scale, a 1 / 2 downsampled scale, and a 1 / 4 downsampled scale. Features at each scale are then processed by global average pooling to convert them into fixed-length feature vectors. Next, the feature vectors from multiple scales are concatenated to form the final multi-scale fused feature vector. The dimension of the feature vector is determined by the number of feature channels and the number of scales. For example, when the original feature channels are 64 and 3 scales are used, the final fused feature vector has a dimension of 64 × 3 = 192. The feature vector undergoes L2 normalization to ensure that the value of each feature is between 0 and 1, facilitating subsequent target recognition model processing.

[0051] To verify the effectiveness of feature fusion, the discriminative performance metrics of the features before and after fusion were calculated. Linear discriminant analysis was used to calculate the ratio of inter-class divergence to intra-class divergence of the features, requiring that the ratio of the fused features be at least 30% higher than that before fusion. Simultaneously, the signal-to-noise ratio (SNR) of the features was calculated, requiring that the SNR of the fused features be at least 20% higher than that of the original features. For regions where the performance improvement was insufficient, the generation parameters of the attention map or the fusion weight calculation method were adjusted, and feature fusion was repeated until the performance requirements were met. The final generated fused feature vector contains rich multimodal information, providing high-quality feature input for subsequent core target mining.

[0052] S6. Based on the fused feature vector, a target identification model is used to mine core targets for muscle fiber specialization, and a list of core targets is output. Specific implementation includes: When mining core targets for myofiber specialization based on fused feature vectors, the fused feature vectors are first input into a pre-trained random forest classifier. This random forest classifier contains 100 decision trees, each with a maximum depth of 10 layers and a minimum of 5 leaf node samples. The classifier's training data comes from a multimodal dataset of muscle tissue obtained from historical experiments, where positive samples are known myofiber specialization-related targets, and negative samples are randomly selected non-specifically expressed genes. Out-of-bag error estimation is used for model validation during training, with an out-of-bag error rate requirement of less than 0.15. Feature importance is evaluated using the average Gini impurity reduction method, where the importance score of each feature is obtained by calculating the average reduction in impurity across all decision trees.

[0053] When ranking candidate targets based on feature importance assessment results, a descending order is used. Importance scores are standardized to a relative score between 0 and 100, and the score calculation comprehensively considers the number of splits in the decision tree and the information gain resulting from those splits. A stability check is implemented during the ranking process, with the importance score repeatedly calculated 100 times using bootstrap sampling. The coefficient of variation for the top 10 targets is required to be less than 0.2. For targets with insufficient stability, the number of decision trees is increased to 200 before recalculation to ensure the reliability of the ranking results.

[0054] When selecting candidate targets with importance scores higher than a set threshold as core targets, the threshold is determined using statistical distribution analysis. First, the mean and standard deviation of the importance scores for all candidate targets are calculated, and the threshold is set as the mean plus twice the standard deviation. Simultaneously, biological significance is considered, requiring the number of core targets to be no less than 5 and no more than 50, ensuring the results are both representative and not overly broad. Targets near the threshold undergo manual review, with their known biological functions and expression patterns used for final determination.

[0055] When generating the core target list by ranking targets from highest to lowest importance score, a standardized data format is used. The list includes target identifiers, importance scores, standard error of the score, and functional annotations. The standard error of the score is calculated using bootstrap sampling to reflect the stability of the importance score. Functional annotation information is obtained from authoritative databases and includes gene names, protein function descriptions, and known disease associations. The list output format uses comma-separated value files and generates a visual bar chart showing the importance score distribution of the top 20 core targets.

[0056] To verify the reliability of the core target list, leave-one-out cross-validation was used to evaluate the model's stability. After each sample was removed, the random forest classifier was retrained and a new core target list was generated. The consistency index between the lists was calculated. The Jaccard similarity coefficient of the core target lists was required to be no less than 0.8 to ensure that the mining results were not affected by individual samples. Simultaneously, functional enrichment analysis was performed to verify the enrichment degree of core targets in myofibril specialization-related pathways, with a false discovery rate of less than 0.01. The final output core target list included complete scoring information and validation results.

[0057] Example 2: Figure 2 A schematic diagram of the multimodal learning-based muscle fiber specialization core target mining system of the present invention is provided. The multimodal learning-based muscle fiber specialization core target mining system includes the following modules: The data acquisition module is used to acquire muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window. The registration processing module is used to perform spatial registration processing on muscle tissue imaging data and muscle tissue molecular omics data respectively, generating spatially aligned multimodal data; The timing processing module is used to perform timing alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; The correlation calculation module is used to calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and to determine whether the temporal correlation coefficient exceeds the preset correlation threshold. The feature fusion module is used to generate a biological pathway activation intensity attention map based on muscle tissue molecular omics data if the threshold is exceeded, and to perform feature fusion on muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector. The target mining module is used to mine core targets for muscle fiber specialization based on the fused feature vector and the target recognition model, and outputs a list of core targets.

[0058] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0059] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0060] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0062] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations 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. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0064] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for discovering core target points for muscle fiber specialization based on multimodal learning, characterized in that, Includes the following steps: S1. Obtain muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window; S2. Spatial registration processing is performed on the muscle tissue imaging data and muscle tissue molecular omics data to generate spatially aligned multimodal data. S3. Perform temporal alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; S4. Calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and determine whether the temporal correlation coefficient exceeds the preset correlation threshold. S5. If the preset relevant threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data, and the muscle tissue imaging data is fused according to the biological pathway activation intensity attention map to generate a fused feature vector. S6. Based on the fusion feature vector, the target recognition model is used to mine the core target of muscle fiber specialization and output a list of core targets.

2. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, Acquire muscle tissue imaging and molecular omics data from live geese collected within the same time window, including: Ultrasound scanning of live goose muscle tissue was used to obtain imaging data of the muscle tissue; The muscle tissue of the same live goose was sampled by puncture biopsy and molecular omics data of the muscle tissue was obtained by mass spectrometry analysis. Record the timestamps of ultrasound scans and biopsy sampling to ensure that the data are within the same time window.

3. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, Spatial registration was performed on muscle tissue imaging data and muscle tissue molecular omics data respectively to generate spatially aligned multimodal data, including: Anatomical landmarks were extracted from muscle tissue imaging data as spatial registration references. Determine the sampling location coordinates of muscle tissue molecular omics data in muscle tissue; Map the sampling location coordinates of muscle tissue molecular omics data to the anatomical landmark coordinate system of muscle tissue imaging data; Spatially aligned multimodal data is generated based on coordinate mapping relationships.

4. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, Temporal alignment is performed on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data, including: Obtain the timestamp information of muscle tissue imaging data and muscle tissue molecular omics data in spatially aligned multimodal data; Interpolation methods were used to align the time series data of muscle tissue imaging data and muscle tissue molecular omics data. Establish a correspondence between muscle tissue imaging data and muscle tissue molecular omics data on a unified time axis; Spatiotemporally synchronized multimodal data is generated based on the correspondence on a unified time axis.

5. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, Calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and determine whether the temporal correlation coefficient exceeds a preset correlation threshold, including: Extracting time series of muscle tissue imaging data and time series of muscle tissue molecular omics data from spatiotemporally synchronized multimodal data; The Pearson correlation coefficient between the time series of muscle tissue imaging data and the time series of muscle tissue molecular omics data was calculated as the time series correlation coefficient. The calculated time-series correlation coefficient is compared with a preset correlation threshold. Based on the comparison results, determine whether the time-series correlation coefficient exceeds the preset correlation threshold.

6. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, If the threshold is exceeded, a biological pathway activation intensity attention map is generated based on the muscle tissue molecular omics data. Then, feature fusion is performed on the muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector, including: Gene set enrichment analysis was performed on muscle tissue molecular omics data to identify significantly activated biological pathways; Calculate the activation intensity value of each biological pathway based on gene expression levels; An attention map of biological pathway activation intensity is generated based on the activation intensity values ​​of each biological pathway. Weighted fusion processing of muscle tissue imaging data was performed using biological pathway activation intensity attention maps. The weighted and fused features are combined to generate a fused feature vector.

7. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 6, characterized in that, Gene set enrichment analysis was performed on muscle tissue molecular omics data to identify significantly activated biological pathways, including: Molecular omics data of muscle tissue were analyzed using gene set enrichment analysis. Compare gene expression profiles with gene sets of known biological pathways; Calculate the enrichment score and statistical significance of each biological pathway; Biological pathways with enrichment scores exceeding the significance threshold were selected as significantly activated biological pathways.

8. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 6, characterized in that, The activation intensity values ​​of each biological pathway were calculated based on gene expression levels, including: Extract the expression levels of all genes contained in significantly activated biological pathways; Calculate the weighted average of the expression levels of all genes within this biological pathway; The activation intensity of each biological pathway was determined based on a weighted average. The activation intensity value is normalized to a preset range.

9. The method for discovering core target points for muscle fiber specialization based on multimodal learning according to claim 1, characterized in that, Based on the fusion feature vector, a target identification model is used to mine core targets for myofibril specialization, and a list of core targets is output, including: The fused feature vectors are input into a pre-trained random forest classifier for feature importance evaluation; Candidate targets are ranked based on the feature importance assessment results; Candidate targets with importance scores higher than a set threshold are selected as core targets; The core targets are ranked from highest to lowest importance to generate a list of core targets.

10. A multimodal learning-based muscle fiber specialization core target mining system, used to implement the multimodal learning-based muscle fiber specialization core target mining method according to any one of claims 1-9, characterized in that, Includes the following modules: The data acquisition module is used to acquire muscle tissue imaging data and muscle tissue molecular omics data collected from live geese within the same time window. The registration processing module is used to perform spatial registration processing on muscle tissue imaging data and muscle tissue molecular omics data respectively, generating spatially aligned multimodal data; The timing processing module is used to perform timing alignment processing on spatially aligned multimodal data to generate spatiotemporally synchronized multimodal data; The correlation calculation module is used to calculate the temporal correlation coefficient between muscle tissue imaging data and muscle tissue molecular omics data in spatiotemporally synchronized multimodal data, and to determine whether the temporal correlation coefficient exceeds the preset correlation threshold. The feature fusion module is used to generate a biological pathway activation intensity attention map based on muscle tissue molecular omics data if the threshold is exceeded, and to perform feature fusion on muscle tissue imaging data based on the biological pathway activation intensity attention map to generate a fused feature vector. The target mining module is used to mine core targets for muscle fiber specialization based on the fused feature vector and the target recognition model, and outputs a list of core targets.

Images