Genome variation cold and hot spot region prediction method and device

Abstract

The invention discloses a genome variation cold and hot spot region prediction method and device, and the method comprises the steps: S1, carrying out the slicing of a target genome region according to a preset sliding window length, and constructing a multi-modal input tensor corresponding to a window; s2, inputting the multi-modal input tensor into a pre-trained deep learning prediction model, and outputting a cold and hot spot prediction score of each site in the window through a full connection layer; and S3, according to the cold and hot spot prediction scores, identifying a variation cold spot region and a variation hot spot region in the target genome region. According to the technical scheme provided by the invention, the dependence on the existing variation data density is eliminated, and non-blind area coverage in the whole exon group range is realized; meanwhile, the structured output based on the preset transcript coordinates can directly support clinical variation interpretation, a quantitative basis is provided for PM1 and cold spot evidence, and the proportion of unclear significance variation is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information, and in particular to a method and device for predicting genomic variation hot and cold spot regions. BACKGROUND

[0002] With the rapid development of high-throughput sequencing technology, human genetic variation data is growing exponentially. At present, the guidelines jointly issued by the American Society of Medical Genetics and Genomics and the Society of Molecular Pathologists are the general framework for genetic variation interpretation, and the PM1 standard clearly lists the variation hot spot regions or key functional domains as the medium strength evidence for determining pathogenicity. However, in clinical practice, due to the lack of a comprehensive and systematic database of variation hot and cold spot regions, the application of the PM1 standard is greatly limited; at the same time, the cold spot regions rich in benign variations have not been effectively included in the evidence system so far, resulting in a large number of variations being classified as variations of unknown significance due to the lack of regional evidence, which seriously hinders the accuracy and efficiency of genetic variation classification.

[0003] Although there are tools such as HCSeeker in the prior art that attempt to identify hot and cold spot regions, they have technical limitations in actual application. Statistical modeling based on the number of reported variations in public databases (such as kernel density estimation) is highly dependent on existing variation accumulation. The existing technology can only identify a few gene regions with sufficient variation reports, and there is a huge prediction blind area for regions lacking historical data, which cannot achieve full exome coverage. In addition, the existing method only uses the number of variations, without integrating the deep information contained in the DNA sequence itself, as well as the multi-dimensional biological prior characteristics such as evolutionary conservation and functional tolerance, resulting in insufficient feature extraction capability in low variation density regions, poor biological rationality and robustness of the prediction results, and difficulty in meeting the clinical demand for high-precision, full-coverage variation interpretation support tools. How to properly solve the above problems has become an urgent topic in the industry. SUMMARY

[0004] The present application provides a method and device for predicting genomic variation hot and cold spot regions, which can accurately capture long-range sequence dependencies and achieve full exome blind area prediction, providing effective quantitative evidence for clinical variation interpretation and improving diagnostic efficiency.

[0005] To solve the above technical problems, the technical solution adopted by the present application is as follows: A method for predicting genomic variation hot and cold spot regions is provided, comprising: S1, slice the target genomic region according to a preset sliding window length, extract a DNA sequence vector corresponding to each window, and match a multidimensional biological prior feature vector for each site in the window, splice the DNA sequence vector and the biological prior feature vector in the feature dimension to construct a multi-modal input tensor corresponding to the window; S2, input the multi-modal input tensor into a pre-trained deep learning prediction model, extract local features of the sequence through a one-dimensional convolutional neural network layer in the model, capture long-distance dependence of the sequence through a self-attention mechanism layer, and output a cold-hot spot prediction score of each site in the window through a fully connected layer; S3, according to the cold-hot spot prediction score, filter out a continuous abnormal score interval through a preset cold-hot spot determination threshold, and map the abnormal score interval to a preset transcript coordinate, so as to identify a variant cold spot region and a variant hot spot region in the target genomic region.

[0006] Further, the training data of the pre-trained deep learning prediction model is also constructed in S2, including: S21, collect genomic region data from at least two sources, wherein the genomic region data at least includes an identified cold-hot spot region, an expert-planned region applicable to PM1 evidence, and key functional domain information; S22, establish a linear mapping rule to convert the qualitative information of the genomic region data into continuous numerical labels, linearly map a pathogenic ratio or evidence strength of the hot spot region to a first high score interval, and linearly map a benign ratio of the cold spot region to a second low score interval; S23, compose a training sample set by the converted continuous numerical labels, corresponding DNA sequences and prior features, and supervise the training of the deep learning prediction model through a mean square error loss function.

[0007] Further, the biological prior feature vector in S1 includes a position score feature, including: S11, the position score feature obtains missense variant data from a variant database, and gives an initial score according to a confidence level and a classification standard, wherein a pathogenic variant is given a positive initial score, and a benign variant is given a negative initial score; S12, a random forest regression model is trained by taking a chromosome number and a genomic coordinate of the missense variant data as input and taking the initial score as output target; S13, the trained random forest regression model is used to predict each potential missense variant site in the target genomic region, and a continuous position score is calculated.

[0008] Further, in S2, the multi-modal input tensor is processed by the deep learning prediction model, including: S24, receiving the multi-modal input tensor, performing feature transformation by a one-dimensional convolution block, and processing via a batch normalization layer and an activation function layer to generate an intermediate feature map containing local patterns; S25, inputting the intermediate feature map into a stacked Transformer encoder block, calculating the correlation weight between different sites in each encoder block by a multi-head self-attention mechanism, combining position encoding information and Dropout mechanism, and outputting a high-level feature representation fused with global dependency; S26, passing the high-level feature representation into an output layer, performing convolution operation and fully connected mapping to generate a one-dimensional score vector consistent with the input window length.

[0009] Further, in S3, the mutation cold spot region and the mutation hot spot region in the target genomic region are identified, including: S31, scanning the cold-hot spot prediction score, determining a region corresponding to nucleotide sites of consecutive amino acids not less than 2 and having a score greater than or equal to a first threshold value as a hot spot region, and determining a region corresponding to nucleotide sites of consecutive amino acids not less than 2 and having a score less than or equal to a second threshold value as a cold spot region; S32, obtaining an annotation file of a standard human reference genome and MANE database recommended transcript information, and converting nucleotide coordinates of the hot spot region and the cold spot region into corresponding transcript amino acid coordinates; S33, merging regions adjacent to each other and meeting the same threshold condition on the amino acid coordinates, and outputting a structured list containing chromosome, transcript number, start and end position, and region type.

[0010] Further, after S3, including: S4, grading the evidence strength, including; S41, counting the maximum and minimum values of all site scores in the identified cold spot region or hot spot region, and dividing the score value domain into three equal length score intervals based on the maximum and minimum values; S42, corresponding different evidence strength levels to the three equal length score intervals respectively, for the hot spot region, the highest score value interval is defined as the strong confidence level, the middle score value interval is defined as the moderate confidence level, and the lowest score value interval is defined as the supportive confidence level, for the cold spot region, the highest score value interval is defined as the supportive confidence level, the middle score value interval is defined as the moderate confidence level, and the lowest score value interval is defined as the strong confidence level; S43, according to the specific cold-hot spot prediction score of the to-be-tested variation site falling into an interval, assigning a corresponding PM1 or cold spot evidence intensity label to the variation site.

[0011] The embodiment also provides a genomic variation cold-hot spot region prediction device, comprising a processor and a computer readable storage medium connected to each other, and the computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize the steps of the genomic variation cold-hot spot region prediction method.

[0012] The embodiment also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the genomic variation cold-hot spot region prediction method.

[0013] The embodiment also provides a computer program product, comprising a computer program, and the computer program is executed by a processor to realize the steps of the genomic variation cold-hot spot region prediction method.

[0014] Compared with the prior art, the embodiment has the following advantages: The application can capture local patterns and long-range dependencies of the genome by fusing DNA sequences and multi-dimensional biological prior features and combining CNN and self-attention mechanisms, improves prediction accuracy and biological rationality, and is free from dependence on existing variation data density, realizes no-blind-area coverage in the whole exome range, and at the same time, structured output based on preset transcript coordinates can directly support clinical variation interpretation, provides quantitative basis for PM1 and cold point evidence, and effectively reduces the proportion of variants of uncertain significance.

[0015] Other features and advantages of the present application will be further described in the following description, and some will become apparent from the description, or will be learned through implementation of the present application. The objectives and other advantages of the present application can be achieved and obtained through the structures specifically pointed out in the written description and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a genomic variation cold-hot spot region prediction method provided for an embodiment of the present application is shown in the figure; Figure 2 A flowchart of another genomic variation cold-hot spot region prediction method provided for an embodiment of the present application is shown in the figure; Figure 3 A flowchart of another genomic variation cold-hot spot region prediction method provided for an embodiment of the present application is shown in the figure; Figure 4 A flowchart of another genomic variation cold-hot spot region prediction method provided for an embodiment of the present application is shown in the figure; Figure 5 A flowchart of another genomic variation cold-hot spot region prediction method provided for an embodiment of the present application; Figure 6 A flowchart of another genomic variation cold-hot spot region prediction method provided for an embodiment of the present application; Figure 7 A structural diagram of a genomic variation cold-hot spot region prediction device provided for an embodiment of the present application; Figure 8 A structural diagram of an electronic device provided for an embodiment of the present application; Figure 9 A schematic diagram of a cold-hot spot score assignment and linear mapping rule provided for an embodiment of the present application; Figure 10 A flowchart of position score calculation provided for an embodiment of the present application; Figure 11 An architecture schematic diagram of a deep learning prediction model provided for an embodiment of the present application; Figure 12 A comparison diagram of different region variation pathogenicity prediction score distribution provided for an embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present application is not limited thereby.

[0018] The technical solution adopted in the present embodiment is as shown in Figure 1 , comprising the following steps: S1, slice processing is performed on a target genomic region according to a preset sliding window length, a DNA sequence vector corresponding to each window is extracted, and a multi-dimensional biological prior feature vector is matched for each site in the window, the DNA sequence vector and the biological prior feature vector are spliced in the feature dimension, and a multi-modal input tensor corresponding to the window is constructed; S2, the multi-modal input tensor is input into a pre-trained deep learning prediction model, local features of the sequence are extracted through a one-dimensional convolutional neural network layer in the model, long-distance dependence relationships of the sequence are captured through a self-attention mechanism layer, and cold-hot spot prediction scores of each site in the window are output through a fully connected layer; S3, according to the cold-hot spot prediction scores, a continuous abnormal score interval is screened out through a preset cold-hot spot determination threshold, and the abnormal score interval is mapped to a preset transcript coordinate, so as to identify variation cold spot regions and variation hot spot regions in the target genomic region.

[0019] The execution of step S1 realizes the preprocessing of data and the construction of multi-modal input tensor. Specifically, taking the standard human reference genome (GRCh38 / hg38) as the benchmark, the exon regions of all protein-coding genes are extracted as the target genomic region. In order to adapt to the input requirements of the deep learning prediction model, a sliding window strategy is adopted in this embodiment to slice the target region, and the preset sliding window length is 100 bp (base pairs). For each 100 bp window divided, the original DNA sequence is extracted from the reference genome FASTA file, and One-hot encoding is performed on it (covering A, C, G, T and unknown base N), generating a sequence matrix with a dimension of 100x5. At the same time, for each nucleotide position in the window, a multi-dimensional biological prior feature is matched through a pre-constructed position score calculation model and an Annovar annotation tool. This biological prior feature not only contains information based on the DNA sequence itself, but also integrates evolutionary and functional genomics knowledge, including the average position score of three possible missense changes at the nucleotide position, the conservation score (such as PhastCons100way), the tolerance score (such as Jarvis_score) and the gene intolerance score (such as LOEUF), etc., totaling 24-dimensional features. The sequence matrix of 100x5 and the prior feature matrix of 100x24 are spliced in the feature dimension to construct a complete multi-modal input tensor with a dimension of 100x29 corresponding to the sliding window, which comprehensively represents the integrated information from micro-sequence to macro-biological significance.

[0020] After the construction of the input data is completed, step S2 is performed to perform feature extraction and score generation by the deep learning prediction model. In this embodiment, the pre-trained deep learning prediction model (named HCSeer2) adopts a hybrid architecture design. The 100x29 multi-modal input tensor described above is batched into the input layer of the model. The data passes through a convolutional block composed of a one-dimensional convolutional neural network (1D-CNN), which is mainly used to extract local feature patterns (such as specific sequence motifs) in the sequence. In order to ensure the stability of the data propagation in the network and introduce nonlinear features, a batch normalization (Batch Normalization) layer and a ReLU activation function layer are immediately connected after the convolution operation. The data extracted by the local features enters the core component of the model, the Transformer layer. The Transformer layer contains stacked Transformer encoder blocks, which use a multi-head self-attention mechanism (Multi-head Self-Attention, for example, 8 attention heads are set). The self-attention mechanism can dynamically calculate the correlation weights between the 100 base positions in the window, effectively capturing the long-range dependencies (Global dependencies) within the sequence, overcoming the problem of limited receptive field of traditional convolutional networks. After processing by the feedforward neural network and the Dropout layer (dropout rate is set to 0.1 to prevent overfitting), the high-dimensional features are sent to the fully connected layer for dimension reduction and mapping, which can output a 100-dimensional vector. Each value in the vector is a continuous floating-point number between 0 and 1, representing the probability prediction score of the corresponding nucleotide site in the input sequence becoming a mutation hotspot (the closer the score is to 1) or a mutation cold spot (the closer the score is to 0).

[0021] Step S3 is performed to preliminarily screen and determine the hot and cold spot regions according to the prediction score. Since the deep learning prediction model outputs continuous scores for each base site in the whole exome, in order to convert it into clinically usable structured data, this embodiment sets threshold determination rules. Specifically, the system scans the continuous score map, and when it detects that the corresponding cold and hot spot prediction scores of consecutive nucleotide sites are all greater than or equal to a first threshold (preferably 0.66 in this embodiment), and the length covered by these cold and hot spots is not less than 2 consecutive amino acids, the region is determined as a potential mutation hotspot region; on the contrary, when it detects that the corresponding cold and hot spot prediction scores of consecutive nucleotide sites are all less than or equal to a second threshold (preferably 0.33 in this embodiment), and also meet the length requirement of not less than 2 consecutive amino acids, the region is determined as a potential mutation cold spot region. The strategy of combining double thresholds with length constraints can effectively filter out random noise and ensure that the identified regions have high biological confidence.

[0022] After the initial determination based on nucleotide coordinates is completed, the latter half of step S3, i.e., coordinate mapping and result structuring, is performed. In order to enable the prediction results to be directly interfaced with clinical variant interpretation standards (such as the ACMG guidelines), a standard transcript is introduced, which is specifically the MANE (Matched Annotation from NCBI and EMBL-EBI) database recommended transcript. According to the gene annotation file, the identified hot and cold spot regions based on reference genome nucleotide coordinates are accurately mapped to the corresponding amino acid coordinates of the MANE transcript. In this process, a merging operation is also performed to fuse regions that are physically adjacent and meet the same threshold conditions to form a list of non-overlapping hot and cold spot regions. In one embodiment, 28,907 variant hot spot regions and 159,970 variant cold spot regions are successfully identified within the whole exome range.

[0023] The present embodiment also includes evidence strength grading and output of the identified regions. The structured data file output by the system details the chromosome number, transcript ID, amino acid start and end position, region type (hot spot / cold spot), and average cold-hot score in the region for each predicted region. Based on the distribution of these average cold-hot scores, the evidence strength can be further refined. For example, for hot spot regions, they are divided into PM1_Supporting (score in the lowest interval), PM1_Moderate (score in the middle interval), and PM1_Strong (score in the highest interval) according to the score interval. For example, for cold spot regions, they are divided into BP (score in the highest interval), BM hot spot (score in the middle interval), and BS (score in the lowest interval) according to the score interval. Graded output enables clinical personnel to directly call evidence of corresponding strength according to the specific score of the variant site when facing variants of unknown significance, thereby improving the accuracy and efficiency of variant classification and achieving the technical effect of transitioning from qualitative judgment to quantitative analysis.

[0024] A multi-scale human genome exon region variant cold and hot spot prediction deep learning framework. It overcomes the limitations of existing technologies through an innovative model architecture and multi-modal data fusion. The following will be described in detail from two aspects of structural composition and working process.

[0025] The overall structure of the present application can be divided into a data processing module, a feature engineering module, a core prediction model, a post-processing and output module, and a verification and analysis module.

[0026] The function of the data processing module is to be responsible for the collection, cleaning and labeling of training data. The data sources of the data processing module include: 1) The cold and hot spot regions identified by the HCSeeker tool.

[0027] 2) Areas in ClinGen VCEPs curated by disease experts that can apply PM1 evidence.

[0028] 3) Key functional domain information provided by AutoPVS1 tool.

[0029] wherein the cold-hot score can convert the qualitative information of the region into continuous numerical values that the model can learn. As shown in FIG. 6A, for the hot region of HCSeeker, the pathogenic odds ratio provided by it is linearly mapped to the interval [0.66, 1] as the score of each site in the region; for the cold region of HCSeeker, the benign odds ratio is linearly mapped to the interval [0, 0.33]; for the key functional domain of AutoPVS1, the proportion of pathogenic variants in the region is linearly mapped to the interval [0, 1]; and for the PM1 region of ClinGen, the PM1 evidence strength is linearly mapped to the interval [0.66, 1]. Figure 9

[0030] The function of the feature engineering module is to input features for model construction, including sequence information and multi-dimensional prior feature information.

[0031] Input sequence processing includes: 1) Length determination: The optimal input sequence length is determined to be 100 bp through comparative experiments (50 bp, 100 bp, 200 bp).

[0032] 2) Sequence encoding: One-hot encoding is used to convert DNA sequences (A, T, C, G) into 5-dimensional vectors (e.g., A = [1, 0, 0, 0, 0]), and [0, 0, 0, 0, 1] is used to represent unknown or filled bases 'N'.

[0033] 3) Data augmentation: Techniques such as window slicing, reverse complementation, and flipping are used to process input sequences to increase the amount of training data and improve the generalization ability and robustness of the model.

[0034] Prior feature extraction and fusion integrates 24-dimensional biological prior knowledge in four categories, which together with sequence information forms the input of the model (N x 100 x 29-dimensional tensor, where N is the number of samples).

[0035] Feature knowledge includes: 1) Position score: calculated by a random forest model, which takes the chromosome and coordinates of the variant on the genome as input and outputs scores based on ClinVar credibility levels and classifications. The score is used to reveal key functional regions where variants are clustered. As shown in FIG. 6B, the position score is calculated for each variant in the training set, and the score of the variant is assigned to the corresponding position in the input sequence. Figure 10 ​As shown, the position score is calculated as follows: first, obtain missense variant data from the ClinVar database, and assign the corresponding initial score according to its credibility level and classification standard. Specifically, for pathogenic variants rated as 2 stars and above, an initial score of 1 is assigned; and for benign variants of the same level, an initial score of -1 is assigned. For pathogenic variants rated as 1 star, an initial score of 0.5 is assigned; for benign variants rated as 1 star, an initial score of -0.5 is assigned. All variants used for training can be divided into four types, namely pathogenic variants, possibly pathogenic variants, benign variants and possibly benign variants. Subsequently, using the chromosome number and genomic coordinates as input parameters, a random forest model is applied to calculate a continuous position score for each potential missense variant site in the human exome region.

[0036] 2) Variant-level features (14 types): including variant conservation scores (such as PhastCons, PhyloP) and variant tolerance scores (such as jarvis_score, CCR_score), which reflect the conservation of the variant in evolution and the potential impact on gene function, respectively.

[0037] 3) Gene-level features (10 types): mainly gene intolerance scores (such as syn_oe, syn_z_score), which reflect the tolerance of the entire gene to the variant.

[0038] Dealing with missing values: for cases where the sequence length is less than 100 bp, 'N' is used for padding, and the corresponding prior feature score is filled with a neutral value (such as 0.5).

[0039] The function of the core prediction model is to learn from the input composite features and predict the cold-hot spot score of each site. Its architecture is shown in Figure 11 .

[0040] 1) Input layer: receives the fused 29-dimensional features (5-dimensional sequence + 24-dimensional prior features). First, a one-dimensional convolutional block is used to extract local features (such as motifs) from the input sequence. Then, a batch normalization layer is used to speed up training and improve stability, and finally a ReLU activation function is used to introduce nonlinearity.

[0041] 2) Transformer layer: used to capture long-range dependencies and global features in the input data. This layer consists of multiple stacked Transformer encoder blocks.

[0042] 3) Self-attention mechanism: 8 attention heads are set to allow the model to simultaneously focus on information at different positions in the sequence and dynamically calculate the correlation between different sites.

[0043] 4) Feed-forward neural network: further transformation on attention output.

[0044] 5) Dropout mechanism: dropout (dropout rate = 0.1) is applied in both positional encoding and each Transformer layer, which effectively prevents overfitting and improves the generalization ability of the model.

[0045] 6) Output layer: convert the high-level features extracted by the Transformer layer into the final prediction result. It contains another one-dimensional convolutional block to further extract features, a fully connected layer to map the features to the output of each site, and again uses the Dropout layer (rate = 0.1) and the ReLU activation function, and finally outputs a cold-hot spot score matrix with a shape of (N x 100).

[0046] The function of the post-processing and output module is to interpret and apply the original scores output by the model.

[0047] Cold-hot spot region determination refers to the cold-hot spot score (between 0 and 1) output by the model for each site. The amino acid interval with a continuous cold-hot spot score >= 0.66 is defined as a hot spot region, and the amino acid interval with a continuous score <= 0.33 is defined as a cold spot region. All coordinates are positioned to amino acid coordinates based on the MANE database recommended transcript.

[0048] The function of the verification and analysis module is to reasonably evaluate the prediction results of the model. Its core is to verify the effectiveness of the model output by comparing the intrinsic pathogenic potential of variants in different regions. Extract all missense variants from the cold spot region, hot spot region, and neither cold spot nor hot spot region predicted by the model, respectively, to form three sets of variants to be analyzed as data sources.

[0049] Nine authoritative pathogenicity prediction tools based on different algorithm principles are used as integrated annotation tools, including: AlphaMissense, EVE, PrimateAI, CADD, ClinPred, ESM1b, gMVP, MutFormer, and MutScore. These tools evaluate the harmfulness of variants from the perspectives of evolutionary conservation, physical and chemical properties, deep learning, etc. By using nine variant pathogenicity prediction tools, batch annotation is performed on the three sets of variants to obtain the prediction scores of each variant under the nine tools. The core analysis lies in systematically comparing the pathogenicity prediction score distribution of variants in three regions. By analyzing the score distribution in each region, it is proved that the variants in the predicted hot spot region indeed have higher pathogenicity scores, while the variants in the predicted cold spot region have lower pathogenicity scores, see Figure 12 .

[0050] The complete workflow of this embodiment is as follows: Training phase: a. Data preparation: The data processing module collects training data from multiple sources and assigns a cold-hot spot score label for each site.

[0051] b. Feature construction: The feature engineering module extracts DNA one-hot encoding and 24-dimensional prior features for each 100bp sequence and performs data augmentation.

[0052] c. Model training: The constructed (sequence, prior feature, label) data is input into the core prediction model. The AdamW optimizer (learning rate = 1e-5) is used to minimize the mean square error loss function, which calculates the difference between the model's predicted score and the true score.

[0053] d. Validation and tuning: Monitor model performance on the validation set and use early stopping to prevent overfitting. Training is completed on a single NVIDIA GeForce RTX 4060 GPU, and after 28 training cycles, the model's loss (MSE) on the test set is reduced to 0.0146.

[0054] Prediction and application phase: a. Whole-genome exome prediction: The entire human exome is divided into 100bp sliding windows, and the corresponding prior features are input into the trained HCSeer2 model.

[0055] b. Generate scores: The model outputs a cold-hot spot prediction score for each site in the exome.

[0056] c. Region identification: The post-processing module identifies continuous cold and hot spot regions based on preset thresholds (0.66 and 0.33). Finally, 28,907 variant hot spot regions and 159,970 variant cold spot regions are successfully identified within the entire exome range.

[0057] d. Variant interpretation assistance: The generated cold-hot spot regions and scores can be directly used for ACMG guideline-based variant interpretation, providing quantitative and site-specific support for PM1 (hot spot) and cold spot evidence, significantly improving the accuracy and efficiency of VUS variant classification.

[0058] The technical effects of this embodiment include: (1) The prediction coverage and the number of capacity implementation are improved by an order of magnitude, solving the "blind area" problem of the prior art. The prior art HCSeeker highly depends on the number of reported variants in the ClinVar database, resulting in its ability to identify only a limited number of hot and cold spots in genes. The technology provided by the present application does not depend on the variant density of a specific database, but directly analyzes the sequence and prior features through a deep learning model, thereby having the ability to predict any region of the whole exome.

[0059] Specific data proves that: HCSeeker only identifies 988 hot spot regions in 384 genes and 683 cold spot regions in 505 genes. By applying the present application, 28,907 hot spot regions and 159,970 cold spot regions are successfully identified in the whole exome. The number of hot spot regions is increased by about 29 times, and the number of cold spot regions is increased by about 234 times, greatly expanding the region resources that can be used for ACMG guideline PM1 and cold spot evidence, and solving the problem that a large number of regions cannot be evaluated due to lack of data.

[0060] (2) The prediction accuracy and biological relevance are significantly improved, and the output results have clear clinical significance. An excellent prediction model must be consistent with the real biological phenomenon. If the hot spot regions predicted by the present application are indeed functionally critical regions, then the variants located in these regions should have a higher pathogenicity; otherwise, the variants in the cold spot regions should be more benign.

[0061] Specific data proves that: through the verification module of the present application, 9 kinds of authoritative pathogenicity prediction tools (AlphaMissense, EVE, etc.) are used to analyze the predicted regions, and the results show that the pathogenicity prediction score of missense variants located in the hot spot regions predicted by the present application is significantly higher than that of the cold spot regions and non-cold and hot spot regions. This result is consistent in all 9 kinds of tools, which strongly proves from the calculation level that the prediction results of the present application are highly related to the potential pathogenicity of the variants, and the model output has a solid biological basis rather than random guess.

[0062] (3) Direct and efficient evidence support is provided for clinical variant interpretation, improving the accuracy of interpretation. The cold and hot spot regions and site-specific scores can be directly integrated into the variant interpretation process of ACMG / AMP guidelines, providing quantitative support for PM1 (hot spot) and cold spot evidence.

[0063] Specific data proves that: in a case study of 2048 VUS (variants of uncertain significance) in the ClinGen database, after applying the prediction results of the present application: Among the 295 VUS located in the hot spot regions, 56 obtained new evidence for upgrading to "likely pathogenic".

[0064] Of the 1,396 VUS located in the cold spot region, 146 received new evidence of being downgraded to "potentially benign".

[0065] Independent validation of these reclassified variants showed that their pathogenicity prediction scores were highly consistent with the new classification orientation. This demonstrates that the present invention can effectively reduce the number of VUS (variables under different classification systems), providing a powerful and reliable new tool for clinical gene interpretation.

[0066] (4) It enables a refined grading of evidence strength, promoting the interpretation of variations from qualitative to quantitative. Traditional regional evidence (such as PM1) is usually used as a qualitative "present" or "absent" piece of evidence. Each site output by this invention has a continuous hot and cold spot score (0-1), which allows us to classify the evidence strength more finely.

[0067] Specific data demonstrates that, based on the distribution of genome-wide locus scores, this invention can further subdivide hotspot evidence into PM1_Strong, PM1_Moderate, and PM1_Supporting; and coldspot evidence into BS coldspots, BM coldspots, and BP coldspots. This refined intensity grading allows evidence values ​​contributing different weights to a Bayesian-based quantitative variant classifier, thereby achieving more accurate and personalized pathogenicity assessments—something that existing technologies cannot achieve.

[0068] The technical solution in this embodiment, by fusing DNA sequences with multidimensional biological prior features and combining CNN and self-attention mechanisms, can simultaneously capture local patterns and long-range dependencies in the genome, improving prediction accuracy and biological rationality. It also eliminates dependence on existing variant data density, achieving blind-spot-free coverage across the entire exome; simultaneously, the structured output based on preset transcript coordinates can directly support clinical variant interpretation, providing quantitative evidence for PM1 and cold spot evidence, effectively reducing the proportion of variants of unknown significance.

[0069] Preferred, such as Figure 2 As shown, the training data for the pre-trained deep learning prediction model is constructed in S2, including the following steps S21-S23: S21, collect genomic region data from at least two sources, the genomic region data including at least identified hot and cold regions, regions where expert-designed evidence for PM1 can be applied, and key functional domain information; S22, establish a linear mapping rule to convert the qualitative information of the genomic region data into continuous numerical labels, linearly map the pathogenicity ratio or evidence strength of hot spot regions to the first high score interval, and linearly map the benignity ratio of cold spot regions to the second low score interval. S23, the continuous numerical label after conversion is combined with the corresponding DNA sequence and prior features to form a training sample set, and a deep learning prediction model is supervised trained by a mean square error loss function.

[0070] In the embodiment, the pre-trained deep learning prediction model can realize the construction of high-quality training data and the supervision training strategy. Step S21 can realize the collection and integration of multi-source genomic region data. The embodiment is not limited to a single data source, and real value (Ground Truth) data is obtained through multiple channels. Specifically, the collected data sources include at least three aspects. First, the cold and hot spot regions identified by existing technical tools (such as HCSeeker) through statistical methods are used as the basic data set. Second, the ClinGen VCEPs (Clinical Genomic Resource Variation Exhibition Expert Group) are manually planned by disease experts, and the regions with PM1 evidence are clear and applicable, which have high biological confidence. Third, the key functional domain information provided by the AutoPVS1 tool.

[0071] Step S22 is performed to establish a linear mapping rule from qualitative region information to quantitative continuous label. Since the original data is mostly binary qualitative description, direct use for regression prediction effect is poor, therefore, the embodiment converts qualitative evidence into continuous numerical label that can be learned by the model through the design of scoring mapping mechanism. In specific implementation, for the regions showing hot spot attributes, such as HCSeeker hot spots or ClinGen PM1 regions, the system extracts the associated Odds Ratio or PM1 evidence strength level, and maps them to the first high score interval through linear transformation. The value of the first high score interval is preferably set to [0.66, 1], and the first high score interval is used as the target hot spot score of the site in the region. For the regions showing cold spot attributes, the system extracts the benign Odds Ratio and linearly maps it to the second low score interval, which is preferably set to [0, 0.33]. For the key functional domain of the AutoPVS1 tool, the pathogenic variation proportion in the region is mapped to the [0, 1] interval. The differences in evidence strength between different regions are preserved, and smooth gradient information is also provided, so that the prediction result better reflects the tendency degree of cold and hot spots.

[0072] The step S23 of assembling the training sample set and supervising the training of the model is performed. The system takes the continuous numerical label as the training target, splices the DNA sequence features (after One-hot encoding) of the corresponding genomic location with the 24-dimensional biological prior features (including position score, conservation score, tolerance score, etc.) to construct paired training samples. These training samples are batched into the deep learning prediction model, and the mean squared error (MSE) is configured as the loss function to measure the Euclidean distance between the predicted score output by the model and the real score generated by the mapping. In the training process, the AdamW optimizer (learning rate preferably 1e-5) is used to iteratively update the network weights, and the early stopping strategy is implemented to prevent overfitting by monitoring the performance of the validation set. After sufficient iterative training, for example, 28 cycles, when the deep learning prediction model converges to the preset minimum value on the test set, the model training is considered complete, that is, it has the ability to score cold and hot spots with high precision.

[0073] Preferably, as shown in S1, the biological prior feature vector in S1 includes a position score feature, including the following steps S11-S13: Figure 3 S11, the position score feature is obtained from the missense variant data in the variant database, and an initial score is given according to the confidence level and classification standard, wherein pathogenic / possibly pathogenic variants are given a positive initial score, and benign / possibly benign variants are given a negative initial score; S12, training a random forest regression model with the chromosome number and genomic coordinates of the missense variant data as input and the initial score as output target; S13, by the trained random forest regression model, each potential missense variant site in the target genomic region is predicted, and a continuous position score is calculated.

[0074] ​Step S11 is performed, that is, the acquisition of basic data and the establishment of a quantitative scoring system. Specifically, the system extracts full-amount missense variant data from the ClinVar database as training corpus. In order to convert the discrete clinical interpretation in the database into a numerical signal that can be calculated by the model, the embodiment formulates a binary opposite scoring strategy based on the confidence level. For those “pathogenic” or “likely pathogenic” variants marked by the ClinVar rating system as 2 stars and above, the system assigns them a positive initial score (preferably set to +1) to represent the potential functional sensitivity and high pathogenic risk of the site; on the contrary, for “benign” or “likely benign” variants with the same confidence level, the system assigns them a negative initial score (preferably set to -1). For “pathogenic” or “likely pathogenic” variants marked by the ClinVar rating system as 1 star, the system assigns them a positive initial score (preferably set to +0.5) to represent the potential functional sensitivity and pathogenic risk of the site; on the contrary, for “benign” or “likely benign” variants with the same confidence level, the system assigns them a negative initial score (preferably set to -0.5). Through this positive and negative assignment mechanism, the biological and clinical classification of the variant is effectively converted into a numerical supervised label.

[0075] Step S12 is performed, that is, the construction and training of the spatial mapping model. The embodiment selects a random forest regression algorithm with strong anti-noise ability and nonlinear fitting ability as the basic model. In the training stage, the system takes the chromosome number (Chromosome) and specific genomic coordinates (Genomic Coordinate) corresponding to the above assigned variant records as the input features of the model, and takes the initial score generated in step S11 as the target variable of regression prediction. Through supervised learning on a large amount of labeled data, the random forest model can capture and remember the spatial distribution patterns of pathogenic variants and benign variants in different regions of the genome, thereby establishing an internal mapping relationship from “physical coordinates” to “pathogenicity tendency”, and extracting the regional functional constraint features hidden behind the coordinates.

[0076] Step S13, namely the generation and application of continuous position scores across the entire genome, is executed. Using the trained random forest regression model, the system predicts each potential missense mutation site within the target genomic region (i.e., the whole exome). At this point, the model no longer relies on whether the site is recorded in the database; instead, it outputs a continuous predicted value, the position score, based solely on its coordinates. This position score is a continuous floating-point value, and its magnitude directly reflects whether the position and its neighboring region tend to be enriched with pathogenic variants (high score) or benign variants (low score). In practical feature engineering, to comprehensively evaluate the impact of a nucleotide site, the system calculates the arithmetic mean of the predicted scores corresponding to the three possible missense mutations at that position, using this as the position score feature for that site. This enhances the model's ability to distinguish between hot and cold hotspot attributes in unknown regions.

[0077] Preferred, such as Figure 4 As shown, in S2, the multimodal input tensor is processed by the deep learning prediction model, including the following steps S24-S26: S24, receive the multimodal input tensor, perform feature transformation through a one-dimensional convolutional block, and process it through a batch normalization layer and an activation function layer to generate an intermediate feature map containing local patterns; S25, the intermediate feature map is input into the stacked Transformer encoder blocks. In each encoder block, the correlation weights between different sites are calculated through a multi-head self-attention mechanism. Combined with positional encoding information and the Dropout mechanism, a high-level feature representation that incorporates global dependencies is output. S26, the high-level feature representation is fed into the output layer, and after convolution and fully connected mapping, a one-dimensional score vector with the same length as the input window is generated.

[0078] The multi-modal input tensor is processed by the pre-trained deep learning prediction model to perform step S24, i.e., preliminary extraction and transformation of local features. The input layer of the deep learning model receives a multi-modal input tensor with a dimension of 100x29, which carries sequence encoding information within the window and biological prior knowledge. The data enters a one-dimensional convolution block (1D-Convolution Block), and the convolution kernel slides in the sequence dimension to capture short-range local patterns (Local Patterns) in the DNA sequence, such as specific transcription factor binding sites or sequence motifs. The feature map output by the convolution operation is then processed by a batch normalization (Batch Normalization) layer to standardize the feature distribution, which can solve the gradient dispersion problem in deep network training and accelerate convergence; the ReLU activation function is introduced to introduce nonlinearity, enabling the model to fit complex biological nonlinear relationships.

[0079] Step S25 is performed to capture and deepen global dependencies. The intermediate feature map generated in step S24 is input into the stacked Transformer encoder layer of the deep learning prediction model. This Transformer encoder layer discards the serial computation limitation of traditional recurrent neural networks and uses a multi-head self-attention mechanism to process information within the entire window in parallel. Within each encoder block, the model dynamically calculates the pairwise correlation weights between the 100 nucleotide sites within the window. Regardless of the physical distance between two sites in the sequence, the attention mechanism can establish a direct connection, effectively capturing the crucial long-range dependencies and global context information in the genomic sequence. To preserve the positional order information of the sequence, the deep learning prediction model superimposes position encoding (Positional Encoding) at the input end; at the same time, to prevent the deep learning prediction model from overfitting on the training set and improve the generalization ability, the Dropout mechanism (dropout rate preferably set to 0.1) is integrated into the self-attention layer and the feedforward network layer.

[0080] Step S26 can realize the mapping output from high-dimensional features to prediction scores. The high-level feature representation after the deep abstraction processing of the Transformer layer is transmitted to the output layer of the model. The data is subjected to one-dimensional convolution operation for feature integration and dimension reduction, and then enters the fully connected layer. The fully connected layer linearly maps the abstract high-dimensional feature space back to a 100-dimensional output space consistent with the input window length. After activation processing, a one-dimensional score vector containing 100 floating-point values is generated. Each element in the one-dimensional score vector accurately corresponds to each nucleotide site in the input sequence window, and the numerical size quantitatively represents the probability strength of the site belonging to the mutation hot spot or the mutation cold spot.

[0081] Preferably, as shown in S3, the mutation cold spot region and the mutation hot spot region in the target genomic region are identified, including steps S31-S33: Figure 5 S31, scanning the cold-hot spot prediction scores, and determining that a region corresponding to nucleotide sites of not less than 2 consecutive amino acids and having scores greater than or equal to a first threshold value is a hot spot region, and determining that a region corresponding to nucleotide sites of not less than 2 consecutive amino acids and having scores less than or equal to a second threshold value is a cold spot region; S32, obtaining an annotation file of a standard human reference genome and MANE database recommended transcript information, and converting nucleotide coordinates of the hot spot region and the cold spot region into corresponding transcript amino acid coordinates; S33, merging regions adjacent to each other and meeting the same threshold condition on the amino acid coordinates, and outputting a structured list containing a chromosome, a transcript number, a start position, an end position, and a region type.

[0082] ​Step S3 transforms the micro-site scores output by the deep learning model into regional information with clear biological and clinical significance. This mainly includes threshold determination of hot and cold regions, standardized mapping of the coordinate system, and structured integration of the results. Step S31 is the preliminary screening of regions based on threshold rules. The system performs a linear scan of the predicted hot and cold regions at each site of the whole exome output in step S2. To ensure the biological robustness of the identification results and filter random noise, this embodiment sets dual constraints on score strength and sequence length. Specifically, for the determination of hot regions, the system selects continuous sequence fragments that are long enough to cover at least 2 amino acids (corresponding to at least 6 nucleotides), requiring that the predicted score of each nucleotide site within these fragments is greater than or equal to a first preset threshold (preferably set to 0.66 in this embodiment). Similarly, for the determination of cold regions, the system selects sequence intervals that also meet the requirement of a length of at least 2 amino acid spans, and whose predicted scores for all sites within the fragment are less than or equal to a second preset threshold (preferably set to 0.33 in this embodiment). Steps S32 to S33 are the standardized transformation of the coordinate system and the establishment of regions. Since the original predictions are based on absolute physical coordinates of the genome (genomic coordinates at the nucleotide level), to align with the interpretation standards of the ACMG / AMP clinical guidelines, the system calls upon annotation files from standard human reference genomes (such as GRCh38 / hg38) and transcript information recommended by the MAME (Matched Annotation from NCBI and EMBL-EBI) database. The nucleotide coordinates of the candidate regions selected above are mapped to the corresponding standard transcript amino acid coordinates. Based on this, region merging logic is executed, fusing physically adjacent regions of the same type (both hot spots or both cold spots) on the amino acid coordinates to eliminate fragmented and redundant predictions. The system generates and outputs a standardized structured list, which details the chromosome number, corresponding MAME transcript number, start and end positions at the amino acid level, and type interpretation (cold spot or hot spot) for each identified region.

[0083] Preferred, such as Figure 6 As shown, after S3, S4 grades the strength of evidence, and S4 includes S41-S43: S41, Statistically calculate the maximum and minimum scores of all sites within the identified cold or hot areas, and divide the score range into three equal-length score intervals based on the maximum and minimum scores; S42, the three equal length score intervals correspond to different evidence intensity levels respectively, for the hotspot region, the interval with the highest score value is defined as the strong confidence level, the interval with the middle score value is defined as the moderate confidence level, and the interval with the lowest score value is defined as the supportive confidence level, for the cold spot region, the interval with the highest score value is defined as the supportive confidence level, the interval with the middle score value is defined as the moderate confidence level, and the interval with the lowest score value is defined as the strong confidence level; S43, according to the interval in which the specific cold-hot point prediction score of the to-be-tested variant site falls, the variant site is assigned a corresponding PM1 or cold spot evidence intensity label.

[0084] After the identification and coordinate mapping of the cold and hot spot regions of variation are completed (i.e., after step S3), in order to further improve the practicality and accuracy of the prediction results in clinical interpretation of variations, step S4 can be used to finely classify the evidence strength. The limitation of using PM1 evidence as a single moderate strength in the traditional ACMG / AMP guidelines can be broken, and a transition from qualitative judgment to quantitative classification is achieved. In specific implementation, step S41 is statistical analysis and interval division of data. The system extracts the original prediction scores of all nucleotide sites within each identified independent hot spot region or cold spot region. The statistical distribution characteristics of the scores in the region are calculated, and the maximum value Max_Score and the minimum value Min_Score of the scores in the region are determined. Taking the maximum value and the minimum value as boundaries, the score value range of the region is cut into three continuous and equal length score intervals using an equal distance division strategy. Step S42 can establish a mapping relationship between the score interval and the evidence strength. According to the classification system of evidence strength in the ACMG / AMP guidelines, the highest score interval divided above is defined as the strong confidence level interval (Strong Level), which corresponds to the PM1_Strong (for hot spots) or BP (for cold spots) label in clinical interpretation, indicating that the sites in this interval have a very high pathogenic or benign tendency. The interval with a score value in the middle is defined as the moderate confidence level interval (Moderate Level), which corresponds to the PM1_Moderate or BM label, representing moderate evidence strength. The lowest score interval is defined as the supporting confidence level interval (Supporting Level), which corresponds to the PM1_Supporting (for hot spots) or BS (for cold spots) label, which ensures that biological signals of different strengths can be differentiated and applied in the pathogenicity assessment model. For hot spots, the highest score interval divided above is defined as the strong confidence level interval (Strong Level), which corresponds to the PM1_Strong in clinical interpretation, indicating that the sites in this interval have a very high pathogenic or benign tendency. The interval with a score value in the middle is defined as the moderate confidence level interval (Moderate Level), which corresponds to the PM1_Moderate, representing moderate evidence strength. The lowest score interval is defined as the supporting confidence level interval (Supporting Level), which corresponds to the PM1_Supporting label, which ensures that biological signals of different strengths can be differentiated and applied in the pathogenicity assessment model.For cold spots, the interval with the lowest score value defined as the strong confidence level interval (Strong Level), corresponding to the BS in clinical interpretation, indicating that the site in this interval has a very high pathogenic or benign tendency; the interval with the middle score value is defined as the moderate confidence level interval (Moderate Level), corresponding to BM, representing moderate evidence strength; the interval with the highest score value is defined as the supporting confidence level interval (Supporting Level), corresponding to the BP label, which ensures that biological signals of different strengths can be differentiated and applied to the pathogenicity evaluation model.

[0085] Step S43 can realize the label assignment and application for the variant site to be tested. When a specific variant site in the clinical sample needs to be interpreted, the system first retrieves the specific cold-hot spot prediction score of the site in the whole exome prediction result. The system determines which of the above preset evidence strength intervals the specific score falls into. Once the interval is determined, the system automatically assigns the corresponding evidence strength label to the variant site. For example, if a site is in the hot spot region and the score is very high, falling into the first interval, the system will automatically add the "PM1_Strong" label to it, which can be directly integrated into the automated annotation process (such as VCF file annotation). When facing a large number of variants of uncertain significance (VUS), clinical personnel can obtain accurate quantitative support based on the specific site score, thereby effectively assisting in the reclassification of variants (such as upgrading to pathogenic or downgrading to benign), and improving the scientificity and efficiency of genetic diagnosis.

[0086] Figure 7 is a block diagram of a genomic variant cold-hot spot region prediction device according to an exemplary embodiment. As shown in Figure 7 The genomic variant cold-hot spot region prediction device includes a construction module 71, an output module 72, and an identification module 73.

[0087] The construction module 71 is configured to slice the target genomic region according to a preset sliding window length, extract a DNA sequence vector corresponding to each window, and match a multi-dimensional biological prior feature vector for each site in the window. The DNA sequence vector and the biological prior feature vector are spliced in the feature dimension to construct a multi-modal input tensor corresponding to the window. The output module 72 inputs the multi-modal input tensor into a pre-trained deep learning prediction model, extracts local features of the sequence through a one-dimensional convolutional neural network layer in the model, captures long-distance dependency relationships of the sequence through a self-attention mechanism layer, and outputs a cold-hot spot prediction score for each site in the window through a fully connected layer. The identification module 73 screens out a continuous abnormal score interval through a preset cold-hot spot judgment threshold according to the cold-hot spot prediction score, and maps the abnormal score interval to a preset transcript coordinate, so as to identify a variant cold spot region and a variant hot spot region in the target genomic region.

[0088] The modules 71, 72 and 73 included in the block diagram of the genomic variant cold-hot spot region prediction device are controlled to perform the genomic variant cold-hot spot region prediction method described in any of the embodiments.

[0089] As shown in Figure 8 The present application provides an electronic device 800, which comprises a communication interface, a processor 801, and a memory 802. The memory 802 is configured to store program instructions, and the program instructions are executed by the processor 801 connected with the memory 802 through the communication interface.

[0090] The present application provides a computer readable storage medium, which stores computer program instructions. When the computer program instructions are executed by a processor, a target genomic region is sliced according to a preset sliding window length, a DNA sequence vector corresponding to each window is extracted, and a multi-dimensional biological prior feature vector is matched for each site in the window. The DNA sequence vector and the biological prior feature vector are spliced in the feature dimension to construct a multi-modal input tensor corresponding to the window. The multi-modal input tensor is input into a pre-trained deep learning prediction model, local features of the sequence are extracted by a one-dimensional convolutional neural network layer in the model, long-distance dependency of the sequence is captured by a self-attention mechanism layer, and cold-hot spot prediction scores of each site in the window are output by a fully connected layer. According to the cold-hot spot prediction scores, a continuous abnormal score interval is screened out through a preset cold-hot spot judgment threshold, and the abnormal score interval is mapped to a preset transcript coordinate, so as to identify a variant cold spot region and a variant hot spot region in the target genomic region.

[0091] Compared with the prior art, the present embodiment has the following advantages: The present embodiment can capture local patterns and long-range dependencies of the genome simultaneously by fusing DNA sequences and multi-dimensional biological prior features and combining CNN and self-attention mechanism, thereby improving prediction accuracy and biological rationality. The present embodiment also eliminates the dependence on existing variant data density and realizes no-blind-zone coverage in the whole exome range. Meanwhile, the structured output based on the preset transcript coordinate can directly support clinical variant interpretation, provide quantitative evidence for PM1 and cold spot evidence, and effectively reduce the proportion of variants of uncertain significance.

[0092] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, apparatus, or computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer readable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage devices, etc.) embodying computer readable program code. The application is described in reference to the flow diagrams and / or block diagrams of the methods, apparatus (systems) and computer program products according to embodiments of the application illustrated in the attached drawings, which flow diagrams and / or block diagrams Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagram and / or block diagram. Figure 1 one or more functions specified in the flow diagram and / or block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagram and / or block diagram.

[0093] The above description is only preferred embodiments of the present application, the protection scope of the present application is not limited to the above-mentioned embodiments, any technical solutions falling within the concept of the present application shall be considered as falling within the protection scope of the present application. It should be noted that, for ordinary skilled in the art, some improvements and refinements without departing from the principles of the present application shall be considered as falling within the protection scope of the present application.

Claims

1. A method for predicting hot and cold regions of genomic variation, characterized in that, The method comprises the following steps: S1, slicing a target genomic region according to a preset sliding window length, extracting a DNA sequence vector corresponding to each window, and matching a multidimensional biological prior feature vector for each site in the window, splicing the DNA sequence vector and the biological prior feature vector in the feature dimension to construct a multi-modal input tensor corresponding to the window; S2, inputting the multi-modal input tensor into a pre-trained deep learning prediction model, extracting local features of the sequence through a one-dimensional convolutional neural network layer in the model, capturing long-distance dependence of the sequence through a self-attention mechanism layer, and outputting a cold-hot spot prediction score of each site in the window through a fully connected layer; S3, according to the cold-hot spot prediction score, filtering out a continuous abnormal score interval through a preset cold-hot spot judgment threshold, and mapping the abnormal score interval to a preset transcript coordinate, so as to identify a variant cold spot region and a variant hot spot region in the target genomic region.

2. The method of claim 1, wherein the method is performed by a computer system. The training data for constructing the pre-trained deep learning prediction model in S2 comprises: S21, collecting genomic region data from at least two sources, wherein the genomic region data at least comprises an identified cold-hot spot region, an expert-planned applicable PM1 evidence region and key functional domain information; S22, establishing a linear mapping rule to convert the qualitative information of the genomic region data into continuous numerical labels, and linearly mapping a pathogenic ratio or evidence strength of the hot spot region to a first high score interval, and linearly mapping a benign ratio of the cold spot region to a second low score interval; S23, forming a training sample set by combining the converted continuous numerical labels with corresponding DNA sequences and prior features, and supervising the training of the deep learning prediction model through a mean square error loss function. 3.The method of claim 1, wherein, The biological prior feature vector in S1 comprises a position score feature, which comprises: S11, the position score feature obtains missense variant data from a variant database, and assigns an initial score according to a confidence level and a classification standard, wherein a pathogenic variant is assigned a positive initial score, and a benign variant is assigned a negative initial score; S12, training a random forest regression model by taking a chromosome number and a genomic coordinate of the missense variant data as input and taking the initial score as output target; S13, predicting each potential missense variant site in the target genomic region through the trained random forest regression model to obtain a continuous position score. 4.The method of claim 1, wherein, The processing of the multi-modal input tensor by the deep learning prediction model in S2 comprises: S24, receiving the multi-modal input tensor, performing feature transformation through a one-dimensional convolutional block, and processing through a batch normalization layer and an activation function layer to generate an intermediate feature map containing local patterns; S25, inputting the intermediate feature map into a stacked Transformer encoder block, calculating the correlation weight between different sites through a multi-head self-attention mechanism in each encoder block, combining position encoding information and a Dropout mechanism, and outputting a high-level feature representation fused with global dependence; S26, the high-level feature representation is transmitted into an output layer to generate a one-dimensional score vector consistent with the input window length through convolution operation and full connection mapping.

5. The method of claim 1, wherein the method is performed by a computer system. In S3, the cold and hot spot regions in the target genomic region are identified, including: S31, the cold and hot spot prediction scores are scanned, and a region corresponding to nucleotide sites of consecutive amino acids not less than 2 and having scores greater than or equal to a first threshold value is determined as a hot spot region, and a region corresponding to nucleotide sites of consecutive amino acids not less than 2 and having scores less than or equal to a second threshold value is determined as a cold spot region; S32, an annotation file of a standard human reference genome and MANE database recommended transcript information are obtained, and nucleotide coordinates of the hot and cold spot regions are converted into corresponding transcript amino acid coordinates; S33, regions adjacent to each other and meeting the same threshold condition on the amino acid coordinates are merged, and a structured list containing a chromosome, a transcript number, a start and end position, and a region type is output.

6. The method of predicting genomic variant cold-hot spot regions according to claim 5, wherein, After S3, including: S4, the evidence strength is graded, including: S41, the maximum and minimum values of all site scores in the identified cold or hot spot region are counted, and the score value domain is divided into three equal length score intervals based on the maximum and minimum values; S42, the three equal length score intervals correspond to different evidence strength levels, for the hot spot region, the interval with the highest score value is defined as the strong confidence level, the interval with the intermediate score value is defined as the intermediate confidence level, and the interval with the lowest score value is defined as the supporting confidence level, for the cold spot region, the interval with the highest score value is defined as the supporting confidence level, the interval with the intermediate score value is defined as the intermediate confidence level, and the interval with the lowest score value is defined as the strong confidence level; S43, according to the interval into which the specific cold and hot spot prediction score of the to-be-tested variation site falls, a corresponding PM1 or cold spot evidence strength label is assigned to the variation site.

7. A genomic variant cold-hot spot region prediction apparatus, characterized by, Including: The construction module is configured to slice the target genomic region according to a preset sliding window length, extract a DNA sequence vector corresponding to each window, match a multidimensional biological prior feature vector for each site in the window, splice the DNA sequence vector and the biological prior feature vector in the feature dimension, and construct a multi-modal input tensor corresponding to the window; The output module inputs the multi-modal input tensor into a pre-trained deep learning prediction model, extracts local features of the sequence through a one-dimensional convolutional neural network layer in the model, captures long distance dependence of the sequence through a self-attention mechanism layer, and outputs cold and hot spot prediction scores of each site in the window through a full connection layer; The identification module screens out continuous abnormal score intervals through a preset cold and hot spot determination threshold according to the cold and hot spot prediction scores, maps the abnormal score intervals to preset transcript coordinates, and identifies cold and hot spot regions in the target genomic region.

8. The genomic variant cold-hot spot region prediction apparatus of claim 7, wherein: The construction module, the output module and the identification module are controlled to perform the genomic variation cold-hot spot region prediction method of any one of claims 1 to 6.

9. An electronic device, comprising: Comprise: A communication interface, a processor, a memory; Wherein, the memory is used for storing program instructions, the program instructions when being executed by the processor which is communicated with the memory through the communication interface make the electronic equipment realize the genomic variation cold-hot spot region prediction method of any one of claims 1 to 6.

10. A computer-readable storage medium having stored thereon program instructions, wherein, The program instructions when being executed by the computer make the computer realize the genomic variation cold-hot spot region prediction method of any one of claims 1 to 6.

Images