Rapid processing method for HPV (human papillomavirus) gene sequence data

Abstract

The invention provides a rapid processing method for HPV (human papillomavirus) gene sequence data. The method relates to the technical field of data processing, and comprises the following steps: acquiring and integrating HPV sequence data, host transcriptome data and clinical association data to form a comprehensive input matrix; performing feature screening on the comprehensive input matrix by adopting a bee colony optimization algorithm based on reinforcement learning to obtain an optimal feature subset; according to the bee colony optimization algorithm, global search of the bee colony algorithm is combined with a dynamic decision-making mechanism of reinforcement learning, and a feature selection strategy is dynamically adjusted; and constructing an HPV subtype accurate typing and variation detection model based on multi-task deep learning, performing deep coding and weighted fusion on the optimal feature subset through an unstable attention mechanism, and outputting two key prediction results of subtype probability and mutation site probability. According to the method, the accuracy and biological interpretability of virus typing and variation detection are improved, and efficient extraction and multi-task accurate recognition of high-dimensional features are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a rapid processing method for HPV virus gene sequence data. BACKGROUND

[0002] Human papillomavirus (HPV) is a class of DNA viruses with high genetic heterogeneity, which is closely related to the occurrence of various reproductive system tumors. Among them, high-risk types 16 and 18 HPV are considered to be the main pathogenic factors of cervical cancer. With the development of high-throughput sequencing technology, researchers can obtain HPV virus whole genome sequence, host transcriptome expression profile and its clinical pathological information, which provides rich data basis for virus typing, variation detection and infection mechanism research. However, HPV virus and host gene data have problems such as high dimension, strong heterogeneity and obvious dynamic non-stationarity. The traditional analysis method has the following limitations: (1) Low efficiency of feature selection and weak biological correlation: traditional feature selection algorithms such as ReliefF and recursive feature elimination can select key variables through feature importance ranking, but they have low search efficiency and are prone to local optimization in high-dimensional space. Taking HPV-host joint features containing 50,000 dimensions as an example, the traditional ReliefF algorithm needs to iterate for several hours to complete the screening, and the screening results often miss the functional features of key oncogenes such as E6 / E7, which greatly reduces the biological interpretability of subsequent modeling.

[0003] (2) Insufficient dynamic non-stationary mutation signal capture ability: the traditional attention mechanism (such as self-attention in Transformer) adopts a fixed weight allocation strategy, which can only capture the global association between sequences statically and cannot adapt to the dynamic changes of mutation signals: when E6 / E7 gene appears short-term mutation peak, the model cannot timely increase the attention weight of this area, resulting in that the key mutation signal is submerged by the stable signal of the conservative region; while in the conservative region, a small variation (such as synonymous mutation of L1 gene) is prone to over-attention to non-pathogenic signals, causing false positive detection results. This "static attention" mechanism is seriously mismatched with the "dynamic non-stationarity" characteristics of HPV mutation, which is a core obstacle restricting the accuracy of variation detection.

[0004] (3) The tasks of typing and variation detection are fragmented: HPV subtype identification and variation detection are two highly related tasks. However, existing models generally adopt a "task-fragmented" modeling approach: subtype typing models such as BLAST-based sequence alignment methods and support vector machine classification models only focus on global features of the virus genome (such as whole sequence homology), ignoring local variation information, which leads to the inability to distinguish the pathogenicity difference within the same subtype (such as HPV16 type, the carcinogenic risk of the strain containing E6 mutation is significantly higher than that of the wild type). Variation detection models: such as GATK, VarScan and other tools, need to rely on known subtype information to pre-limit the detection range. If the subtype information is not accurate, it will lead to an increase in the false positive rate of mutation detection (such as misjudging the unique sequence of HPV18 type as a mutation of HPV16 type). This "fragmented modeling" approach not only causes insufficient feature sharing during training (such as the virus genome features learned by the subtype typing model cannot be used for variation detection), but also produces redundant predictions (such as two models need to process the same batch of sequence data), ultimately leading to difficulty in improving the overall performance of the model (such as typing accuracy, variation detection sensitivity), and failing to meet the clinical demand for "subtype-variation" joint analysis.

[0005] Therefore, there is an urgent need for a HPV virus gene sequence rapid processing method that combines intelligent feature screening, dynamic attention, and task collaboration to achieve efficient extraction of high-dimensional features and accurate multi-task recognition. SUMMARY

[0006] The present application proposes a rapid processing method based on deep learning and reinforcement optimization fusion to address the problems of high-dimensional redundancy, dynamic instability, and low task coupling in existing HPV virus gene sequence data processing, aiming to improve the accuracy and biological interpretability of virus typing and variation detection.

[0007] To achieve the above-mentioned purposes, the following technical solutions are implemented: The present application provides a rapid processing method for HPV virus gene sequence data, comprising the following steps: Obtain HPV virus sequence data, host transcriptome data, and clinical association data and integrate them to form a comprehensive input matrix; Perform feature screening on the comprehensive input matrix using a reinforcement learning-based bee colony optimization algorithm to obtain an optimal feature subset; wherein the bee colony optimization algorithm combines the global search of the bee colony algorithm with the dynamic decision mechanism of reinforcement learning, and uses virus subtype classification performance and mutation prediction accuracy as reward feedback to dynamically adjust the feature selection strategy; The HPV subtype accurate typing and variation detection model based on multi-task deep learning is constructed, the optimal feature subset is deeply encoded and weighted fused through an unstable attention mechanism, and a global feature representation capable of sensitively reflecting the key mutation region of the HPV sequence is generated; based on the global feature, the HPV subtype typing task and the virus variation detection task are simultaneously performed, and finally two key prediction results of subtype probability and mutation site probability are output.

[0008] Further, the swarm optimization algorithm based on reinforcement learning is specifically: Each feature subset is regarded as a swarm search direction; In each iteration, the performance of each feature subset is evaluated according to a fitness function, and the fitness function comprehensively considers the classification accuracy of the validation set, the mutation prediction error, and the biological correlation score based on the importance of biological pathways; According to the fitness ranking, the feature subsets with low fitness are eliminated, the neighborhood around the feature subsets with high search fitness is copied and strengthened, and the retention or non-retention of the features is dynamically decided through the Q-learning strategy of reinforcement learning, and finally the optimal feature subset that maximizes the fitness function is output.

[0009] Further, the structure of the HPV subtype accurate typing and variation detection model comprises: a feature encoder for encoding the input optimal feature subset; an unstable attention mechanism module that receives the output of the feature encoder and dynamically adjusts the feature weights of different sequence positions, and by introducing a time stationarity control variable that is adaptively adjusted according to the fluctuation degree of the sliding window, the attention focusing on the key gene fragment is automatically enhanced when the mutation peak value appears in the region; a dual-task output layer that receives the global fused feature representation processed by the unstable attention mechanism module to simultaneously perform the HPV subtype typing task and the virus variation detection task.

[0010] Further, the working mode of the unstable attention mechanism module is: calculate the Prompt feature according to the optimal feature subset as the input of Query in the attention mechanism; introduce a time stationarity control variable and an offset correction vector to calculate the similarity of Query and Key; wherein the time stationarity control variable is obtained by calculating the average value of sequence signal fluctuations under different size sliding windows; normalize the adjusted similarity through a Softmax function to calculate the attention weight, thereby realizing adaptive capture of non-stationary mutation signals; The Value vectors are weighted and summed according to the attention weights to obtain weighted output features, and the attention-weighted output features of all positions are averaged and pooled to form a global fusion feature representation, which is used for subsequent double-task prediction.

[0011] Further, the Prompt feature is calculated according to the optimal feature subset as the input of Query in the attention mechanism, including: After normalizing the optimal feature subset to remove overall bias and capture sequence signal fluctuations, the Query input of the unstable attention is generated by combining the learnable scaling and offset parameters to adapt the dynamic distribution of different samples.

[0012] Further, the double-task output layer shares the global fusion feature, including: The subtype typing head is composed of a fully connected layer connected with a Softmax function, and a Softmax classifier is used to predict the subtypes of HPV samples, and the probability distribution of all preset HPV subtypes is output , and the subtype classification loss is calculated. The variant detection head is composed of a fully connected layer connected with a Sigmoid function, and a Sigmoid function is used to predict the position and type of gene mutation, and the probability of mutation at each site is output , and the variant detection loss is calculated.

[0013] Further, in the model training stage, the HPV subtype precise typing and variant detection model are optimized end-to-end by a weighted joint loss function to ensure that the two tasks promote each other and improve the performance together. The weighted joint loss function is the weighted sum of the subtype typing task loss, the variant detection task loss and the regularization term to realize the joint optimization of the two tasks.

[0014] Further, the weighted joint loss function is represented as:

[0015] wherein, is the loss of the subtype typing task, is the loss of the variant detection task, is the L2 regularization term of the model parameters, , and are the corresponding weight coefficients, respectively. The final optimization goal is:

[0016] wherein, The optimal parameter set obtained after joint task optimization.

[0017] Further, the HPV virus sequence data includes whole genome sequence and gene fragment sequence, and is formed into an initial virus feature matrix through k-mer frequency coding and base physicochemical feature vectorization.

[0018] Further, the host transcriptome data is derived from RNA-Seq sequencing results of cervical tissue, and significant differential genes are screened through differential expression analysis to construct a host expression matrix.

[0019] Compared with the prior art, the present application has the following beneficial effects: 1. The present application proposes a bee colony optimization algorithm (Bee-Q) based on reinforcement learning, which combines the global distributed search of the bee colony algorithm with the dynamic decision mechanism of reinforcement learning to realize intelligent exploration of the high-dimensional virus-host joint feature space. The Bee-Q algorithm introduces Q-learning strategy and reward feedback function, takes virus subtype classification performance and mutation prediction accuracy as reward indicators, and dynamically adjusts the feature selection strategy to obtain the optimal feature subset, thereby realizing rapid screening of the feature subset and strengthening the biological significance.

[0020] 2. The present application designs an unstable attention mechanism, which designs a time-dynamically adjustable attention structure for the non-stationary mutation characteristics of HPV gene sequence. By introducing a time stationarity control variable and an offset correction vector in similarity calculation, the attention weight is adaptively adjusted according to the fluctuation degree of the sliding window. When the key gene fragments such as E6 / E7 have mutation peaks, the value of tau automatically increases, and the model enhances the attention focus in this area, thereby effectively capturing the non-stationary genetic pattern and suppressing noise interference.

[0021] 3. The present application constructs a double-task output layer and designs a weighted joint optimization mechanism, constructs a double-task output layer composed of a subtype typing head and a variation detection head, uses a shared encoder to extract a unified feature representation, and respectively completes virus subtype identification and mutation type prediction. At the same time, a weighted joint loss function is designed to realize joint optimization of the typing task and the mutation task, so that the model guarantees classification accuracy while considering mutation detection sensitivity and parameter stability.

[0022] Through the above innovative design, the present application realizes a fast, efficient and interpretable processing flow of HPV virus gene sequence data, and has a breakthrough advantage in feature screening, attention modeling and multi-task optimization, thereby providing a new technical path for molecular diagnosis and risk prediction of HPV-related tumors.

[0023] It is to be understood that the description in the summary is not intended to identify key or essential features of embodiments of the application, nor is it intended to limit the scope of the application. Other features, aspects, and advantages of the application will become apparent from the following description, when taken in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0024] The above and other features, aspects, and advantages of embodiments of the present application will become more apparent from the following description when taken in conjunction with the accompanying drawings. The drawings are intended to better illustrate the present application and are not intended to limit the scope of the application. In the drawings, like reference numerals refer to like elements throughout: Figure 1 is a flowchart of a rapid processing method for HPV virus gene sequence data according to an embodiment of the present application; Figure 2 is a flowchart of feature screening of a comprehensive input matrix using a bee colony optimization algorithm based on reinforcement learning according to an embodiment of the present application; Figure 3 is a structural diagram of an HPV subtype accurate typing and variation detection model according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of embodiments of the present application clearer, the following will be used to clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of the present application.

[0026] In addition, the term "and / or" herein merely describes an association relationship of associated objects, and indicates that there can be three relationships, for example, A and / or B can represent three cases of existence of A alone, existence of A and B together, and existence of B alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0027] Figure 1 is a flowchart of a rapid processing method for HPV virus gene sequence data according to an embodiment of the present application. As shown in Figure 1 A rapid processing method 100 for HPV virus gene sequence data provided by the present application includes the following steps: S110: Obtain HPV virus sequence data, host transcriptome data, and clinical association data and integrate them to form a comprehensive input matrix; To realize efficient processing of HPV virus and host genome association features, this step integrates three types of core data sources to construct a unified input data .

[0028] (1) HPV virus sequence data sources include the whole genome sequences and gene fragment sequences (FASTA / FASTQ format) of different subtypes of HPV (such as high-risk types 16, 18 and low-risk types 6, 11). The features include E6, E7 oncogene mutation sites, CpG island methylation patterns, and L1 / L2 gene conservative region alignment scores. The initial virus feature matrix is formed by using k-mer (short sequence fragment) frequency coding and base physicochemical feature vectorization (such as GC content, entropy value) . Among them, the k value in k-mer frequency coding is an integer between 3 and 7, and preferably k = 5.

[0029] (2) The host transcriptome data is derived from RNA-Seq sequencing results (FPKM expression matrix) of cervical adenocarcinoma and adjacent normal tissues. By differential expression analysis (DESeq2), significant differential genes (DEGs) are screened to construct the host expression matrix . Among them, the differential expression analysis uses the DESeq2 software package, and the standard for screening significant differential genes is: the adjusted p value is less than 0.05, and the absolute log2 fold change (|log2FoldChange|) is greater than 1.

[0030] (3) Clinical association data includes patient infection subtypes, pathological stages (Stage I-IV), treatment responses (PR, CR, PD), and follow-up survival periods, etc. Clinical feature matrix .

[0031] Then, the clinical feature matrix is mapped to the HPV sequence and host transcriptome data through unified sample ID mapping, to form a comprehensive input matrix X:

[0032] Among them, X is the input matrix; represents the feature splicing operation; is the initial virus feature matrix; is the host expression matrix; is the clinical feature matrix.

[0033] S120: Use a reinforcement learning-based bee colony optimization algorithm to perform feature selection on the comprehensive input matrix to obtain an optimal feature subset; wherein the bee colony optimization algorithm combines the global search of the bee colony algorithm with the dynamic decision mechanism of reinforcement learning, and uses virus subtype classification performance and mutation prediction accuracy as reward feedback to dynamically adjust the feature selection strategy; To cope with the characteristics of high dimensionality and strong correlation between HPV virus sequences and host transcriptome features, the step S120 proposes a Bee-Q algorithm based on reinforcement learning to realize fast screening and biological significance reinforcement of feature subsets. Bee-Q combines the global distributed search capability of the bee colony algorithm and the dynamic decision and reward feedback mechanism of reinforcement learning. When searching in a high-dimensional biological feature space, each bee colony represents a feature subset search direction, which is continuously optimized through the Q-learning strategy of the agent to improve the screening efficiency and feature representativeness.

[0034] Figure 2 is a process schematic diagram of feature screening of a comprehensive input matrix by the embodiment of the application based on the Bee-Q algorithm based on reinforcement learning. As shown in Figure 2 , specifically, the Bee-Q algorithm based on reinforcement learning is specifically: S121: Each feature subset is regarded as a bee colony search direction; S122: In each iteration, the performance of each feature subset is evaluated according to an adaptive function, which comprehensively considers the validation set classification accuracy, mutation prediction error, and biological correlation score (BioScore) based on the importance of biological pathways; wherein the agent selects an action (i.e., whether to retain the feature) in the tthiteration, and the bee colony with the lowest fitness is eliminated in each iteration; the feature subset of the bee colony with the highest fitness is copied; the search range of the neighborhood thereof is strengthened to ensure the diversity of feature distribution. The virus subtype classification performance is used as a reward feedback :

[0035] wherein, : reward feedback; : validation set classification accuracy, used to measure the subtype discrimination ability; : virus mutation frequency prediction error; : biological correlation score, calculated according to the weight of the E6 / E7, p53, etc. pathway corresponding to the feature; α, β, λ: weight hyperparameters, used to balance the model performance and biological significance, and satisfy α + β + λ = 1, and α, β, λ are all greater than 0. In the preferred embodiment, α = 0.5, β = 0.3, and λ = 0.2 can be taken.

[0036] wherein the biological correlation score (BioScore) is a core reward item driving the feature screening towards a direction with biological significance. The biological correlation score is calculated in the following manner: a gene pathway database related to HPV carcinogenicity is constructed in advance, which at least includes key pathways such as the E6 / p53 pathway and the E7 / pRb pathway and the gene set therein. For each feature subset each feature, if the feature corresponds to a gene located in any of the above pathways, then score 1, otherwise score 0. BioScore is the sum of scores of all features in the feature subset BioScore = (number of features in the feature subset belonging to the predefined key pathways) / (total number of features in the feature subset) . .

[0037] In the bee colony optimization algorithm, multiple candidate feature subsets are maintained and iterated simultaneously (e.g., there are C=20 feature subsets), c represents different subsets, for example, bee colony 1, bee colony 2, …, bee colony C, and the value range of c is the number of bee colonies.

[0038] Each bee colony corresponds to a feature subset , and the fitness of the bee colony is defined as:

[0039] is the average reward of the subset in the classification task; is a subset size penalty term (encouraging to reduce the feature set); , is a weighting coefficient for balancing performance and size, usually . , for example, preferably, =0.8, =0.2.

[0040] In this step S122, the implementation process of the Q-learning strategy of reinforcement learning is: the binary representation of each feature subset , wherein 1 represents retaining the feature and 0 represents eliminating, as a state; the action of the agent is to flip the value of a bit in the state, that is, to add or delete a feature; the Q-table is used to store the long-term value expectation of performing a specific action in a specific state.

[0041] S123: According to the fitness ranking, eliminate the feature subsets with low fitness, copy and strengthen the neighborhood around the feature subsets with high fitness, and dynamically decide whether to retain the features through the Q-learning strategy of reinforcement learning, and finally output the optimal feature subset that maximizes the fitness function.

[0042] In this step S123, the reinforcement of its surrounding neighborhood search range specifically means: for the feature subset with the highest fitness (leading bee), a certain proportion (for example, 10%-20%) of bits in its binary representation is randomly selected to flip, to generate a new candidate feature subset (following bee), so as to simulate the process of local exploration around the excellent solution.

[0043] The final output optimal feature subset The expression is:

[0044] Among them, : The final output optimal feature subset, which comprehensively evaluates the performance of the feature set in classification accuracy, stability and biological interpretability. : The fitness of the bee colony.

[0045] S130: Constructing an HPV subtype accurate typing and variation detection model based on multi-task deep learning, deep encoding and weighted fusion of the optimal feature subset through unstable attention mechanism, generating a global feature representation that can sensitively reflect the key mutation region of HPV sequence; based on the global feature, the HPV subtype typing task and the virus variation detection task are simultaneously performed, and finally two key prediction results of subtype probability and mutation site probability are output.

[0046] After the feature selection in step S120 is completed, the high-dimensional optimal feature subset Input to the constructed HPV subtype accurate typing and variation detection model, wherein the optimal feature subset Contains k-mer structural features of HPV virus sequences, host gene expression signals and clinical state features. This step S130 aims to establish a deep neural network with multi-task processing capability to realize simultaneous learning of subtype typing and variation detection.

[0047] Figure 3 is the structure diagram of the HPV subtype accurate typing and variation detection model of the embodiment of the present application. As shown in Figure 3 The structure of the HPV subtype accurate typing and variation detection model based on multi-task deep learning includes: (1) Feature encoder, used for encoding the input optimal feature subset; In the embodiments of the present application, the feature encoder can be composed of a multi-layer perception mechanism. In a preferred embodiment, the feature encoder includes two fully connected layers, wherein the first layer uses a ReLU activation function, and the second layer uses a linear activation. The role is to map the input high-dimensional features to a more representative hidden space.

[0048] (2) Unstable attention mechanism module, which receives the output of the feature encoder and dynamically adjusts the feature weight of different sequence positions, by introducing a time stationarity control variable that is adaptively adjusted with the fluctuation degree of the sliding window, so that when a mutation peak of a key gene fragment appears, the attention focus on this area is automatically enhanced.

[0049] In order to capture the non-stationary mutation signal in the HPV gene sequence (such as the short-term fluctuation of the E6 / E7 gene mutation), an unstable attention mechanism is introduced, the core of which is to dynamically adjust the feature weight of different sequence positions. The working mode of the unstable attention mechanism module is: 2.1 Calculate Prompt features as the input of Query in the attention mechanism according to the optimal feature subset ; Query generation: first, perform layer normalization on the encoded feature sequence, and introduce learnable parameters γ and δ for scaling and offset to generate dynamic Prompt features , and then obtain the Query vector Q through linear transformation.

[0050] Specifically, after normalizing the optimal feature subset to remove overall offset and capture sequence signal fluctuations, the learnable scaling and offset parameters are combined to adapt the dynamic distribution of different samples to generate the Query input of unstable attention. The Query input of unstable attention is :

[0051] : the optimal feature subset obtained after step S120; : the mean in the sample sequence position dimension, used to remove overall offset; : sequence position variance, used to capture sequence signal fluctuations; : a small constant to prevent the denominator from being zero; , : learnable scaling and offset parameters for adapting the dynamic distribution of different samples; : Prompt features generated after normalization, used to construct the Query input of unstable attention , whose expression is:

[0052] 2.2 Introduce time stationary control variable and an offset correction vector to calculate the similarity of Query and Key; wherein the time stationary control variable is obtained by calculating the average value of the fluctuation of the sequence signal under different size sliding windows; In the embodiment of the application, when calculating the similarity of Query and Key , a time stationary control variable is introduced. The calculation method of is as follows: taking position i as the center, respectively taking sliding windows with lengths of 5, 15 and 30 bases, calculating the standard deviations of the features in the windows , , , and then calculating the arithmetic mean value thereof. In this way, when the model detects that the fluctuation of a certain region (such as the E6 / E7 gene) is greatly increased, the value of can be automatically increased, and the attention to the segment is enhanced.

[0053] Specifically: the dynamic attention weight is calculated by introducing the time stationary control variable and the offset correction vector, so as to realize the dynamic adjustment of the attention with the sequence position:

[0054] , wherein, : Query and Key vectors of the i and j sequence positions, respectively; : Similarity of Query and Key, used to measure the feature correlation degree between different sequence positions; : Attention feature dimension; : Offset correction term, which is a learnable scalar parameter, whose initial value is set based on the relative distance of the sequence positions i and j (|i-j|), and is optimized through gradient descent in the model training process, and is used to capture the position offset of the mutation point relative to the sequence trunk region; : Control variable, defined by the stability indicators of different windows, representing the average fluctuation of the HPV gene sequence under different sliding windows (such as 5, 15 and 30 base sliding windows), when the mutation peak appears in the E6 / E7 gene segment, , the system automatically enhances the attention response of the region:

[0055] : Standard deviation calculated by taking 5 bases as the sliding window, capturing short-range mutations or minor structural variations; ​​: Standard deviation calculated with a window of 15 bases, used to reflect the medium-range mutation trend (e.g. local instability of repetitive regions or deletion fragments); : Standard deviation calculated with a window of 30 bases, used to depict the long-term fluctuation characteristics within larger gene fragments (e.g. L1, L2 regions).

[0056] wherein, in the calculation of the time stationary control variable , the sliding window slides on the sequence in a center-aligned manner. For positions that cannot satisfy the complete window size at both ends of the sequence, zero padding or repeating the boundary value is used to fill in, so as to ensure that the corresponding value can be calculated for all positions i.

[0057] According to the above embodiment of the present application, the time stationary control variable is introduced, and the attention mechanism can dynamically adapt to the non-stationary mutation characteristics of the HPV gene sequence. When a short-term mutation peak occurs in a key gene fragment such as E6 / E7, the value automatically increases, thereby significantly increasing the proportion of this region in the attention weight, effectively solving the technical problem in the background art that the static attention mechanism cannot timely increase the weight of the key region, resulting in the key mutation signal being submerged.

[0058] 2.3 Calculate the attention weight based on the adjusted similarity, so as to realize adaptive capture of non-stationary mutation signals.

[0059] Specifically, the adjusted similarity is normalized by a Softmax function, and the attention weight is calculated by the following formula:

[0060] wherein, : Attention weight, which makes the model focus on the high fluctuation fragment (such as mutation site) and suppress the response in the stable fragment (such as conservative region), thereby improving the sensitivity to non-stationary genetic characteristics, : Normalization function, which maps the prediction result to a probability distribution of .

[0061] 2.4 According to the attention weight , the value vector V is weighted and summed to calculate the weighted output feature , and the attention-weighted output features of all positions are averaged and pooled to form a global fusion feature representation , which is used for the subsequent step (3) of double-task prediction, wherein, the expression is:

[0062] wherein, : the attention-weighted output feature, representing the comprehensive biological representation of the HPV gene sequence at the i-th position, used to capture the local mutation signal and context-dependent relationship. : the Value vector corresponding to the feature representation of the Key, containing the local context information extracted by the encoder (such as base mutation pattern, GC content, fragment stability, etc.). L: the total length of the HPV gene sequence (i.e. the number of bases in the input sequence), used to control the range of attention weight and feature aggregation. : the global fusion feature representation, obtained by aggregating all : the global fusion feature representation, obtained by aggregating all

[0063] (3) a dual-task output layer, which receives the global feature representation processed by the unstable attention mechanism module, to simultaneously perform the HPV subtype typing task and the virus mutation detection task.

[0064] wherein, the dual-task output layer shares the global fusion feature , including: a subtype typing head composed of a fully connected layer connected with a Softmax function, which uses a Softmax classifier to predict the subtype of the HPV sample, and outputs the probability distribution of all preset HPV subtypes , and calculates the subtype classification loss; a mutation detection head composed of a fully connected layer connected with a Sigmoid function, which uses a Sigmoid function to predict the position and type of gene mutation, and outputs the probability of mutation at each site , and calculates the mutation detection loss. The subtype typing head and the mutation detection head both share the fusion feature vector obtained above , realizing the joint learning of HPV gene sequence typing and mutation prediction. Specifically: 3.1 Constructing the subtype typing head: Using a Softmax classifier to predict the subtype of the HPV sample:

[0065] wherein, : the global fusion feature vector after attention weighting and average pooling, containing the high-dimensional semantic and fluctuation feature information of the HPV sequence; : the learnable weight matrix of the subtype classification layer, projecting the feature space to the class probability space; : the bias term of the subtype classification layer, used for offset adjustment after linear transformation; : The output HPV subtype probability distribution vector, such as HPV16, HPV18, HPV52, HPV58, etc. The corresponding loss function is defined as cross-entropy loss:

[0066] Where, M: the number of subtype categories (i.e., the number of HPV types); : One-hot encoding of the true label; : The probability of the m-th class predicted by the model; : Classification error of the typing task, used to measure the difference between the model prediction and the true label.

[0067] 3.2 Building the variant detection head: The variant detection head is used to predict the location and type of gene mutations (SNV / InDel):

[0068] Where, : Weight matrix of the variant detection layer, which maps the fused features to the mutation detection space; : Bias term of the variant detection layer; : Sigmoid function, which limits the output to the interval, used for multi-label mutation prediction; : Predicted mutation probability vector, corresponding to the possibility of each site having a mutation (1) or not having a mutation (0).

[0069] The loss function for variant detection is binary cross-entropy loss:

[0070] Where, N: the number of mutation sites or the total number of samples; : The n-th mutation label (0 or 1); : The mutation probability of the n-th site predicted by the model; : Mutation detection error, used to measure the accuracy of the model's mutation prediction.

[0071] Further, during the model training phase, a weighted joint loss function is used to optimize the HPV subtype accurate typing and variant detection model end-to-end, ensuring that the two tasks promote each other and improve performance together; Where, the weighted joint loss function is the weighted sum of the subtype typing task loss, the variant detection task loss, and the regularization term, to achieve joint optimization of the two tasks. Specifically, the weighted joint loss function is represented as:

[0072] Where, : total loss function, jointly optimizing subtype classification and variant calling tasks; : loss for subtype classification task, : loss for variant calling task; : network parameters : L2 norm penalty term, limiting model complexity; : set of all trainable parameters in the model; : subtype classification task weight coefficient, used to adjust the proportion of the influence of subtype classification in joint optimization; : variant calling task weight coefficient, used to balance the loss contribution of the two tasks; : L2 regularization coefficient, used to constrain the size of network parameters to prevent overfitting. In one specific embodiment, the weight coefficients are set to = 0.6, = 0.4, = 0.001. By gradient descent algorithm to minimize the total loss, the model parameters are optimized, and finally a model that can simultaneously complete high-precision classification and variant calling is obtained.

[0073] The final optimization goal is:

[0074] wherein, is the optimal parameter set obtained after joint task optimization.

[0075] The optimization process is realized by gradient descent and back propagation algorithm, so that the model achieves optimal performance in both HPV subtype identification and mutation detection.

[0076] According to the above embodiment of the present application, a fast processing method for HPV virus gene sequence data is provided. The intelligent exploration and efficient screening of high-dimensional virus-host joint feature space are realized by using the fusion reinforcement learning of the bee colony optimization algorithm, which significantly reduces the feature dimension under the premise of ensuring biological significance. By designing an unstable attention mechanism, introducing a time stationarity control variable that adaptively adjusts with sequence fluctuations, dynamic focusing of key gene fragment mutation peaks such as E6 / E7 is realized, effectively capturing non-stationary genetic patterns. Through double-task joint optimization, shared coding features and weighted loss function, synchronous and accurate identification of HPV subtype classification and variant calling is realized. Finally, the present application realizes efficient processing of HPV virus data, sensitive detection of key mutation signals, and accurate implementation of subtype-variant joint analysis, providing a new technical path for molecular diagnosis of HPV-related tumors.

[0077] It should be noted that the various embodiments described in the specification are intended to be exemplary only and that the scope of the application is not intended to be limited to the embodiments described in the specification. Rather, the scope of the application is intended to cover all modifications and variations of these embodiments and other embodiments that fall within the scope of the application. It is intended that the scope of the application shall, therefore, be limited by the claims.

[0078] It should also be noted that, in the specification, relational terms such as first and second, and the like can be used solely to distinguish one from another entity or action without necessarily implying a required order of such entities or actions. Moreover, the terms "include," "have," or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. In addition, the word "comprising" does not exclude the presence of elements or steps other than those listed in a process, method, article, or apparatus.

[0079] The above description of disclosed embodiments is intended to be illustrative only and not limiting of the application. Many modifications and variations of the described embodiments are possible and will be apparent to those of ordinary skill in the art without departing from the scope of the application. Accordingly, the scope of the application should be determined not with reference to the above description, but with reference to the appended claims, along with their full scope of equivalents.

Claims

1. A rapid processing method for HPV viral gene sequence data, characterized in that, Includes the following steps: HPV virus sequence data, host transcriptome data, and clinical correlation data are acquired and integrated to form a comprehensive input matrix; A bee colony optimization algorithm based on reinforcement learning is used to select features from the comprehensive input matrix to obtain the optimal feature subset. The bee colony optimization algorithm combines the global search of the bee colony algorithm with the dynamic decision-making mechanism of reinforcement learning, and uses the performance of virus subtype classification and the accuracy of mutation prediction as reward feedback to dynamically adjust the feature selection strategy. A precise HPV subtype classification and mutation detection model based on multi-task deep learning is constructed. The optimal feature subset is deeply encoded and weighted through an unstable attention mechanism to generate a global feature representation that can sensitively reflect the key mutation regions of HPV sequences. Based on this global feature, HPV subtype typing and viral mutation detection tasks are executed simultaneously, ultimately outputting two key prediction results: subtype probability and mutation site probability.

2. The method according to claim 1, characterized in that, in, The specific bee colony optimization algorithm based on reinforcement learning is as follows: Each feature subset is considered as a bee colony search direction; In each iteration, the performance of each feature subset is evaluated based on a fitness function that integrates validation set classification accuracy, mutation prediction error, and biological relevance score based on the importance of biological pathways. Based on fitness ranking, feature subsets with low fitness are eliminated, while feature subsets with high fitness are replicated and reinforced in their neighborhoods. The Q-learning strategy of reinforcement learning is used to dynamically decide whether to retain features, and finally, the optimal feature subset that maximizes the fitness function is output.

3. The method according to claim 1, characterized in that, The structure of the HPV subtype accurate typing and variant detection model includes: A feature encoder is used to encode the optimal subset of features from the input; The unstable attention mechanism module receives the output of the feature encoder and dynamically adjusts the feature weights at different sequence positions. By introducing a time stationarity control variable that is adaptively adjusted according to the fluctuation of the sliding window, it automatically enhances attention focus on the region when a mutation peak occurs in a key gene fragment. The dual-task output layer receives the global fusion feature representation processed by the unstable attention mechanism module to simultaneously execute the HPV subtype typing task and the virus mutation detection task.

4. The method according to claim 3, characterized in that, in, The unstable attention mechanism module works as follows: The Prompt feature is calculated based on the optimal feature subset and used as the input to the Query in the attention mechanism; A time-stationarity control variable and an offset correction vector are introduced to calculate the similarity between the query and the key; wherein, the time-stationarity control variable is obtained by calculating the average value of the sequence signal fluctuation under different sliding window sizes; The adjusted similarity is normalized by using the Softmax function, and the attention weight is calculated to achieve adaptive capture of non-stationary abrupt change signals. The Value vector is weighted and summed according to the attention weights to obtain the weighted output features. The attention-weighted output features at all positions are then averaged to form a global fusion feature representation, which is used for subsequent dual-task prediction.

5. The method according to claim 4, characterized in that, The step of calculating the Prompt feature based on the optimal feature subset, and using it as input to the Query in the attention mechanism, includes: By normalizing the optimal feature subset, removing the overall offset, and capturing the fluctuations in the sequence signal, the query input for the unstable attention is generated by combining learnable scaling and offset parameters to adapt to the dynamic distribution of different samples.

6. The method according to claim 3, characterized in that, in, The dual-task output layer shares the global fusion feature, including: The subtype classification header consists of a fully connected layer connected to a Softmax function. It uses a Softmax classifier to predict HPV subtypes from HPV samples and outputs the probability distribution of all preset HPV subtypes. And calculate the subtype classification loss; The mutation detection head consists of a fully connected layer connected to a sigmoid function. It uses the sigmoid function to predict the location and type of gene mutations and outputs the probability of mutation at each site. And calculate the mutation detection loss.

7. The method according to claim 6, characterized in that, During the model training phase, the HPV subtype accurate typing and variant detection model is optimized end-to-end using a weighted joint loss function to ensure that the two tasks promote each other and improve performance together. The weighted joint loss function is a weighted sum of the subtyping task loss, the mutation detection task loss, and the regularization term, to achieve joint optimization of the two tasks.

8. The method according to claim 7, characterized in that, The weighted joint loss function is expressed as follows: in, The loss due to subtyping tasks, For the loss of mutation detection tasks, For the L2 regularization term of the model parameters, and These are the corresponding weight coefficients; The final optimization goal is: in, This is the optimal parameter set obtained after joint task optimization.

9. The method according to claim 1, characterized in that, in, The HPV virus sequence data includes whole genome sequences and gene fragment sequences, and an initial virus feature matrix is ​​formed by k-mer frequency encoding and base physicochemical feature vectorization.

10. The method according to claim 1, characterized in that, in, The host transcriptome data were obtained from RNA-Seq sequencing results of cervical tissue, and significantly differentially expressed genes were screened out through differential expression analysis to construct a host expression matrix.

Images