A snv-cancer association prediction system based on multi-output hierarchical perception

Abstract

The application provides a SNV-cancer correlation prediction system based on multi-output hierarchical perception, and belongs to the technical field of bioinformatics and artificial intelligence. The system comprises an input module, a feature coding module, a hierarchical multi-output prediction module and a double-path hierarchical perception loss calculation module. The input module receives original data of SNV and converts the original data into an input vector, which is sent to the feature coding module. The feature coding module converts the original input into a joint representation vector that can be learned by the model, and inputs the joint representation vector into the hierarchical multi-output prediction module. The hierarchical multi-output prediction module performs collaborative prediction based on the joint representation vector, and outputs the correlation prediction probability of SNV and each cancer type. The prediction probability and the real label are jointly input into the double-path hierarchical perception loss calculation module, and the training optimization of the system is driven through the composite loss. The application solves the technical problems of the lack of hierarchical logic consistency in multi-dimensional correlation prediction and the label noise caused by sparse data labeling.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and artificial intelligence, specifically, it relates to a SNV-cancer association prediction system based on multi-output hierarchical perception. Background Technology

[0002] With the rapid development and widespread application of high-throughput sequencing technology, a vast number of single nucleotide variants (SNVs) have been identified. Of these SNVs, only a small number have been confirmed to be associated with specific cancers, while most are classified as "variables of indeterminate significance" (VUS). To advance precision oncology and personalized treatment, deciphering the potential associations between these VUSs and specific cancers is crucial.

[0003] Currently, traditional statistical methods (such as genome-wide association studies (GWAS)) are widely used to reveal the association between variants and diseases, but they suffer from statistical power bottlenecks when faced with a large volume of potential associations. In recent years, methods based on machine learning and deep learning have gradually gained popularity. However, existing methods are mostly limited to the single-dimensional prediction of macroscopic pathogenicity (i.e., only determining whether a variant is pathogenic), making it difficult to directly apply to the association analysis between SNVs and multiple specific cancer types, and failing to meet the needs of precision oncology for SNV-cancer type specific matching.

[0004] Predicting the association between SNVs and multiple specific cancer types can be modeled as a multi-output prediction problem, which can be solved using a multi-output modeling framework. However, when directly applying the multi-output framework to this task, the following two key issues need to be addressed: 1. Logical Consistency Constraints of Cancer Hierarchy: A clear biological hierarchical relationship exists between cancers and their subtypes (e.g., "non-small cell lung cancer" is a subclass of "lung cancer"). Ignoring this inherent hierarchical logic during multi-dimensional correlation prediction can easily lead to biological contradictions, such as prediction results showing that "SNVs are related to subclasses but not to parent classes." Therefore, effectively integrating the hierarchical structure into the prediction process while ensuring logical consistency across levels is a critical structural challenge that needs to be addressed.

[0005] 2. Label noise caused by sparse data annotation: Existing SNV-cancer association data suffers from incomplete coverage. If an SNV is proven to be associated with cancer in its offspring, it should theoretically also be associated with cancer in its ancestors, but such ancestral associations are often not explicitly labeled in databases. When performing multi-dimensional modeling, directly treating such unlabeled samples as negative examples will introduce severe label noise. Therefore, how to avoid erroneous penalties and improve the ability to learn true potential associations when data is incomplete is a data-level challenge for this task.

[0006] Therefore, there is an urgent need for an SNV-cancer association prediction model that can fully utilize the cancer hierarchy, has logical consistency constraints, and can cope with data noise. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a predictive model for analyzing the association between single nucleotide variants (SNVs) and cancer based on a multi-output hierarchical perception mechanism. This model aims to solve the technical problems of lack of hierarchical logical consistency and label noise caused by sparse data annotation in multi-dimensional association prediction.

[0008] This invention is achieved through the following technical solution: A SNV-cancer association prediction system based on multi-output hierarchical perception: The system includes an input module, a feature encoding module, a hierarchical multi-output prediction module, and a dual-path hierarchical perception loss calculation module. The input module receives the raw data from the SNV and converts it into an input vector, which is then fed into the sequence encoding submodule and the annotation encoding submodule of the feature encoding module. The feature encoding module transforms the original input into a joint representation vector that the model can learn, and the input is directed to the hierarchical multi-output prediction module. The hierarchical multi-output prediction module performs collaborative prediction based on joint representation vectors and outputs the association prediction probability between SNV and each cancer type. The predicted probability and the true label are input together into the dual-path hierarchical perception loss calculation module, and the training and optimization of the system are driven by the composite loss.

[0009] Furthermore, the input module includes a sequence information acquisition submodule and an annotation information acquisition submodule; The sequence information acquisition submodule extracts the reference sequence and the mutated replacement sequence from the reference genome and outputs them to the sequence encoding submodule of the feature encoding module. The annotation information acquisition submodule receives the mutation consequence type and other functional annotation information, and outputs it to the annotation encoding submodule of the feature encoding module.

[0010] Furthermore, the feature encoding module includes a sequence encoding submodule, an annotation encoding submodule, and a feature concatenation unit; The sequence encoding submodule encodes the reference sequence and the substitution sequence and extracts the sequence feature vector. The annotation encoding submodule encodes the annotation information and extracts the annotation feature vector. The feature concatenation unit concatenates the extracted sequence feature vector with the annotation feature vector to obtain a joint representation vector.

[0011] Furthermore, the hierarchical multi-output prediction module constructs an independent prediction pathway for each cancer type, adopting a hierarchical multi-gated hybrid expert architecture, including several shared expert networks and task-specific gating networks corresponding to each cancer type. The shared expert network receives the joint representation vector in parallel and outputs its respective expert representation; The task-specific gating network receives the joint representation vector and outputs the normalized weights of each shared expert under the corresponding cancer type.

[0012] Furthermore, the hierarchical multi-output prediction module also includes a hierarchical representation aggregation unit and a task-specific prediction network corresponding to each cancer type; The hierarchical representation aggregation unit uses the normalized weights of the task-specific gating network output to weight and fuse the outputs of each shared expert network to generate a task-specific intermediate representation corresponding to the cancer type. The task-specific prediction network receives task-specific intermediate representations and outputs the associated prediction probability of the cancer type. The prediction probabilities of all cancer types together constitute the final prediction probability vector.

[0013] Furthermore, in the dual-path hierarchical perception loss calculation module, the composite loss is the total loss obtained by summing the data denoising loss and the weighted structural constraint loss; The data denoising loss is calculated using a hierarchical awareness focus loss. By constructing a tolerable false positive mask, unlabeled ancestor associations that should be exempted from punishment due to hierarchical inheritance relationships are identified, thereby reducing the erroneous punishment of unlabeled ancestor associations. The structural constraint loss is based on the topological structure of the cancer type hierarchical tree, which constrains the predicted probability distribution between parent and child nodes, forces the predicted probability of the child class to not exceed the predicted probability of the parent class, and generates a penalty value only when the predicted probability of the child class exceeds the predicted probability of the parent class. Furthermore, the method for constructing the tolerable false positive mask is as follows: for each SNV, if the true label of a certain cancer type is negative and the cancer type is an ancestor node of at least one positively labeled cancer, then the cancer type is marked as a tolerable false positive. The effective loss mask is defined based on the true label and the tolerable false positive mask. The loss is calculated only for cancer types with a true positive label and cancer types with a true negative label that do not belong to the tolerable false positive category.

[0014] A method for predicting SNV-cancer association based on multi-output hierarchical perception; The method specifically includes the following steps: S1: Data acquisition and preprocessing: Obtain variant-disease association data from the database and filter high-confidence records, limiting the variant type to SNV; map disease names to the disease ontology to obtain hierarchical information, retain only the "cancer" node and all its subtype nodes, and construct the dataset; S2: Dataset partitioning. The dataset is randomly divided into a training set, a validation set, and a test set. The training set is used for iterative updates of model parameters; the validation set is used for monitoring model performance, hyperparameter tuning, and early stopping detection; and the test set is used for the final evaluation of the model's generalization ability. S3: Model building and initialization, building an SNV-cancer association prediction model based on multi-output hierarchical perception; S4: Model training and performance monitoring; Input the training set into the model and construct the optimization objective based on the total loss; Evaluate the current performance of the model using the validation set; S5: Model prediction and hierarchical correction evaluation.

[0015] Furthermore, S3 includes: S3.1: Configure the input module and set the sequence input channel and annotation input channel; S3.2: Configure the feature encoding module, load the pre-trained HyenaDNA genome language model, and initialize the CNN adapter and annotation embedding layer; S3.3: Configure the hierarchical multi-output prediction module, and initialize the shared expert network, the dedicated gating network for each cancer type, and the corresponding prediction head network; S3.4: Configure the dual-path hierarchical perception loss calculation module, and set the balance coefficient, the category balance factor of the focus loss, and the focus parameter.

[0016] Furthermore, S5 includes: S5.1: Association Prediction; Input the sequence features and annotation information of the SNV to be predicted, and output the association prediction probability vector between the SNV and the cancer type through the trained model; S5.2: Test set hierarchical correction evaluation; During the testing phase, an ancestor expansion strategy based on the DO hierarchy tree is performed on the test set labels to obtain hierarchically corrected test labels, and the model performance is evaluated based on these labels. S5.3: Evaluation metrics calculation; Since the predicted probabilities are globally comparable, and the label "0" represents "unlabeled" rather than a clear negative example, the micro-average AUPRC, F1-Score, and Recall are used to measure the overall performance.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieve fine-grained association modeling between SNVs and cancer subtypes: By fusing semantic features and functional annotation information of variant sequences and co-predicting the association probability between SNVs and multi-level cancer types under a unified framework, the model can simultaneously capture the semantic and functional influences of the sequence context of variant sites, thereby establishing a specific mapping relationship between SNVs and specific cancer subtypes, overcoming the limitation of traditional methods that can only make macroscopic pathogenicity judgments.

[0018] 2. Enhance the robustness of the model to sparse labeled data: By using the masking mechanism in the data denoising loss, the ancestor associations that are not explicitly labeled are handled with fault tolerance, avoiding the erroneous gradient penalty caused by incomplete labels, so that the model can still converge stably and maintain high generalization ability in extreme class imbalance and sparse labeling scenarios.

[0019] 3. Ensure the hierarchical logical consistency of prediction results: By using structural constraint loss, the predicted probabilities output by the model are forced to follow the parent-child topological relationship of cancer taxonomy. This eliminates the biological contradiction that "SNVs are related to subclasses but not to parent classes" in the probability space, ensuring the scientific reliability of the prediction results in clinical interpretation. Attached Figure Description

[0020] Figure 1 This is a hierarchical architecture diagram of a multi-output hierarchical perception-based SNV-cancer association prediction model proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Unless otherwise specified, the experimental methods used in the following examples are conventional methods. Unless otherwise specified, the materials, reagents, methods, and instruments used are all conventional materials, reagents, methods, and instruments in the art, and can be obtained commercially by those skilled in the art.

[0023] like Figure 1 As shown, the model of this invention adopts a bottom-up abstract hierarchical organization method, which is displayed from bottom to top as follows: input module, feature encoding module, hierarchical multi-output prediction module and dual-path hierarchical perception loss calculation module.

[0024] The modules work collaboratively according to the following data flow: the input module feeds the raw data into the sequence encoding submodule and the annotation encoding submodule of the feature encoding module, respectively; the joint representation vector output by the feature encoding module is fed into the hierarchical multi-output prediction module, which serves as the core prediction unit and outputs the association prediction probability between SNV and each cancer type; the prediction probability and the true label are jointly input into the dual-path hierarchical perceptual loss calculation module to drive the training and optimization of the model. The specific implementation of each module is as follows: Input module: Configured to receive multimodal feature data of single nucleotide variants (SNVs). Specifically includes: (1) Sequence: Configured to receive a reference sequence and a replacement sequence centered on the variant site, and output them to the sequence encoding submodule; (2) Annotation: Configured to receive the mutation consequence type (including missense mutation, nonsense mutation, etc.) and other functional annotation information, and output it to the annotation encoding submodule.

[0025] Feature encoding module: configured to transform the raw input into a joint representation that the model can learn, including sequence encoding submodule, annotation encoding submodule and feature concatenation unit.

[0026] (1) Sequence encoding submodule: configured to encode the reference sequence and the alternative sequence and extract the sequence feature vector; (2) Annotation encoding submodule: configured to encode the annotation information and extract annotation feature vectors; (3) Feature splicing unit: configured to splice the extracted sequence feature vector with the annotation feature vector to obtain a joint representation vector.

[0027] The hierarchical multi-output prediction module is configured to collaboratively predict the association probabilities of SNVs with multiple cancer types based on joint representations. This module provides a flexible joint prediction space to complement the subsequent dual-path hierarchical perceptual loss, thereby transmitting hierarchical prior knowledge of cancer to the entire model and ensuring that the prediction results maintain logical consistency between levels.

[0028] Dual-path hierarchical perception loss calculation module: configured to optimize model parameters through composite loss. The composite loss is the total loss obtained by summing the data denoising loss and the weighted structural constraint loss, i.e.:

[0029] in, Indicates the total loss; This is the balance coefficient for structural constraint losses; To reduce the data denoising loss, we configure it to be based on the conventional classification loss and introduce a masking mechanism to process sparse labeled data, thereby reducing the erroneous penalty for associating unlabeled ancestors. The structural constraint loss is configured as a topology based on a cancer type hierarchical tree, which constrains the predicted probability distribution between parent and child nodes and forces the model to learn hierarchical logical consistency.

[0030] Example 1: A SNV-Cancer Association Prediction Model Based on Multi-Output Hierarchical Perception This embodiment provides a specific architecture implementation of the prediction model. The model mainly consists of four cascaded parts: an input module, a feature encoding module, a hierarchical multi-output prediction module, and a dual-path hierarchical perceptual loss calculation module.

[0031] In actual deployment, the modules work together using the following specific configurations: 1. Input module The input module is configured to receive the raw data from the SNV and transform it into an input vector that the model can process. Specifically, it includes the following two parts: (1) Sequence information acquisition: Centered on the mutation site, extract the reference sequence and the replacement sequence after mutation from the reference genome. In this embodiment, the sequence length is preferably 1001 bp (i.e., 500 bp upstream and downstream of the mutation site) to fully capture the local sequence context and potential regulatory element information on both sides of the mutation site. All the above sequences are represented in one-hot encoding form, and the encoding dimension corresponds to the size of the nucleotide alphabet.

[0032] (2) Obtaining annotation information: Obtain the annotation information of the SNV, including but not limited to missense mutations, nonsense mutations, frameshift mutations, and other mutation types. In this embodiment, the annotation information is represented by a one-hot vector containing 25 mutation consequence types, with each mutation consequence type corresponding to an independent dimension of the vector.

[0033] 2. Feature encoding module The feature encoding module aims to transform the raw input into feature vectors that can be learned by the deep learning model. It includes a sequence encoding submodule, an annotation encoding submodule, and a feature concatenation unit.

[0034] (1) Sequence coding submodule This submodule is responsible for encoding the reference sequence and the substitution sequence, and extracting the sequence feature vector. The specific implementation is as follows: GLM Selection: Traditional Transformer-based GLMs are limited by the quadratic computational complexity of self-attention, making them inefficient for processing long genome sequences. In contrast, HyenaDNA employs long-range convolutions with subquadratic complexity, improving computational efficiency while maintaining feature representation capabilities. Therefore, this embodiment selects HyenaDNA as the basic encoder. It should be noted that other genome language models with long-range processing capabilities can be used instead, and this invention is not limited thereto.

[0035] Two-stage fine-tuning of GLM: While pre-trained HyenaDNA can extract general genomic sequence features, these features lack specificity for distinguishing cancer-associated SNVs, thus requiring fine-tuning of HyenaDNA. Considering that in multi-label classification scenarios, supervised end-to-end fine-tuning of the backbone GLM could easily lead to gradient conflicts or catastrophic forgetting between different label optimization objectives due to the large amount of parameter updates, this invention designs a two-stage fine-tuning paradigm of "unsupervised contrastive learning fine-tuning + supervised feature adapter fine-tuning": In the unsupervised contrastive learning fine-tuning stage, using the reference sequence as an anchor, positive samples are brought closer and negative samples are pushed away within the HyenaDNA embedding space. The negative samples are cancer-associated SNVs presented as 1001 bp sequences; the positive samples, based on the reference sequences corresponding to the negative samples, generate simulated unassociated background SNVs (also 1001 bp in length) through random single-base substitutions. The substitution sites are located within ±500 bp of the variant sites and avoid the original variant sites. This process shifts the HyenaDNA embedding space towards the cancer-associated context, providing a suitable feature base for subsequent supervised fine-tuning.

[0036] In the supervised feature adapter fine-tuning stage, the parameters of the backbone HyenaDNA are kept frozen, and three parallel lightweight CNN adapters are added for fine-tuning. Each lightweight CNN adapter contains three cascaded units of "one-dimensional convolutional layer + one-dimensional average pooling layer" and a global average pooling layer at the end. The kernel sizes of the three one-dimensional convolutional layers are 7, 7, and 9, respectively, the number of output channels are 32, 64, and 128, respectively, and the dilation values ​​are 1, 2, and 3, respectively. It plays a dual role here: first, as a fine-tuning adapter, it reduces gradient conflicts during multi-label joint optimization by updating only a small number of parameters; second, it performs dimensionality reduction mapping on the high-dimensional features output by HyenaDNA.

[0037] GLM-based sequence encoding: First, high-dimensional embeddings of the reference and substitution sequences are extracted using the fine-tuned HyenaDNA. and And calculate the difference embedding between the two. This is done to explicitly model mutation perturbations. Then, the data is received via the three parallel lightweight CNN adapters. , and Map it to the corresponding feature representation , and Finally, the feature maps output by each CNN are flattened into feature vectors. , and .

[0038] (2) Annotation coding submodule The acquired annotation information is input into the learnable embedding layer, which is then mapped to a dimension of [dimensional value missing]. Continuous annotation feature vector Preferred dimensions in this embodiment .

[0039] (3) Feature splicing unit The sequence feature vector and the annotation feature vector are concatenated to obtain the joint representation vector. :

[0040] 3. Hierarchical multi-output prediction module This module is designed for joint modeling and hierarchical prediction of 360 cancer type labels. It is configured to work in conjunction with the hierarchical perceptual loss, achieving adaptive utilization of upstream shared features through a specific network architecture. Its core principle is to construct an independent prediction pathway for each cancer type, enabling the model to capture shared mechanisms and specific features among different cancers based on shared features, while adhering to the hierarchical logic of the disease ontology.

[0041] As an optimal implementation, this module specifically adopts a hierarchical multi-gated hybrid expert architecture (MMoE). The specific implementation is as follows: (1) Shared Experts: Joint representation vector (in (The concatenated feature dimensions) are input in parallel to... A shared expert network Each expert network is a two-layer fully connected neural network (ReLU activation), with an output dimension of... (In this embodiment) ).Right now:

[0042] In this embodiment, the preferred number of expert networks is... .

[0043] (2) Task-specific gating: Regarding the first For each cancer type, set up a separate task-specific gating network. Its structure is a single-layer fully connected neural network (Softmax activation), receiving the same input. Output 3D normalized weight vector:

[0044] in, , These are learnable parameters.

[0045] (3) Hierarchical representation aggregation: The expert outputs are weighted and fused using gating weights to generate the first... Task-specific intermediate representations for each cancer type :

[0046] This process enables the construction of a "flexible collaborative space," allowing different cancer tasks to dynamically select expert combinations based on their own characteristics, thereby propagating hierarchical priors.

[0047] (4) Task-specific prediction: intermediate representation Enter to the number Task-specific prediction networks for each cancer type (In this embodiment, a two-layer fully connected neural network is used, with the first layer activated by ReLU and the second layer activated by Sigmoid.) The final association probability is generated as follows:

[0048] (5) Output: For all Perform steps (2)–(4) above for each cancer type to obtain the final predicted probability vector:

[0049] 4. Dual-path hierarchical perception loss calculation module This module addresses the dual technical challenges of cancer label noise and biological logic conflict by jointly optimizing data fit and hierarchical structure consistency. A composite loss is calculated based on the predicted results and the true labels, and the model parameters are updated via backpropagation. This loss is the sum of the data denoising loss and the weighted structural constraint loss.

[0050] (1) Data denoising loss. Hierarchical awareness focus loss is adopted. By constructing a tolerable false positive mask, unlabeled ancestor associations that should be exempted from penalty due to hierarchical inheritance are identified. Specifically: Build a False Positive Mask for Each SNV ( (Number of cancer types): If true label And cancer type If it is at least one ancestral node with a positive marker for cancer, then let Otherwise, it is 0. This mask identifies unmarked ancestor associations that should be exempt from penalties due to hierarchical inheritance.

[0051] Define effective loss mask This is used to shield against incorrect penalties for not labeling ancestor associations.

[0052] Focus loss for computational hierarchical perception :

[0053] in, This represents the standard focus loss, which has two hyperparameters: As a category balance factor, To focus the parameters, this embodiment sets... , ; ; These are real labels; It is the predicted probability of the model; It is the number of effective cancer types, when hour, Set to 0.

[0054] (2) Structural Constraint Loss. A hierarchical-aware structural loss method is used. All parent-child pairs in the cancer hierarchy tree are traversed, and the prediction probability of a subclass is forced to not exceed that of its parent class. Specifically: Traverse all parent-child pairs in the cancer hierarchy tree ,in As the parent class, It is a subclass.

[0055] Calculate structural loss :

[0056] This loss only applies to the predicted probability of the subclass. Exceeding the parent class's predicted probability A positive penalty is applied at certain times to force the prediction results to conform to the biological hierarchical logic that "subclass association implies parent class association".

[0057] (3) Total loss. Total loss is defined as:

[0058] in, For the balance coefficient, this embodiment preferably uses... .

[0059] Example 2: A method for predicting SNV-cancer association based on multi-output hierarchical perception.

[0060] This embodiment, based on the prediction model described in Embodiment 1, provides a complete association prediction and training method. The method includes the following steps: Step S1: Data acquisition and preprocessing.

[0061] Step S1.1: Data acquisition.

[0062] Variance-disease association (VDA) data were obtained from the DisGeNet v7.0 database, and records with a score_vda≥0.7 were filtered for high confidence, i.e., records from ClinVar, UniProt, GWASdb, or GWAS Catalog.

[0063] Step S1.2: Data filtering and preprocessing.

[0064] The mutation type is limited to SNV. Disease names are uniformly mapped to the disease ontology to obtain hierarchical information, retaining only the "cancer" node and all its subtype nodes.

[0065] Step S1.3: Construct the dataset.

[0066] Following steps S1.1 and S1.2, a dataset containing 73,852 associated records was obtained, covering 20,157 SNVs and 360 cancer types. This dataset was then used to construct a multi-label supervision matrix. ,in, for A binary matrix, elements Indicates the first The SNV and the first There is a correlation between different cancer types. This indicates that there are no explicitly associated records in the current dataset.

[0067] Step S2: Dataset partitioning.

[0068] The dataset was randomly divided into training, validation, and test sets in a 7:1:2 ratio. The division was based on 20,157 SNVs as the basic unit, with the same division result shared by 360 cancer association prediction tasks. The training set was used for iterative updates of model parameters; the validation set was used for monitoring model performance, hyperparameter tuning, and early stopping detection; and the test set was used for final evaluation of the model's generalization ability.

[0069] Step S3: Model construction and initialization.

[0070] Constructing the SNV-cancer association prediction model based on multi-output hierarchical perception as described in Example 1 specifically includes: Step S3.1: Configure the input module and set the sequence input channel and annotation input channel; Step S3.2: Configure the feature encoding module, load the pre-trained HyenaDNA genome language model, and initialize the CNN adapter and annotation embedding layer; Step S3.3: Configure the hierarchical multi-output prediction module, and initialize the shared expert network, the dedicated gating network for each cancer type, and the corresponding prediction head network; Step S3.4: Configure the dual-path hierarchical perception loss calculation module, and set the balance coefficient, the category balance factor of the focus loss, and the focus parameter.

[0071] Step S4: Model training and performance monitoring.

[0072] Step S4.1: Model training based on the training set The training set is input into the model, and an optimization objective is constructed based on the total loss. The Adam optimizer is used, with an initial learning rate of [value missing]. The batch size is 256, and the maximum training epochs are 100. Network parameters are updated iteratively using the backpropagation algorithm. During training, the loss is calculated using the original sparse labels to simulate incomplete data coverage in the real world.

[0073] Step S4.2: Performance monitoring and early stopping based on the validation set.

[0074] The model's current performance is evaluated using the validation set. During evaluation, an ancestor expansion strategy based on the DO hierarchical tree is applied to the validation set labels (if the SNV is related to descendant cancer, all ancestral cancer labels are padded to 1) to obtain hierarchically corrected validation labels. The micro-average AUPRC is calculated based on these hierarchically corrected labels as a monitoring metric. If this metric does not improve within 5 consecutive epochs, training is stopped and the optimal model weights are saved.

[0075] Step S5: Model prediction and hierarchical correction evaluation.

[0076] Step S5.1: Association prediction.

[0077] Input the sequence features and annotation information of the SNV to be predicted, and output the association prediction probability vector between the SNV and 360 cancer types through the trained model to help determine the cancer-specific association of the SNV.

[0078] Step S5.2: Test set hierarchical correction evaluation.

[0079] During the testing phase, the same ancestor expansion strategy based on the DO hierarchical tree as in step S4.2 is applied to the test set labels to obtain hierarchically corrected test labels. The model performance is then evaluated using these labels as a benchmark to verify the model's robustness to sparse, noisy data and its consistency with biological hierarchies.

[0080] Step S5.3: Calculation of evaluation indicators.

[0081] Since the predicted probabilities are globally comparable, the micro-average is used to measure overall performance. Considering that the label "0" represents "unlabeled" rather than a clear negative example, the micro-average AUPRC is used as the core evaluation metric, supplemented by F1-Score and Recall to evaluate the ability to identify positive samples.

[0082] Performance comparison under different multi-tasking architectures (verification of best practice) 1. Experimental setup Given the current lack of dedicated prediction models for fine-grained SNV-cancer association, the applicant has set up the following two sets of comparative experiments to verify the effectiveness of the proposed SNV-cancer association prediction model based on multi-output hierarchical perception: (1) External comparison model: The heterogeneous graph network MSDRP (originally used for drug-target association prediction) previously proposed by the team was migrated to this task and renamed MVDAP. This model does not have the multi-output prediction module and hierarchical perception mechanism specific to this invention.

[0083] (2) Comparative Examples of Multi-Output Architectures: All models in this group adopted the core feature extraction backbone of this invention and were configured with a multi-output prediction module and hierarchical perception loss. The only difference lay in the internal network architecture of the multi-output prediction module, to verify which architecture best leverages the hierarchical perception mechanism. Specifically, this included: Single-task learning architecture (STL); A multi-task model based on a shared-bottom. Single-gated model based on gated hybrid expert (OMoE); Multi-gated model based on progressive hierarchical extraction (PLE).

[0084] 2. Experimental Results The performance comparison of each model on the test set is shown in Table 1: Table 1. Performance of different embodiments of the present invention compared with external models.

[0085]

[0086] 3. Results Analysis (1) The criticality of multi-output prediction module architecture selection: Experimental results show that, under the premise of using the hierarchical perceptual loss and feature extraction backbone of the present invention, the choice of architecture for the multi-output prediction module has a significant impact on the final performance. The preferred embodiment of the present invention (HAN-MMoE) adopts the MMoE architecture and achieves the best performance, with an AUPRC of 0.8255. Compared with the comparative embodiment using the PLE architecture (AUPRC 0.7949), the performance is improved by approximately 3.9%. This demonstrates that the multi-expert hybrid mechanism of the MMoE architecture and the hierarchical perceptual loss have the best synergistic effect, and can most effectively capture the complex associations between different cancer types.

[0087] (2) Overall advantages of the technical solution of the present invention: Compared to the heterogeneous graph network model MVDAP (AUPRC 0.5527), which lacks the multi-output prediction module and hierarchical awareness mechanism of this invention, all embodiments employing the technical solution of this invention (including HAN-MMoE, HAN-PLE, HAN-STL, etc.) show significant performance improvements. This strongly demonstrates that the overall technical framework of "feature extraction + multi-output prediction + hierarchical awareness" proposed in this invention is key to solving the problem of fine-grained association prediction between SNVs and cancer.

[0088] (3) Model stability: The preferred embodiment of the present invention (HAN-MMoE) not only has the highest accuracy, but also has a standard deviation (±0.0073) that is much lower than that of architectures such as HAN-OMoE (±0.0341), indicating that the MMoE architecture selected in this invention has stronger robustness and convergence stability when used in conjunction with hierarchical perception loss for multi-task optimization.

[0089] Technical module ablation experiment: To verify the effectiveness of each technical module in this invention and its contribution to the overall prediction performance, the applicant conducted an ablation experiment based on the test environment described in Example 1.

[0090] 1. Experimental Results The performance of each model variant on the test set is shown in Table 2: Table 2. Ablation test results.

[0091]

[0092] 2. Results Analysis: (1) The criticality of hierarchical perception loss: As shown in Table 2, when the core "dual-path hierarchical perceptual loss" of this invention was replaced with a conventional classification loss (such as focus loss), the model performance experienced a precipitous drop. AUPRC plummeted from 0.8255 to 0.4224, and both F1-Score and Recall also decreased significantly. This significant difference fully demonstrates that in SNV-cancer association prediction tasks with sparse data annotations and hierarchical logical constraints, conventional classification losses alone cannot effectively address the problems of label noise and logical inconsistencies. The hierarchical perceptual loss designed in this invention plays a decisive role in ensuring the biological logical consistency of the model's predictions and improving overall prediction accuracy; it is an indispensable core feature of this technical solution.

[0093] (2) The necessity of a two-stage fine-tuning strategy: Comparing the experimental results of the "complete model" and the "replacing the two-stage fine-tuning with end-to-end fine-tuning" approach reveals that replacing the "unsupervised contrastive learning + supervised adapter" two-stage fine-tuning strategy of this invention with the conventional end-to-end fine-tuning strategy results in a severe degradation in model performance (AUPRC drops to 0.6522, and F1-Score drops to 0.2774). This indicates that in multi-label cancer association prediction scenarios, simultaneously optimizing the large number of GLM parameters and the entire complex architecture can easily lead to conflicts between optimization objectives for different labels, resulting in serious convergence problems. The two-stage fine-tuning strategy proposed in this invention effectively alleviates the above-mentioned conflicts, strongly demonstrating its irreplaceable necessity in stabilizing the training process, releasing the model's representational capabilities, and ensuring performance.

[0094] (3) Contributions of the Genome Language Model (GLM): Comparing the experimental results of the "complete model" and the "GLM removed" model, we can see that the AUPRC of the model dropped from 0.8255 to 0.8142 after removing GLM. This performance drop indicates that although annotation information and shallow networks can achieve certain results, the GLM introduced in this invention provides crucial deep sequence semantic representations. It can capture implicit regulatory information in the context of long sequence sequences, which cannot be replaced by traditional handmade features or shallow encoding, thus laying a solid feature foundation for high-precision prediction.

[0095] (4) Complementary effects of variation consequence characteristics: Experimental data show that removing the "variation consequence feature" leads to a decrease in AUPRC of approximately 4.4% (0.8255 → 0.7813). This indicates that functional annotation information provides important biological prior knowledge beyond sequence features, forming a good complement to sequence semantic features, and the combination of the two further enhances the model's discriminative ability.

[0096] (5) Robustness of sequence features: Removing any one of the reference sequence, the alternative sequence, or the difference embedding individually only results in a slight decrease in model performance (e.g., AUPRC remains above 0.823), maintaining a high level. This demonstrates that the model architecture of this invention has strong robustness, can flexibly handle different forms of sequence input features, and can maintain stable prediction performance even when some sequence features are missing.

[0097] A computer device system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method. A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0098] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the method described above.

[0099] The above provides a detailed description of the SNV-cancer association prediction system based on multi-output hierarchical perception proposed in this invention, and elucidates the principles and implementation methods of this invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A SNV-cancer association prediction system based on multi-output hierarchical perception, characterized in that: The system includes an input module, a feature encoding module, a hierarchical multi-output prediction module, and a dual-path hierarchical perception loss calculation module. The input module receives the raw data from the SNV and converts it into an input vector, which is then fed into the sequence encoding submodule and the annotation encoding submodule of the feature encoding module. The feature encoding module transforms the original input into a joint representation vector that the model can learn, and the input is directed to the hierarchical multi-output prediction module. The hierarchical multi-output prediction module performs collaborative prediction based on joint representation vectors and outputs the association prediction probability between SNV and each cancer type. The predicted probability and the true label are input together into the dual-path hierarchical perception loss calculation module, and the training and optimization of the system are driven by the composite loss.

2. The system according to claim 1, characterized in that: The input module includes a sequence information acquisition submodule and an annotation information acquisition submodule; The sequence information acquisition submodule extracts the reference sequence and the mutated replacement sequence from the reference genome and outputs them to the sequence encoding submodule of the feature encoding module. The annotation information acquisition submodule receives the mutation consequence type and other functional annotation information, and outputs it to the annotation encoding submodule of the feature encoding module.

3. The system according to claim 2, characterized in that: The feature encoding module includes a sequence encoding submodule, an annotation encoding submodule, and a feature splicing unit; The sequence encoding submodule encodes the reference sequence and the substitution sequence and extracts the sequence feature vector. The annotation encoding submodule encodes the annotation information and extracts the annotation feature vector. The feature concatenation unit concatenates the extracted sequence feature vector with the annotation feature vector to obtain a joint representation vector.

4. The system according to claim 3, characterized in that: The hierarchical multi-output prediction module constructs an independent prediction pathway for each cancer type, adopting a hierarchical multi-gated hybrid expert architecture, including several shared expert networks and task-specific gating networks corresponding to each cancer type. The shared expert network receives the joint representation vector in parallel and outputs its respective expert representation; The task-specific gating network receives the joint representation vector and outputs the normalized weights of each shared expert under the corresponding cancer type.

5. The system according to claim 4, characterized in that: The hierarchical multi-output prediction module also includes a hierarchical representation aggregation unit and a task-specific prediction network corresponding to each cancer type. The hierarchical representation aggregation unit uses the normalized weights of the task-specific gating network output to weight and fuse the outputs of each shared expert network to generate a task-specific intermediate representation corresponding to the cancer type. The task-specific prediction network receives task-specific intermediate representations and outputs the associated prediction probability of the cancer type. The prediction probabilities of all cancer types together constitute the final prediction probability vector.

6. The system according to claim 5, characterized in that: In the dual-path hierarchical perception loss calculation module, the composite loss is the total loss obtained by summing the data denoising loss and the weighted structural constraint loss; The data denoising loss is calculated using a hierarchical awareness focus loss. By constructing a tolerable false positive mask, unlabeled ancestor associations that should be exempted from punishment due to hierarchical inheritance relationships are identified, thereby reducing the erroneous punishment of unlabeled ancestor associations. The structural constraint loss is based on the topological structure of the cancer type hierarchical tree, which constrains the predicted probability distribution between parent and child nodes, forces the predicted probability of the child class to not exceed the predicted probability of the parent class, and generates a penalty value only when the predicted probability of the child class exceeds the predicted probability of the parent class.

7. The system according to claim 6, characterized in that: The method for constructing the tolerable false positive mask is as follows: for each SNV, if the true label of a certain cancer type is negative and the cancer type is an ancestor node of at least one positively labeled cancer, then the cancer type is marked as a tolerable false positive. The effective loss mask is defined based on the true label and the tolerable false positive mask. The loss is calculated only for cancer types with a true positive label and cancer types with a true negative label that do not belong to the tolerable false positive category.

8. A method for predicting SNV-cancer association based on multi-output hierarchical perception, characterized in that, The method is used to execute the SNV-cancer association prediction system based on multi-output hierarchical perception as described in any one of claims 1 to 7; The method specifically includes the following steps: S1: Data acquisition and preprocessing: Obtain variant-disease association data from the database and filter high-confidence records, limiting the variant type to SNV; map disease names to the disease ontology to obtain hierarchical information, retain only the "cancer" node and all its subtype nodes, and construct the dataset; S2: Dataset partitioning. The dataset is randomly divided into a training set, a validation set, and a test set. The training set is used for iterative updates of model parameters; the validation set is used for monitoring model performance, hyperparameter tuning, and early stopping detection; and the test set is used for the final evaluation of the model's generalization ability. S3: Model building and initialization, building an SNV-cancer association prediction model based on multi-output hierarchical perception; S4: Model training and performance monitoring; Input the training set into the model and construct the optimization objective based on the total loss; Evaluate the current performance of the model using the validation set; S5: Model prediction and hierarchical correction evaluation.

9. The prediction method according to claim 8, characterized in that, S3 includes: S3.1: Configure the input module and set the sequence input channel and annotation input channel; S3.2: Configure the feature encoding module, load the pre-trained HyenaDNA genome language model, and initialize the CNN adapter and annotation embedding layer; S3.3: Configure the hierarchical multi-output prediction module, and initialize the shared expert network, the dedicated gating network for each cancer type, and the corresponding prediction head network; S3.4: Configure the dual-path hierarchical perception loss calculation module, and set the balance coefficient, the category balance factor of the focus loss, and the focus parameter.

10. The prediction method according to claim 8, characterized in that, S5 includes: S5.1: Association Prediction; Input the sequence features and annotation information of the SNV to be predicted, and output the association prediction probability vector between the SNV and the cancer type through the trained model; S5.2: Test set hierarchical correction evaluation; During the testing phase, an ancestor expansion strategy based on the DO hierarchy tree is performed on the test set labels to obtain hierarchically corrected test labels, and the model performance is evaluated based on these labels. S5.3: Evaluation metrics calculation; Since the predicted probabilities are globally comparable, and the label "0" represents "unlabeled" rather than a clear negative example, the micro-average AUPRC, F1-Score, and Recall are used to measure the overall performance.

Images