Gene mutation function influence prediction method based on multi-scale deep learning fusion

Abstract

The present application relates to a gene variation function influence prediction method based on multi-scale deep learning fusion, comprising: data preprocessing of gene sequence data, constructing a data set; inputting the data in the data set into a preset deep learning model, training the model, and obtaining a prediction model; inputting the preprocessed gene sequence data to be detected into the prediction model for gene variation function prediction. The present application combines the feature extraction advantages of convolutional neural network and bidirectional long short-term memory network, realizes accurate prediction of gene variation function through multi-scale feature fusion and dynamic optimization strategy, and is suitable for sequence analysis and genome function research in bioinformatics research.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of bioinformatics and artificial intelligence, and in particular to a method for predicting the functional impact of gene variations based on multi-scale deep learning fusion. Background Technology

[0002] In the field of gene variant prediction, traditional computational methods such as SIFT and PolyPhen-2 rely on artificially designed evolutionary conservation and protein structure features. In non-coding region variant analysis, they may struggle to fully capture complex regulatory signals, often achieving prediction accuracy below 50% for rare variants with a frequency of less than 0.1% in the population. Single deep learning models (such as CNNs or RNNs) have limitations in feature extraction: while CNNs excel at local feature recognition, they may be insufficient in modeling long-range dependencies, such as the contextual association between splice signals and upstream / downstream sequences exceeding 300 bp; while RNNs can handle temporal information, they face challenges in efficiently extracting short-scale functional elements. Existing databases typically contain less than 5% of functional variants, and existing models suffer from data imbalance to some extent, with recall rates for rare functional variants generally below 60%. More importantly, most deep learning models lack interpretable design, and the correlation between their prediction logic and biological functions such as promoters and enhancers may be ambiguous, which limits their application value to some extent. Summary of the Invention

[0003] The purpose of this invention is to develop a deep learning model architecture capable of integrating multi-scale feature fusion. This architecture addresses the limitations of traditional methods in non-coding region variation analysis by collaboratively capturing local functional elements and long-range regulatory dependencies in gene sequences. It aims to significantly improve the accuracy of identifying rare variants, particularly functional variants with a frequency below 0.1% in the population, by designing a dynamically weighted loss function that adaptively adjusts based on the misclassification rate of validation set samples. Simultaneously, by optimizing the model architecture and training process, it achieves computational efficiency sufficient for real-time analysis while maintaining prediction accuracy, thereby advancing deep learning technology in genome functional research – a key objective of this invention.

[0004] To achieve the above objectives, the present invention provides the following solution: A method for predicting the functional impact of gene variations based on multi-scale deep learning fusion, characterized by comprising: Gene sequence data are preprocessed to construct a dataset; The data in the dataset is input into a preset deep learning model for model training to obtain a prediction model; wherein, the preset deep learning model is a multi-scale fusion deep learning model that integrates multi-scale CNN and BiLSTM bidirectional attention mechanism. The preprocessed gene sequence data to be detected is input into the prediction model to predict gene variation function.

[0005] Optionally, data preprocessing of gene sequence data includes: Filtering variants, one-hot encoding, Z-score normalization, and SMOTE oversampling.

[0006] Optionally, the preset deep learning model includes: a feature extraction layer, a feature fusion module, and a classifier module; The feature extraction layer is used to capture multi-scale key features of gene sequences; The feature fusion module is used to fuse the multi-scale key features; The classifier module is used to convert the fused features into prediction results.

[0007] Optionally, the feature extraction layer includes: CNN module and BiLSTM module; The CNN module employs parallel multi-scale convolutional layers, covering sequence patterns of different scales from short to long, and is specifically designed to detect known conserved functional element regions in gene sequences; wherein, the functional element regions include: transcription factor binding sites, exon-intron splicing boundaries, CpG islands, etc. The BiLSTM module, combined with a bidirectional attention mechanism, models the long-range dependencies and regulatory logic of sequences, and dynamically identifies key base sites that have the greatest impact on the function of the current variant (such as branch points that affect splicing efficiency) through attention weights, thereby achieving comprehensive modeling of complex regulatory patterns of gene sequences. The BiLSTM module shares a single-hot encoded input with the CNN module.

[0008] The feature fusion module achieves dynamic coupling and integration of local functional element features extracted by CNN and global context-dependent features modeled by BiLSTM through a regularization strategy.

[0009] Optionally, a two-layer fully connected network is used for nonlinear transformation, mathematically expressed as: in, x To fuse feature vectors, W and b are trainable parameters. s Here, z is the Sigmoid function, and z is the pathogenicity probability of the gene variant predicted by the model.

[0010] Optionally, model training includes: We use weighted binary cross-entropy as the loss function to train the model. Based on the preset weight calculation method, we assign high weights to rare mutations and low weights to common mutations. We also introduce a weight smoothing constraint term during the training process.

[0011] Optionally, the loss function is: in, For sample labels, To predict probabilities, For loss function, For the sample i , As a weighting factor, This represents the total number of samples used to train the model.

[0012] Optionally, the preset weight calculation method is as follows: in, β For balancing parameters, As weights, MAF i For the sample i Secondary allele frequencies.

[0013] Optionally, the weight smoothing constraint term is: in, For weighted smoothing constraints, l For harmonic parameters, This represents the change in weight between adjacent epochs.

[0014] The beneficial effects of this invention are as follows: This invention addresses the limitations of traditional methods in non-coding region variation analysis by synergistically capturing local functional elements and long-range regulatory dependencies in gene sequences. Its aim is to significantly improve the accuracy of identifying rare variants, particularly functional variants with a frequency below 0.1% in the population, by designing a dynamically weighted loss function that adaptively adjusts based on the misclassification rate of validation set samples. Simultaneously, by optimizing the model architecture and training process, it achieves computational efficiency sufficient for real-time analysis while maintaining prediction accuracy, thereby advancing deep learning technology in genome functional research – a key objective of this invention. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the gene variation functional impact prediction method based on multi-scale deep learning fusion according to an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this embodiment proposes a method for predicting the functional impact of gene variations based on multi-scale deep learning fusion, including: Gene sequence data are preprocessed to construct a dataset; The data in the dataset is input into a preset deep learning model for model training to obtain a prediction model; the preset deep learning model is a multi-scale fusion deep learning model that integrates multi-scale CNN and BiLSTM bidirectional attention mechanism. The preprocessed gene sequence data to be detected is input into the prediction model to predict the function of gene variations.

[0020] Furthermore, the pre-defined deep learning model includes: a feature extraction layer, a feature fusion module, and a classifier module; Feature extraction layer, used to capture key multi-scale features of gene sequences; The feature fusion module is used to fuse key features at multiple scales; The classifier module is used to transform the fused features into prediction results.

[0021] Furthermore, the feature extraction layer includes: CNN module and BiLSTM module; The CNN module employs parallel multi-scale convolutional layers, covering sequence patterns of different scales from short to long, and is specifically designed to detect known conserved functional element regions in gene sequences. These functional element regions include transcription factor binding sites, exon-intron splicing boundaries, CpG islands, etc. The BiLSTM module, combined with a bidirectional attention mechanism, models the long-range dependencies and regulatory logic of sequences, and dynamically identifies key base sites that have the greatest impact on the function of the current variant (such as branch points that affect splicing efficiency) through attention weights, thereby achieving comprehensive modeling of complex regulatory patterns of gene sequences. The BiLSTM module shares a single-hot encoded input with the CNN module.

[0022] Furthermore, the feature fusion module achieves dynamic coupling and integration of local functional element features extracted by CNN and global context-dependent features modeled by BiLSTM through a regularization strategy.

[0023] Furthermore, the classifier module employs a two-layer fully connected network for nonlinear transformation.

[0024] Furthermore, model training includes: We use weighted binary cross-entropy as the loss function to train the model. Based on the preset weight calculation method, we assign high weights to rare mutations and low weights to common mutations. We also introduce a weight smoothing constraint term during the training process.

[0025] Specifically, in this embodiment, a gene variation functional impact prediction method based on multi-scale deep learning fusion is proposed. The essence of a gene sequence is a linear arrangement of bases (A / T / C / G), and its functional characteristics are often determined by local sequence patterns, such as regulatory elements and protein binding sites. In terms of system architecture, the input layer adopts a one-hot encoding method for gene sequences, transforming them into an L×4 matrix (L is the sequence length), laying the foundation for subsequent feature extraction.

[0026] The feature extraction layer comprises a parallel multi-scale CNN module and a BiLSTM module. The core advantage of CNN lies in extracting hierarchical features through local receptive fields. However, gene regulation mechanisms exhibit multi-scale feature fusion characteristics: short-range patterns (such as transcription factor binding motifs, typically 5-15 bp) coexist with long-range patterns (such as chromatin loops, reaching several kb). Therefore, this embodiment innovatively employs parallel multi-scale convolutional layers (kernel sizes: 5, 10, 15) to precisely locate key regions such as transcription factor binding sites, exon boundaries, and CpG islands.

[0027] The CNN module extracts features primarily based on local, spatially correlated sequence patterns. Utilizing parallel multi-scale convolutional kernels (sizes 5, 10, and 15), this module is specifically designed to identify conserved functional elements at different scales within gene sequences, such as: Nuclear size 5: Matches and captures short-range regulatory signals such as transcription factor binding motifs.

[0028] Nuclear size 10: Covers and identifies variant signals associated with exon-intron boundaries (such as splicing donor / acceptor sites).

[0029] Nuclear size 15: Captures longer fragment patterns associated with epigenetic modifications such as CpG island methylation.

[0030] In short, the CNN module is responsible for "seeing" local, key regional features with specific biological functions from the sequence.

[0031] The BiLSTM module, combined with a bidirectional attention mechanism, not only captures long-range dependencies in sequences but also enhances the contributions of key positions such as splice sites, achieving comprehensive modeling of complex patterns in gene sequences. Sharing a one-hot encoded input (dimension: L×4) with the CNN module, BiLSTM analyzes the functional semantics of the sequence base by base based on its temporal characteristics. The bidirectional structure contains two independent LSTM layers, one forward and one backward, traversing the sequence from 5'→3' and 3'→5' respectively. The BiLSTM module primarily extracts long-range, time-series-related context dependencies. Combining a bidirectional structure and an attention mechanism, this module can: By scanning sequences from both the forward and reverse directions, we can understand long-range dependencies between bases, such as enhancer-promoter interactions spanning several kb.

[0032] The attention mechanism dynamically strengthens the weights of key base positions that contribute significantly to prediction, regardless of where these positions are located in the sequence.

[0033] In short, the BiLSTM module is responsible for "understanding" the global story and contextual logic of the sequence, capturing complex regulatory relationships that are difficult to capture with local features.

[0034] The feature fusion module receives the 128-dimensional "local feature vector" output by the CNN module and the 128-dimensional "global context vector" output by the BiLSTM module. Its core structure is as follows: The concatenation operation combines two 128-dimensional vectors into a single 256-dimensional fused feature vector. This method avoids information dilution and preserves the original semantics of both features.

[0035] Regularization and Transformation: The concatenated features undergo nonlinear transformation and dimensionality reduction (256-dimensional → 64-dimensional) through a network containing two fully connected layers. Key regularization strategies are applied during this process: Dropout randomly masks some neurons during training (rate=0.5) to prevent overfitting; L2 weight decay constrains model parameters and improves generalization ability.

[0036] Output: The probability of pathogenicity of the mutation is finally output through a Sigmoid function.

[0037] The loss function uses a dynamic weighted optimization algorithm. The rationale for using a weighted loss function is that assigning weights to each sample maximizes model learning, thus treating more critical samples. This embodiment uses weighted binary cross-entropy as the loss function: in ∈{0,1} represents the sample label. To predict probabilities, As a weighting factor, This represents the total number of samples used to train the model. Compared to a fixed-weight strategy, dynamic adjustment allows the model to focus on overall convergence in the early stages of training (low-weight phase) and gradually shift its focus to hard example discovery in the later stages (high-weight phase).

[0038] For minority classes (rare mutations), assign higher weights. w i For the majority class (common mutations), lower weights are assigned. The specific calculation method for the weights is as follows: in, β It is a balance parameter, MAF i It is a sample i The minor allele frequency. In this way, the model is ensured to pay more attention to rare variants.

[0039] To prevent gradient explosion caused by a surge in weights, a weight smoothing constraint is introduced: in, This represents the change in weight between adjacent epochs. l=0.1 is the harmonic parameter. Combining the Adam optimizer (learning rate = 0.001) and the early stopping mechanism (patience = 10 epochs), the model converges stably within 50 epochs. The dynamic weighted loss function updates the weights based on the recall rate of rare mutations on the validation set after each training epoch. The initial weights are set according to the inverse probability of class frequency (the weight of rare mutations is 50 times that of the majority class), and the adjustment range is limited to ±10% of the current weight. Optimization stops when the F1 score on the validation set improves by less than 0.01 after 3 consecutive epochs or when the number of weight adjustments reaches 5.

[0040] The parameter configurations for each layer of the model are shown in Table 1 below.

[0041] Table 1 In addition, this embodiment completes the semantic consistency and decision optimization of cross-scale features based on multi-scale feature fusion and classification model design. The main innovation lies in achieving dynamic coupling between CNN local feature learning and BiLSTM contextual information modeling through a regularization strategy. This strategy avoids information dilution caused by early fusion, such as element-wise addition, while preserving the physical meaning of the original features. The fused features undergo nonlinear transformation through two fully connected network layers (256 → 64 units), the mathematical form of which is: in x To fuse feature vectors, W and b are trainable parameters. s For the Sigmoid function, This represents the pathogenicity probability of the gene variant predicted by the model.

[0042] The final classification layer outputs a variant function score (0-1) via a Sigmoid function. The threshold setting (default 0.5) can be adjusted according to research needs. The design and interaction logic of each module in this invention are as follows: Figure 1 As shown.

[0043] The innovations of this embodiment are mainly reflected in the following aspects. First, the multi-scale convolutional kernel collaborative mechanism, through parallel CNN layers with kernel sizes of 5, 10, and 15, covers sequence patterns of different scales from short to long, accurately matching key regions such as transcription factor binding sites (5bp), exon boundaries (10bp), and CpG islands (15bp), achieving comprehensive capture of local features of gene sequences. Second, spatiotemporal attention fusion, the BiLSTM module introduces time-step attention, assigning higher weights to key positions in the sequence (such as splice sites), strengthening their influence on prediction results, and improving the model's ability to model long-range dependencies. Third, the dynamic weight allocation algorithm, dynamically adjusting the weights of the loss function based on the recall rate of the validation set, allows the model to gradually focus on rare variants that are difficult to classify during training, significantly improving the ability to identify high-risk functional variants.

[0044] From a theoretical perspective, the multi-scale feature fusion architecture has, for the first time, achieved hierarchical modeling of "local elements – global regulation" in gene variant prediction. This breakthrough challenges the simplistic assumption in traditional methods that treat sequences as independent sets of features. Through the parallel operation of 5bp, 10bp, and 15bp convolutional kernels, the model captures regulatory mechanisms at different scales, such as transcription factor binding, splicing signals, and epigenetic modifications, confirming the multi-scale synergistic effect of gene regulatory networks. The design of the dynamic weighted loss function provides a new approach for training on imbalanced biomedical data. This mechanism, by simulating the clinical priority given to rare variants, increases the recall rate of variants with MAF < 0.1% from 52% to 80%. This finding may provide methodological reference for the further development of genome function research.

[0045] In terms of practical application value, the robustness of the model in cross-population generalization tests (AUC fluctuation ≤5%) makes it a promising tool for gene detection in multi-ethnic populations. In cross-population tests, the traditional method SIFT showed an AUC decrease of 10%-15%, and DeepVariant showed 8%, while the model of this invention exhibited significantly lower AUC fluctuations in East Asian (4%), African (5%), and European (3.5%) populations—this advantage stems from the ability of multi-scale feature fusion to capture population-specific regulatory elements. However, in tests for extremely rare variants, the model still exhibits a prediction variance of approximately 10%, suggesting that multi-source information, including metadata background, should be considered when conducting genomic function studies.

[0046] Compared with previous studies, this embodiment achieves innovative breakthroughs in three dimensions: At the architectural level, the hybrid design of CNN and BiLSTM overcomes the insufficient modeling of long-range dependencies by DeepVariant (pure CNN), and ablation experiments show that removing BiLSTM leads to a decrease in the recall rate of rare variants; At the algorithm level, the dynamic weighting mechanism improves the recall rate compared to the fixed weight strategy and avoids the problem of training instability; At the interpretability level, SHAP value analysis directly links model decisions with biological concepts such as CpG islands (P=4.5e-8) and splice signals (P=1.2e-6), which is something that black box models such as SpliceAI cannot achieve.

[0047] The main steps of the method in this embodiment include: Input: FASTA format gene sequence data (SNPs / Indels screened by ClinVar, without VUS) Output: Sequence variation function score (0-1 continuous value) Step 1: Preprocess the data: filter for mutations, encode one-hot data into an L×4 matrix, normalize with Z-score, oversample to 1:3 using SMOTE, and split the dataset.

[0048] Step 2: Model building: CNN multi-scale convolutional kernels + BiLSTM bidirectional attention, and output probabilities after feature fusion through fully connected layers.

[0049] Step 3: Train the model: Initialize hyperparameters, dynamically adjust weights using the Adam optimizer, and implement an early stopping strategy to prevent overfitting.

[0050] Step 4: Evaluate performance: Calculate AUC, F1 score, and recall for the test set, and use SHAP analysis to verify the biological association of features.

[0051] Step 5: Application Prediction: Preprocess the sequence to be detected, input it into the model to generate a variant function score and output a report.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the functional impact of gene variations based on multi-scale deep learning fusion, characterized in that, include: Gene sequence data are preprocessed to construct a dataset; The data in the dataset is input into a preset deep learning model for model training to obtain a prediction model; wherein, the preset deep learning model is a multi-scale fusion deep learning model that integrates multi-scale CNN and BiLSTM bidirectional attention mechanism. The preprocessed gene sequence data to be detected is input into the prediction model to predict gene variation function.

2. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 1, characterized in that, Data preprocessing for gene sequence data includes: Filtering variants, one-hot encoding, Z-score normalization, and SMOTE oversampling.

3. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 1, characterized in that, The preset deep learning model includes: a feature extraction layer, a feature fusion module, and a classifier module; The feature extraction layer is used to capture multi-scale key features of gene sequences; The feature fusion module is used to fuse the multi-scale key features; The classifier module is used to convert the fused features into prediction results.

4. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 1, characterized in that, The feature extraction layer includes: CNN module and BiLSTM module; The CNN module employs parallel multi-scale convolutional layers, covering sequence patterns of different scales from short to long, and is specifically designed to detect known conserved functional element regions in gene sequences; wherein, the functional element regions include: transcription factor binding sites, exon-intron splicing boundaries, and CpG islands; The BiLSTM module, combined with a bidirectional attention mechanism, models the long-range dependence and regulatory logic of sequences, and dynamically identifies the key base sites that have the greatest impact on the function of the current variant through attention weights, thereby achieving comprehensive modeling of complex regulatory patterns of gene sequences. The BiLSTM module shares a single-hot encoded input with the CNN module.

5. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 3, characterized in that, The feature fusion module achieves dynamic coupling and integration of local functional element features extracted by CNN and global context-dependent features modeled by BiLSTM through a regularization strategy.

6. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 3, characterized in that, The classifier module employs a two-layer fully connected network for nonlinear transformation, mathematically expressed as: in, x To fuse feature vectors, W and b are trainable parameters. σ Here, z is the Sigmoid function, and z is the pathogenicity probability of the gene variant predicted by the model.

7. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 1, characterized in that, Model training includes: We use weighted binary cross-entropy as the loss function to train the model. Based on the preset weight calculation method, we assign high weights to rare mutations and low weights to common mutations. We also introduce a weight smoothing constraint term during the training process.

8. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 7, characterized in that, The loss function is: in, For sample labels, To predict probabilities, For loss function, For the sample i , As a weighting factor, This represents the total number of samples used to train the model.

9. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 7, characterized in that, The default weight calculation method is as follows: in, β For balancing parameters, As weights, MAF i For the sample i Secondary allele frequencies.

10. The method for predicting the functional impact of gene variations based on multi-scale deep learning fusion according to claim 7, characterized in that, The weight smoothing constraint term is: in, For weighted smoothing constraints, λ For harmonic parameters, This represents the change in weight between adjacent epochs.

Images