Splice variant pathogenicity prediction method and system based on comparative learning

Abstract

The invention relates to the technical field of bioinformatics, and particularly provides a splicing variant pathogenicity prediction method and system based on comparative learning, which only uses DNA sequence information before and after the variation of a splicing variant, and greatly simplifies the feature input of a model. Embedding of sequence information is obtained through a DNA pre-training model, then processing is carried out through a contrast learning module, and for variants showing pathogenicity after variation, the contrast learning module processes the variants into embedding with larger difference; and for the variants which are benign after variation, the difference between the variants is reduced, so that the classifier can more easily capture the subtle difference in the variants, and further efficient pathogenicity prediction is realized. According to the method, learning is directly carried out from the sequence, model deviation caused by different feature selection is reduced, additional errors introduced in biological feature calculation are avoided, the performance on a blind test set shows that the model can accurately predict pathogenic variation, and meanwhile balance, good generalization and robustness are achieved.

Description

Technical Field

[0001] The invention relates to the technical field of bioinformatics, and in particular to a method and system for predicting the pathogenicity of splicing variants based on comparative learning. Background Art

[0002] RNA splicing is a process that removes introns from pre-mRNA and connects exons to obtain mature mRNA. This process is controlled by multiple cis-acting and trans-acting regulatory elements. When these regulatory elements are affected by mutations, the spliceosome may misrecognize the splicing site and produce abnormal splicing variants. Splicing variants may be associated with rare pathogenic variants, cancer, or a variety of genetic diseases. The prediction of the pathogenicity of splicing variants is of great significance for genetic counseling and clinical treatment.

[0003] Traditional biological methods have difficulty simulating the splicing environment when identifying the pathogenicity of splicing variants, and are unable to process large amounts of data information, making it difficult to make accurate assessments.

[0004] Some existing prediction models based on machine learning need to consider multiple complex biological characteristics such as sequence conservation, splicing site strength and secondary structure at the same time. These biological characteristics are obtained through biological experiments or other bioinformatics tools, which takes a lot of time and is prone to introducing noise. Summary of the invention

[0005] The present invention provides a method and system for predicting the pathogenicity of splicing variants based on comparative learning, which at least solves the problem that the prediction of pathogenicity of splicing variants requires reference to a large number of complex biological features, takes a lot of time and is very easy to introduce noise.

[0006] In a first aspect, the present invention provides a method for predicting the pathogenicity of splicing variants based on comparative learning, comprising: Inputting original sequence information and variant sequence information, respectively obtaining a first original embedding vector and a first variant embedding vector; wherein the first original embedding vector is obtained according to the original sequence information, and the first variant embedding vector is obtained according to the variant sequence information; Inputting the first original embedding vector and the first variant embedding vector into a contrastive learning module, adjusting the embedding distance between the first original embedding vector and the first variant embedding vector in the contrastive learning module, and outputting a second original embedding vector and a second variant embedding vector; wherein the second original embedding vector is obtained according to the first original embedding vector, and the second variant embedding vector is obtained according to the first variant embedding vector; Performing feature fusion on the second original embedding vector and the second mutated embedding vector to obtain a fused tensor; The fused tensor is input into a classifier for classification prediction, and a pathogenicity prediction result is output.

[0007] The splicing variant pathogenicity prediction method based on comparative learning provided by the embodiment of the present invention, before inputting the original sequence information and the variant sequence information, the splicing variant pathogenicity prediction method based on comparative learning further includes: Creating a classification prediction model based on the DNA pre-training model, the contrastive learning module and the classifier; wherein the DNA pre-training model is used to extract a corresponding embedding representation based on sequence information; Without access to a classifier, training the DNA pre-training model and the comparative learning module according to training data; wherein the training data includes sample pairs consisting of multiple groups of sequence information; The DNA pre-training model and the comparative learning module are frozen, the classifier is connected, and the classifier is trained according to the training data.

[0008] The method for predicting pathogenicity of splicing variants based on comparative learning provided by the present invention is to train the DNA pre-training model and the comparative learning module according to training data, including: Input the training data into the DNA pre-training model and perform maximum pooling to obtain an embedded representation group; wherein the DNA pre-training model includes initial pre-training parameters; Inputting the embedding representation group into the contrastive learning module to obtain an embedding vector group; In the contrastive learning module, a contrastive learning loss function is used to calculate the contrastive learning loss of the embedding vector group; wherein the formula of the contrastive learning loss function is: ; ; ; is the total contrastive learning loss; is the contrastive learning loss for positive pairs, is the contrastive learning loss for negative sample pairs; is the label of the sample pair; is the Euclidean distance between the embedding vectors before and after mutation in the embedding vector group, is the maximum allowed Euclidean distance among the negative sample pairs, is the number of sample pairs in the training data; The DNA pre-training model and the contrastive learning module are trained until the total contrastive learning loss converges; and the trained DNA pre-training model and the contrastive learning module are obtained.

[0009] The method for predicting pathogenicity of splicing variants based on contrastive learning provided by the present invention is to train the classifier according to the training data, comprising: Input the training data into the trained DNA pre-training model and perform maximum pooling; Inputting the maximum pooling result into the trained contrastive learning module; Inputting the output result of the trained contrastive learning module into the classifier; wherein the classifier uses a cross entropy loss function; The classifier is trained until the classification prediction model loss converges, and the trained classifier is obtained.

[0010] The splicing variant pathogenicity prediction method based on contrastive learning provided by the embodiment of the present invention inputs original sequence information and variant sequence information, and obtains a first original embedding vector and a first variant embedding vector respectively, including: Inputting the original sequence information and the variant sequence information into the trained DNA pre-training model to obtain a first original embedding representation and a first variant embedding representation respectively; Max pooling is performed on the first original embedding representation and the first mutated embedding representation to obtain the first original embedding vector and the first mutated embedding vector, respectively; wherein the first original embedding vector and the first mutated embedding vector are both 768-dimensional vectors.

[0011] The splicing variant pathogenicity prediction method based on contrastive learning provided by the embodiment of the present invention, in the contrastive learning module, adjusting the embedding distance between the first original embedding vector and the first variant embedding vector, and outputting the second original embedding vector and the second variant embedding vector comprises: Adjusting the embedding distance between the first original embedding vector and the first variant embedding vector according to different labels; wherein the embedding distance is set as the Euclidean distance; If the splice variant is a pathogenic variant, making the Euclidean distance between the second original embedding vector and the second variant embedding vector greater than the Euclidean distance between the first original embedding vector and the first variant embedding vector; If the splice variant is a benign variant, the Euclidean distance between the second original embedding vector and the second variant embedding vector is made smaller than the Euclidean distance between the first original embedding vector and the first variant embedding vector.

[0012] The splicing variant pathogenicity prediction method based on contrastive learning provided by the embodiment of the present invention performs feature fusion on the second original embedding vector and the second variant embedding vector to obtain a fusion tensor including: The second original embedding vector and the second mutated embedding vector are stacked to complete feature fusion to obtain the fused tensor; wherein the second original embedding vector and the second mutated embedding vector are both 128-dimensional vectors, and the fused tensor is a two-dimensional tensor of (2, 128).

[0013] The splicing variant pathogenicity prediction method based on contrastive learning provided by the present invention is to input the fusion tensor into a classifier for classification prediction, and output the pathogenicity prediction result including: Input the fused tensor into the trained classifier for binary classification prediction, perform nonlinear division through the activation function of the classifier, and output the classification result; wherein the binary classification prediction result of the classifier is 0 or 1; If the binary prediction result is 1, the pathogenicity prediction result is that the splicing variant has a pathogenic variation; If the binary prediction result is 0, the pathogenicity prediction result is that the splicing variant has a benign variation.

[0014] In a second aspect, the present invention also provides a splicing variant pathogenicity prediction system based on contrastive learning, comprising: a user input module, a sequence information matching module, a splicing variant pathogenicity prediction module and a result feedback and notification module; wherein the splicing variant pathogenicity prediction module comprises the splicing variant pathogenicity prediction method based on contrastive learning described in any of the above embodiments; The user input module is used to input genome sequence information; The sequence information matching module is used to identify the mutation site in the genome sequence information, and extract the DNA sequence information before and after the mutation, centered on the mutation site, and with a window size of 1024 bases as the original sequence information and the variant sequence information, respectively, and then input the original sequence information and the variant sequence information into the splicing variant pathogenicity prediction module; The splicing variant pathogenicity prediction module outputs a pathogenicity prediction result according to the original sequence information and the variant sequence information, and transmits the pathogenicity prediction result to the result feedback and notification module; The result feedback and notification module is used to collect and record the pathogenicity prediction results and display the pathogenicity prediction results.

[0015] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising: a processor, and a memory for storing programs, wherein the program comprises instructions, which, when executed by the processor, cause the processor to execute the splicing variant pathogenicity prediction method based on comparative learning as described in any of the above embodiments.

[0016] The present invention provides a splicing variant pathogenicity prediction method and system based on comparative learning, which uses sequence information as input to achieve splicing variant pathogenicity prediction based entirely on sequence information. Compared with existing methods, it can avoid introducing other biological features such as structural conservation and splicing site strength. On the one hand, learning directly from the sequence can reduce the model bias caused by different feature selection, and thus better adapt to new variant types and data sets, and have better model generalization and robustness. On the other hand, it greatly simplifies the feature input of the model, does not rely on complex feature engineering, avoids the cumbersome process of screening and optimization in multiple biological features, and makes the model more automated and efficient. It reduces the computing time and feature engineering complexity, and is suitable for processing large-scale genetic data.

[0017] The splicing variant pathogenicity prediction method and system based on contrastive learning provided by the present invention apply contrastive learning to the variant prediction task, and use the natural relative characteristics between variants to achieve more accurate pathogenicity prediction. Through contrastive learning, the model enhances its learning ability of key features when the variant characteristics are similar, thereby surpassing the traditional classification prediction model, effectively identifying the subtle differences between different pathogenic variants, and improving the prediction effect of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other embodiments can be obtained based on these drawings without creative work.

[0019] Figure 1 It is a schematic diagram of the principle of the classification prediction model in the splicing variant pathogenicity prediction method based on comparative learning in Example 1 created by the present invention.

[0020] Figure 2 It is a flow chart of the method for predicting pathogenicity of splicing variants based on comparative learning in Example 1 created by the present invention.

[0021] Figure 3 It is a schematic diagram of the splicing variant pathogenicity prediction system based on comparative learning in Example 2 created by the present invention. DETAILED DESCRIPTION

[0022] The embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention. Embodiment 1

[0023] The genes of eukaryotic organisms are chimeric, with coding regions separated by non-coding regions, where coding regions are called exons and non-coding regions are called introns. Removing introns and connecting exons together in mature mRNA is called RNA splicing. Splicing variations mainly include exon skipping, 5' splice site changes, 3' splice site changes, intron retention, and exon mutual exclusion, depending on the form.

[0024] Existing technologies have confirmed that splicing variants are associated with the occurrence of a variety of diseases, including a variety of genetic diseases, neurological diseases and cancer. In cancer, splicing variants may lead to oncogene activation or tumor suppressor gene inactivation, promoting tumor growth and spread. In some neurodegenerative diseases, abnormal splicing processes affect the normal function of neurons, thereby aggravating the progression of the disease.

[0025] Some pathogenic variants will affect the function of cis-regulatory elements, thereby disrupting the normal splicing process. About 15-60% of rare pathogenic variants are located in these cis-regulatory regions, causing abnormal gene expression by generating new splicing sites or interfering with the normal use of splicing sites.

[0026] In addition, splicing variants have other uncertain clinical significance in diagnosis, and it is difficult to accurately determine their pathogenicity. These uncertainties pose challenges to genetic diagnosis and genetic screening, and may affect the formulation of genetic counseling and treatment plans.

[0027] Therefore, splicing variants pose a great threat to human health, and predicting the pathogenicity of splicing variants is crucial for disease prevention and treatment.

[0028] Traditional predictions of the pathogenicity of splicing variants are made through biological methods. Traditional biological methods include reverse transcription (RT-PCR), RNA sequencing (RNA-seq), and minigene assays. Although these biological methods can provide direct experimental evidence for the pathogenicity of splicing variants, they are usually time-consuming and costly, and are difficult to promote in large-scale genetic screening. Traditional methods are difficult to perform high-throughput variant screening and analysis, and cannot quickly process a large number of samples. With the development of genome sequencing technology, the generation of massive data requires more efficient analysis methods, and traditional methods are obviously difficult to meet this demand.

[0029] In addition, the splicing process is tissue or cell type specific, and the splicing patterns in different tissues may be different. Traditional biological methods are difficult to fully simulate the splicing environment of specific tissues or cell types in in vitro experiments, resulting in experimental results that may not represent the actual situation in vivo. In particular, some key disease-related splicing events may only occur in specific tissues. Many splicing variants are located in deep intron regions or non-classical splicing sites, and traditional experimental methods are difficult to accurately assess their effects on splicing. These variants are generally considered to be benign or clinically uncertain variants (Variant of Uncertain Significance, VUS). Traditional biological methods are difficult to interpret these variants and cannot provide clear pathogenicity judgments.

[0030] In view of the limitations of traditional biological methods in processing power and experimental results, existing technologies have introduced machine learning into the prediction of pathogenicity of splicing variants. The prediction model CADD uses up to 60 biological features, covering conservation and structural features, etc., and uses machine learning algorithms such as support vector machines (SVM) to train the model to integrate these different types of biological features and finally generate a prediction score.

[0031] dbscSNV uses an integrated learning algorithm, including the iterative algorithm AdaBoost and random forest RandomForest, and combines multiple features such as splice site strength, sequence conservation, secondary structure information, etc. However, this algorithm is only suitable for identifying the impact of single nucleotide variations (SNV) on splice sites.

[0032] These machine learning-based algorithms need to calculate a variety of complex biological features in advance, such as sequence conservation, splice site strength, secondary structure, and regional guanine and cytosine ratios (GC content). The calculation and selection of these features not only takes a lot of time, but also requires high professionalism in feature engineering, usually requiring the support of genome annotation databases and complex computing resources. Some of these biological features are not directly derived from experimental results, but rely on other bioinformatics tools. A large number of splicing-related biological features also need to be calculated in advance, introducing additional noise and affecting the model's prediction efficiency and accuracy.

[0033] With the continuous development of deep learning, architectures such as convolutional neural networks, recurrent neural networks, and residual networks ResNet have begun to be applied to the prediction of pathogenicity of splicing variants. SpliceAI, developed by the DeepVariant team, is a splicing variant prediction tool based on deep learning. It uses convolutional neural networks to process the nucleotide sequence information of the variant context. The maximum range of the nucleotide sequence information of the variant context is 10kb. Although this can extract local features of the sequence to a certain extent, it is difficult to predict and judge long-distance dependencies, which destroys the relevance and integrity of the information.

[0034] Accordingly, the first embodiment of the present invention provides a method for predicting the pathogenicity of splicing variants based on contrastive learning, which only uses the DNA sequence before and after the splicing variant is mutated as input. Based on the DNA pre-training model, embeddings are obtained respectively, and the prior knowledge of the DNA pre-training model is used to directly extract high-dimensional, multi-level nucleotide local features and global information from the original sequence, and the influence of the splicing process before and after the mutation on the nucleotide sequence and the existence state of biological elements in the DNA sequence are fully considered. The embeddings obtained by the DNA pre-training model are further processed by a contrastive learning module. For variants that are pathogenic after mutation, the contrastive learning module processes them into embeddings with greater differences; for variants that are benign after mutation, the difference between the two is reduced, so that the classifier can more easily capture the difference, thereby achieving efficient pathogenicity prediction.

[0035] Compared with traditional biological detection methods, the present invention can realize the analysis and prediction of massive data and meet the requirements of large-scale gene screening. Compared with machine learning prediction models, it can avoid calculating a large number of biological features related to shearing and mutation, is time-saving and does not introduce additional noise, and has higher prediction efficiency and accuracy.

[0036] The method for predicting pathogenicity of splicing variants based on contrastive learning provided in Example 1 of the present invention comprises the following steps.

[0037] Step S100, creating a classification prediction model and training the classification prediction model.

[0038] As an implementable method, step S100 specifically includes the following steps: Step S110, creating a classification prediction model based on the DNA pre-training model, the comparative learning module and the classifier.

[0039] The DNA pre-training model is used to extract the corresponding embedding representation according to the sequence information. In the first embodiment, the DNA pre-training model is preferably set to DNABert-2.

[0040] In order to solve the problem that traditional languages ​​are difficult to capture information between DNA semantics, Professor Ramana V. Davuluri of Northwestern University and others published an article in the journal Bioinfomatics titled "DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome". The authors proposed a pre-trained bidirectional encoder representation DNABERT to perform global or migration analysis of DNA sequences through contextual information. DNABERT can directly rank the importance of nucleotide molecules and analyze the relationship between the context of the input sequence, thereby obtaining better visualization information and accurate motif extraction. DNABert-2 is a follow-up study of DNABERT that has been made public, which improves word segmentation efficiency and computational efficiency, and reduces sampling redundancy and inefficiency.

[0041] In other embodiments of the present invention, the DNA pre-training model may also use other models that meet the requirements for sequence information extraction and is not limited to DNABert-2.

[0042] Step S120: training the DNA pre-training model and the comparative learning module according to the training data without accessing the classifier.

[0043] For the classification prediction model, in step S120, only the DNA pre-training model and the comparative learning module are connected, and the classifier is not connected.

[0044] The training data includes sample pairs composed of multiple sets of sequence information. Among the sample pairs, the sample pairs with pathogenicity after splicing variation are set as positive sample pairs, and the expected output result corresponding to the positive sample pairs is pathogenicity. The label of the positive sample pairs is In the sample pairs, the sample pairs that are benign after splicing variation are set as negative sample pairs, and the expected output result corresponding to the negative sample pairs is benign. The label of the negative sample pairs is .

[0045] The training data is input into the DNA pre-training model and the maximum pooling is performed to obtain an embedded representation group of the training data; wherein the DNA pre-training model includes initial pre-training parameters.

[0046] The embedding representation group of the training data is input into the contrastive learning module to obtain the embedding vector group of the training data.

[0047] In the contrastive learning module, the contrastive learning loss function is used to calculate the contrastive learning loss of the embedding vector group of the training data. The contrastive learning loss function includes the loss of positive sample pairs and the loss of negative sample pairs.

[0048] The loss formula for positive sample pairs is: ; ; in, is the contrastive learning loss for positive pairs, is the label of the sample pair, for the positive sample pair .

[0049] is the Euclidean distance between the embedding vectors before and after mutation in the embedding vector group of the training data, is the embedding vector of the original sequence before mutation in the training data, is the embedding vector of the mutated sequence in the training data.

[0050] The loss calculation formula for negative sample pairs is: ; in, is the contrastive learning loss for negative sample pairs, is the label of the sample pair, for the negative sample pair . is the maximum allowed Euclidean distance between the embedding vectors before and after mutation in a negative sample pair.

[0051] The total contrastive learning loss is calculated as: .

[0052] in, is the total contrastive learning loss. is the number of sample pairs in the training data.

[0053] The DNA pre-training model and the contrastive learning module are trained until the total contrastive learning loss converges, and the training of the DNA pre-training model and the contrastive learning module is completed. The initial pre-training parameters in the DNA pre-training model are changed, and the trained DNA pre-training model and the contrastive learning module are obtained as training results.

[0054] Step S130, freezing the DNA pre-training model and the comparative learning module according to the training result of step S120, and then connecting to the classifier after freezing to train the classifier. Step S130 specifically includes: The training data in step S120 is input into the DNA pre-training model trained in step S120 and the maximum pooling is performed; the result of the maximum pooling is input into the comparative learning module trained in step S120.

[0055] The output result of the contrastive learning module after training in step S120 is input into the classifier, wherein the classifier uses a cross entropy loss function.

[0056] Preferably, the cross entropy loss function used by the classifier is a binary cross entropy loss (BCELoss) function, which is one of the most commonly used loss functions for binary classification problems. The binary cross entropy loss function is used to calculate the difference between the probability distribution of the model output and the probability distribution of the true label. The formula of the binary cross entropy loss function is: ; in, is the true label, is the predicted probability. When the predicted probability is close to the true label, the value of binary cross entropy decreases; otherwise, it increases. This enables the classification prediction model to improve the accuracy of the prediction by minimizing the loss function.

[0057] Train the classifier until the loss of the classification prediction model converges, and obtain the trained classifier.

[0058] Step S140, the DNA pre-training model and the comparative learning module trained in step S120 are connected to the classifier trained in step S130 to complete the training of the classification prediction model.

[0059] In this embodiment, a step-by-step freezing training framework is designed in step S100, and the DNA pre-training model and the contrastive learning module are trained in stages. After the loss function of the contrastive learning converges, the relevant parameters are frozen, and then the classifier is connected to perform the classification task. Only the classifier loss is considered for back propagation, and the training is performed until the model reaches the optimal performance, so that the training effect and convergence speed of the entire classification prediction model are improved.

[0060] The step-by-step freezing training strategy can effectively ensure that the pre-trained model and the contrastive learning module can stably generate differentiated embedding representations under the guidance of labels, thereby realizing the functional decoupling and complementarity of the contrastive learning module and the classifier within the model. By fixing the model parameters of the pre-trained model and the contrastive learning part, this strategy reduces the risk of overfitting of the pre-trained model for downstream tasks, while enhancing the sensitivity of the contrastive learning module to label information. It not only improves the overall prediction accuracy of the model, but also achieves a better performance balance between positive and negative samples, which is suitable for pathogenicity prediction tasks of various complex variant types.

[0061] Step S200, inputting original sequence information and variant sequence information, and obtaining a first original embedding vector and a first variant embedding vector respectively; wherein the first original embedding vector is obtained according to the original sequence information, and the first variant embedding vector is obtained according to the variant sequence information.

[0062] As an implementable method, step S200 specifically includes the following steps: Step S210: input the original sequence information before mutation and the mutated sequence information after mutation into the DNA pre-training model to obtain a first original embedding representation and a first mutated embedding representation, respectively.

[0063] The DNA pre-training model is the classification prediction model obtained in step S100, specifically the result obtained after training in step S120. The first original embedding representation is obtained by outputting the original sequence information before mutation through the DNA pre-training model, and the first variant embedding representation is obtained by outputting the variant sequence information after mutation through the DNA pre-training model.

[0064] Step S220, performing maximum pooling on the first original embedding representation and the first variant embedding representation to obtain the first original embedding vector and the first variant embedding vector respectively.

[0065] The first original embedding vector is a 768-dimensional vector obtained by performing a maximum pooling operation on the first original embedding representation. The first variant embedding vector is a 768-dimensional vector obtained by performing a maximum pooling operation on the first variant embedding representation.

[0066] Max pooling compresses the embedding representation of the entire sequence into an embedding vector of fixed dimension. In Example 1, the window width of the sequence information before and after the mutation is 1024 bases, and the max pooling operation compresses (1024, 768) to 768 dimensions.

[0067] Through step S200, the nucleotide sequence information can be converted into a vector form composed of real numbers that can be processed by deep learning by capturing potential relationships and structures.

[0068] Step S300: input the first original embedding vector and the first variant embedding vector into a contrastive learning module, adjust the embedding distance between the first original embedding vector and the first variant embedding vector in the contrastive learning module, and output a second original embedding vector and a second variant embedding vector. The second original embedding vector is obtained based on the first original embedding vector, and the second variant embedding vector is obtained based on the first variant embedding vector.

[0069] As an implementable method, step S300 specifically includes: The first original embedding vector and the first variant embedding vector are input into a contrastive learning module; wherein the contrastive learning module is the classification prediction model obtained in step S100, specifically the result obtained after training in step S120. The contrastive learning module includes a Transformer encoder layer and a linear layer.

[0070] According to the difference in sample pair labels, the embedding distance between the first original embedding vector and the first variant embedding vector is adjusted in the contrastive learning module, wherein the embedding distance is set as the Euclidean distance.

[0071] The adjusted results obtained in the contrastive learning module are: The Euclidean distance of the embedding vector before and after the mutation of the pathogenic variant is made as far as possible, and the Euclidean distance of the embedding vector before and after the mutation of the benign variant is made as close as possible.

[0072] That is to say, if the splice variant is a pathogenic variant, the Euclidean distance between the first original embedding vector and the first variant embedding vector is enlarged so that the Euclidean distance between the output second original embedding vector and the second variant embedding vector is at least greater than the Euclidean distance between the first original embedding vector and the first variant embedding vector, and the larger the better.

[0073] If the splice variant is a benign variant, the Euclidean distance between the first original embedding vector and the first variant embedding vector is reduced so that the Euclidean distance between the output second original embedding vector and the second variant embedding vector is at least smaller than the Euclidean distance between the first original embedding vector and the first variant embedding vector, and the smaller the better.

[0074] In the contrastive learning module, the embedding distance is adjusted to make the embedding distance differentiation between pathogenic variants and benign variants more significant, which helps to distinguish the two variants more efficiently and easily in subsequent classification tasks, make judgments on whether they are pathogenic, and improve classification accuracy and efficiency.

[0075] It should be noted that both the second original embedding vector and the second variant embedding vector are 128-dimensional vectors. In the processing of the contrastive learning module, not only the distance relationship between the embedding vectors is adjusted, but also the dimension of the embedding vector is reduced from 768 to 128 through dimensionality reduction. The dimensionality reduction process can effectively extract the most relevant information in the sequence for the prediction task, while removing noise and irrelevant variables, thereby simplifying the learning task of the classification prediction model and achieving a significant noise reduction effect. In addition, dimensionality reduction also reduces the computational complexity of the classification layer, significantly improves the training speed of the model, and reduces the risk of overfitting caused by high-dimensional features, further improving the generalization performance of the model.

[0076] Step S400, feature fusion is performed on the second original embedding vector and the second variant embedding vector output after adjustment in step S300 to obtain a fused tensor. The fused tensor is a two-dimensional tensor of (2, 128).

[0077] Preferably, the feature fusion operation in this embodiment is to stack two vectors. In other embodiments, feature fusion may also use other fusion operations that meet prediction requirements, but is not limited thereto.

[0078] Step S500, input the fusion tensor outputted from step S400 into the classifier for classification prediction, and output the pathogenicity prediction result. The classifier is the classification prediction model obtained from step S100, specifically the result obtained after training in step S130. The classifier consists of a convolutional layer and two linear layers.

[0079] As an implementable method, step S500 specifically includes: The fused tensor is input into the trained classifier for binary classification prediction, and nonlinear division is performed through the activation function of the classifier to output the classification result; the activation function is the PReLU activation function; the binary classification prediction result of the classifier is 0 or 1.

[0080] If the binary prediction result is 1, the pathogenicity prediction result is that the splicing variant has a pathogenic variant; If the binary prediction result is 0, the pathogenicity prediction result is that the splicing variant has a benign variation.

[0081] The SPiCE model, dbscSNV_RF model, SpliceAI model and kipoiSplice model in the prior art were used as four comparative examples, and their performance was compared with the classification prediction model used in the splicing variant pathogenicity prediction method based on contrastive learning provided in Example 1 of the present invention, and the same test data was used for testing.

[0082] Specifically, the test data is set to the standard test set of SPCards, which has eliminated the training data that may be used by other splicing prediction tools and has high reliability. The obtained performance test results are shown in Table 1, where the classification prediction model used in the splicing variant pathogenicity prediction method based on contrastive learning provided in Example 1 is represented by the ACE model.

[0083] Table 1

[0084]

[0085] The results of the performance test include six indicators, namely, positive predictive value (PPV), negative predictive value (NPV), specificity (indicated by Spec in Table 1), sensitivity (indicated by Sens in Table 1), accuracy, Matthews correlation coefficient (MCC) and coverage, all of which are commonly used indicators in binary classification tasks.

[0086] The positive prediction value is also called the positive prediction value. Specifically, it refers to the proportion of samples whose actual values ​​are also "positive" among the samples whose model prediction results are "positive". The calculation formula of the positive prediction value is: ; In the above formula, TP (True Positive) represents the number of actual "positive" and predicted "positive" results; FP (False Positive) represents the number of actual "negative" results but predicted "positive" results.

[0087] Negative predictive value is also called negative predictive value. Specifically, it refers to the proportion of samples whose actual values ​​are also "negative" among the samples whose model predicts the results to be "negative". The calculation formula of negative predictive value is: ; In the above formula, TN (True Negative) represents the number of examples that are actually negative and predicted as negative; FN (False Negative) represents the number of examples that are actually "positive" but predicted as "negative".

[0088] Specificity refers to the proportion of samples that are correctly predicted as "negative" among all samples that are actually "negative". The specificity calculation formula is: .

[0089] Sensitivity refers to the proportion of samples that are correctly predicted as "positive" among all samples that are actually "positive". The calculation formula for sensitivity is: .

[0090] Accuracy refers to the proportion of correct predictions to the total number of predictions. The calculation formula for accuracy is: .

[0091] The Matthews correlation coefficient is an indicator for comprehensively evaluating classification performance. The higher the value, the more robust the model is to imbalanced data. The calculation formula of the Matthews correlation coefficient is: .

[0092] As shown in Table 1, the classification prediction model in Example 1 is superior to other existing prediction models in six indicators: positive prediction value PPV, negative prediction value NPV, specificity Spec, accuracy, Matthews correlation coefficient MCC and coverage, and also shows suboptimal performance in sensitivity Sens.

[0093] Furthermore, the positive prediction value PPV of the classification prediction model increased by 3.1% compared to the second best, the negative prediction value NPV increased by 7.1%, and the Matthews correlation coefficient MCC achieved a 9.8% performance improvement. Due to the imbalance of positive and negative samples in the SPCards standard test set, the higher MCC value shows the overall stability and comprehensiveness of the model's accurate prediction in different categories, which further indicates that the model has good robustness in dealing with data imbalance problems.

[0094] The prediction coverage of the classification prediction model provided in the first embodiment of the present invention reaches 100%. Other models are limited by the calculation defects of biological features and can only predict the pathogenicity of 26.1% to 87.9% of some variants, with poor versatility. The pathogenicity prediction method provided in the first embodiment of the present invention is not limited by the calculation of biological features, and can give high-precision predictions for all variants, with high versatility and high accuracy. Embodiment 2

[0095] Embodiment 2 of the present invention provides a splicing variant pathogenicity prediction system based on contrastive learning. The system comprises: a user input module, a sequence information matching module, a splicing variant pathogenicity prediction module and a result feedback and notification module.

[0096] The splicing variant pathogenicity prediction module includes the splicing variant pathogenicity prediction method based on contrastive learning provided in Example 1, and specifically applies the classification prediction model provided in Example 1.

[0097] Specifically, the user input module is used to input genome sequence information. Users can upload sequence information before and after mutation according to their own needs.

[0098] The uploaded sequence information is determined based on the amount of predicted variant data. The specific upload format is the standard format for variants established by the Human Genome Variation Society (HGVS), which can be used for computer recognition. Users paste the standard format sequence information corresponding to the variant into the text box to upload, or directly upload a csv format file.

[0099] Specifically, the sequence information matching module is used to process and extract sequence information.

[0100] The sequence information matching module matches the sequence information of the variant in the reference genome sequence stored in the cloud based on the user's uploaded data, identifies the mutation site in the genome sequence information, and extracts at least part of the bases centered on the mutation site before and after the mutation, and temporarily stores them as the original sequence information and variant sequence information respectively. The temporarily stored information will be forwarded to the splicing variant pathogenicity prediction module.

[0101] It should be noted that the window width of the extracted base sequence for temporary storage is preferably set to 1024 bases. In other embodiments, the number of bases in the window can also be 128, 256, 512 or 2048, but is not limited thereto. The inventors of this case have found that the setting of the window width affects the time required for the DNA pre-training model to obtain sequence information and output embedded representation. When the window width is set to 1024 bases, the training time and prediction time can be controlled within the acceptable range of the model.

[0102] Specifically, the splicing variant pathogenicity prediction module outputs a pathogenicity prediction result based on the original sequence information and variant sequence information output by the sequence information matching module.

[0103] When the splicing variant pathogenicity prediction module is working, if there is a graphics acceleration processing unit GPU, the GPU is used first to accelerate the prediction task of the classification prediction model; if there is no GPU or the GPU is occupied, the central processing unit CPU is used to load the classification prediction model for pathogenicity prediction.

[0104] The process of determining and outputting the pathogenicity prediction result applies the splicing variant pathogenicity prediction method. Since the splicing variant pathogenicity prediction method has been described in Example 1, the system in this embodiment has all the technical effects and contents of the method in Example 1, which will not be repeated here.

[0105] The splicing variant pathogenicity prediction module obtains the pathogenicity prediction result and transmits it to the result feedback and notification module.

[0106] Specifically, the result feedback and notification module is used to collect and record the pathogenicity prediction results. On the one hand, according to the user's choice, the prediction results are displayed to the user in the form of email attachments or online displays. On the other hand, the calculation records and prediction result files are saved to the cloud for the user. Embodiment 3

[0107] Embodiment 3 of the present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to enable the computer to execute the splicing variant pathogenicity prediction method based on comparative learning provided in Embodiment 1 of the present invention. Embodiment 4

[0108] Embodiment 4 of the present invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a processor of a computer, is used to enable the computer to execute the splicing variant pathogenicity prediction method based on comparative learning provided in Embodiment 1 of the present invention. Embodiment 5

[0109] The fifth embodiment of the invention also provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor. The memory stores a computer program that can be executed by the at least one processor, and when the at least one processor executes the computer program, the electronic device executes the splicing variant pathogenicity prediction method based on contrast learning provided in the first embodiment.

[0110] Electronic devices are intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the inventions described and / or claimed herein.

[0111] The computer program for implementing the method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the computer program, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The computer program can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0112] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0113] It should be noted that the term "including" and its variations used in the embodiments of the present invention are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "multiple" mentioned in the embodiments of the present invention are illustrative and not restrictive. Those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0114] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0115] The various steps described in the method implementation methods provided by the embodiments of the present invention can be performed in different orders and / or in parallel. In addition, the method implementation methods may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0116] The term "embodiment" in this specification refers to specific features, structures or characteristics described in conjunction with the embodiment that can be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. The various embodiments in this specification are described in a related manner, and the same and similar parts between the various embodiments refer to each other. In particular, for the device, equipment, and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiment.

[0117] The above-described embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of protection. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting pathogenicity of splicing variants based on contrastive learning, characterized in that: include: Inputting original sequence information and variant sequence information, respectively obtaining a first original embedding vector and a first variant embedding vector; wherein the first original embedding vector is obtained according to the original sequence information, and the first variant embedding vector is obtained according to the variant sequence information; Inputting the first original embedding vector and the first variant embedding vector into a contrastive learning module, adjusting the embedding distance between the first original embedding vector and the first variant embedding vector in the contrastive learning module, and outputting a second original embedding vector and a second variant embedding vector; wherein the second original embedding vector is obtained according to the first original embedding vector, and the second variant embedding vector is obtained according to the first variant embedding vector; Performing feature fusion on the second original embedding vector and the second mutated embedding vector to obtain a fused tensor; The fused tensor is input into a classifier for classification prediction, and a pathogenicity prediction result is output.

2. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 1, characterized in that: Before inputting the original sequence information and the variant sequence information, the splicing variant pathogenicity prediction method based on comparative learning further includes: Creating a classification prediction model based on the DNA pre-training model, the contrastive learning module and the classifier; wherein the DNA pre-training model is used to extract a corresponding embedding representation based on sequence information; Without access to a classifier, training the DNA pre-training model and the comparative learning module according to training data; wherein the training data includes sample pairs consisting of multiple groups of sequence information; The DNA pre-training model and the comparative learning module are frozen, the classifier is connected, and the classifier is trained according to the training data.

3. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 2, characterized in that: Training the DNA pre-training model and the comparative learning module according to the training data includes: Input the training data into the DNA pre-training model and perform maximum pooling to obtain an embedded representation group; wherein the DNA pre-training model includes initial pre-training parameters; Inputting the embedding representation group into the contrastive learning module to obtain an embedding vector group; In the contrastive learning module, a contrastive learning loss function is used to calculate the contrastive learning loss of the embedding vector group; wherein the formula of the contrastive learning loss function is: ； ； ； is the total contrastive learning loss; is the contrastive learning loss for positive pairs, is the contrastive learning loss for negative sample pairs; is the label of the sample pair; is the Euclidean distance between the embedding vectors before and after mutation in the embedding vector group, is the maximum allowed Euclidean distance among the negative sample pairs, is the number of sample pairs in the training data; The DNA pre-training model and the contrastive learning module are trained until the total contrastive learning loss converges; and the trained DNA pre-training model and the contrastive learning module are obtained.

4. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 3, characterized in that: Training the classifier according to the training data comprises: Input the training data into the trained DNA pre-training model and perform maximum pooling; Inputting the maximum pooling result into the trained contrastive learning module; Inputting the output result of the trained contrastive learning module into the classifier; wherein the classifier uses a cross entropy loss function; The classifier is trained until the classification prediction model loss converges, and the trained classifier is obtained.

5. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 1, characterized in that: Inputting the original sequence information and the variant sequence information to obtain the first original embedding vector and the first variant embedding vector respectively includes: Inputting the original sequence information and the variant sequence information into the trained DNA pre-training model to obtain a first original embedding representation and a first variant embedding representation respectively; Max pooling is performed on the first original embedding representation and the first mutated embedding representation to obtain the first original embedding vector and the first mutated embedding vector, respectively; wherein the first original embedding vector and the first mutated embedding vector are both 768-dimensional vectors.

6. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 1, characterized in that: Adjusting the embedding distance between the first original embedding vector and the first variant embedding vector in the contrastive learning module, and outputting the second original embedding vector and the second variant embedding vector comprises: Adjusting the embedding distance between the first original embedding vector and the first variant embedding vector according to different labels; wherein the embedding distance is set as the Euclidean distance; If the splice variant is a pathogenic variant, making the Euclidean distance between the second original embedding vector and the second variant embedding vector greater than the Euclidean distance between the first original embedding vector and the first variant embedding vector; If the splice variant is a benign variant, the Euclidean distance between the second original embedding vector and the second variant embedding vector is made smaller than the Euclidean distance between the first original embedding vector and the first variant embedding vector.

7. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 1, characterized in that: The second original embedding vector and the second variant embedding vector are subjected to feature fusion to obtain a fused tensor including: The second original embedding vector and the second mutated embedding vector are stacked to complete feature fusion to obtain the fused tensor; wherein the second original embedding vector and the second mutated embedding vector are both 128-dimensional vectors, and the fused tensor is a two-dimensional tensor of (2, 128).

8. The method for predicting pathogenicity of splicing variants based on contrastive learning according to claim 1, characterized in that: The fusion tensor is input into the classifier for classification prediction, and the output pathogenicity prediction results include: Input the fused tensor into the trained classifier for binary classification prediction, perform nonlinear division through the activation function of the classifier, and output the classification result; wherein the binary classification prediction result of the classifier is 0 or 1; If the binary prediction result is 1, the pathogenicity prediction result is that the splicing variant has a pathogenic variation; If the binary prediction result is 0, the pathogenicity prediction result is that the splicing variant has a benign variation.

9. A splicing variant pathogenicity prediction system based on contrastive learning, characterized in that: include: A user input module, a sequence information matching module, a splicing variant pathogenicity prediction module and a result feedback and notification module; wherein the splicing variant pathogenicity prediction module comprises the splicing variant pathogenicity prediction method based on comparative learning according to any one of claims 1 to 8; The user input module is used to input genome sequence information; The sequence information matching module is used to identify the mutation site in the genome sequence information, and extract the DNA sequence information before and after the mutation, centered on the mutation site, and with a window size of 1024 bases as the original sequence information and the variant sequence information, respectively, and then input the original sequence information and the variant sequence information into the splicing variant pathogenicity prediction module; The splicing variant pathogenicity prediction module outputs a pathogenicity prediction result according to the original sequence information and the variant sequence information, and transmits the pathogenicity prediction result to the result feedback and notification module; The result feedback and notification module is used to collect and record the pathogenicity prediction results and display the pathogenicity prediction results.

10. An electronic device, comprising: A processor, and a memory storing a program, characterized in that the program includes instructions, which, when executed by the processor, cause the processor to perform the splicing variant pathogenicity prediction method based on contrast learning according to any one of claims 1 to 8.

Images