Base recognition method, system, medium and program based on full convolutional network

Abstract

The application discloses a base recognition method and system based on a full convolution network, a medium and a program, and belongs to the technical field of base recognition. The method comprises the following steps: constructing a base calling model based on a DenseNet convolutional neural network; constructing a data set to train the base calling model; collecting an electrical signal of a base to be recognized; inputting the electrical signal into the trained base calling model for recognition to obtain a sequence of the base to be recognized. The base calling model provided by the application significantly improves the base calling accuracy of direct RNA sequencing through an innovative full convolution network architecture and species-specific model training, overcomes many defects of the prior art, and has a wide application prospect and remarkable beneficial effects.

Description

[0001] Priority Application

[0002] This application claims priority to Chinese Invention Patent Application 【202410895365.9】 “

Base Recognition Method, System, Medium and Program Based on Full Convolutional Network

[0003] The present application belongs to the technical field of base recognition, and particularly relates to a base recognition method, system, medium and program based on a full convolutional network. BACKGROUND

[0004] The development of third-generation sequencing technology has completely changed the ability to perform long-read sequencing of genomes and transcriptomes. It can directly sequence DNA and RNA and their modifications without the need for fragmentation and amplification. Oxford Nanopore Technologies (ONT) Direct RNA Sequencing (also referred to as DRS for short) technology can directly sequence RNA molecules on a synthetic membrane through a series of protein nanopores. When the motor enzyme slowly passes a single RNA molecule through the nanopore, different nucleotides block the nanopore, causing current fluctuations, which can be recorded by base calling algorithms to determine the sequence of the RNA molecule.

[0005] Direct RNA sequencing technology can produce long-read sequences (up to 21 kb) aimed at obtaining complete transcripts or near-complete RNA genomes. Complete sequencing makes it powerful in measuring the length of the RNA poly(A) tail and analyzing the regulatory function of the poly(A) tail on gene expression regulation. In addition, direct RNA sequencing can directly detect modifications of RNA molecules, such as m6A, 7-methylguanosine (m7G), pseudo-uracil (Ψ) and 2'-O-methylation (Nm), because these modifications will produce different current fluctuations.

[0006] In addition, the non-amplification library preparation method can avoid amplification or RT bias, and is particularly suitable for identifying exitrons that are likely to be derived from reverse transcription artifacts. These advantages make direct RNA sequencing technology particularly advantageous for determining the structure of specific subtypes of RNA, identifying tRNA and rRNA, and sequencing the genomes of organisms such as RNA viruses, prokaryotes, and parasites. Recent studies have also used direct RNA sequencing technology to perform transcriptome or epitranscriptome analysis in a wide range of species, including humans, animals (nematodes, insects), plants, zooplankton, and pathogens. Another application of direct RNA sequencing technology is to analyze RNA metabolism kinetics by labeling nascent RNA with base analogs such as 5-ethynyluracil and 4-thiouracil, followed by nanopore direct RNA sequencing.

[0007] Although direct RNA sequencing (DRS) technology has shown great potential in genome and transcriptome sequencing, the base calling process still faces many challenges.

[0008] The applicant found that DRS at least has the following technical problems:

[0009] 1) The speed of RNA passing through the nanopore is slow

[0010] Due to the slow speed of RNA passing through the nanopore, only 70 bases per second, compared with the rate of 450 bases per second for DNA, the signal sampling rate is low, the signal noise is high, and the accuracy of base calling is affected.

[0011] 2) When applied to RNA sequencing of different species, the accuracy fluctuates greatly (i.e., poor stability)

[0012] Especially in some specific species research, the base calling accuracy is low. The base calling accuracy of DRS is currently about 86%, which has not yet reached the level required for high-precision base calling (more than 90%), especially in identifying key genetic elements such as short exons and exon junctions. In addition, the existing technology also has significant shortcomings in identifying and analyzing the complex current fluctuation characteristics generated by the modification of RNA molecules such as m6A, 7-methylguanosine, pseudo-uracil, and 2'-O-methylation.

[0013] 3) Insufficient base calling accuracy: the most advanced DRS base caller RODAN has an accuracy of about 86%, which has not yet reached the level required for high-precision base calling (more than 90%). Especially for the study of RNA viruses and pathogens with high mutation rates, the existing base calling accuracy is insufficient to reliably record genetic elements such as short exons and exon junctions, affecting the accuracy and depth of the study.

[0014] 4) Species-specific challenges: While Convolutional Neural Networks (CNNs) have improved base calling accuracy to some extent, no species-specific base calling method has been developed to address the low accuracy problem in species-specific research. Existing methods cannot fully consider the unique signal patterns of different species of RNA data, resulting in low base calling accuracy in specific species research.

[0015] 5) Difficulty in identifying current fluctuation features: Direct RNA sequencing technology can detect modifications of RNA molecules (such as m6A, 7-methylguanosine, pseudo-uracil, and 2'-O-methylation), but these modifications produce different current fluctuations, increasing the complexity of base calling. Existing technologies have significant deficiencies in identifying and analyzing these complex current fluctuation features, limiting their application in RNA modification detection.

[0016] The prior art Chinese patent application CN202010026283.2 discloses a base sequence recognition method, comprising the steps of: reading the data file output by the Oxford Nanopore sequencer and extracting the current signal corresponding to the DNA / RNA molecule to be tested; cutting a plurality of current signal segments of a predetermined length from the current signal according to a predetermined overlap rate; inputting each current signal segment into a predetermined time series convolution network model for time series modeling to generate a corresponding base probability matrix for each current signal segment; wherein the base probability matrix is the probability distribution of the occurrence of a base at each sampling time point for the current signal segment; decoding the corresponding base sequence segment from each base probability matrix, and generating the base sequence from each base sequence segment.

[0017] However, the above-mentioned prior art also has the problems of low recognition accuracy and poor universality. SUMMARY

[0018] The purpose of the present application is to provide a base recognition method and system based on a full convolutional network, which partially solves or alleviates the above-mentioned deficiencies in the prior art, and can efficiently and stably recognize base sequences.

[0019] To solve the above-mentioned technical problems, the present application specifically adopts the following technical solutions:

[0020] A base recognition method based on a full convolutional network, comprising:

[0021] S101, constructing a base calling model based on a DenseNet convolutional neural network; the base calling model comprises in sequence:

[0022] a first convolutional layer for denoising the electrical signal;

[0023] a max-pooling layer for down-sampling the electrical signal;

[0024] a plurality of dense blocks for extracting features of the electrical signal;

[0025] a transition layer interspersed between the dense blocks for reducing dimensionality of the features output by the dense blocks;

[0026] a fully connected layer for integrating global features of the feature maps processed by the dense blocks and the transition layer and making a classification prediction;

[0027] a CTC decoder for decoding the prediction result output by the fully connected layer to obtain a sequence of bases; wherein the dense block comprises a plurality of dense layers, and each dense layer comprises a batch normalization layer, a first Silu nonlinear unit layer, a 1x1 second convolutional layer, a second batch normalization layer, a second Silu nonlinear unit layer, and an nxl third convolutional layer, wherein the setting step of n comprises:

[0028] calculating a reference value of the third convolutional layer of the i-th layer according to an incremental model, wherein the incremental model is:

[0029] k = kernel_size + i * kernel_size_increment;

[0030] k is the reference value, kernel_size is the initial convolution kernel size, i is the current dense layer number, and kernel_size_increment is the convolution kernel increment;

[0031] when k < the set convolution size K MAX and k is even, then n = k + 1;

[0032] when k < the set convolution size K MAX and k is odd, then n = k;

[0033] when k ≥ the set convolution size K MAX , then n = K MAX ;

[0034] S102, training the base calling model according to the current training condition using the data set to obtain a base calling model; wherein the training condition comprises: the number of dense layers, the third convolution kernel size of the third convolutional layer;

[0035] S103, determining whether the prediction accuracy of the current obtained base calling model is greater than a set accuracy threshold; if yes, output the corresponding base calling model, and if no, update the training condition and return to S102;

[0036] S104, collecting an electrical signal of a base to be identified;

[0037] S105, inputting the electrical signal into the base calling model for recognition to obtain a sequence of the to-be-recognized base.

[0038] In some embodiments, the step of updating the training condition comprises:

[0039] The number of dense layers of at least one dense block is modified, and the third convolution kernel size of each dense layer is updated according to the modified number of dense layers.

[0040] In some embodiments, the plurality of dense blocks comprises a first dense block, a second dense block, a third dense block, and a fourth dense block; and the step of updating the training condition comprises:

[0041] calculating a first prediction accuracy of the base calling model obtained in the last training process;

[0042] calculating a second prediction accuracy of the base calling model obtained in the current training process;

[0043] calculating an accuracy growth rate according to the first prediction accuracy and the second prediction accuracy;

[0044] when the accuracy growth rate is greater than or equal to a first growth threshold, a first modification operation is recommended for the number of dense layers of the first set of dense blocks, wherein the first set comprises the first dense block, the second dense block, the third dense block, and the fourth dense block;

[0045] when the accuracy growth rate is less than the first growth threshold, a second modification operation is recommended for the number of dense layers of the second set of dense blocks, wherein the second set comprises the third dense block and the fourth dense block.

[0046] In some embodiments, the method further comprises the step of:

[0047] when the accuracy growth rate is still less than the first growth threshold after at least one second modification operation, a modification is recommended for the pooling size of the max pooling layer.

[0048] In some embodiments, the first modification operation or the second modification operation satisfies the following conditions:

[0049] the number of dense layers of the second dense block is less than a first set value;

[0050] the number of dense layers of the third dense block is greater than a second set value.

[0051] In some embodiments, K MAX is 99.

[0052] As an improvement, the pooling size of the max pooling layer is 10.

[0053] As an improvement, the transition layer is a 1x1 convolutional layer.

[0054] As an improvement, the data set includes a training set, a validation set and a test set.

[0055] The training set and the validation set include base electrical signals of Arabidopsis thaliana, Homo sapiens, Caenorhabditis elegans and Escherichia coli.

[0056] The test set includes base electrical signals of Homo sapiens, Arabidopsis thaliana, Mus musculus, Saccharomyces cerevisiae, Populus nigra, Plasmodium berghei, Senecavirus, Porcine epidemic diarrhea virus and Porcine reproductive and respiratory syndrome virus.

[0057] As an improvement, the collected electrical signals of the to-be-identified base are preprocessed, including:

[0058] The electrical signals are divided into signal blocks, and each signal block contains 4096 signal values.

[0059] The signal values are normalized by using the median absolute deviation method.

[0060] As an improvement, the data set of a certain organism is used for base calling model training to obtain a species-specific base calling model of the organism.

[0061] As an improvement, the species-specific base calling model is trained by using the DEMINERS training function; the training parameters are a learning rate of 0.002, a batch size of 32, and training for 30 epochs.

[0062] The application also provides a base recognition system based on a full convolutional network, comprising: a model construction module, configured to construct a base calling model based on a DenseNet convolutional neural network; the base calling model comprises in sequence:

[0063] A first convolutional layer, configured to denoise the electrical signals.

[0064] A max-pooling layer, configured to down-sample the electrical signals.

[0065] A plurality of dense blocks, configured to extract features of the electrical signals.

[0066] Transition layers interspersed between the dense blocks, configured to reduce the dimension of the features output by the dense blocks.

[0067] A fully connected layer, configured to integrate global features of the feature maps processed by the dense blocks and the transition layers, and to perform classification prediction.

[0068] The CTC decoder is configured to decode the prediction result of the full connection layer output to obtain a sequence of bases. The dense block includes a plurality of dense layers, and each dense layer includes a batch normalization layer, a first Silu nonlinear unit layer, a 1*1 second convolution layer, a second batch normalization layer, a second Silu nonlinear unit layer, and an n*1 third convolution layer.

[0069] The setting step of n includes:

[0070] The reference value of the third convolution layer of the i-th layer is calculated according to an incremental model, and the incremental model is:

[0071] k = kernel_size + i * kernel_size_increment

[0072] k is the reference value, kernel_size is the initial convolution kernel size, i is the current dense layer number, and kernel_size_increment is the convolution kernel increment.

[0073] When k < the set convolution size K MAX , and k is even, then n = k + 1.

[0074] When k < the set convolution size K MAX , and k is odd, then n = k.

[0075] When k >= the set convolution size K MAX , then n = K MAX .

[0076] The first training module is configured to train the base calling model according to a current training condition and using a data set to obtain a base calling model.

[0077] The second training module is configured to determine whether the prediction accuracy of the obtained base calling model is greater than a set accuracy threshold, and if so, output the corresponding base calling model, and if not, update the training condition and return to the first training module.

[0078] The data acquisition module is configured to acquire an electrical signal of a base to be identified.

[0079] The recognition module is configured to input the electrical signal into the base calling model for recognition to obtain a sequence of the base to be identified.

[0080] Beneficial technical effects: In the existing RNA sequencing process, due to low RNA signal sampling rate, large signal noise, limited sample size and other problems, it is difficult to improve the prediction accuracy of RNA.

[0081] To this end, for the one-dimensional data characteristics of RNA, the present application selects to introduce a DenseNet convolutional neural network for model training, and preferably sets 4 dense blocks, and the size of the convolution kernel of the third convolutional layer is set in a layer-by-layer increasing mode between adjacent dense blocks and between the dense layers inside the dense blocks, to gradually increase the in-depth learning of the correlation degree of one-dimensional data before and after (i.e., gradually increase the field of view of the third convolutional layer). In addition, in order to coordinate the operation pressure of the model, the size of the convolution kernel of the dense layer in the back-end part will also be limited, such as once the convolution kernel size reaches 99, it will no longer be increased layer by layer.

[0082] In other words, for this one-dimensional data RNA, the present application proposes a training mode of limiting layer-by-layer increasing of the convolution kernel between dense blocks and inside the dense blocks, which can coordinate to a certain extent between training accuracy and operation pressure. And through verification, this limiting layer-by-layer increasing convolution kernel setting can significantly improve the prediction accuracy.

[0083] It can be understood that the parameters of the model may change with different types of prediction data sets. For example, for different types of objects (such as humans, mice, etc.), targeted training may be needed, and in order to improve the training efficiency of the model in the repeated debugging process, the present application also proposes an iterative training mode for adjusting the learning details of the sequence data and the gradient transmission of the details. That is, the present application focuses on adjusting the learning details of the model for sequence data and other information by adjusting the number of dense layers of the dense block, so as to quickly correct the prediction performance of the model.

[0084] Specifically, the scheme of the present application for limiting the size of the third convolution kernel can gradually expand the field of view between multiple dense layers. This limiting layer-by-layer increasing scheme can not only improve the field of view, that is, in-depth learning of the correlation of the sequence of electrical signals, but also help maintain effective gradient transmission to ensure that the network can be trained stably. In addition, this layer-by-layer increasing and limiting the size of the field of view in the back-end (i.e., limiting the size of the convolution kernel) can also control the operation pressure during training.

[0085] In summary, the present application can solve the problems in the prior art by identifying bases based on a base calling model of a fully convolutional network:

[0086] 1. Improve the signal processing capability of the sampling rate: the base calling model architecture design in the present application can effectively process high-noise signals caused by low sampling rate and improve the accuracy of signal analysis.

[0087] 2. Improved base calling accuracy: By introducing the DenseNet convolutional neural network architecture and adjusting the convolutional field of view of the convolutional neural network, the base calling model in the present application achieves the latest performance level on various data sets of animals, plants, and viruses, and the base calling accuracy is significantly improved.

[0088] Species-specific model training: The base calling model in the present application supports species-specific model training, which can optimize the unique signal patterns of different species and significantly improve the base calling accuracy in specific species research.

[0089] In addition to solving the problems in the prior art, the present application has the following advantages:

[0090] 1. Advantages of architecture design.

[0091] The base calling model provided by the present application realizes efficient parameter utilization and overfitting control through the innovative design of full convolutional network and dense block. The design of dense connection enhances feature reuse, reduces the number of parameters, and improves the generalization ability of the model. These advantages in architecture make it perform well in processing complex electrical signals.

[0092] Efficient feature reuse: The dense block design in the base calling model provided by the present application enables each layer to receive feature maps from all previous layers, thereby enhancing feature reuse and information flow. This design not only reduces the number of parameters, but also significantly improves the performance and stability of the model.

[0093] Parameter optimization and computational efficiency: The base calling model provided by the present application increases the number of channels and the size of convolution kernels in the dense layer to capture more extensive base position information. The number of channels in the last dense block reaches 1024, and the size of the convolution kernel is 99. Through this optimization, the model ensures high accuracy while improving computational efficiency, suitable for large-scale data processing.

[0094] 2. Wide range of applications.

[0095] The base calling model provided by the present application not only performs well on single-species data sets, but also performs well on multi-species mixed data sets. Its versatility and adaptability make it have wide application potential in different research fields.

[0096] Wide range of application scenarios: suitable for RNA sequencing of various biological samples, including humans, plants, animals, viruses, etc. In various RNA sequencing applications, whether for basic research or clinical diagnosis, Densecall performs outstanding performance and reliability.

[0097] In the research of high-mutation-rate RNA viruses (such as SARS-CoV-2, PRRSV, SVV and PEDV), the high accuracy and low error rate of the base calling model provided by the present application can more reliably identify and record the genetic information of the virus, providing strong support for the research of virus evolution and transmission.

[0098] In summary, the base calling model provided by the present application significantly improves the base calling accuracy of direct RNA sequencing through the innovative fully convolutional network architecture and species-specific model training, overcomes many shortcomings of the prior art, and exhibits a wide application prospect and significant beneficial effects. BRIEF DESCRIPTION OF DRAWINGS

[0099] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0100] Figure 1 Structure diagram of the base calling model in an exemplary embodiment of the present application;

[0101] Figure 2 Overall accuracy effect diagram of the base calling method in an exemplary embodiment of the present application;

[0102] Figure 3 Result diagram showing the accuracy and error rate of Densecall (the base calling model in the present application), Guppy and RODAN on different species;

[0103] Figure 4 It is shown that the base calling accuracy of Densecall (the base calling model in the present application) on mouse transcriptome is improved from 87.84% to 90.44%, which is higher than 88.16% of ONT and 84.06% of Guppy;

[0104] Figure 5 It is shown that the mapping rate and base calling length of different identification methods are different;

[0105] Figure 6 It is shown that the first diagram of the model training interface in an exemplary embodiment of the present application;

[0106] Figure 7A second schematic diagram of a model training interface in an example embodiment of the present application is shown.

[0107] Figure 8 A third schematic diagram of a model training interface in an example embodiment of the present application is shown.

[0108] Figure 9 A flowchart of a base recognition method in an example embodiment of the present application is shown. DETAILED DESCRIPTION

[0109] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0110] Herein, the suffix such as "module", "component" or "unit" used to represent an element is only for the convenience of description of the present application, and has no specific meaning by itself. Therefore, "module", "component" or "unit" can be used mixedly.

[0111] Herein, the terms "upper", "lower", "inner", "outer", "front", "back", "one end", "the other end" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of description of the present application and simplification of the description, and do not indicate or imply that the indicated device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0112] Herein, unless otherwise explicitly specified and limited, the terms "mount", "provided with", "connected" and the like should be understood broadly, for example, "connected" can be fixedly connected, can be detachably connected, or integrally connected; can be mechanically connected, can be directly connected, or indirectly connected through an intermediate medium, can be the communication inside two elements. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0113] Herein, "and / or" includes any and all combinations of one or more listed associated items.

[0114] Herein, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0115] Embodiment one

[0116] In order to improve the processing capability of the low sampling rate RNA signal and improve the accuracy of base calling, the application provides a base recognition method based on a full convolution network, as shown in Figure 1 and Figure 9 , comprising:

[0117] S101, constructing a base calling model based on a DenseNet convolutional neural network (also referred to as Densecall in this paper).

[0118] DenseNet is a deep convolutional neural network structure, and its core idea is to establish direct connection between each layer of the network, so as to maximize the information transmission between layers.

[0119] In this embodiment, as shown in Figure 1 , the specific structure of the base calling model (Densecall) constructed based on DenseNet comprises in turn:

[0120] 1. The first convolutional layer is used to denoise the electrical signal (which is one-dimensional data).

[0121] The first convolutional layer denoises the input electrical signal through three convolutional layers. The convolutional layer can effectively extract the local features of the signal and reduce noise.

[0122] 2. The maximum pooling layer is used to downsample the electrical signal.

[0123] Preferably, the pooling size of the maximum pooling layer in this embodiment is 10. By downsampling the electrical signal, the dimension of the signal can be reduced, thereby reducing the computational complexity and further removing redundant information.

[0124] 3. A plurality of dense blocks are used to extract features of the electrical signal.

[0125] Preferably, a dense block includes a plurality of dense layers, and the dense layer includes: a batch normalization layer, a first Silu nonlinear unit layer, a 1x1 second convolutional layer, a second batch normalization layer, a second Silu nonlinear unit layer, and a n x 1 third convolutional layer, where n is the size of the convolution kernel of the third convolutional layer (also referred to as: the third convolution kernel size).

[0126] In this embodiment, the batch normalization layer is used to accelerate training and reduce internal covariant drift. The Silu nonlinear unit is a nonlinear activation function. The 1x1 second convolutional layer is used to reduce the feature dimension and reduce the computational complexity. The second batch normalization and SiLU layer repeat the batch normalization and SiLU operation. The n x 1 third convolutional layer is used to extract local features.

[0127] In this embodiment, the setting step of n includes:

[0128] The reference value of the third convolutional layer of the i-th layer is calculated according to an incremental model, wherein the incremental model is:

[0129] k = kernel_size + i * kernel_size_increment;

[0130] k is the reference value, kernel_size is the initial convolution kernel size, i is the current dense layer number, and kernel_size_increment is the convolution kernel increment;

[0131] When k < the set convolution size K MAX , and k is even, then n = k + 1.

[0132] When k < the set convolution size K MAX , and k is odd, then n = k.

[0133] When k >= the set convolution size K MAX , then n = K MAX .

[0134] In this embodiment, on the one hand, the third convolution kernel size of the dense layer is set to be increased layer by layer by using an incremental model, and the field of view of the third convolutional layer is expanded layer by layer in a layer-by-layer increasing manner, so as to ensure that the model can learn one-dimensional data in depth; on the other hand, the maximum value of the convolution kernel is limited by setting the convolution size K MAX , so as to further balance the operation pressure in the model training process. That is to say, the present application proposes a training method for limiting the incremental convolution kernel between the dense layers, and the training method for limiting the incremental convolution kernel can balance the training accuracy and the training data amount.

[0135] 4. Three transition layers interlaced between the four dense blocks, used for dimension reduction of the features output by the dense blocks.

[0136] After each dense block, except the last one, a 1x1 convolutional layer is used as a transition layer for feature compression to speed up processing.

[0137] 5. A fully connected layer, used for global feature integration of the feature maps processed by the dense blocks and the transition layers, and for classification prediction.

[0138] The feature maps processed by the dense blocks and the transition layers are input into the fully connected layer, the multi-dimensional feature maps are flattened into one-dimensional vectors, and the features are integrated through the fully connected operation.

[0139] Classification is performed using a log softmax activation. Log-softmax is an activation function used for multi-classification problems, typically in the output layer of a neural network. Its role is to convert the model's raw output (logits) into a probability distribution and take the logarithm for convenience in computation.

[0140] 6. A CTC decoder is used to decode the prediction results of the full connection layer output to obtain the sequence of bases.

[0141] CTC (Connectionist Temporal Classification) is an algorithm for sequence recognition, especially suitable for handling cases where the length of input and output is inconsistent. The CTC decoder eliminates the blank label in the sequence and combines the continuous repeated labels through the dynamic programming algorithm to obtain the final sequence output.

[0142] The loss function of the base calling model uses CTC loss for gradient descent, and the base calling uses beam search with a size of 5. The input of the base calling model is: electrical signal; the output is: the base sequence corresponding to the electrical signal.

[0143] S102, according to the current training condition, the base calling model is trained using the data set, to train and update the base calling model; wherein, the training condition includes: the number of dense layers, the third convolution kernel size of the third convolution layer;

[0144] S103, judge whether the prediction accuracy of the current obtained base calling model is greater than the set accuracy threshold;

[0145] If yes, output the corresponding base calling model, and recommend the user to use the current base calling model to predict the sequence of RNA. If not, update the training condition and return to S102.

[0146] Generally, the training condition needs to be modified several times (i.e. S102 may be executed several times) to train the base calling model that meets the requirements.

[0147] For example, in some embodiments, the way to update the training condition is to modify the number of dense layers. For example, in some embodiments, the way to modify the number of dense layers is to double the number of dense layers of each dense block. Of course, at this time, the third convolution kernel size will also change adaptively according to the change of the number of dense layers.

[0148] For example, in some embodiments, the way to update the training condition is to modify the third convolution kernel size, for example, the initial convolution kernel size can be increased.

[0149] S104, collecting an electrical signal of a to-be-identified base;

[0150] S105, inputting the electrical signal into the base calling model for identification to obtain a sequence of the to-be-identified base.

[0151] In some embodiments, the step of updating the training condition comprises:

[0152] modifying the number of dense layers of at least one dense block, and updating the third convolution kernel size of each dense layer according to the modified number of dense layers.

[0153] In some embodiments, the plurality of dense blocks comprise, in sequence, a first dense block, a second dense block, a third dense block, and a fourth dense block. Preferably, the number of dense layers of the first, second, and third dense blocks increases in sequence, and the number of dense layers of the fourth dense block is less than that of the third dense block.

[0154] In this embodiment, the number of dense layers of the first and second dense blocks is small, which is mainly used for learning low-level features of data, and the number of dense layers of the third and fourth dense blocks is slightly larger, which is mainly used for learning high-level features of data.

[0155] Further, in some embodiments, the step of updating the training condition further comprises:

[0156] calculating a first prediction accuracy of the base calling model obtained in the last training process;

[0157] calculating a second prediction accuracy of the base calling model obtained in the current training process;

[0158] calculating an accuracy growth rate according to the first prediction accuracy and the second prediction accuracy, wherein the accuracy growth rate = (the second prediction accuracy - the first prediction accuracy) / the first prediction accuracy;

[0159] when the accuracy growth rate is greater than or equal to a first increment threshold, a first modification operation is recommended for the number of dense layers of the first set of dense blocks, wherein the first set comprises the first dense block, the second dense block, the third dense block, and the fourth dense block;

[0160] For example, in some embodiments, when the accuracy growth rate is greater than or equal to the first increment threshold, it is considered that the current proportion of the number of dense layers of each dense block is relatively appropriate, and the number of dense layers can be adjusted on the basis of the current proportion. For example, the number of dense layers of each dense block is doubled.

[0161] When the accuracy increment rate is less than the first increment threshold, a second modification operation is recommended for the number of dense layers of the second set of dense blocks, wherein the second set includes a third dense block and a fourth dense block.

[0162] For example, in some embodiments, when the accuracy increment rate is less than the first increment threshold, it is recommended to directly adjust the proportion of the number of dense layers of each dense block, and preferably to adjust the number of dense layers of the third dense block and the fourth dense block to adjust the learning ability of the model for high-level features.

[0163] In this embodiment, during the model training process, the learning dimension of the base calling model for one-dimensional data is adjusted by focusing on the adjustment of the number of dense layers.

[0164] It is worth noting that the sequence of the electrical signal of the RNA is usually long, and there is often a dependency relationship in the sequence. Increasing the size of the convolution kernel also means increasing the field of view of the convolution, which helps to improve the learning ability of the convolution layer for the dependency relationship.

[0165] To this end, the scheme of the present application for setting the third convolution kernel size by limiting incremental can gradually expand the field of view between multiple layers of dense layers. This limiting incremental scheme not only improves the field of view, i.e. deep learning of the relevance of the sequence of the electrical signal, but also helps to maintain the effective transmission of the gradient to ensure that the network can be stably trained. In addition, this way of gradually increasing layer by layer and limiting the size of the field of view at the back end (i.e. limiting the size of the convolution kernel) can also control the operation pressure during the training process.

[0166] Further, in order to improve the training efficiency of the model during repeated debugging, the present application proposes an iterative training mode for adjusting the learning details of the sequence data and the gradient transmission of the details. That is, the present application focuses on adjusting the learning details of the model for the sequence data by adjusting the number of dense layers of the dense block, so as to quickly correct the prediction performance of the model.

[0167] In some embodiments, the method further comprises the steps of:

[0168] When the accuracy increment rate is still less than the first increment threshold after at least one second modification operation, a modification operation is recommended for the pooling size of the max pooling layer.

[0169] In some embodiments, the first modification operation or the second modification operation satisfies the following conditions:

[0170] The number of dense layers of the second dense block is less than a first set value;

[0171] The number of dense layers of the third dense block is greater than a second set value.

[0172] In this embodiment, the number of dense layers of the second dense block for learning low-level features and the third dense block for learning high-level features is preferably limited respectively to ensure the complete learning of low-level and high-level features, and thus to ensure that the overall details formed by the connection of low-level and high-level features can fully display the characteristics of the sequence.

[0173] In some embodiments, K MAX is 99.

[0174] As a preferred base calling model, the number of dense blocks is 4, i.e., the first dense block, the second dense block, the third dense block, and the fourth dense block. The first dense block includes 6 dense layers, the second dense block includes 12 dense layers, the third dense block includes 24 dense layers, and the fourth dense block includes 16 dense layers. Each layer is connected by a feedforward manner and receives the concatenated feature maps of all previous layers to enhance feature reuse and reduce the number of parameters.

[0175] In this embodiment, the shallow layers of the network are usually responsible for extracting low-level features, while the deep layers are responsible for extracting higher-level semantic features. By gradually increasing the number of dense layers of the dense block, the complexity of the features can be gradually increased, so that the network can better capture the features from low-level to high-level. The present application preferably configures the dense layers of the four dense blocks of Densecall as [6, 12, 24, 16], and after a large number of experimental verification on the training data set, it can be proved that this special configuration can indeed achieve a good balance between model performance and computational efficiency. In the first few dense blocks, the number of layers is gradually increased, which can gradually increase the number of feature maps, thereby capturing more details and complex features. For example, the first dense block has 6 layers, the second dense block has 12 layers, and the third dense block has 24 layers. Reducing the number of layers in the last dense block can reduce the amount of calculation while maintaining the performance of the model. For example, the fourth dense block has 16 layers.

[0176] The existing DenseNet convolutional neural network is widely used in image classification tasks, in which the size of the convolution kernel is usually a two-dimensional structure of n x n, which is suitable for processing two-dimensional image data. However, in this scheme, the DenseNet convolutional neural network is applied to one-dimensional base electrical signal recognition to realize the calling of base sequences. In order to adapt to this change, we correspondingly adjust the convolution kernel to a one-dimensional convolution kernel of n x 1.

[0177] This modification is based on the flexibility and powerful feature learning ability of DenseNet. DenseNet enhances feature transmission and reuse by connecting each layer directly to all previous layers through its unique dense connection method, reducing feature redundancy and improving model performance while reducing parameter quantity. When processing one-dimensional signals, this design can also take advantage of its strengths, effectively extracting local features in sequences and enhancing the expression ability of features through dense connections.

[0178] Specifically, in this embodiment, the convolutional layers within each Dense Block (i.e., dense block) not only receive the output of the previous layer, but also receive the output of all previous layers and concatenate them as input. This dense connection method allows each layer to access all previous layer feature information, improving feature utilization. In one-dimensional signal processing, this helps capture timing features and local patterns in signals.

[0179] In addition, the dense connection structure of DenseNet (i.e., the concept of increasing convolutional kernels layer by layer) helps alleviate the gradient vanishing problem, which is particularly important for training deep networks. In this solution, since the base electrical signal sequence can be long, a deep network helps capture dependencies in long sequences, and the special structure used in DenseNet helps maintain effective gradient transmission, ensuring that the network can be trained stably.

[0180] In practical applications, the two-dimensional convolutional kernel of DenseNet is adjusted to a one-dimensional convolutional kernel to adapt to the data characteristics of the base electrical signal. This includes modifying the parameter settings of the convolutional layer to make it suitable for processing one-dimensional input data. At the same time, in some embodiments, other parts of the network, such as the transition layer (Transition Layer), can also be adjusted to ensure the rationality and effectiveness of the network structure.

[0181] In summary, applying DenseNet convolutional neural network to one-dimensional base electrical signal recognition, by modifying the convolutional kernel to a one-dimensional convolutional kernel of n x 1, and using its dense connection characteristics, is expected to achieve good performance in base sequence calling tasks. This innovative application demonstrates the flexibility and universality of DenseNet and provides new ideas and methods for one-dimensional signal processing.

[0182] In this implementation, the base calling model increases the number of channels and the size of the convolutional kernel used in the dense layer to capture more extensive base position information. In the last dense block, the number of channels reaches 1024 and the convolutional kernel size is 99.

[0183] More specifically, each dense block includes a plurality of dense layers, each of the dense layers includes a one-dimensional convolution kernel of n x 1, and the size of the one-dimensional convolution kernel increases with the number of layers until n = K MAX , and the one-dimensional convolution kernel of the subsequent dense layer is n = K MAX , and n is an odd number. In this embodiment, the size of the third convolution kernel of the first layer of the first dense block (i.e., the initial convolution kernel size) is set to an initial 7 x 1, and the reference value of the current layer can be calculated by the formula k = kernel_size + i * kernel_size_increment, where k is the reference value, kernel_size is the initial convolution kernel size, i is the number of the current dense layer, and kernel_size_increment is the convolution kernel increment (kernel_size_increment = 3 in this embodiment); for example, for the 10th layer of the second dense block (total number of layers i = 6 + 10 = 16), k is 7 + 3 * 16 = 55, i.e., the size of the convolution kernel is 55 x 1.

[0184] In this embodiment, the size of the convolution kernel is also required to be an odd number, so if the final result k is even, n = k + 1. For example, for the 11th layer of the second dense block (total number of layers i = 6 + 11 = 17), k is 7 + 3 * 17 = 58, n = 58 + 1 = 59, i.e., the size of the convolution kernel is 59 x 1, and so on.

[0185] In this embodiment, the maximum size of the convolution kernel is limited, which aims to prevent excessive consumption of computing resources and possible performance problems caused by too large convolution kernel. Specifically, in this embodiment, the maximum size threshold of the convolution kernel is 99, i.e., the size of the one-dimensional convolution kernel increases with the number of layers until n = 99, and the one-dimensional convolution kernel of the subsequent dense layer is 99.

[0186] In addition, the performance evaluation of the base calling model in this embodiment includes:

[0187] 1. Evaluation criteria: evaluate using sequence consistency (accuracy = M / (M + S + I + D)), where M is the number of matching bases, S is the number of mismatches, I is the number of insertions, and D is the number of deletions.

[0188] 2. Alignment and evaluation: use minimap2 (v2.15, -cs) for sequence alignment and accuracy evaluation, including quantifying the number of mismatches, insertions, and deletions compared to the reference genome.

[0189] Figures 6-8 The training interface schematic diagram of the base calling model in an exemplary embodiment of the present application is shown.

[0190] In some embodiments, the dataset includes a training set, a validation set, and a test set.

[0191] To improve the generalization and accuracy of the base calling model, the training set in the embodiment includes data from Arabidopsis thaliana, Homo sapiens, Caenorhabditis elegans, and Escherichia coli, a total of one million signal blocks are used for training, 10 million for validation.

[0192] The sequencing dataset is trained using PyTorch (v2.0.1) in the embodiment, the batch size is set to 32, and the training is performed for 30 epochs. The test set consists of the following parts:

[0193] 1) RODAN research dataset:

[0194] Including Homo sapiens, Arabidopsis thaliana, Mus musculus, S. cerevisiae S288C, and Populus trichocarpa.

[0195] 2) Published SARS-CoV-2 dataset.

[0196] 3) Dataset generated by the present application: including Homo sapiens, P. berghei, Seneca Valley Virus (SVV), Porcine Epidemic Diarrhea Virus (PEDV), and Porcine Reproductive and Respiratory Syndrome Virus (PRRSV).

[0197] On these datasets, the performance of the base calling model provided by the present application is significantly better than the prior art, as shown in Table 1 (DEMINERS in Table 1 is the base calling model in the present application).

[0198] Table 1 Comparison of performance of base calling on RNA-seq data using different base callers

[0199]

[0200] As can be seen, the base calling model in the present application significantly improves the base calling accuracy on various datasets. For example, on the test set, the accuracy is significantly higher than Guppy (a tool kit provided by ONT), and is also better than RODAN on most species. By evaluating on datasets containing multiple species, it not only performs well in overall accuracy, but also shows lower error rates in mismatch, insertion, and deletion rates, such asFigure 2 As shown.

[0201] Compared with Guppy and RODAN, the base calling model in the present application shows higher accuracy and lower error rate on most species, as shown in Table 1. Figure 3 For example, among the 10 species in the test set, Densecall has higher base calling accuracy on 9 species than RODAN, showing a significant advantage (see Table 1). In high mutation rate RNA viruses such as SARS-CoV-2, PRRSV, SVV and PEDV, the base calling accuracy of the base calling model in the present application is significantly better than the prior art, and can more accurately identify and record the genetic elements of these viruses.

[0202] It is worth noting that in order to improve the accuracy of base identification in a specific species, the base identification model in the present embodiment supports species-specific optimization. By training the base calling model using the data set of a certain organism, a species-specific base calling model of the organism can be obtained, which can be optimized for the unique signal patterns of different species, significantly improving the base calling performance in specific species research.

[0203] For example, train a species-specific model using the mouse data set (Mus musculus) of RODAN, the training set includes 20,000 electrical signals, and the validation and test sets each contain 4,000 electrical signals. HDF5 data is generated using Taiyaki (v5.3.0), and the species-specific model is trained using the DEMINERS training function. The training parameters are set to a learning rate of 0.002, a batch size of 32, and training for 30 epochs.

[0204] After specific training, the base calling accuracy of the model on the mouse transcriptome is improved from 87.84% to 90.44%, which is higher than the 88.16% of RODAN and the 84.06% of Guppy, as shown in Table 1. Figure 4 In addition, although the mapping rate is comparable, the base calling length of the mouse-specific model of the present model is 11 bp longer than that of RODAN, as shown in Table 1. Figure 5 This species-specific optimization significantly improves the base calling performance in specific species research, indicating that the present model has high adaptability and accuracy in processing RNA sequencing data of different species.

[0205] For example, the present embodiment can read the data file output by the Oxford Nanopore sequencer and extract the current signal corresponding to the DNA / RNA molecule to be tested.

[0206] Specifically, the Oxford Nanopore sequencing method is a third-generation single-molecule real-time sequencing technology based on an electrical signal, which can directly read the double-stranded DNA / RNA molecule and capture the current signal. In the sequencing process, the DNA / RNA double-strand is first connected with the motor protein enzyme and combined with the nanopore protein inlaid in the biological membrane and helix, and the motor protein enzyme controls the movement of the DNA / RNA double-strand through the nanopore. The ion current in the nanopore fluctuates with the movement of the nucleic acid in the pore, so as to capture the fluctuating current signal and store it in a data file. By connecting with the data file storing the current signal, the current signal corresponding to the DNA / RNA molecule to be tested in the data file is obtained to perform the subsequent base sequence identification process.

[0207] Further, before inputting the data into the model, the training and test data sets can also be preprocessed by Taiyaki (Oxford Nanopore Technologies, v5.3.0). Each signal chunk contains 4096 signal values, and the median absolute deviation method is used for normalization.

[0208] Splitting the long electrical signal sequence into multiple signal chunks can improve processing efficiency, reduce memory usage, and allow parallel processing. 4096 is a power of 2, which facilitates efficient processing in a computer. In actual use, the GPU memory usage can be reduced from 6628MB to 4700MB, and the batch size is 32, which highlights the efficiency and effectiveness of the method.

[0209] It can be understood that the data in the data set during the training of the model can also be preprocessed by the above method.

[0210] Embodiment two:

[0211] The application also provides a base recognition system based on a full convolutional network, comprising:

[0212] A model construction module is configured to construct a base calling model based on a DenseNet convolutional neural network. The base calling model comprises, in sequence:

[0213] A first convolutional layer is configured to denoise the electrical signal.

[0214] A max-pooling layer is configured to down-sample the electrical signal.

[0215] A plurality of dense blocks are configured to extract features of the electrical signal.

[0216] Transition layers interspersed between the dense blocks are configured to reduce the dimensionality of the features output by the dense blocks.

[0217] a full connection layer, configured to integrate global features of the feature map processed by the dense block and the transition layer, and perform classification prediction;

[0218] a CTC decoder, configured to decode the prediction result output by the full connection layer to obtain a sequence of bases; wherein the dense block comprises a plurality of dense layers, and each dense layer comprises a batch normalization layer, a first Silu nonlinear unit layer, a 1*1 second convolution layer, a second batch normalization layer, a second Silu nonlinear unit layer, and an n*1 third convolution layer.

[0219] The setting step of n comprises:

[0220] The reference value of the third convolution layer of the i-th layer is calculated according to an incremental model, wherein the incremental model is:

[0221] k = kernel_size + i * kernel_size_increment

[0222] k is the reference value, kernel_size is the initial convolution kernel size, i is the current dense layer number, and kernel_size_increment is the convolution kernel increment.

[0223] When k < the set convolution size K MAX , and k is even, then n = k + 1.

[0224] When k < the set convolution size K MAX , and k is odd, then n = k.

[0225] When k >= the set convolution size K MAX , then n = K MAX .

[0226] The first training module is configured to train the base calling model according to a current training condition and a data set to obtain a base calling model; wherein the training condition comprises the number of dense layers and the third convolution kernel size of the third convolution layer.

[0227] The second training module is configured to determine whether the prediction accuracy of the current base calling model is greater than a set accuracy threshold; if yes, the corresponding base calling model is output, and if no, the training condition is updated and returned to the first training module.

[0228] The data acquisition module is configured to acquire an electrical signal of a to-be-identified base.

[0229] The recognition module is configured to input the electrical signal into the base calling model for recognition to obtain a sequence of the to-be-identified base.

[0230] It can be understood that the system in the embodiment can also implement the method or steps in any of the above embodiments, which will not be described here.

[0231] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method or steps in any of the embodiments of the application.

[0232] The application also provides a computer program product, which is tangibly stored on a non-transitory computer readable medium and includes machine executable instructions for performing the method or steps in any of the embodiments of the application.

[0233] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles, or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or further include elements inherent to such processes, methods, articles, or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article, or device that includes the element.

[0234] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and necessary general hardware platforms, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a computer terminal (which can be a mobile phone, computer, server, or network device, etc.) execute the methods described in various embodiments of the application.

[0235] The embodiments of the application are described above in combination with the drawings, but the application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative, not restrictive, and those skilled in the art can make many forms under the inspiration of the application without departing from the purpose of the application and the scope protected by the claims, which are all within the protection of the application.

Claims

1. A base recognition method based on a fully convolutional network, characterized in that, include: S101, A base calling model is constructed based on the DenseNet convolutional neural network. The base calling model includes, in sequence: The first convolutional layer is used to reduce noise in the electrical signal; Max pooling layer is used to downsample electrical signals; Several dense blocks are used to extract features from electrical signals; Transition layers, interspersed between dense blocks, are used to reduce the dimensionality of the features output by the dense blocks; Fully connected layers are used to integrate the feature maps processed by dense blocks and transition layers globally and perform classification prediction. The CTC decoder is used to decode the prediction results output by the fully connected layer to obtain the sequence of bases. The dense block comprises multiple dense layers, which include: a batch normalization layer, a first Silu nonlinear unit layer, a 1×1 second convolutional layer, a second batch normalization layer, a second Silu nonlinear unit layer, and an n×1 third convolutional layer. The steps for setting n include: The reference value for the third convolutional layer of the i-th layer is calculated based on an incremental model, wherein the incremental model is: k= kernel_size + i * kernel_size_increment; k is a reference value, kernel_size is the initial convolutional kernel size, i is the current number of dense layers, and kernel_size_increment is the convolutional kernel increment; When k < the set convolution size K MAX If k is even, then n = k + 1; When k < the set convolution size K MAX If k is odd, then n = k; When k ≥ the set convolution size K MAX Then n=K MAX ; S102, the base calling model is trained using the dataset according to the current training conditions to obtain the updated base calling model; wherein, the training conditions include: the number of dense layers and the size of the third convolution kernel of the third convolutional layer; S103, determine whether the prediction accuracy of the currently obtained base calling model is greater than the set accuracy threshold; if yes, output the corresponding base calling model; if no, update the training conditions and return to S102. The plurality of dense blocks includes: a first dense block, a second dense block, a third dense block, and a fourth dense block; the first dense block and the second dense block are used to learn low-level features of the data, and the third dense block and the fourth dense block are used to learn high-level features of the data; correspondingly, the step of updating the training conditions further includes: Calculate the first prediction accuracy of the base calling model obtained from the previous training process; Calculate the second prediction accuracy of the base calling model obtained in the current training process; Calculate the accuracy growth rate based on the first prediction accuracy rate and the second prediction accuracy rate; When the accuracy increment rate is greater than or equal to the first increment threshold, it is recommended to perform a first modification operation on the number of dense layers in the dense blocks of the first set, wherein the first set includes: a first dense block, a second dense block, a third dense block and a fourth dense block. When the accuracy increment rate is less than the first increment threshold, it is recommended to perform a second modification operation on the number of dense layers in the dense blocks of the second set, wherein the second set includes: the third dense block and the fourth dense block; S104, collects the electrical signal of the base to be identified; S105, the electrical signal is input into the base calling model for identification to obtain the sequence of the base to be identified.

2. The base recognition method based on a fully convolutional network according to claim 1, characterized in that, The steps for updating the training conditions include: The number of dense layers in at least one dense block is modified, and the size of the third convolution kernel of each dense layer is updated based on the modified number of dense layers.

3. The base recognition method based on a fully convolutional network according to claim 2, characterized in that, It also includes the following steps: If the accuracy increment rate is still less than the first increment threshold after at least one second modification operation, it is recommended to modify the pooling size of the max pooling layer.

4. The base recognition method based on a fully convolutional network according to claim 3, characterized in that, The first or second modification operation must meet the following conditions: The number of dense layers in the second dense block is less than the first set value; The number of dense layers in the third dense block is greater than the second set value.

5. The base recognition method based on a fully convolutional network according to claim 1, characterized in that, K MAX The value is 99; and / or, the transition layer comprises a 1×1 convolutional layer.

6. The base recognition method based on a fully convolutional network according to claim 1, characterized in that, The dataset includes a training set, a validation set, and a test set; Furthermore, the training and validation sets include one or more of the following: base electrical signals from Arabidopsis thaliana, humans, Caenorhabditis elegans, and Escherichia coli; The test set data includes one or more of the following: base electrical signals from humans, Arabidopsis thaliana, mice, Saccharomyces cerevisiae, Populus alba, Plasmodium berghei, Seneca Valley virus, porcine epidemic diarrhea virus, and porcine reproductive and respiratory syndrome virus. And / or, the method further includes: preprocessing the electrical signal of the acquired base to be identified, including: The electrical signal is divided into signal blocks, each containing 4096 signal values; the signal values ​​are normalized using the median absolute bias method.

7. A base recognition system based on a fully convolutional network, characterized in that, include: The model building module is used to build a base calling model based on the DenseNet convolutional neural network. The base calling model includes, in sequence: The first convolutional layer is used to reduce noise in the electrical signal; Max pooling layer is used to downsample electrical signals; Several dense blocks are used to extract features from electrical signals; Transition layers, interspersed between dense blocks, are used to reduce the dimensionality of the features output by the dense blocks; Fully connected layers are used to integrate the feature maps processed by dense blocks and transition layers globally and perform classification prediction. The CTC decoder is used to decode the prediction results output by the fully connected layer to obtain the sequence of bases. The dense block includes multiple dense layers, which include: a batch normalization layer, a first Silu nonlinear unit layer, a 1×1 second convolutional layer, a second batch normalization layer, a second Silu nonlinear unit layer, and an n×1 third convolutional layer. The steps for setting n include: The reference value for the third convolutional layer of the i-th layer is calculated based on an incremental model, wherein the incremental model is: k= kernel_size + i * kernel_size_increment; k is a reference value, kernel_size is the initial convolutional kernel size, i is the current number of dense layers, and kernel_size_increment is the convolutional kernel increment; When k < the set convolution size K MAX If k is even, then n = k + 1; When k < the set convolution size K MAX If k is odd, then n = k; When k ≥ the set convolution size K MAX Then n=K MAX ; The first training module is used to train the base calling model using the dataset according to the current training conditions, so as to train and update the base calling model; wherein, the training conditions include: the number of dense layers and the size of the third convolution kernel of the third convolutional layer; The second training module is used to determine whether the prediction accuracy of the currently obtained base calling model is greater than a set accuracy threshold; if yes, it outputs the corresponding base calling model; if no, it updates the training conditions and returns to the first training module; wherein, the plurality of dense blocks include: a first dense block, a second dense block, a third dense block, and a fourth dense block; the first dense block and the second dense block are used to learn low-level features of the data, and the third dense block and the fourth dense block are used to learn high-level features of the data; correspondingly, the step of updating the training conditions further includes: Calculate the first prediction accuracy of the base calling model obtained from the previous training process; Calculate the second prediction accuracy of the base calling model obtained in the current training process; Calculate the accuracy growth rate based on the first prediction accuracy rate and the second prediction accuracy rate; When the accuracy increment rate is greater than or equal to the first increment threshold, it is recommended to perform a first modification operation on the number of dense layers in the dense blocks of the first set, wherein the first set includes: a first dense block, a second dense block, a third dense block and a fourth dense block. When the accuracy increment rate is less than the first increment threshold, it is recommended to perform a second modification operation on the number of dense layers in the dense blocks of the second set, wherein the second set includes: the third dense block and the fourth dense block; The data acquisition module is used to acquire the electrical signals of the bases to be identified; The identification module is used to input the electrical signal into the base calling model for identification to obtain the sequence of the base to be identified.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

9. A computer program product, said computer program product being tangibly stored on a non-transitory computer-readable medium and comprising machine-executable instructions, characterized in that, The machine-executable instructions are used to perform the method according to any one of claims 1-6.

Images