CN111370055A - A method for establishing an intron retention prediction model and its prediction method - Google Patents

Abstract

The invention discloses a method for establishing an intron retention prediction model, which includes collecting simulation data and real data related to intron retention; defining all independent intron sets in a genome and taking them as standard templates; The set intron sequence reading distribution pattern image data set is processed and processed to obtain the processed data set; the processed data set is divided into training set and test set according to the set ratio; the training set is used to train the neural network model to obtain the final establishment A neural network intron retention prediction model. The invention also discloses a prediction method including the method for establishing the intron retention prediction model. The invention can visualize and predict introns based on the distribution pattern of intron retention readings, and has high reliability and accuracy.

Description

A method for establishing an intron retention prediction model and its prediction method

technical field

The present invention specifically designs a method for establishing an intron retention prediction model and a method for predicting the same.

Background technique

Intron retention is a type of alternative splicing, which means that the intron in the precursor mRNA is not spliced out and remains in the mature mRNA. Intron retention was previously thought to be the result of incorrect splicing and received less attention. Many recent studies have shown that intron retention is related to gene expression regulation and complex diseases (such as Alzheimer's disease); and with the development of high-throughput sequencing technology, there are now many available for intron retention detection. The method proposed, iREAD and IRFinder are more prominent. Among them, iREAD detects intron retention by assuming that the reads of intron retention are evenly distributed, and calculates the entropy value, and the corresponding filter index is relatively strict. IRFinder detects intron retention by calculating the IR-ratio indicating the proportion of introns present in the transcript.

Although the above methods have been successfully applied in real environments, the analysis based on sequence features is more or less limited by the biases that may be caused by intron retention, resulting in insufficient robustness of the method, which makes the current The reliability of the method is not high, which restricts the development of related technologies.

SUMMARY OF THE INVENTION

One of the objectives of the present invention is to provide a method for establishing an intron retention prediction model with high reliability and accuracy.

Another object of the present invention is to provide a prediction method including the method for establishing the intron retention prediction model.

This intron retention prediction model establishment method provided by the invention comprises the following steps:

S1. Collect simulated data and real data related to intron retention;

S2. Define all independent intron sets in the genome and use them as standard templates;

S3. Obtain the intron sequence reading distribution pattern picture data set set in the simulation data obtained in step S1, and perform preprocessing to obtain the processed data set;

S4. Divide the processed data set obtained in step S3 into a training set and a test set according to a set ratio;

S5. Use the training set obtained in step S4 to train the neural network model, thereby obtaining the finally established neural network intron retention prediction model.

The method for establishing an intron retention prediction model further comprises the following steps:

S6. According to the neural network intron retention prediction model obtained in step S5, the evaluation parameters of the neural network intron retention prediction model are calculated on the test set obtained in step S4;

S7. Obtain the intron sequence reading distribution pattern picture test set of the real data obtained in step S1;

S8. according to the neural network intron retention prediction model obtained in step S5, predict the intron retention result on the test set obtained in step S7, thereby obtaining the predicted intron retention set;

S9. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of W1 bases on the exon side and N1 bases on the intron side, a total of W1+N1 bases 5'-end sequence;

S10. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of the exon side W2 bases, the intron side N2 bases, a total of W2+N2 bases 3'-end sequences;

S11. According to the 5'-end sequence of W1+N1 bases obtained in step S9 and the 3'-end sequence of W2+N2 bases obtained in step S10, calculate the splice site intensity, thereby obtaining the average splice site at the 5' end Intensity value and average splice site intensity value at 3' end;

S12. According to the average splice site intensity value at the 5' end and the average splice site intensity value at the 3' end obtained in step S11, evaluate the neural network intron retention prediction model established in step S5.

The described collection of intron retention related simulation data and real data in step S1 is specifically to adopt the BEER algorithm to generate a simulation data sequence file SIMU30 containing a certain number of introns; the sequencing depth of the simulated data sequence file SIMU30 is 3,000 10,000, read length is 100 bases, set to generate 15,000 genes, 69,338 introns; and a real data sequence file APP from the Tau and APP mouse model research of the Alzheimer's Disease Accelerated Drug Collaboration Project , the sequencing depth is 100 million, and the read length is 101 bases.

The set of all independent introns in the defined genome described in step S2 is combined as a standard template, specifically, the following steps are used to define:

A. From the annotated gtf file of the release-75 version of the GRCm38 mouse genome, extract all independent intron sets Independent_intron; the independent introns are defined as introns that do not overlap with any homotypic exons;

B. In the independent intron set Independent_intron obtained in step A, take the gene as the unit, merge the overlapping introns in the coordinate interval, and obtain the final independent intron set intron cluster.

The extraction of all the independent intron sets Independent_intron described in step A, specifically, is to merge all the exons in a chromosome, and then delete all the exons from the gene region, thereby obtaining all the independent introns.

The step S3 obtains the intron sequence reading distribution pattern picture data set set in the simulation data obtained in the step S1, and performs preprocessing to obtain the processed data set. Specifically, the following steps are used to obtain the data set and analyze the data. :

a. Perform IGV visualization on each intron in the simulation data sequence file SIMU30 obtained in step S1 to obtain a preliminary visualization image;

b. Save the visual images of two sequences of 20 bases on the left and right sides of the 5' end and 3' end of each intron, with a total length of 40 bases; the height of the visual image is 100mm, and the representative base The bar graph height of base abundance is normalized;

c. For the image obtained in step b, crop the part with a vertical length of 131-231 pixels and a part with a horizontal length of 280-1070 pixels;

d. Merge the images cropped in step c horizontally to obtain the final processed data set.

Described in step S4, the processed data set obtained in step S3 is divided into a training set and a test set according to a set ratio, specifically, in the simulated data sequence file SIMU30 obtained in step S1, it is defined that the total reading of the sequence is greater than the first set. Value, FPKM (fragments per kilobase per million reads matched to the gene, Fragments Per KilobaseMillion) is greater than the second set value and the introns with consecutive reads greater than the third set value are positive samples, and the remaining The introns are negative samples; then, in the positive and negative samples, X2 positive samples and X2 negative samples are randomly selected to form the final data set; then the data set is divided into training set and test set according to the set ratio; X2 is a positive integer.

The neural network model described in step S5 is specifically a VGG16 network structure model.

Described in step S5, using the training set obtained in step S4 to train the neural network model, so as to obtain the finally established neural network intron retention prediction model, specifically adopting the following steps to train the model:

(1) Obtain the VGG16 network structure model that has been trained on the ImageNet task and the corresponding weight parameter file; the network structure model includes 13 convolutional layers;

(2) Load the network and weight obtained in step (1) as a pre-training network, but freeze the network to ensure that the network does not participate in training;

(3) Define a two-class network, and perform training on the training set obtained in step S4; the two-class network has three layers in total, the first two layers are fully connected layers, and the number of neurons is 256 and 64 respectively. A dropout layer is followed to prevent overfitting, and the probability of randomly dropping neurons is set to 0.5 and 0.3, respectively. The last layer is the sigmoid layer for binary classification;

(4) After the classification network is trained, unfreeze the last three convolution layers of the pre-training network, use the training set obtained in step S4 again to train the classification network and the pre-training network together, and adjust the weights;

(5) The parameters of the model training process are set as follows:

The total number of parameters for model training is 33 million, of which the number of trainable parameters is 26 million, and the number of non-trainable parameters is 7 million;

The loss function is the two-category cross entropy loss, and the calculation formula is

where i is each sample, t _i is the real label of sample i; y _i is the predicted label of sample i;

The optimizer is RMSprop, the learning rate is 2e- ⁵ , and the number of iterations is 30;

The evaluation index is accuracy, and the calculation formula is:

Among them, Truepositive is the number of samples that are predicted to be positive and are actually positive; Truenegative is the number of samples that are predicted to be negative and are actually negative; Allsamples is the total number of samples;

Set ReduceLROnPlateau to monitor the learning rate every 2 iterations. If the monitoring learning rate does not decrease, adjust the learning rate to decrease by 50%;

Set if the evaluation index accuracy does not decrease in 10 iterations, stop the iteration in advance.

The calculation of the evaluation parameters of the neural network intron retention prediction model on the test set obtained in step S4 in step S6 is specifically calculating the AUC value of the neural network intron retention prediction model on the test set obtained in step S4.

The step S7 obtaining the intron sequence reading distribution pattern picture test set of the real data obtained in the step S1 is specifically inputting the sequence file APP of the real data obtained in the step S1 into the prediction tool iREAD and the prediction tool IRFinder, and obtaining respectively Two sets of intron retention prediction sets IR1 and IR2; map IR1 and IR2 to the independent intron set intron cluster according to the rule with the largest matching coordinate interval length, and then take the intersection of the two to obtain the intersection IC; then, the intersection IC IGV visualization, image cropping and merging operations are performed on each intron coordinate in the image, so as to obtain the image test set real_test of the intron sequence read distribution pattern of the real data.

According to the 5'-end sequence of W1+N1 bases obtained in step S9 and the 3'-end sequence of W2+N2 bases obtained in step S10 described in step S11, the splice site strength is calculated, so as to obtain the average 5'-end sequence. The splice site intensity value and the average splice site intensity value at the 3' end are input into the MaxEntScan model by inputting the 5' end sequence score5ss sequence set obtained in step S9 and the 3' end sequence score3ss sequence set obtained in step S10 into the MaxEntScan model. The entropy model is used to score to obtain a given splice site intensity value; then the average splice site intensity corresponding to the 5'-end sequence and the 3'-end sequence is averaged to obtain the final 5'-end average splice site intensity value and the average splice site intensity value at the 3' end.

In step S12, the neural network intron retention prediction model established in step S5 is evaluated according to the 5'-end average splice site intensity value and the 3'-end average splice site intensity value obtained in step S11. The smaller the 5'-end average splice site intensity value and the 3'-end average splice site intensity value of the network intron retention prediction model, the better the prediction effect of the neural network intron retention prediction model.

The present invention also provides a prediction method comprising the above-mentioned intron retention prediction model establishment method, which specifically further comprises the following steps:

S13. Use the neural network intron retention prediction model obtained in step S5 to predict the intron retention result.

The intron retention prediction model establishment method and prediction method provided by the present invention, the intron retention deep learning prediction method based on the intron retention read distribution pattern can predict intron retention more generally and easily; The intron retention read distribution pattern, combined with deep learning model construction and transfer learning applications, transfers the knowledge structure of large-scale image classification tasks, and completes and improves the learning effect of intron retention prediction tasks; at the same time, there is no gold standard. The prediction effect was evaluated on the real data set, and it was proposed to calculate the average splice site strength of the 5' and 3' end sequences of the predicted intron-reserved sequences to measure the overall prediction effect; therefore, the method of the present invention can be based on introns. Preserve read distribution patterns for visualization and prediction of introns with high reliability and accuracy.

Description of drawings

FIG. 1 is a schematic flowchart of a method for establishing an intron retention prediction model according to the present invention.

Figure 2 is a schematic diagram of the visualization results of the distribution pattern of intron retention reads of the present invention.

FIG. 3 is a schematic structural diagram of the deep learning model VGG16 of the present invention.

FIG. 4 is a schematic flowchart of the prediction method of the present invention.

Detailed ways

As shown in Figure 1, the method flow diagram of the intron retention prediction model establishment method of the present invention: this intron retention prediction model establishment method provided by the present invention comprises the following steps:

S1. collect simulation data and real data related to intron retention; specifically, adopt the BEER algorithm to generate a simulation data sequence file SIMU30 containing a certain number of introns; the sequencing depth of the simulated data sequence file SIMU30 is 30 million, and the number of reads is 30 million. The length is 100 bases, and it is set to generate 15,000 genes and 69,338 introns; and a real data sequence file APP from the Tau and APP mouse model research of the Alzheimer's Disease Accelerated Drug Cooperation Project, sequencing depth is 100 million, and the read length is 101 bases;

S2. Define all independent intron sets in the genome and use them as standard templates; the present invention can be applied to mice, so the genome can be a mouse genome; specifically, the following steps are used to define:

A. From the annotated gtf file of the release-75 version of the GRCm38 mouse genome, extract all independent intron sets Independent_intron; the independent introns are defined as introns that do not overlap with any homotypic exons;

Among them, extract all independent intron sets Independent_intron, specifically, merge all exons in a chromosome, and then delete all exons from the gene region to obtain all independent introns;

B. In the independent intron set Independent_intron obtained in step A, take the gene as the unit, merge the overlapping introns in the coordinate interval, and obtain the final independent intron set intron cluster;

S3. Obtain the intron sequence read distribution pattern picture data set set in the simulation data obtained in step S1, and perform preprocessing to obtain the processed data set; specifically, the following steps are used to obtain the data set and carry out the data:

a. Perform IGV visualization on each intron in the simulation data sequence file SIMU30 obtained in step S1 to obtain a preliminary visualization image;

b. Since the length of each intron is indeterminate and very different, two sections of 20 bases on the left and right of the 5' end and 3' end of each intron are respectively saved, with a total length of 40 bases. Sequence visualization image; the height of the visualization image is 100mm, and the height of the bar graph representing base abundance is normalized;

c. For the image obtained in step b, the original visual image of a single-segment sequence is 621 pixels long and 1150 pixels horizontally, so the part of the entire image with a vertical length of 131 to 231 pixels and a horizontal length of 280 to 1070 pixels are cropped. part;

d. Merge the images cropped in step c horizontally to obtain the final processed data set; the visualization result is shown in Figure 2;

S4. Divide the processed data set obtained in step S3 into a training set and a test set according to a set ratio; specifically, in the simulated data sequence file SIMU30 obtained in step S1, the total reading of the defined sequence is greater than the first set value (such as 10), FPKM (the number of fragments matching each kilobase in the gene per million reads, Fragments Per KilobaseMillion) is greater than the second set value (such as 0.3) and consecutive reads are greater than the third set value (such as 1) The introns are positive samples, and the remaining introns are negative samples; then, in the positive and negative samples, randomly select X2 (such as 5000) positive samples and X2 negative samples to form the final data set; then follow the setting The ratio (such as 7:3) divides the dataset into training set and test set; X2 is a positive integer.

S5. use the training set obtained in step S4 to train the neural network model, thereby obtaining the finally established neural network intron retention prediction model; during specific implementation, the prediction model is preferably the VGG16 model; and when VGG16 is selected as the prediction model, you can Use the following steps to train the model:

(1) Obtain the VGG16 network structure model that has been trained on the ImageNet task (as shown in Figure 3) and the corresponding weight parameter file; the network structure model includes 13 convolution layers;

(2) Load the network and weight obtained in step (1) as a pre-training network, but freeze the network to ensure that the network does not participate in training;

(3) Define a two-class network, and perform training on the training set obtained in step S4; the two-class network has three layers in total, the first two layers are fully connected layers, and the number of neurons is 256 and 64 respectively. A dropout layer is followed to prevent overfitting, and the probability of randomly dropping neurons is set to 0.5 and 0.3, respectively. The last layer is the sigmoid layer for binary classification;

(4) After the classification network is trained, unfreeze the last three convolution layers of the pre-training network, use the training set obtained in step S4 again to train the classification network and the pre-training network together, and adjust the weights;

(5) The parameters of the model training process are set as follows:

The total number of parameters for model training is 33 million, of which the number of trainable parameters is 26 million, and the number of non-trainable parameters is 7 million;

The loss function is the two-category cross entropy loss, and the calculation formula is

where i is each sample, t _i is the real label of sample i; y _i is the predicted label of sample i;

The optimizer is RMSprop, the learning rate is 2e- ⁵ , and the number of iterations is 30;

The evaluation index is accuracy, and the calculation formula is:

Among them, Truepositive is the number of samples that are predicted to be positive and are actually positive; Truenegative is the number of samples that are predicted to be negative and are actually negative; Allsamples is the total number of samples;

Set ReduceLROnPlateau to monitor the learning rate every 2 iterations. If the monitoring learning rate does not decrease, adjust the learning rate to decrease by 50%;

Set if the evaluation index accuracy does not decrease in 10 iterations, stop the iteration in advance

S6. according to the neural network intron retention prediction model obtained in step S5, calculate the evaluation parameter (preferably AUC value) of the neural network intron retention prediction model on the test set obtained in step S4;

S7. Obtain the intron sequence reading distribution pattern picture test set of the real data obtained in step S1; specifically, input the sequence file APP of the real data obtained in step S1 into the prediction tool iREAD and the prediction tool IRFinder, and obtain two groups of introns respectively. Intron retention prediction sets IR1 and IR2; map IR1 and IR2 to the independent intron set intron cluster according to the rule with the largest matching coordinate interval length, and then take the intersection of the two to obtain the intersection IC; The intron coordinates are subjected to IGV visualization, image cropping and merging, etc., to obtain the image test set real_test of the intron sequence read distribution pattern of the real data;

S8. according to the neural network intron retention prediction model obtained in step S5, predict the intron retention result on the test set obtained in step S7, thereby obtaining the predicted intron retention set;

S9. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of W1 bases on the exon side and N1 bases on the intron side, a total of W1+N1 bases 5'-end sequence;

S10. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of the exon side W2 bases, the intron side N2 bases, a total of W2+N2 bases 3'-end sequences;

S11. According to the 5'-end sequence of W1+N1 bases obtained in step S9 and the 3'-end sequence of W2+N2 bases obtained in step S10, calculate the splice site intensity, thereby obtaining the average splice site at the 5' end The intensity value and the 3'-end average splice site intensity value; specifically, the 5'-end sequence score5ss sequence set obtained in step S9 and the 3'-end sequence score3ss sequence set obtained in step S10 are input into the MaxEntScan model, and the maximum entropy model is used to carry out Score to obtain a given splice site intensity value; then average the splice site intensities corresponding to the 5'-end sequence and the 3'-end sequence to obtain the final 5'-end average splice site intensity value and 3 ' end average splice site strength value;

S12. Evaluate the neural network intron retention prediction model established in step S5 according to the average splice site intensity value at the 5' end and the average splice site intensity value at the 3' end obtained in step S11; specifically, if the neural network contains The smaller the 5'-end average splice site intensity value and the 3'-end average splice site intensity value of the intron retention prediction model, the better the prediction effect of the neural network intron retention prediction model.

The method of the present invention is verified as follows:

The invention is evaluated on simulated data SIMU30 and real data set APP, and the tools compared with the invention are iREAD and IRFinder.

1) Experimental analysis of SIMU30 simulation data set

For 3000 test set samples of SIMU30 simulation data, the prediction Accuracy of the present invention on it reaches 0.925, and the AUC reaches 0.975;

2) APP real data set experimental analysis

Due to the lack of gold standards for real data, on the one hand, the predicted labels of other methods can only be used as real labels to test the gap between the AUC of the VGG16 model of the present invention and other methods; on the other hand, other evaluation indicators can be customized to verify the performance of the present invention effectiveness. In terms of AUC evaluation, after predicting the real data picture test set real_test, the VGG16 model of the present invention is compared with iREAD and IRFinder in Table 1. The real_test has a total of 68326 samples. When iREAD is the gold standard, the number of positive samples is 2816 and the number of negative samples is 65510. At this time, the AUC of the VGG16 model of the present invention is better than that of IRFinder. When taking IRFinder as the gold standard, the number of positive samples is 19044 and the number of negative samples is 49282. At this time, the present invention is also better than iREAD.

Table 1 Schematic representation of the AUC evaluation results of the present invention, iREAD and IRFinder

In addition, the present invention also defines the 5'-end and 3'-end splice site strengths to measure the prediction effect of the VGG16 model. The lower the average splice site intensity, the better the overall prediction effect of the model. The average splice site strength evaluation results are shown in Table 2.

Table 2 Schematic table of the average splice site strength evaluation results between the present invention and iREAD and IRFinder

From the results in Table 2, although the results of the present invention are slightly worse than IRFinder and iREAD in terms of average splice site intensity, it is noted that as the number of introns involved in calculating the average splice site intensity increases, IRFinder and iREAD The mean splice site strength of α increases with it, while it decreases for the present invention. This reflects that the VGG16 model designed by the present invention is superior to IRFinder and iREAD in terms of robustness.

Figure 4 is a schematic flow chart of the prediction method of the present invention: the prediction method provided by the present invention including the above-mentioned intron retention prediction model establishment method specifically includes the following steps:

S1. collect simulation data and real data related to intron retention; specifically, adopt the BEER algorithm to generate a simulation data sequence file SIMU30 containing a certain number of introns; the sequencing depth of the simulated data sequence file SIMU30 is 30 million, and the number of reads is 30 million. The length is 100 bases, and it is set to generate 15,000 genes and 69,338 introns; and a real data sequence file APP from the Tau and APP mouse model research of the Alzheimer's Disease Accelerated Drug Cooperation Project, sequencing depth is 100 million, and the read length is 101 bases;

S2. Define all independent intron sets in the genome and use them as standard templates; specifically, the following steps are used to define:

A. From the annotated gtf file of the release-75 version of the GRCm38 mouse genome, extract all independent intron sets Independent_intron; the independent introns are defined as introns that do not overlap with any homotypic exons;

Among them, extract all independent intron sets Independent_intron, specifically, merge all exons in a chromosome, and then delete all exons from the gene region to obtain all independent introns;

B. In the independent intron set Independent_intron obtained in step A, take the gene as the unit, merge the overlapping introns in the coordinate interval, and obtain the final independent intron set intron cluster;

S3. Obtain the intron sequence read distribution pattern picture data set set in the simulation data obtained in step S1, and perform preprocessing to obtain the processed data set; specifically, the following steps are used to obtain the data set and carry out the data:

a. Perform IGV visualization on each intron in the simulation data sequence file SIMU30 obtained in step S1 to obtain a preliminary visualization image;

b. Since the length of each intron is indeterminate and very different, two sections of 20 bases on the left and right of the 5' end and 3' end of each intron are respectively saved, with a total length of 40 bases. Sequence visualization image; the height of the visualization image is 100mm, and the height of the bar graph representing base abundance is normalized;

c. For the image obtained in step b, the original visual image of a single-segment sequence is 621 pixels long and 1150 pixels horizontally, so the part of the entire image with a vertical length of 131 to 231 pixels and a horizontal length of 280 to 1070 pixels are cropped. part;

d. Merge the images cropped in step c horizontally to obtain the final processed data set; the visualization result is shown in Figure 2;

S4. Divide the processed data set obtained in step S3 into a training set and a test set according to a set ratio; specifically, in the simulated data sequence file SIMU30 obtained in step S1, the total reading of the defined sequence is greater than the first set value (such as 10), FPKM (the number of fragments matching each kilobase in the gene per million reads, Fragments Per KilobaseMillion) is greater than the second set value (such as 0.3) and consecutive reads are greater than the third set value (such as 1) The introns are positive samples, and the remaining introns are negative samples; then, in the positive and negative samples, randomly select X2 (such as 5000) positive samples and X2 negative samples to form the final data set; then follow the setting The ratio (such as 7:3) divides the dataset into training set and test set; X2 is a positive integer.

S5. use the training set obtained in step S4 to train the neural network model, thereby obtaining the finally established neural network intron retention prediction model; during specific implementation, the prediction model is preferably the VGG16 model; and when VGG16 is selected as the prediction model, you can Use the following steps to train the model:

(1) Obtain the VGG16 network structure model that has been trained on the ImageNet task (as shown in Figure 3) and the corresponding weight parameter file; the network structure model includes 13 convolution layers;

(2) Load the network and weight obtained in step (1) as a pre-training network, but freeze the network to ensure that the network does not participate in training;

(3) Define a two-class network, and perform training on the training set obtained in step S4; the two-class network has three layers in total, the first two layers are fully connected layers, and the number of neurons is 256 and 64 respectively. A dropout layer is followed to prevent overfitting, and the probability of randomly dropping neurons is set to 0.5 and 0.3, respectively. The last layer is the sigmoid layer for binary classification;

(4) After the classification network is trained, unfreeze the last three convolution layers of the pre-training network, use the training set obtained in step S4 again to train the classification network and the pre-training network together, and adjust the weights;

(5) The parameters of the model training process are set as follows:

The total number of parameters for model training is 33 million, of which the number of trainable parameters is 26 million, and the number of non-trainable parameters is 7 million;

The loss function is the two-category cross entropy loss, and the calculation formula is

where i is each sample, t _i is the real label of sample i; y _i is the predicted label of sample i;

The optimizer is RMSprop, the learning rate is 2e- ⁵ , and the number of iterations is 30;

The evaluation index is accuracy, and the calculation formula is:

Among them, Truepositive is the number of samples that are predicted to be positive and are actually positive; Truenegative is the number of samples that are predicted to be negative and are actually negative; Allsamples is the total number of samples;

Set ReduceLROnPlateau to monitor the learning rate every 2 iterations. If the monitoring learning rate does not decrease, adjust the learning rate to decrease by 50%;

Set if the evaluation index accuracy does not decrease in 10 iterations, stop the iteration in advance

S6. according to the neural network intron retention prediction model obtained in step S5, calculate the evaluation parameter (preferably AUC value) of the neural network intron retention prediction model on the test set obtained in step S4;

S7. Obtain the intron sequence reading distribution pattern picture test set of the real data obtained in step S1; specifically, input the sequence file APP of the real data obtained in step S1 into the prediction tool iREAD and the prediction tool IRFinder, and obtain two groups of introns respectively. Intron retention prediction sets IR1 and IR2; map IR1 and IR2 to the independent intron set intron cluster according to the rule with the largest matching coordinate interval length, and then take the intersection of the two to obtain the intersection IC; The intron coordinates are subjected to IGV visualization, image cropping and merging, etc., to obtain the image test set real_test of the intron sequence read distribution pattern of the real data;

S8. according to the neural network intron retention prediction model obtained in step S5, predict the intron retention result on the test set obtained in step S7, thereby obtaining the predicted intron retention set;

S9. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of W1 bases on the exon side and N1 bases on the intron side, a total of W1+N1 bases 5'-end sequence;

S10. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of the exon side W2 bases, the intron side N2 bases, a total of W2+N2 bases 3'-end sequences;

S11. According to the 5'-end sequence of W1+N1 bases obtained in step S9 and the 3'-end sequence of W2+N2 bases obtained in step S10, calculate the splice site intensity, thereby obtaining the average splice site at the 5' end The intensity value and the 3'-end average splice site intensity value; specifically, the 5'-end sequence score5ss sequence set obtained in step S9 and the 3'-end sequence score3ss sequence set obtained in step S10 are input into the MaxEntScan model, and the maximum entropy model is used to carry out Score to obtain a given splice site intensity value; then average the splice site intensities corresponding to the 5'-end sequence and the 3'-end sequence to obtain the final 5'-end average splice site intensity value and 3 ' end average splice site strength value;

S12. Evaluate the neural network intron retention prediction model established in step S5 according to the average splice site intensity value at the 5' end and the average splice site intensity value at the 3' end obtained in step S11; specifically, if the neural network contains The smaller the 5'-end average splice site intensity value and the 3'-end average splice site intensity value of the intron retention prediction model, the better the prediction effect of the neural network intron retention prediction model;

S13. Use the neural network intron retention prediction model obtained in step S5 to predict the intron retention result.

Claims (14)

1. a method for establishing an intron retention prediction model, comprising the steps: S1. Collect simulated data and real data related to intron retention; S2. Define all independent intron sets in the genome and use them as standard templates; S3. Obtain the intron sequence reading distribution pattern picture data set set in the simulation data obtained in step S1, and perform preprocessing to obtain the processed data set; S4. Divide the processed data set obtained in step S3 into a training set and a test set according to a set ratio; S5. Use the training set obtained in step S4 to train the neural network model, thereby obtaining the finally established neural network intron retention prediction model. 2. intron retention prediction model establishment method according to claim 1 is characterized in that also comprising the steps: S6. According to the neural network intron retention prediction model obtained in step S5, the evaluation parameters of the neural network intron retention prediction model are calculated on the test set obtained in step S4; S7. Obtain the intron sequence reading distribution pattern picture test set of the real data obtained in step S1; S8. according to the neural network intron retention prediction model obtained in step S5, predict the intron retention result on the test set obtained in step S7, thereby obtaining the predicted intron retention set; S9. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of W1 bases on the exon side and N1 bases on the intron side, a total of W1+N1 bases 5'-end sequence; S10. Obtain the predicted intron retention set obtained in step S8, revealing the coordinates of the exon side W2 bases, the intron side N2 bases, a total of W2+N2 bases 3'-end sequences; S11. According to the 5'-end sequence of W1+N1 bases obtained in step S9 and the 3'-end sequence of W2+N2 bases obtained in step S10, calculate the splice site intensity, thereby obtaining the average splice site at the 5' end Intensity value and average splice site intensity value at 3' end; S12. According to the average splice site intensity value at the 5' end and the average splice site intensity value at the 3' end obtained in step S11, evaluate the neural network intron retention prediction model established in step S5. 3. intron retention prediction model establishment method according to claim 2 is characterized in that the described collection of step S1 retains relevant simulated data and real data, and is specially for adopting BEER algorithm to generate and contain certain introns The number of simulated data sequence files SIMU30; the sequencing depth of the simulated data sequence file SIMU30 is 30 million, the read length is 100 bases, and it is set to generate 15,000 genes and 69,338 introns; A real data sequence file APP in the Tau and APP mouse model research of the Disease Accelerating Drug Cooperation Project, with a sequencing depth of 100 million and a read length of 101 bases. 4. intron retention prediction model establishment method according to claim 3 is characterized in that all independent intron sets in the defined genome described in step S2 are combined as standard template, specifically adopt the following steps to define: A. From the annotated gtf file of the release-75 version of the GRCm38 mouse genome, extract all independent intron sets Independent_intron; the independent introns are defined as introns that do not overlap with any homotypic exons; B. In the independent intron set Independent_intron obtained in step A, take the gene as the unit, merge the overlapping introns in the coordinate interval, and obtain the final independent intron set intron cluster. 5. The method for establishing an intron retention prediction model according to claim 4, wherein the extraction of all independent intron sets Independent_intron described in step A is specifically merging all exons in a chromosome, and then extracting all the exons from the chromosome. The gene region deletes all exons, resulting in all independent introns. 6. The method for establishing an intron retention prediction model according to claim 5, characterized in that the intron sequence reading distribution pattern picture data set set in the simulation data obtained by the obtaining step S1 of the step S3, and The processed data set is obtained by preprocessing, specifically, the following steps are used to obtain the data set and perform the data: a. Perform IGV visualization on each intron in the simulation data sequence file SIMU30 obtained in step S1 to obtain a preliminary visualization image; b. Save the visual images of two sequences of 20 bases on the left and right sides of the 5' end and 3' end of each intron, with a total length of 40 bases; the height of the visual image is 100mm, and the representative base The bar graph height of base abundance is normalized; c. For the image obtained in step b, crop the part with a vertical length of 131-231 pixels and a part with a horizontal length of 280-1070 pixels; d. Merge the images cropped in step c horizontally to obtain the final processed data set. 7. intron retention prediction model establishment method according to claim 6 is characterized in that described in step S4, the processed data set obtained in step S3 is divided into training set and test set according to a set ratio, and is specifically In the simulated data sequence file SIMU30 obtained in step S1, define the introns whose total sequence readings are greater than the first set value, the FPKM is greater than the second set value, and the continuous readings are greater than the third set value as positive samples, and the remaining introns are defined as positive samples. Contain is a negative sample; then, in the positive and negative samples, randomly select X2 positive samples and X2 negative samples to form the final data set; then divide the data set into training set and test set according to the set ratio; X2 is positive integer. 8 . The method for establishing an intron retention prediction model according to claim 7 , wherein the neural network model described in step S5 is specifically a VGG16 network structure model. 9 . 9. intron retention prediction model establishment method according to claim 8 is characterized in that the training set training neural network model that adopts step S4 to obtain described in step S5, thereby obtains the neural network intron retention prediction finally established The model is specifically trained by the following steps: (1) Obtain the VGG16 network structure model that has been trained on the ImageNet task and the corresponding weight parameter file; the network structure model includes 13 convolutional layers; (2) Load the network and weight obtained in step (1) as a pre-training network, but freeze the network to ensure that the network does not participate in training; (3) Define a two-class network, and perform training on the training set obtained in step S4; the two-class network has three layers in total, the first two layers are fully connected layers, and the number of neurons is 256 and 64 respectively. A Dropout layer is followed to prevent overfitting, and the probability of randomly discarding neurons is set to 0.5 and 0.3 respectively; the last layer is a sigmoid layer for binary classification; (4) After the classification network is trained, unfreeze the last three convolution layers of the pre-training network, use the training set obtained in step S4 again to train the classification network and the pre-training network together, and adjust the weights; (5) The parameters of the model training process are set as follows: The total number of parameters for model training is 33 million, of which the number of trainable parameters is 26 million, and the number of non-trainable parameters is 7 million; The loss function is the two-category cross entropy loss, and the calculation formula is

where i is each sample, t _i is the real label of sample i; y _i is the predicted label of sample i; The optimizer is RMSprop, the learning rate is 2e- ⁵ , and the number of iterations is 30; The evaluation index is accuracy, and the calculation formula is:

Among them, Truepositive is the number of samples that are predicted to be positive and are actually positive; Truenegative is the number of samples that are predicted to be negative and are actually negative; Allsamples is the total number of samples; Set ReduceLROnPlateau to monitor the learning rate every 2 iterations. If the monitoring learning rate does not decrease, adjust the learning rate to decrease by 50%; Set if the evaluation index accuracy does not decrease in 10 iterations, stop the iteration in advance. 10. The method for establishing an intron retention prediction model according to claim 9, wherein the step S6 calculates the evaluation parameter of the neural network intron retention prediction model on the test set obtained in the step S4, specifically in the The AUC value of the neural network intron retention prediction model is calculated on the test set obtained in step S4. 11. The method for establishing an intron retention prediction model according to claim 10, characterized in that the intron sequence reading distribution pattern picture test set of the real data obtained in step S7 of obtaining step S1 is specifically the step S1. The sequence file APP of the obtained real data is input into the prediction tool iREAD and the prediction tool IRFinder, and two sets of intron retention prediction sets IR1 and IR2 are obtained respectively; IR1 and IR2 are mapped to independent introns according to the rule with the longest matching coordinate interval. On the sub-set intron cluster, take the intersection of the two to obtain the intersection IC; then, perform IGV visualization, image cropping and merging operations on the coordinates of each intron in the intersection IC to obtain the intron sequence read distribution pattern of the real data. Image test set real_test. 12. The method for establishing an intron retention prediction model according to claim 11, wherein the 5'-end sequence of W1+N1 bases obtained in step S11 according to step S9 and W2+N2 obtained in step S10 The 3' end sequence of each base is calculated, and the splice site intensity is calculated to obtain the average splice site intensity value of the 5' end and the average splice site intensity value of the 3' end. Specifically, the 5' end sequence score5ss sequence obtained in step S9 is obtained. The set and the 3'-end sequence score3ss sequence set obtained in step S10 are input into the MaxEntScan model, and the maximum entropy model is used to score, thereby obtaining a given splice site strength value; then the 5'-end sequence and the 3'-end sequence corresponding to The average splice site intensity of , to obtain the final 5' end average splice site intensity value and 3' end average splice site intensity value. 13. The method for establishing an intron retention prediction model according to claim 12, characterized in that the 5'-end average splice site intensity value and the 3'-end average splice site intensity value obtained according to step S11 in step S12 , evaluate the neural network intron retention prediction model established in step S5, specifically, if the 5'-end average splice site intensity value and the 3'-end average splice site intensity value of the neural network intron retention prediction model are smaller , the prediction effect of the neural network intron retention prediction model is better. 14. A prediction method comprising the method for establishing an intron retention prediction model according to one of claims 1 to 13, further comprising the steps of: S13. Use the neural network intron retention prediction model obtained in step S5 to predict the intron retention result.

Images