A method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model

Abstract

The present invention relates to the field of computer technology, and discloses a method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model, comprising the steps of: screening tumor information of historical nasopharyngeal carcinoma patients before and after neoadjuvant chemotherapy, and screening differential genes; constructing a training data set based on differential genes, calculating the connection weights between differential gene feature data of any two patients, constructing a loss function of a sensitivity prediction model, and training the sensitivity prediction model; after the sensitivity prediction model is trained, the patient's gene data is input into the sensitivity prediction model to obtain a prediction result of whether the patient belongs to a chemotherapy sensitive group or a chemotherapy resistant group. The present invention constructs a sensitivity prediction model by learning the mutual relationship between differential gene feature data between nasopharyngeal carcinoma patients. After the patient's genes are input into the sensitivity prediction model, the model can predict the patient's sensitivity to neoadjuvant chemotherapy according to various combinations of the patient's genes, which is helpful for individualized treatment decisions.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model. Background Art

[0002] According to the 2020 data released by the International Agency for Research on Cancer (IARC), there were 133,000 new cases of nasopharyngeal carcinoma (NPC) worldwide in 2020. The site of NPC onset is hidden, and the early clinical symptoms are not obvious. More than 70% of NPC patients are already locally advanced when diagnosed. At present, the recommended treatment for locally advanced NPC is neoadjuvant chemotherapy combined with induction chemotherapy and synchronous chemoradiotherapy. Studies have shown that this neoadjuvant chemotherapy method has benefits in terms of overall survival (OS), progression-free survival (PFS), distant metastasis-free survival (DMFS), and local recurrence-free survival (LRRFS). However, about 20% to 30% of patients are insensitive to neoadjuvant chemotherapy and cannot benefit from it, and the treatment efficacy is also affected due to the delay in radiotherapy. With the development of molecular biology research and high-throughput sequencing technology, people have gradually realized that differences at the genetic level may affect chemotherapy sensitivity. Although some studies have tried to find a single biomarker or clinical feature for prediction, the overall accuracy is poor, clinical promotion is limited, the human judgment cycle is long, the sample is small, the accuracy is low, and the prediction is difficult. Therefore, there is an urgent need for a neural network prediction model constructed based on multi-gene expression characteristics to effectively and quickly identify whether patients are sensitive to neoadjuvant chemotherapy and help make individualized treatment decisions. Summary of the invention

[0003] The purpose of the present invention is to effectively and quickly identify whether a patient is sensitive to neoadjuvant chemotherapy, and to provide a method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model.

[0004] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0005] The method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model includes the following steps:

[0006] Step 1: Screen the tumor information of historical nasopharyngeal carcinoma patients before and after neoadjuvant chemotherapy to screen out differentially expressed genes;

[0007] Step 2: construct a training data set based on differentially expressed genes, calculate the connection weight between differentially expressed gene feature data of any two patients, construct a loss function of the sensitivity prediction model, and train the sensitivity prediction model;

[0008] Step 3, after the sensitivity prediction model is trained, the patient's genetic data is input into the sensitivity prediction model to obtain a prediction result of whether the patient belongs to the chemotherapy-sensitive group or the chemotherapy-resistant group.

[0009] The step 2 specifically includes the following steps:

[0010] Step 2-1, construct differential gene signature data for each patient, thereby forming a training data set for all patients;

[0011] Step 2-2, construct feature matrix;

[0012] Step 2-3, calculate the connection weight of each element in the feature matrix based on the flow learning algorithm;

[0013] Step 2-4, based on the connection weights, construct the loss function of the sensitivity prediction model.

[0014] Compared with the prior art, the present invention has the following beneficial effects: the present invention constructs a sensitivity prediction model by learning the mutual relationship between differential gene feature data among nasopharyngeal carcinoma patients. After the patient's genes are input into the sensitivity prediction model, the model can predict the patient's sensitivity to neoadjuvant chemotherapy based on various combinations of the patient's genes (including differential genes and non-differential genes), which is helpful for individualized treatment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 is a flow chart of the method of the present invention;

[0017] Figure 2 Schematic diagram of differential gene expression in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0019] It should be noted that similar numbers and letters represent similar items in the following figures, so once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance, or implying any such actual relationship or order between these entities or operations. In addition, the terms "connected", "connected", etc. can be directly connected between elements, or indirectly connected via other elements.

[0020] Embodiment 1:

[0021] The present invention is achieved through the following technical solutions: Figure 1 As shown, the method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model includes the following steps:

[0022] Step 1: Screen the tumor information of historical nasopharyngeal carcinoma patients before and after neoadjuvant chemotherapy to screen out differentially expressed genes.

[0023] A retrospective analysis was conducted on 500 patients who underwent nasopharyngeal carcinoma treatment in multiple hospitals from January 2020 to December 2021. By comparing the imaging examinations before and after neoadjuvant chemotherapy, n chemotherapy-sensitive patients and m chemotherapy-resistant patients were screened out, m+n<500 or m+n≤500. The division criteria were: ① Chemotherapy-sensitive group: the tumor volume after chemotherapy was reduced by more than 75% compared with that before chemotherapy; ② Chemotherapy-resistant group: the tumor volume after chemotherapy was reduced by less than 25% compared with that before chemotherapy, and the heart function was normal, and there was no significant difference in age, gender, and education level between the experimental group and the experimental group.

[0024] Exclusion criteria: 1. Patients under 18 years old; 2. Patients with other tumors at the same time or who had other tumors within 5 years; 3. Patients who had not completed at least one efficacy evaluation during data collection; 4. Patients without complete follow-up data; 5. Patients with severe mental illness or psychological disorders and unable to cooperate with treatment.

[0025] Based on the division of chemotherapy-sensitive group and chemotherapy-resistant group, the gene expression of patients was analyzed:

[0026] 1. Library construction and sequencing: RNA is extracted from the patient's tumor tissue, and transcripts mainly composed of nRNA are obtained from the total RNA. Through fragmentation, reverse transcription, adding adapters and indexing, a stable library that can be used for sequencing is obtained. After the library inspection is qualified, different libraries are pooled according to the effective concentration and the target amount of data required for the machine, and then gene sequencing is performed on the Illumina sequencing platform. The basic principle of gene sequencing is sequencing while synthesis. Fluorescently labeled dNTPs, DNA polymerases, and adapter primers are added to the sequencing flow absorption pool for extension. When each sequencing cluster extends the complementary chain, each fluorescently labeled dNTP added releases a corresponding fluorescent signal. The Illumina sequencing platform captures the fluorescent signal and converts the light signal into a sequencing peak through computer software, thereby obtaining the original sequence information of the fragment to be tested.

[0027] 2. Data quality control: The original sequence information obtained by sequencing contains low-quality, short sequences, and adapter sequences, which will have a great impact on subsequent information analysis. In order to ensure the quality of information analysis, the original data (raw reads) must be horizontally filtered and sample filtered to obtain high-quality data (clean reads), and clean reads are used for subsequent analysis.

[0028] The criteria for horizontal filtering of raw reads using fastp software are as follows: (1) remove adapters; (2) remove the first 3 bp bases of two reads; (3) if the proportion of low-quality bases in a read exceeds 40%, the entire double-end read is deleted; (4) if the number of N in a read is higher than 5, the entire double-end read is discarded; (5) low_complexity_filter, reads with low complexity will be discarded, and complexity is defined as the percentage of bases that are different from the next base; (6) read length must be greater than 50 bp. After fastp software filtering, bowtie2 software is used to align to the rRNA sequence database containing humans and microorganisms to remove rRNA. After filtering at the read level, filtering at the sample level was performed. First, quality control indicators such as the Q20 / Q30 base ratio, data volume, and deduplication alignment rate were calculated after alignment with the reference genome. Then, sample filtering was performed according to the following standards: (1) Base quality filtering: Q20 base ratio > 90%, Q30 base ratio > 80%, otherwise it was judged as an unqualified sample; (2) Data volume after filtering: Clean Bases > 3G after removing adapters, low-quality reads, and rRNA, otherwise it was judged as an unqualified sample; (3) Deduplication alignment rate: The ratio of deduplication reads aligned to the human reference genome > 60%, otherwise it was judged as an unqualified sample.

[0029] 3. Transcript quantification: Use featureCounts software to perform quantitative analysis of gene expression levels for each sample to obtain the expression (readcount) of each gene in each sample. This expression can effectively indicate whether the region is actually expressed and the true expression abundance, and can approximate the true expression situation. The readcount is also converted into FPKM (Fragments Per Kilobase of exon model per Million mapped fragments) and TPM (Transcripts Per Kilobase ofexon model per Million mapped reads). In layman's terms, the number of fragments of a gene being compared is divided by the length of its gene, and the ratio is then divided by the total length of all genes. The calculation formula is as follows:

[0030] ;

[0031] Among them, total exon fragments represents the number of gene fragments, mapped reads (million reads) represents the gene length, and exon length (KB) represents the total length of all genes.

[0032] ;

[0033] Among them, Ni represents the number of reads of the ith gene obtained by comparison, Li represents the length of the ith gene; sum(N1 / L1+N2 / L2+N3 / L3+...+Nn / Ln) represents the sum of the values ​​of all genes after being standardized by length, and n is the number of genes.

[0034] After analyzing the gene expression, the genes with significantly different expression levels in different states were screened. First, the original readcount was standardized, mainly to correct the depth of the sequenced genes. Then the statistical model calculated the hypothesis test probability (Pvalue), and finally multiple hypothesis test correction was performed to obtain the FDR value (false discovery rate, padj is its common form). Generally speaking, if the expression difference of a gene in the chemotherapy-sensitive group and the chemotherapy-resistant group is more than twice, such a gene is considered to have expression differences. In order to determine whether the expression difference between the two samples is caused by various errors or essential differences, it is necessary to perform hypothesis tests on the expression data of all genes in the two samples. Transcriptome analysis is performed on thousands of genes, which will lead to the accumulation of false positives. The more genes there are, the higher the degree of false positive accumulation of the hypothesis test will be. Therefore, padj is introduced to correct the P-value of the hypothesis test to control the proportion of false positives. The DESeq2 software was used to perform significant analysis of gene expression differences, and the threshold standard used was FoldChange greater than or equal to 2 and P-value less than or equal to 0.05 to determine differential genes.

[0035] See also Figure 2 The horizontal axis is the fold change of gene expression between the chemotherapy-sensitive group and the chemotherapy-resistant group (lg2(FoldChange)), and the vertical axis is the significance level of gene expression difference between the chemotherapy-sensitive group and the chemotherapy-resistant group (-log 10 P value), red dots indicate up-regulated genes, blue dots indicate normal genes, and green dots indicate down-regulated genes.

[0036] The method of selecting differentially expressed genes in step 1 may be selected by other existing technologies, and this scheme does not limit it.

[0037] Step 2: construct a training data set based on the differentially expressed genes, calculate the connection weight between the differentially expressed gene feature data of any two patients, construct the loss function of the sensitivity prediction model, and train the sensitivity prediction model.

[0038] Assume that there are N patients in the jth patient screened in step 1 j The differential gene feature data of the patient is expressed as ,in Represents the i-th differential gene feature data, i=1,2,...,N j , such as N j =20.

[0039] N j The differentially expressed genes are randomly shuffled, and K differentially expressed gene feature data are selected as training data sets each time to train the sensitivity prediction model. When training the sensitivity prediction model, the K differentially expressed gene feature data are used as the input of the sensitivity prediction model, and the sample label of patient j is used as the output of the deep learning model. In an practicable manner, the sample label of the patient is the label of whether the patient belongs to the chemotherapy sensitive group or the chemotherapy resistant group. j The differential gene feature data are sampled M times, and K differential gene feature data are selected each time. However, the K value of each sampling can be changed, as long as K≤N j As an example, the differential gene feature data of M samplings can be expressed as:

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] It can be seen that for patient j, N j After sampling M times the differential gene feature data, the training data set can be obtained , these training data sets correspond to a sample label, that is, patient j is in the chemotherapy sensitive group or chemotherapy resistant group. If there are P patients in total, P training data sets can be collected , j=1,2,...,P.

[0046] Randomly select two patients' training data sets from P patients for feature learning, first construct the feature matrix :

[0047]

[0048] in, represents the differential gene feature data of the m-th sampling of patient j1 and the differential gene feature data of the m-th sampling of patient j2, j1=1,2,...,P, j2=1,2,...,P, m=1,2,...,M.

[0049] The feature matrix Each element in is regarded as a set of features with a connection relationship, and the connection weight of each set of features is calculated based on the flow learning algorithm:

[0050]

[0051] in, represents the connection weight between the differential gene feature data of the first sampling of patient j1 and the differential gene feature data of the mth sampling of patient j2; represents the two-norm; Represents the Gaussian distribution variance of the first sampling of differential gene feature data of patient j1. By connecting weights, differential gene feature data can learn information from each other, which is of great help to the sensitivity prediction model to predict the sensitivity of patients to neoadjuvant chemotherapy based on differential genes.

[0052] Based on the connection weights of each group of features , construct the loss function LOSS:

[0053]

[0054] Among them, Y represents the true label of patient j1 (Y=1 means that patient j1 is in the chemotherapy-sensitive group, and Y=0 means that patient j1 is in the chemotherapy-resistant group); represents the prediction result of patient j1 ( The closer it is to 1, the patient j1 is in the chemotherapy sensitive group. The closer it is to 0, the patient j1 belongs to the chemotherapy resistance group); Represents the connection weight between the differential gene feature data of the m1th sampling of patient j1 and the differential gene feature data of the m2th sampling of patient j2; m1=1,2,...,M, m2=1,2,...,M, M is the number of samplings; j1=1,2,...,P, j2=1,2,...,P and j2≠j1, P is the total number of patients.

[0055] When calculating the connection weights, only the relationship between the differential gene feature data is considered, without contacting the non-differential gene feature data, which can reduce the difficulty of model training. When the patient's non-differential gene is input, it means that the non-differential gene will not affect the sensitivity of neoadjuvant chemotherapy. Among the differential genes, it is necessary to consider the impact of the combination of multiple differential genes on the sensitivity of neoadjuvant chemotherapy, and it cannot be judged only from the impact of a single differential gene. Therefore, this scheme can achieve the judgment effect of multiple genes.

[0056] In order to verify the feasibility of the sensitivity prediction model, paraffin sections of patient tumor tissue were used for immunohistochemical staining, and quantitative analysis was performed using ImageJ software. The sections were first immersed in xylene, and then dehydrated from low to high concentrations of ethanol; antigen repair solution was added, the sections were placed in a water-soluble pot for heating, and the temperature was continuously measured with a thermometer. When the temperature of the antigen repair solution reached the effective degree (92°C), it was maintained for 40 minutes, then the staining box was taken out and cooled to room temperature; 3% hydrogen peroxide solution was added and allowed to stand at room temperature for 10 minutes; the sections were washed with PBS and soaked twice, 5 minutes / time; the blocking solution was added to the tissue sections, and then the sections were placed in a wet box and incubated at room temperature for 30 minutes; after the blocking was completed, the sections were washed with PBS; the primary antibody was diluted with PBS, and the diluted The primary antibody was evenly coated on the blocked tissue sections, placed in a humidified box, and incubated at room temperature for 1 hour; the secondary antibody was diluted with blocking solution in an appropriate proportion, and the diluted secondary antibody was added to the tissue sections that had been incubated with the primary antibody and washed, and incubated at room temperature for 30-60 minutes. After the incubation was completed, the sections were washed by slowly shaking on a shaker with washing solution; DAB was added to observe the color development degree and color development time of the sections; 80-100 μL of hematoxylin staining solution was added until the tissue was completely covered, incubated at room temperature for 5 minutes, differentiated with 1% hydrochloric acid alcohol for 1 second, washed with tap water, and the tap water turned blue for 5 minutes; finally, the sections were sealed with sealing solution and observed under a microscope.

[0057] Take 3-5 pictures of each slice under the microscope and process them through ImageJ software. According to the needs, the average optical density value, number of positive cells, staining intensity score, etc. can be selected to quantitatively analyze the staining results. According to the results of immunohistochemical quantitative analysis, the expression differences of differential genes at the protein level are summarized to further support and verify the biological significance of the sensitivity prediction model.

[0058] Step 3, after the sensitivity prediction model is trained, the patient's genetic data is input into the sensitivity prediction model to obtain a prediction result of whether the patient belongs to the chemotherapy-sensitive group or the chemotherapy-resistant group.

[0059] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model, characterized by: The following steps are involved: Step 1: Screen the tumor information of historical nasopharyngeal carcinoma patients before and after neoadjuvant chemotherapy to screen out differentially expressed genes; Step 2: construct a training data set based on differentially expressed genes, calculate the connection weight between differentially expressed gene feature data of any two patients, construct a loss function of the sensitivity prediction model, and train the sensitivity prediction model; The step 2 specifically includes the following steps: Step 2-1, construct differential gene signature data for each patient, thereby forming a training data set for all patients; The step 2-1 is specifically as follows: Assume that there are N patients among the jth patients screened out j The differential gene feature data of the patient is expressed as ,in Represents the i-th differential gene feature data, i=1,2,...,N j ; N j The N differentially expressed genes of patient j were randomly shuffled. j The differential gene feature data are sampled M times, and K differential gene feature data are selected each time, where K is a variable value, K≤N j ; then for patient j, N j After sampling M times the differential gene feature data, the training data set can be obtained , training data set There is a corresponding sample label, that is, patient j is in the chemotherapy-sensitive group or chemotherapy-resistant group; If there are P patients in total, then P training data sets can be collected , j=1,2,...,P; Step 2-2, construct feature matrix; In step 2-2, the constructed feature matrix is: in, represents the differential gene feature data of the m-th sampling of patient j1 and the differential gene feature data of the m-th sampling of patient j2, j1=1,2,...,P, j2=1,2,...,P, m=1,2,...,M, M is the number of sampling times; Step 2-3, calculate the connection weight of each element in the feature matrix based on the flow learning algorithm; Step 2-4, construct the loss function of the sensitivity prediction model based on the connection weight; Step 3, after the sensitivity prediction model is trained, the patient's genetic data is input into the sensitivity prediction model to obtain a prediction result of whether the patient belongs to the chemotherapy-sensitive group or the chemotherapy-resistant group.

2. The method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model according to claim 1, characterized in that: The steps 2-3 are as follows: The feature matrix Each element in is regarded as a set of features with a connection relationship, and the connection weight of each set of features is calculated based on the flow learning algorithm: in, represents the connection weight between the differential gene feature data of the first sampling of patient j1 and the differential gene feature data of the mth sampling of patient j2; represents the two-norm; Represents the Gaussian distribution variance of the differential gene feature data of patient j1 when it is sampled for the first time; by connecting the weights, the differential gene feature data can learn the information between each other, which is of great help for the sensitivity prediction model to predict the sensitivity of patients to neoadjuvant chemotherapy based on differential genes.

3. The method for predicting the sensitivity of nasopharyngeal carcinoma to neoadjuvant chemotherapy based on a neural network model according to claim 2, characterized in that: In the steps 2-4, the loss function LOSS constructed is: Among them, Y represents the true label of patient j1 (Y=1 means that patient j1 is in the chemotherapy-sensitive group, and Y=0 means that patient j1 is in the chemotherapy-resistant group); represents the prediction result of patient j1 ( The closer it is to 1, the patient j1 is in the chemotherapy sensitive group. The closer it is to 0, the patient j1 belongs to the chemotherapy resistance group); Represents the connection weight between the differential gene feature data of the m1th sampling of patient j1 and the differential gene feature data of the m2th sampling of patient j2; m1=1,2,...,M, m2=1,2,...,M; j1=1,2,...,P, j2=1,2,...,P and j2≠j1, P is the total number of patients.

Images