Esophageal cancer immune data prediction device based on tumor markers, equipment and storage medium

Abstract

The application discloses an esophageal cancer immune data prediction device based on tumor markers, equipment and storage medium, relates to the technical field of digital data processing, and solves the problem of poor prediction effect of traditional biomarkers (such as PD-L1 expression and tumor mutation load) by constructing a more comprehensive tumor microenvironment evaluation system through 5hmC epigenome sequencing of ctDNA and analysis of tumor tissue succinyl modification level. Secondly, through the double-channel LSTM-Cox hybrid model, dynamic ctDNA data is processed through a time sequence channel, and succinylated scores are processed through a static channel, so that the advantages of the two kinds of data are complementary. Finally, by quantifying complex immune data into three intuitive good / average / poor score intervals, the application provides clear treatment decision basis for clinicians, and the invalid treatment rate can be obviously reduced by using the scoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric digital data processing, and in particular to an esophageal cancer immune data prediction device based on tumor markers, equipment and a storage medium. BACKGROUND

[0002] Esophageal cancer (EC) is a malignant disease that seriously threatens human life and health, mainly a malignant tumor of esophageal squamous epithelium or glandular epithelium. At present, the prognosis of patients with locally advanced EC is poor, and the 5-year survival rate of patients with esophageal squamous cell carcinoma treated by simple surgical resection is only 20.64%-34.00% for stage II A-III. Even through multidisciplinary treatment, the 5-year survival time of esophageal cancer is only about 30%, and tumor metastasis and recurrence are the main two reasons.

[0003] In summary, the current efficiency of immunotherapy for esophageal cancer is low, the research progress is not much, and the relevant mechanism is limited. On the one hand, the etiology and pathology of esophageal cancer in China are different from those abroad. Esophageal cancer itself is a disease with large heterogeneity. The efficiency of immunotherapy is low. Among them, esophageal squamous cell carcinoma (ESCC) accounts for 90% of global esophageal cancer cases. Due to its highly immunosuppressive tumor microenvironment (TME), the response rate to simple immunotherapy is less than 20%, and about 30% to 50% of patients face drug resistance problems. Traditional biomarkers (such as PD-L1 expression and tumor mutation burden) perform poorly in predicting the efficacy of combined chemotherapy and immunotherapy. Although chemotherapy combined with PD-1 antibody can significantly improve the efficacy of ESCC patients, the mechanism is still not fully understood, resulting in low interpretability and large errors in subsequent efficacy prediction, which cannot be used as effective data for subsequent treatment decisions. SUMMARY

[0004] The main purpose of the present application is to provide an esophageal cancer immune data prediction device based on tumor markers, equipment and a storage medium, to solve the problem that the traditional biomarkers (such as PD-L1 expression and tumor mutation burden) in the prior art perform poorly in predicting the efficacy of combined chemotherapy and immunotherapy. Although chemotherapy combined with PD-1 antibody can significantly improve the efficacy of ESCC patients, the mechanism is still not fully understood, resulting in low interpretability and large errors in subsequent efficacy prediction, which cannot be used as effective data for subsequent treatment decisions.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0006] An esophageal cancer immune data prediction device based on tumor markers, the esophageal cancer immune data prediction device comprises:

[0007] The ctDNA sequencing data acquisition module is configured to acquire ctDNA based on a peripheral blood sample of the esophageal cancer patient, and perform 5hmC epigenome sequencing on the ctDNA to obtain ctDNA sequencing data.

[0008] The succinylation score acquisition module is configured to quantitatively analyze the level of succinylation modification based on a tumor tissue sample of the esophageal cancer patient by immunohistochemical staining and a scoring system to obtain a succinylation score.

[0009] The spatiotemporal alignment module is configured to perform spatiotemporal alignment of the ctDNA sequencing data and the succinylation score by a dynamic time warping algorithm.

[0010] The comprehensive grade score acquisition module is configured to construct a pre-trained dual-channel LSTM-Cox hybrid model, input the aligned ctDNA sequencing data into a time sequence channel of the dual-channel LSTM-Cox hybrid model, input the aligned succinylation score into a static channel of the dual-channel LSTM-Cox hybrid model, and output a comprehensive grade score of esophageal cancer immune data by the dual-channel LSTM-Cox hybrid model.

[0011] The score interval definition module is configured to define continuous low-score intervals, medium-score intervals, and high-score intervals based on a preset score strategy, and each score interval corresponds to good immune effect, general immune effect, and poor immune effect in sequence.

[0012] The immune effect matching and sending module is configured to classify the comprehensive grade score into a corresponding score interval, and send the matched immune effect to an external receiving end.

[0013] As a further improvement of the present application, the ctDNA sequencing data acquisition module comprises:

[0014] The ctDNA fragmentation processing unit is configured to perform fragmentation processing on the ctDNA to obtain ctDNA fragments of a preset length.

[0015] The raw sequencing data acquisition unit is configured to perform high-throughput sequencing on the ctDNA fragments by a preset sequencing platform to obtain raw sequencing data.

[0016] The raw sequencing data processing unit is configured to remove low-quality reads and adapter contamination of the raw sequencing data by FastQC to obtain processed sequencing data.

[0017] The ctDNA sequencing data saving unit is configured to save the processed sequencing data in FASTQ format to obtain the ctDNA sequencing data.

[0018] As a further improvement of the present application, the succinylation score acquisition module comprises:

[0019] An immunohistochemical staining image acquisition unit is configured to acquire a high-resolution immunohistochemical staining image of the tumor tissue sample by a full-automatic digital pathology scanning system;

[0020] An immunohistochemical staining image processing unit is configured to perform denoising processing and contrast enhancement on the high-resolution immunohistochemical staining image to obtain a processed immunohistochemical staining image;

[0021] A binary mask acquisition unit is configured to separate the succinylation positive area from the background by a color threshold segmentation algorithm to obtain a binary mask;

[0022] A staining intensity grade acquisition unit is configured to calculate the average staining intensity in the succinylation positive area by a gray level histogram and grade the average staining intensity according to a standard staining intensity to obtain a staining intensity grade;

[0023] A positive proportion calculation unit is configured to identify the total number of cells of the binary mask and the number of positive cells of the succinylation positive area by a pre-trained SqueezeNet_Residual model and acquire a positive proportion of the number of positive cells based on the total number of cells;

[0024] A succinylation score calculation unit is configured to multiply the staining intensity grade and the positive proportion to obtain the succinylation score.

[0025] As a further improvement of the present application, the spatiotemporal alignment module comprises:

[0026] A time series construction unit is configured to construct the ctDNA sequencing data into a time series X in chronological order, with each time point corresponding to a multi-dimensional feature vector, and construct the succinylation score into a time series Y in chronological order, with each time point corresponding to a single-dimensional score value;

[0027] An Euclidean distance matrix calculation unit is configured to calculate an Euclidean distance matrix D of the time series X and the time series Y, where D[i,j] is the distance between the i-th time point of the time series X and the j-th time point of the time series Y;

[0028] A cumulative distance matrix construction unit is configured to create a cumulative distance matrix C of the same size as the Euclidean distance matrix D and initialize the cumulative distance of the starting point C[0,0]=D[0,0];

[0029] a cumulative distance matrix filling unit configured to fill the cumulative distance matrix C[i,j] = D[i,j] + min{C[i-1,j], C[i,j-1], C[i-1,j-1]} according to the dynamic time warping algorithm, where i > 0 and j > 0, to obtain a filled cumulative distance matrix;

[0030] an optimal path backtracking unit configured to backtrack from the lower right corner of the filled cumulative distance matrix to the upper left corner of the filled cumulative distance matrix with the minimum cumulative distance to obtain an optimal path;

[0031] a data linear interpolation unit configured to perform linear interpolation on data in the optimal path that is not moved along a diagonal line to align the ctDNA sequencing data with the succinylated score in space-time.

[0032] As a further improvement of the present application, the comprehensive grade score obtaining module comprises:

[0033] a dual-channel LSTM-Cox hybrid model constructing unit configured to construct a dual-channel LSTM-Cox hybrid model comprising a time sequence channel and a static channel, wherein the time sequence channel adopts an LSTM network to process time sequence data, the static channel adopts a fully connected network to process static features, and outputs of the two channels are input into a Cox proportional hazards model after being fused by a concatenation layer;

[0034] a dual-channel data training unit configured to input the aligned ctDNA sequencing data into the LSTM network for training according to time steps, and input the aligned succinylated score into the fully connected network for training;

[0035] a comprehensive grade score output unit configured to concatenate output features of the trained time sequence channel and static channel, input the output features into the Cox proportional hazards model, and jointly train the output features through a negative log-likelihood loss function, and output the comprehensive grade score through the Cox proportional hazards model.

[0036] As a further improvement of the present application, the score interval defining module comprises:

[0037] a comprehensive grade score amplification unit configured to repeatedly execute the ctDNA sequencing data obtaining module to the comprehensive grade score obtaining module based on a plurality of peripheral blood samples of esophageal cancer patients to obtain a plurality of comprehensive grade scores;

[0038] a comprehensive grade score clustering unit configured to cluster all comprehensive grade scores into three clusters through a K-means clustering algorithm;

[0039] a score interval midpoint defining unit configured to sequentially take center points of the three clusters as midpoints of the low score interval, the medium score interval and the high score interval.

[0040] The low-score interval obtaining unit obtains the low-score interval by taking a zero value as a left boundary of the low-score interval and obtaining a right boundary of the low-score interval by symmetry of an interval midpoint of the low-score interval.

[0041] The medium-score interval obtaining unit obtains the medium-score interval by taking the right boundary of the low-score interval as a left boundary of the medium-score interval and obtaining a right boundary of the medium-score interval by symmetry of an interval midpoint of the medium-score interval.

[0042] The high-score interval obtaining unit obtains the high-score interval by taking the right boundary of the medium-score interval as a left boundary of the high-score interval and obtaining a right boundary of the high-score interval by symmetry of an interval midpoint of the high-score interval.

[0043] The immune effect giving unit gives the low-score interval, the medium-score interval, and the high-score interval labels of good immune effect, general immune effect, and poor immune effect in sequence.

[0044] As a further improvement of the present application, after the immune effect giving unit, the esophageal cancer immune data prediction device further comprises:

[0045] The comprehensive grade score additional amplification unit repeatedly executes the ctDNA sequencing data obtaining module to the comprehensive grade score obtaining module based on additional peripheral blood samples to obtain additional comprehensive grade scores.

[0046] The score interval iterative updating unit repeatedly executes the comprehensive grade score clustering unit to the high-score interval obtaining unit with the collection period of the additional peripheral blood samples as a step to iteratively update the low-score interval, the medium-score interval, and the high-score interval.

[0047] To achieve the above-mentioned purpose, the present application further provides the following technical solutions:

[0048] An electronic device comprises a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; the processor executes the program instructions stored in the memory to realize the operation of the above-mentioned esophageal cancer immune data prediction device.

[0049] To achieve the above-mentioned purpose, the present application further provides the following technical solutions:

[0050] A storage medium, the storage medium storing program instructions, the program instructions being executed by a processor to realize the operation of the above-mentioned esophageal cancer immune data prediction device.

[0051] Beneficial effects:

[0052] The application obtains ctDNA from a peripheral blood sample of an esophageal cancer patient, and performs 5hmC epigenome sequencing on the ctDNA to obtain ctDNA sequencing data; based on a tumor tissue sample of the esophageal cancer patient, the level of succinyl modification is quantitatively analyzed by immunohistochemical staining and a scoring system to obtain a succinyl score; the ctDNA sequencing data and the succinyl score are spatiotemporally aligned by a dynamic time warping algorithm; a pre-trained dual-channel LSTM-Cox hybrid model is constructed, and the aligned ctDNA sequencing data is input into a time sequence channel of the dual-channel LSTM-Cox hybrid model, and the aligned succinyl score is input into a static channel of the dual-channel LSTM-Cox hybrid model, and a comprehensive grade score of esophageal cancer immune data is output by the dual-channel LSTM-Cox hybrid model; based on a preset scoring strategy, a continuous low score interval, a medium score interval and a high score interval are defined, and each score interval corresponds to good immune effect, general immune effect and poor immune effect in turn; the comprehensive grade score is classified into the corresponding score interval, and the matched immune effect is sent to an external receiving end. First, by 5hmC epigenome sequencing of ctDNA and analysis of the level of succinyl modification of tumor tissue, a more comprehensive tumor microenvironment evaluation system is constructed, solving the problem of poor prediction effect of traditional biomarkers (such as PD-L1 expression and tumor mutation load); secondly, the dual-channel LSTM-Cox hybrid model processes dynamic ctDNA data through a time sequence channel and processes succinyl score through a static channel, realizing the complementary advantages of the two kinds of data, so that the AUC value of immune therapy effect prediction can reach more than 0.90, which is obviously better than the traditional single modality prediction method (AUC≈0.75); finally, by quantifying complex immune data into three intuitive good / general / poor score intervals, the application provides clear treatment decision basis for clinicians, and the scoring system can obviously reduce the rate of ineffective treatment. The technical scheme of the application not only improves the prediction accuracy of the precision of esophageal cancer immunotherapy, but also provides a new research perspective for understanding the mechanism of immunotherapy resistance, which has clinical application value and transformation potential. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The figure is a step flowchart of an embodiment of the esophageal cancer immune data prediction method based on tumor markers of the application.

[0054] Figure 2 The figure is a functional module diagram of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the application.

[0055] Figure 3A functional module schematic diagram of a ctDNA sequencing data acquisition module 1 of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the present application;

[0056] Figure 4 A functional module schematic diagram of an amber score acquisition module 2 of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the present application;

[0057] Figure 5 A functional module schematic diagram of a spatiotemporal alignment module 3 of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the present application;

[0058] Figure 6 A functional module schematic diagram of a comprehensive grade score acquisition module 4 of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the present application;

[0059] Figure 7 A functional module schematic diagram of a score interval definition module 5 of an embodiment of the esophageal cancer immune data prediction device based on tumor markers of the present application;

[0060] Figure 8 A structural schematic diagram of an embodiment of the electronic device of the present application;

[0061] Figure 9 A structural schematic diagram of an embodiment of the storage medium of the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0063] The terms "first", "second", "third" in the present application are only for descriptive purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly and specifically limited. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between the components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0064] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification is not necessarily all referring to the same instance of the embodiment, nor is it necessarily referring to a separate or alternative embodiment. It is explicitly and implicitly understood that the embodiments described herein can be combined with other embodiments.

[0065] As shown in Figure 1 , the present embodiment provides an embodiment of an esophageal cancer immune data prediction method based on tumor markers, which is equivalent to various modules, units in the esophageal cancer immune data prediction device. In a popular understanding, the present embodiment does not repeat the names of various modules and units.

[0066] The esophageal cancer immune data prediction method comprises the following steps:

[0067] Step S1, obtaining ctDNA based on peripheral blood samples of esophageal cancer patients, and performing 5hmC epigenome sequencing on the ctDNA to obtain ctDNA sequencing data.

[0068] Preferably, step S1 involves sample processing of peripheral blood samples, specifically as follows:

[0069] (1) Peripheral blood sample collection and processing:

[0070] ① Sample collection: 7.5-10 ml of peripheral blood of esophageal cancer patients is collected by venipuncture and stored in an EDTA anticoagulation tube.

[0071] ② Plasma separation: centrifuge at 3000 rpm for 10 minutes at 4℃ to separate plasma, transfer to a new EP tube.

[0072] ③ Secondary centrifugation: centrifuge at 16,000 rpm for 10 minutes to remove cell debris, and the supernatant is the source of ctDNA.

[0073] (2) Extract ctDNA by magnetic bead method:

[0074] ① Add 2 times the volume of Binding Buffer to the plasma, which contains high concentration of NaCl.

[0075] ② Add magnetic bead suspension, incubate at room temperature for 10 minutes to allow DNA binding.

[0076] ③ Use a magnetic stand to separate the magnetic beads, discard the supernatant.

[0077] ④ Wash the magnetic beads twice with Wash Buffer.

[0078] ⑤ Elute the DNA with Elution Buffer (10 mM Tris-HCl, pH 8.5).

[0079] ⑥ Quantify by Qubit fluorometer, ctDNA concentration ≥ 0.1 ng / μl.

[0080] (3) Perform 5hmC (methylated cytosine) epigenome sequencing using TET-assisted bisulfite sequencing (TAB-seq) method:

[0081] ① TET oxidation: react ctDNA with TET1 protein to oxidize 5hmC to 5fC (5-formylcytosine).

[0082] ② Bisulfite treatment: use bisulfite to convert unmodified C (cytosine) and 5fC to U (uracil).

[0083] ③ PCR amplification: amplify the treated DNA using special primers.

[0084] ④ Library construction: construct sequencing library using Nextera XT DNA Library Prep Kit.

[0085] ⑤ Sequencing: perform 150 bp paired-end sequencing on the Illumina NovaSeq platform to obtain 5hmC modification map data based on BED format ctDNA.

[0086] Step S2, based on the tumor tissue sample of the esophageal cancer patient, the succinyl modification level is quantitatively analyzed by immunohistochemical staining and scoring system, and the succinyl score is obtained.

[0087] Preferably, the succinylation modification is also biased towards sample processing, as follows:

[0088] (1) Tumor tissue sample processing:

[0089] ①‌Sample acquisition‌: Esophageal cancer tumor tissue samples are obtained by surgical resection or endoscopic biopsy, and the sample size can be set to 1.5 x 1.5 x 0.5 cm³.

[0090] ②‌Fixation processing‌: Immediately place the tissue in a 10% neutral formalin solution for fixation, with a fixation time of 18-24 hours, and the volume of the fixation solution should be 10-20 times the volume of the tissue.

[0091] ③‌Dehydration embedding‌:

[0092] Gradient alcohol dehydration: 75% alcohol for 1 hour → 85% alcohol for 1 hour → 95% alcohol for 1 hour → 100% alcohol for 2 hours each time.

[0093] Xylene transparency: 2 times for 10 minutes each.

[0094] Paraffin embedding: 56-58°C wax immersion, cool to solidification.

[0095] ‌Section preparation‌: Use a microtome to cut 4-6 micron thin sections and attach them to a polylysine-treated glass slide.

[0096] (2) Immunohistochemical staining detection of succinylation modification:

[0097] ①‌Dewaxing to water‌:

[0098] Xylene I, II for 10 minutes each.

[0099] Gradient alcohol: 100% → 95% → 80% → 70% for 2 minutes each.

[0100] Distilled water wash for 5 minutes x 2 times.

[0101] ②‌Antigen retrieval‌:

[0102] Heat repair: Use sodium citrate buffer (pH 6.0) or EDTA buffer (pH 9.0).

[0103] Method selection: High-pressure steam repair (120°C, 20 minutes) or microwave repair (37°C, 10-30 minutes).

[0104] ③‌Blocking non-specific binding‌:

[0105] Use 5% goat serum or BSA blocking solution to cover the sections for 15-30 minutes.

[0106] PBS wash 3 times.

[0107] IV. Primary antibody incubation:

[0108] Anti-succinylated lysine antibody (such as Anti-Succinyl-Lysine antibody) is used.

[0109] Dilution ratio is recommended by the instruction (usually 1:100-1:500).

[0110] Incubation condition: 4℃ overnight or 37℃ for 1-2 hours.

[0111] V. Secondary antibody incubation:

[0112] HRP-labeled secondary antibody (such as anti-rabbit IgG-HRP) is used.

[0113] Incubate for 30-60 minutes.

[0114] PBS wash 3 times.

[0115] VI. Color development reaction:

[0116] DAB color development system is used.

[0117] Color development time is controlled at 5-10 minutes, and monitored under a microscope.

[0118] Tap water is used to terminate the reaction.

[0119] Preferably, the scoring system adopts the existing succinylation scoring system, and the staining intensity score can be evaluated under a 200-fold microscope to assess the staining intensity, see Table 1 below:

[0120] Table 1: Succinylation Scoring System Details:

[0121]

[0122] Preferably, the staining range score is:

[0123] 0 points: <5% of cells are stained.

[0124] 1 point: 5%-25% of cells are stained.

[0125] 2 points: 26%-50% of cells are stained.

[0126] 3 points: >50% of cells are stained.

[0127] Therefore, the comprehensive score can be set as the staining intensity score x the staining range score (total score 0-9).

[0128] Step S3: Temporally and spatially align the ctDNA sequencing data with the succinylation score by the dynamic time warping algorithm.

[0129] Preferably, step S3 can be implemented by the following key code:

[0130] import numpy as np

[0131] from fastdtw import fastdtw

[0132] from scipy.spatial.distance import euclidean

[0133] # Example data

[0134] ctDNA_sequence = np.array([0.1, 0.3, 0.5, 0.7, 0.9]) # Simulated ctDNA data

[0135] succ_score = np.array([0.2, 0.4, 0.6, 0.8]) # Simulated succinyl score

[0136] # DTW alignment

[0137] distance, path = fastdtw(ctDNA_sequence, succ_score, dist=euclidean)

[0138] # Visualize alignment results

[0139] import matplotlib.pyplot as plt

[0140] plt.plot(ctDNA_sequence, label='ctDNA')

[0141] plt.plot(succ_score, label='Succinyl Score')

[0142] plt.legend()

[0143] plt.title('DTW Alignment')

[0144] plt.show()

[0145] Step S4, a pre-trained two-channel LSTM-Cox hybrid model is constructed, and the aligned ctDNA sequencing data is input into the time sequence channel of the two-channel LSTM-Cox hybrid model, and the aligned succinylated score is input into the static channel of the two-channel LSTM-Cox hybrid model, and the comprehensive grade score of the esophageal cancer immune data is output by the two-channel LSTM-Cox hybrid model.

[0146] Preferably, the data preparation stage can use historical esophageal cancer patient data (including ctDNA time sequence data and succinylated score), and the data has been spatiotemporally aligned by the DTW algorithm, and at the same time, the training set (70%), the validation set (15%) and the test set (15%) are divided.

[0147] Preferably, the key code of step S4 is as follows:

[0148] from tensorflow.keras.layers import LSTM, Dense, Input

[0149] from tensorflow.keras.models import Model

[0150] # Time sequence channel input

[0151] timeline_input = Input(shape=(None, ctDNA_feature_dim), name='timeline_input')

[0152] # LSTM layer

[0153] lstm_out = LSTM(64, return_sequences=True)(timeline_input)

[0154] lstm_out = LSTM(32)(lstm_out)

[0155] # Time sequence channel output

[0156] timeline_output = Dense(16, activation='relu')(lstm_out)

[0157] # Static channel input

[0158] static_input = Input(shape=(succ_score_dim,), name='static_input')

[0159] # Fully connected layer

[0160] static_output = Dense(16, activation='relu')(static_input)

[0161] # Merge outputs from both channels

[0162] merged = Concatenate()([timeline_output, static_output])

[0163] # Cox proportional hazards model

[0164] cox_output = Dense(1, activation='linear', name='cox_output')(merged)

[0165] # Define the model

[0166] model = Model(inputs=[timeline_input, static_input], outputs=cox_output)

[0167] from tensorflow.keras.losses import MeanSquaredError

[0168] # Cox proportional hazards loss function

[0169] def cox_loss(y_true, y_pred):

[0170] # Implement Cox proportional hazards loss

[0171] pass

[0172] model.compile(optimizer='adam', loss=cox_loss)

[0173] model.fit([X_train_timeline, X_train_static], y_train, validation_data=([X_val_timeline, X_val_static], y_val))

[0174] Step S5, define continuous low-score intervals, medium-score intervals and high-score intervals based on the preset scoring strategy, and each scoring interval corresponds to good immune effect, general immune effect and poor immune effect in turn.

[0175] Step S6, classify the comprehensive grade score to the corresponding scoring interval, and send the matched immune effect to the external receiving end.

[0176] Further, step S1, based on the peripheral blood sample of the esophageal cancer patient, ctDNA is obtained, and 5hmC epigenome sequencing is performed on the ctDNA to obtain ctDNA sequencing data, which includes the following steps:

[0177] Step S11, fragmenting the ctDNA to obtain ctDNA fragments of a preset length.

[0178] Preferably, through the above processing steps, the ctDNA can be fragmented to 150-300 bp.

[0179] Step S12, high-throughput sequencing of the ctDNA fragments by a preset sequencing platform to obtain raw sequencing data.

[0180] Preferably, the preset sequencing platform can be the Illumina NovaSeq platform.

[0181] Step S13, removing low-quality reads and adapter contamination of the raw sequencing data by FastQC to obtain processed sequencing data.

[0182] Preferably, FastQC is installed by Bash statement of Linux system: conda install -c bioconda fastqc, and the raw sequencing data files are defined as ctDNA_raw_R1.fastq.gz and ctDNA_raw_R2.fastq.gz. Then, FastQC is run for quality evaluation by: fastqc -o. / fastqc_report -t 8 ctDNA_raw_R1.fastq.gz ctDNA_raw_R2.fastq.gz statement, and low-quality reads and adapter contamination can be controlled by the following bash statement:

[0183] java -jar trimmomatic-0.39.jar PE \

[0184] -threads 8 \

[0185] -phred33 \

[0186] ctDNA_raw_R1.fastq.gz ctDNA_raw_R2.fastq.gz \

[0187] ctDNA_clean_R1_paired.fq.gz ctDNA_clean_R1_unpaired.fq.gz \

[0188] ctDNA_clean_R2_paired.fq.gz ctDNA_clean_R2_unpaired.fq.gz \

[0189] ILLUMINACLIP:TruSeq3-PE.fa:2:30:10 \

[0190] LEADING:20 \

[0191] TRAILING:20 \

[0192] SLIDINGWINDOW:4:20 \

[0193] MINLEN:36

[0194] Step S14, save the processed sequencing data in FASTQ format to obtain ctDNA sequencing data.

[0195] Further, in step S2, based on the tumor tissue sample of the esophageal cancer patient, the succinyl modification level is quantitatively analyzed by immunohistochemical staining and a scoring system to obtain a succinyl score, which specifically includes the following steps:

[0196] Step S21, obtain a high-resolution immunohistochemical staining image of the tumor tissue sample by a full-automatic digital pathology scanning system.

[0197] Preferably, the full-automatic digital pathology scanning system is a professional equipment that converts traditional glass pathology sections into high-resolution digital images, and is widely used in the fields of pathological diagnosis, teaching training, drug research and scientific research. Its core functions include full-section rapid imaging, accurate color restoration and multi-dimensional analysis support.

[0198] It is worth noting that the full-automatic digital pathology scanning system is a mature existing technology and existing equipment, and the specific principles will not be described here.

[0199] Step S22, denoising and contrast enhancement are performed on the high-resolution immunohistochemical staining image to obtain a processed immunohistochemical staining image.

[0200] Preferably, the denoising and contrast enhancement of the high-resolution immunohistochemistry staining image are also inherent functions of the fully automated digital pathology scanning system.

[0201] Step S23, separating the succinylation positive area from the background by color threshold segmentation algorithm to obtain a binary mask.

[0202] Preferably, the high-resolution immunohistochemistry staining image is usually a microscope image in RGB format, and the separation of the succinylation positive area from the background can be achieved by the following steps:

[0203] (1) Convert the image from RGB color space to HSV color space.

[0204] (2) Determine the HSV range of DAB coloration, where a typical range of DAB coloration in HSV space is: Hue (H): 10-30 (corresponding to brown / brownish yellow), Saturation (S): 50-255, Lightness (V): 50-255.

[0205] (3) Create a mask based on the inRange function to extract the DAB coloration area.

[0206] (4) Remove small noise: use the open operation (first erosion then dilation) to remove small noise points and false positive areas.

[0207] (5) Fill the holes: use the close operation (first dilation then erosion) to fill small holes and gaps in the positive area.

[0208] (6) Region connectivity analysis: further optimize the segmentation result and retain the main positive area.

[0209] (7) Generate the final binary mask: output the processed binary image, where white (255) represents the DAB positive area and black (0) represents the background.

[0210] Preferably, the key code is as follows:

[0211] import cv2

[0212] import numpy as np

[0213] # Read the image

[0214] image = cv2.imread('ihc_image.jpg')

[0215] # Convert to HSV color space

[0216] hsv_image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)

[0217] # Define HSV range for DAB development

[0218] lower_dab = np.array([10, 50, 50])

[0219] upper_dab = np.array([30, 255, 255])

[0220] # Create color mask

[0221] dab_mask = cv2.inRange(hsv_image, lower_dab, upper_dab)

[0222] # Define morphological operation kernel

[0223] kernel = np.ones((5,5), np.uint8)

[0224] # Open operation to remove noise

[0225] opened_mask = cv2.morphologyEx(dab_mask, cv2.MORPH_OPEN, kernel,iterations=1)

[0226] # Close operation to fill holes

[0227] closed_mask = cv2.morphologyEx(opened_mask, cv2.MORPH_CLOSE, kernel,iterations=2)

[0228] # Connected component analysis

[0229] num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(closed_mask)

[0230] # Filter regions by area (example: keep regions with area greater than 100 pixels)

[0231] final_mask = np.zeros_like(closed_mask)

[0232] for i in range(1, num_labels):

[0233] if stats[i, cv2.CC_STAT_AREA] > 100:

[0234] final_mask[labels == i] = 255

[0235] # Save results

[0236] cv2.imwrite('dab_segmentation_mask.png', final_mask)

[0237] Step S24: Calculate the average staining intensity in the succinylated positive area using a grayscale histogram, and grade the average staining intensity according to the standard staining intensity to obtain the staining intensity level.

[0238] Preferably, the intensity is graded using the 0 (unstained)-1 (weak)-2 (medium)-3 (strong) standard mentioned above.

[0239] Step S25: The total number of cells in the binarized mask and the number of positive cells in the succinylated positive region are identified by the pre-trained SqueezeNet_Residual model, and the positive ratio of the number of positive cells based on the total number of cells is obtained.

[0240] Preferably, the key code for the SqueezeNet_Residual model loading stage in step S25 is as follows:

[0241] import torch

[0242] import torch.nn as nn

[0243] # Define the SqueezeNet_Residual model structure

[0244] class SqueezeNet_Residual(nn.Module):

[0245] def __init__(self, num_classes=2):

[0246] super(SqueezeNet_Residual, self).__init__()

[0247] # Model definition code...

[0248] # Load pre-trained weights

[0249] model = SqueezeNet_Residual()

[0250] model.load_state_dict(torch.load('pretrained_squeezenet_residual.pth'))

[0251] model.eval()

[0252] Preferably, the binary mask can be converted into a format acceptable to the model during the binary mask preprocessing stage, usually requiring resizing and adding channel dimensions, the key code is as follows:

[0253] import cv2

[0254] import numpy as np

[0255] def preprocess_mask(binary_mask):

[0256] # Convert binary mask to three-channel image

[0257] mask_3ch = np.stack([binary_mask]*3, axis=-1)

[0258] # Resize to model input requirements (e.g. 224x224)

[0259] resized_mask = cv2.resize(mask_3ch, (224, 224))

[0260] # Normalize and convert to tensor

[0261] normalized = resized_mask.astype(np.float32) / 255.0

[0262] tensor_mask = torch.from_numpy(normalized).permute(2, 0, 1).unsqueeze(0)

[0263] return tensor_mask

[0264] Preferably, during the model identification stage, the preprocessed binary mask can be input into the model to obtain the prediction result, the key code is as follows:

[0265] # Input pre-processed binary mask

[0266] input_tensor = preprocess_mask(binary_mask)

[0267] # Disable gradient computation

[0268] with torch.no_grad():

[0269] # Model inference

[0270] output = model(input_tensor)

[0271] # Get feature maps

[0272] feature_maps = model.get_feature_maps(input_tensor)

[0273] def parse_model_output(output, feature_maps):

[0274] # Get cell detection heatmap

[0275] detection_heatmap = feature_maps['detection']

[0276] # Get cell classification probability map (positive / negative)

[0277] classification_probs = torch.softmax(output, dim=1)

[0278] return detection_heatmap, classification_probs

[0279] def count_total_cells(detection_heatmap, threshold=0.5):

[0280] # Convert heatmap to binary image

[0281] binary_heatmap = (detection_heatmap > threshold).float()

[0282] # Count using connected component analysis

[0283] num_components, labeled_image = cv2.connectedComponents(

[0284] binary_heatmap.squeeze().cpu().numpy().astype(np.uint8))

[0285] # Subtract background (connected components analysis will label background as 0)

[0286] total_cells = num_components - 1

[0287] return total_cells, labeled_image

[0288] def identify_positive_cells(labeled_image, classification_probs, pos_threshold=0.7):

[0289] positive_cells = 0

[0290] cell_properties = []

[0291] # Iterate over each labeled cell region

[0292] for label in range(1, labeled_image.max()+1):

[0293] # Get the mask for the current cell

[0294] cell_mask = (labeled_image == label)

[0295] # Calculate the mean of the positive probability in this region

[0296] cell_prob = classification_probs[0, 1][cell_mask].mean().item()

[0297] # Determine if it's a positive cell

[0298] if cell_prob > pos_threshold:

[0299] positive_cells += 1

[0300] # Record cell properties (optional)

[0301] cell_properties.append({

[0302] 'label': label,

[0303] 'positive_prob': cell_prob,

[0304] 'is_positive': cell_prob > pos_threshold

[0305] })

[0306] return positive_cells, cell_properties

[0307] def generate_cell_report(binary_mask):

[0308] # 1. Preprocess binary mask

[0309] input_tensor = preprocess_mask(binary_mask)

[0310] # 2. Model inference

[0311] with torch.no_grad():

[0312] output = model(input_tensor)

[0313] feature_maps = model.get_feature_maps(input_tensor)

[0314] # 3. Parse output

[0315] detection_heatmap, classification_probs = parse_model_output(output, feature_maps)

[0316] # 4. Count total cells

[0317] total_cells, labeled_image = count_total_cells(detection_heatmap)

[0318] # 5. Identify positive cells

[0319] positive_cells, cell_properties = identify_positive_cells(

[0320] labeled_image, classification_probs)

[0321] # 6. Generate report

[0322] report = {

[0323] 'total_cells': total_cells,

[0324] 'positive_cells': positive_cells,

[0325] 'positive_ratio': positive_cells / total_cells if total_cells> 0 else 0,

[0326] 'cell_details': cell_properties

[0327] }

[0328] return report

[0329] Step S26, multiply the staining intensity grade with the positive cell ratio to obtain the succinylation score.

[0330] Preferably, the staining intensity grade is 0-3 points, the positive cell ratio is 0%-100%, and the succinylation score is 0-300 points.

[0331] Further, step S3, the ctDNA sequencing data is spatiotemporally aligned with the succinylation score by dynamic time warping algorithm, which specifically includes the following steps:

[0332] Step S31, the ctDNA sequencing data is constructed into a time series X according to the sampling time sequence, each time point corresponds to a multi-dimensional feature vector, and the succinylation score is constructed into a time series Y according to the detection time sequence, each time point corresponds to a single-dimensional score value.

[0333] Step S32, calculate the Euclidean distance matrix D of the time series X and the time series Y, where D[i,j] is the distance between the ith time point of the time series X and the jth time point of the time series Y.

[0334] Step S33, create a cumulative distance matrix C with the same size as the Euclidean distance matrix D, and initialize the starting point cumulative distance C[0,0]=D[0,0].

[0335] Step S34, fill the cumulative distance matrix C according to the dynamic time warping algorithm C[i,j]=D[i,j]+min{C[i-1,j],C[i,j-1],C[i-1,j-1]}, where i>0 and j>0, to obtain the filled cumulative distance matrix.

[0336] Step S35, starting from the lower right corner of the filled cumulative distance matrix, backtracking to the upper left corner of the filled cumulative distance matrix with the smallest cumulative distance to obtain the optimal path.

[0337] Step S36, linear interpolation is performed on the data in the non-diagonal movement of the optimal path to align the ctDNA sequencing data with the succinylated score in space-time.

[0338] Preferably, the position that needs to be interpolated can be determined according to the point pair relationship in the optimal path, wherein when the time points of the two sequences correspond directly (the path is diagonal movement), no interpolation is needed; when the path is horizontal or vertical movement (one time point corresponds to multiple other time points), interpolation is needed.

[0339] Preferably, the interpolation operation is performed on the sequence that needs to be adjusted, and the interpolation formula is:

[0340] y=y1+(x-x1)*(y2-y1) / (x2-x1).

[0341] Where (x1, y1) and (x2, y2) are known points at both ends of the interpolation interval.

[0342] For example, let:

[0343] ctDNA sequencing data (time points t1-t4): [10, 15, 20, 18].

[0344] Succinylated score (time points t1'-t3'): [50, 55, 60].

[0345] Optimal path point pair: (t1, t1'), (t2, t1'), (t3, t2'), (t4, t3').

[0346] Where the t2 time point of the ctDNA needs to correspond to the t1' time point of the succinylated score, but the succinylated score is missing data points between t1' and t2', so linear interpolation is needed to generate new score data points.

[0347] Therefore, for the t2 time point (value = 15) of the ctDNA, the adjacent succinylated score points are t1'(50) and t2'(55), and according to the interpolation formula: y = y1 + (x-x1)*(y2-y1) / (x2-x1), assuming that the time intervals are uniform, t2 is located in the middle position of t1' and t2', the calculation is: y = 50 + (0.5)*(55-50) = 52.5; for the t3 time point (value = 20) of the ctDNA, it directly corresponds to the t2' time point (55) of the succinylated score, and no interpolation is required; for the t4 time point (value = 18) of the ctDNA, the adjacent succinylated score points are t2'(55) and t3'(60), and assuming that t4 is located at 2 / 3 of the position of t2' and t3', the calculation is y = 55 + (2 / 3)*(60-55)≈58.33.

[0348] In summary, the aligned data pairs are (t1, 10) corresponding to (t1', 50), (t2, 15) corresponding to (interpolation point, 52.5), (t3, 20) corresponding to (t2', 55), and (t4, 18) corresponding to (interpolation point, 58.33).

[0349] Further, in step S4, a pre-trained dual-channel LSTM-Cox hybrid model is constructed, and the aligned ctDNA sequencing data is input into the time sequence channel of the dual-channel LSTM-Cox hybrid model, and the aligned succinylated score is input into the static channel of the dual-channel LSTM-Cox hybrid model, and the comprehensive grade score of the esophageal cancer immune data is output by the dual-channel LSTM-Cox hybrid model, which specifically includes the following steps:

[0350] In step S41, a dual-channel LSTM-Cox hybrid model containing a time sequence channel and a static channel is constructed, wherein the time sequence channel uses an LSTM network to process time sequence data, the static channel uses a fully connected network to process static features, and the outputs of the two channels are input into a Cox proportional hazards model after being fused by a splicing layer.

[0351] In step S42, the aligned ctDNA sequencing data is input into the LSTM network for training according to the time step, and the aligned succinylated score is input into the fully connected network for training.

[0352] In step S43, the output features of the trained time sequence channel and static channel are spliced and input into the Cox proportional hazards model, and are jointly trained by a negative log-likelihood loss function, and a comprehensive grade score is output by the Cox proportional hazards model.

[0353] Preferably, the aligned ctDNA sequencing data can be organized into a three-dimensional tensor (sample number x time step x feature dimension) by time step, input into the LSTM network of the time sequence channel, and extract the time sequence feature representation; the aligned succinylated score is organized into a two-dimensional tensor (sample number x feature dimension), input into the fully connected network of the static channel, and extract the static feature representation.

[0354] Further, in step S5, the continuous low-score interval, the medium-score interval, and the high-score interval are defined based on the preset score strategy, and each score interval corresponds to good immune effect, general immune effect, and poor immune effect in turn, and specifically includes the following steps:

[0355] In step S51, steps S1 to S4 are repeatedly performed based on the peripheral blood samples of a plurality of esophageal cancer patients to obtain a plurality of comprehensive grade scores.

[0356] In step S52, all comprehensive grade scores are clustered into three clusters by the K-means clustering algorithm.

[0357] Preferably, K=3 can be set according to the clinical needs, corresponding to the low, medium, and high score intervals, respectively representing good, general, and poor immune effect.

[0358] Preferably, the K-means++ algorithm can be used to initialize three initial center points c1, c2, and c3 to maximize the initial center point distance and avoid local optimal problems caused by random initialization; the Euclidean distance d(x, c1), d(x, c2), and d(x, c3) of each score sample x and the three centers is calculated, x is assigned to the cluster corresponding to the center with the smallest distance, forming three temporary clusters C1, C2, and C3; the center points of each temporary cluster are recalculated: c1'=mean(C1), c2'=mean(C2), and c3'=mean(C3) to obtain the updated center point set {c1', c2', c3'}; the distance change of the new and old center points is compared, and if max(|c1-c1'|, |c2-c2'|, |c3-c3'|)<ε (preset threshold) or the maximum iteration number is reached, the algorithm terminates; finally, after the algorithm converges, the center points of the three final clusters are used as the midpoint of the interval division.

[0359] Wherein, ε is usually taken between 0.01 to 0.1, if the data is standardized to 0-1 range, then 0.01 is selected.

[0360] In step S53, the center points of the three clusters are sequentially used as the interval midpoints of the low-score interval, the medium-score interval, and the high-score interval.

[0361] In step S54, zero is used as the left boundary of the low-score interval, and the right boundary of the low-score interval is obtained by symmetry through the interval midpoint of the low-score interval, and the low-score interval is obtained.

[0362] Step S55, taking the right boundary of the low-score interval as the left boundary of the medium-score interval, obtaining the right boundary of the medium-score interval by symmetry through the interval midpoint of the medium-score interval, to obtain the medium-score interval.

[0363] Step S56, taking the right boundary of the medium-score interval as the left boundary of the high-score interval, obtaining the right boundary of the high-score interval by symmetry through the interval midpoint of the high-score interval, to obtain the high-score interval.

[0364] Step S57, sequentially assigning the low-score interval, the medium-score interval, and the high-score interval with labels of good immune effect, general immune effect, and poor immune effect.

[0365] Further, step S57, sequentially assigning the low-score interval, the medium-score interval, and the high-score interval with labels of good immune effect, general immune effect, and poor immune effect, and then further comprising the following steps:

[0366] Step S10, repeatedly performing steps S1 to S4 based on additional peripheral blood samples to obtain additional comprehensive grade scores.

[0367] Step S20, repeatedly performing steps S52 to S56 to iteratively update the low-score interval, the medium-score interval, and the high-score interval with the collection period of the additional peripheral blood samples as the step size.

[0368] The embodiment is based on ctDNA obtained from peripheral blood samples of esophageal cancer patients, and 5hmC epigenome sequencing is performed on the ctDNA to obtain ctDNA sequencing data; based on tumor tissue samples of esophageal cancer patients, the level of succinyl modification is quantitatively analyzed by immunohistochemical staining and scoring system to obtain a succinyl score; the ctDNA sequencing data and the succinyl score are spatiotemporally aligned by a dynamic time warping algorithm; a pre-trained dual-channel LSTM-Cox hybrid model is constructed, and the aligned ctDNA sequencing data is input into the time sequence channel of the dual-channel LSTM-Cox hybrid model, and the aligned succinyl score is input into the static channel of the dual-channel LSTM-Cox hybrid model, and the comprehensive grade score of the esophageal cancer immune data is output by the dual-channel LSTM-Cox hybrid model; based on a pre-set scoring strategy, a continuous low score interval, a medium score interval and a high score interval are defined, and each score interval corresponds to good immune effect, general immune effect and poor immune effect in turn; the comprehensive grade score is classified into the corresponding score interval, and the matched immune effect is sent to an external receiving end. The embodiment first constructs a more comprehensive tumor microenvironment evaluation system through 5hmC epigenome sequencing of ctDNA and analysis of the level of succinyl modification of tumor tissue, solves the problem of poor prediction effect of traditional biomarkers (such as PD-L1 expression and tumor mutation load), and then realizes the complementary advantages of the two kinds of data through the dual-channel LSTM-Cox hybrid model by processing dynamic ctDNA data through the time sequence channel and processing the succinyl score through the static channel, so that the AUC value of the immune therapy effect prediction can reach more than 0.90, which is obviously better than the traditional single-mode prediction method (AUC≈0.75); finally, the complex immune data is quantified into three intuitive good / average / poor score intervals to provide clear treatment decision basis for clinicians, and the invalid treatment rate can be obviously reduced by the scoring system. The technical scheme of the embodiment not only improves the prediction accuracy of the precision of esophageal cancer immunotherapy, but also provides a new research perspective for understanding the mechanism of immunotherapy resistance, and has clinical application value and conversion potential.

[0369] As Figure 2 shown, the embodiment provides an embodiment of an esophageal cancer immune data prediction device based on tumor markers, and in the embodiment, the esophageal cancer immune data prediction device is applied to the esophageal cancer immune data prediction method in the above embodiment.

[0370] Specifically, the esophageal cancer immune data prediction device comprises a ctDNA sequencing data acquisition module 1, a succinyl score acquisition module 2, a spatiotemporal alignment module 3, a comprehensive grade score acquisition module 4, a score interval definition module 5, and an immune effect matching and sending module 6 which are sequentially electrically or signal connected.

[0371] The ctDNA sequencing data acquisition module 1 is used to acquire ctDNA based on a peripheral blood sample of an esophageal cancer patient, and to perform 5hmC epigenome sequencing on the ctDNA to obtain ctDNA sequencing data; the succinylation score acquisition module 2 is used to quantitatively analyze the succinylation modification level based on a tumor tissue sample of the esophageal cancer patient through immunohistochemical staining and a scoring system to obtain a succinylation score; the spatiotemporal alignment module 3 is used to perform spatiotemporal alignment of the ctDNA sequencing data and the succinylation score through a dynamic time warping algorithm; the comprehensive grade score acquisition module 4 is used to construct a pre-trained dual-channel LSTM-Cox hybrid model, and to input the aligned ctDNA sequencing data into a time sequence channel of the dual-channel LSTM-Cox hybrid model and to input the aligned succinylation score into a static channel of the dual-channel LSTM-Cox hybrid model, so as to output a comprehensive grade score of esophageal cancer immune data through the dual-channel LSTM-Cox hybrid model; the score interval definition module 5 is used to define continuous low-score intervals, medium-score intervals and high-score intervals based on a preset score strategy, and each score interval corresponds to good immune effect, general immune effect and poor immune effect in turn; and the immune effect matching and sending module 6 is used to classify the comprehensive grade score into a corresponding score interval, and to send the matched immune effect to an external receiving end.

[0372] Further, as shown in Figure 3 The ctDNA sequencing data acquisition module 1 specifically includes a ctDNA fragmentation processing unit, an original sequencing data acquisition unit, an original sequencing data processing unit and a ctDNA sequencing data saving unit which are electrically or signal connected in turn; and the ctDNA sequencing data saving unit is electrically or signal connected with the succinylation score acquisition module 2.

[0373] The ctDNA fragmentation processing unit is used to perform fragmentation processing on the ctDNA to obtain ctDNA fragments of a preset length; the original sequencing data acquisition unit is used to perform high-throughput sequencing on the ctDNA fragments through a preset sequencing platform to obtain original sequencing data; the original sequencing data processing unit is used to remove low-quality reads and adapter contamination of the original sequencing data through FastQC to obtain processed sequencing data; and the ctDNA sequencing data saving unit is used to save the processed sequencing data in a FASTQ format to obtain ctDNA sequencing data.

[0374] Further, as shown in Figure 4As shown, the succinylation score acquisition module 2 specifically includes an immunohistochemical staining image acquisition unit, an immunohistochemical staining image processing unit, a binary mask acquisition unit, a staining intensity level acquisition unit, a positive proportion calculation unit, and a succinylation score calculation unit, which are sequentially electrically or signal connected. The immunohistochemical staining image acquisition unit is electrically or signal connected with the ctDNA sequencing data storage unit, and the succinylation score calculation unit is electrically or signal connected with the space-time alignment module 3.

[0375] The immunohistochemical staining image acquisition unit is configured to acquire a high-resolution immunohistochemical staining image of a tumor tissue sample through a full-automatic digital pathology scanning system; the immunohistochemical staining image processing unit is configured to perform denoising processing and contrast enhancement on the high-resolution immunohistochemical staining image to obtain a processed immunohistochemical staining image; the binary mask acquisition unit is configured to separate the succinylation positive area from the background through a color threshold segmentation algorithm to obtain a binary mask; the staining intensity level acquisition unit is configured to calculate the average staining intensity in the succinylation positive area through a gray scale histogram and grade the average staining intensity according to the standard staining intensity to obtain the staining intensity level; the positive proportion calculation unit is configured to identify the total number of cells of the binary mask and the number of positive cells of the succinylation positive area through a pre-trained SqueezeNet_Residual model and obtain the positive proportion of the number of positive cells based on the total number of cells; and the succinylation score calculation unit is configured to multiply the staining intensity level and the positive proportion to obtain the succinylation score.

[0376] Further, as shown in the succinylation score acquisition module 2, Figure 5 The space-time alignment module 3 specifically includes a time series construction unit, an Euclidean distance matrix calculation unit, a cumulative distance matrix construction unit, a cumulative distance matrix filling unit, an optimal path backtracking unit, and a data linear interpolation unit, which are sequentially signal connected. The time series construction unit is electrically or signal connected with the succinylation score calculation unit, and the data linear interpolation unit is electrically or signal connected with the comprehensive grade score acquisition module 4.

[0377] The time series construction unit is configured to construct the ctDNA sequencing data into a time series X in chronological order, each time point corresponding to a multi-dimensional feature vector, and to construct the succination score into a time series Y in chronological order, each time point corresponding to a single-dimensional score value; the Euclidean distance matrix calculation unit is configured to calculate a Euclidean distance matrix D of the time series X and the time series Y, wherein D[i,j] is the distance between the ith time point of the time series X and the jth time point of the time series Y; the cumulative distance matrix construction unit is configured to create a cumulative distance matrix C of the same size as the Euclidean distance matrix D and initialize the cumulative distance of the starting point C[0,0]=D[0,0]; the cumulative distance matrix filling unit is configured to fill the cumulative distance matrix C[i,j]=D[i,j]+min{C[i-1,j],C[i,j-1],C[i-1,j-1]} according to the dynamic time warping algorithm, wherein i>0 and j>0, to obtain a filled cumulative distance matrix; the optimal path backtracking unit is configured to start from the lower right corner of the filled cumulative distance matrix and backtrack to the upper left corner of the filled cumulative distance matrix with the minimum cumulative distance to obtain an optimal path; and the data linear interpolation unit is configured to perform linear interpolation on the data that moves along the non-diagonal line in the optimal path, so that the ctDNA sequencing data and the succination score are spatiotemporally aligned.

[0378] Further, as shown in the figure, Figure 6 The comprehensive grade score acquisition module 4 specifically includes a double-channel LSTM-Cox hybrid model construction unit, a double-channel data training unit, and a comprehensive grade score output unit connected in sequence in an electrical or signal manner; the double-channel LSTM-Cox hybrid model construction unit is electrically or signal connected with the data linear interpolation unit, and the comprehensive grade score output unit is electrically or signal connected with the score interval definition module 5.

[0379] The double-channel LSTM-Cox hybrid model construction unit is configured to construct a double-channel LSTM-Cox hybrid model including a time series channel and a static channel, wherein the time series channel adopts an LSTM network to process time series data, the static channel adopts a fully connected network to process static features, and the outputs of the two channels are fused through a concatenation layer and then input into a Cox proportional hazards model; the double-channel data training unit is configured to input the aligned ctDNA sequencing data into the LSTM network for training at a time step and input the aligned succination score into the fully connected network for training; the comprehensive grade score output unit is configured to concatenate the output features of the trained time series channel and static channel, input them into the Cox proportional hazards model, and jointly train them through a negative log-likelihood loss function, and output a comprehensive grade score through the Cox proportional hazards model.

[0380] Further, as shown in the figure, Figure 7As shown, the score interval definition module 5 specifically includes a comprehensive level score amplification unit, a comprehensive level score clustering unit, a score interval midpoint definition unit, a low score interval acquisition unit, a medium score interval acquisition unit, a high score interval acquisition unit, and an immune effect assignment unit, which are sequentially electrically or signal connected. The comprehensive level score amplification unit is electrically or signal connected with the comprehensive level score output unit, and the immune effect assignment unit is electrically or signal connected with the immune effect matching and sending module 6.

[0381] The comprehensive level score amplification unit is configured to repeatedly execute the ctDNA sequencing data acquisition module to the comprehensive level score acquisition module based on peripheral blood samples of a plurality of esophageal cancer patients to obtain a plurality of comprehensive level scores. The comprehensive level score clustering unit is configured to cluster all the comprehensive level scores into three clusters by using a K-means clustering algorithm. The score interval midpoint definition unit is configured to sequentially take the center points of the three clusters as the interval midpoints of the low score interval, the medium score interval, and the high score interval. The low score interval acquisition unit is configured to take a zero value as the left boundary of the low score interval, and obtain the right boundary of the low score interval by symmetry with respect to the interval midpoint of the low score interval, to obtain the low score interval. The medium score interval acquisition unit is configured to take the right boundary of the low score interval as the left boundary of the medium score interval, and obtain the right boundary of the medium score interval by symmetry with respect to the interval midpoint of the medium score interval, to obtain the medium score interval. The high score interval acquisition unit is configured to take the right boundary of the medium score interval as the left boundary of the high score interval, and obtain the right boundary of the high score interval by symmetry with respect to the interval midpoint of the high score interval, to obtain the high score interval. The immune effect assignment unit is configured to sequentially assign the low score interval, the medium score interval, and the high score interval with labels of good immune effect, general immune effect, and poor immune effect.

[0382] Further, the esophageal cancer immune data prediction device includes a comprehensive level score additional amplification unit and a score interval iterative update unit, which are sequentially electrically or signal connected. The comprehensive level score additional amplification unit is electrically or signal connected with the immune effect assignment unit.

[0383] It should be noted that the present embodiment is a functional module embodiment based on the above-mentioned method embodiment. The preferred, expanded, limited, exemplified, and principle explained parts of the present embodiment can be referred to the above-mentioned embodiment. The present embodiment will not be described again.

[0384] The embodiment is based on ctDNA obtained from peripheral blood samples of esophageal cancer patients, and 5hmC epigenome sequencing is performed on the ctDNA to obtain ctDNA sequencing data; based on tumor tissue samples of esophageal cancer patients, the level of succinyl modification is quantitatively analyzed by immunohistochemical staining and scoring system to obtain a succinyl score; the ctDNA sequencing data and the succinyl score are spatiotemporally aligned by a dynamic time warping algorithm; a pre-trained dual-channel LSTM-Cox hybrid model is constructed, and the aligned ctDNA sequencing data is input into the time sequence channel of the dual-channel LSTM-Cox hybrid model, and the aligned succinyl score is input into the static channel of the dual-channel LSTM-Cox hybrid model, and the comprehensive grade score of the esophageal cancer immune data is output by the dual-channel LSTM-Cox hybrid model; based on a preset scoring strategy, a continuous low score interval, a medium score interval and a high score interval are defined, and each score interval corresponds to good immune effect, general immune effect and poor immune effect in turn; the comprehensive grade score is classified into the corresponding score interval, and the matched immune effect is sent to an external receiving end. The embodiment first constructs a more comprehensive tumor microenvironment evaluation system through 5hmC epigenome sequencing of ctDNA and analysis of the level of succinyl modification of tumor tissue, solves the problem of poor prediction effect of traditional biomarkers (such as PD-L1 expression and tumor mutation load), and then the dual-channel LSTM-Cox hybrid model processes dynamic ctDNA data through the time sequence channel and processes the succinyl score through the static channel, realizes the complementary advantages of the two kinds of data, and makes the AUC value of the immune therapy effect prediction reach more than 0.90, which is obviously better than the traditional single modality prediction method (AUC≈0.75); finally, the complex immune data is quantified into three intuitive good / general / poor score intervals to provide clear treatment decision basis for clinicians, and the invalid treatment rate can be obviously reduced. The technical scheme of the embodiment not only improves the prediction accuracy of the precision of esophageal cancer immunotherapy, but also provides a new research perspective for understanding the mechanism of immunotherapy resistance, and has clinical application value and conversion potential.

[0385] Figure 8 FIG. 1 is a structural schematic diagram of an electronic device according to an embodiment of the present application. As shown in FIG. 1, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71. Figure 8

[0386] The memory 72 stores program instructions for implementing the esophageal cancer immune data prediction method based on tumor markers of any of the above embodiments.

[0387] The processor 71 is configured to execute the program instructions stored in the memory 72 to perform the esophageal cancer immune data prediction based on tumor markers.

[0388] ​The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0389] Furthermore, Figure 9 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 9 In this embodiment of the application, the storage medium 8 stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0390] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0391] In addition, the various functional units in the embodiments of the present application can be integrated in one processing unit, or each can exist physically as a separate unit, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware, or in the form of a software functional unit. The above is only an implementation of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An esophageal cancer immunological data prediction device based on tumor markers, characterized by, The esophageal cancer immune data prediction device comprises: a ctDNA sequencing data acquisition module configured to acquire ctDNA based on a peripheral blood sample of the esophageal cancer patient, and perform 5hmC epigenome sequencing on the ctDNA to obtain ctDNA sequencing data; a succinylation score acquisition module configured to quantitatively analyze a succinylation modification level based on a tumor tissue sample of the esophageal cancer patient by immunohistochemical staining and a scoring system to obtain a succinylation score; a space-time alignment module configured to perform space-time alignment on the ctDNA sequencing data and the succinylation score by a dynamic time warping algorithm; a comprehensive grade score acquisition module configured to construct a pre-trained dual-channel LSTM-Cox hybrid model, input the aligned ctDNA sequencing data into a time sequence channel of the dual-channel LSTM-Cox hybrid model, input the aligned succinylation score into a static channel of the dual-channel LSTM-Cox hybrid model, and output a comprehensive grade score of esophageal cancer immune data by the dual-channel LSTM-Cox hybrid model; a score interval definition module configured to define continuous low-score intervals, medium-score intervals and high-score intervals based on a preset score strategy, and each score interval corresponds to good immune effect, general immune effect and poor immune effect in turn; an immune effect matching and sending module configured to classify the comprehensive grade score into a corresponding score interval, and send the matched immune effect to an external receiving end.

2. The esophageal cancer immunization data prediction device according to claim 1, characterized by, The ctDNA sequencing data acquisition module comprises: a ctDNA fragmentation processing unit configured to perform fragmentation processing on the ctDNA to obtain ctDNA fragments of a preset length; an original sequencing data acquisition unit configured to perform high-throughput sequencing on the ctDNA fragments by a preset sequencing platform to obtain original sequencing data; an original sequencing data processing unit configured to remove low-quality reads and adapter contamination of the original sequencing data by FastQC to obtain processed sequencing data; a ctDNA sequencing data saving unit configured to save the processed sequencing data in a FASTQ format to obtain the ctDNA sequencing data.

3. The esophageal cancer immunization data prediction device according to claim 1, characterized by, The succinylation score acquisition module comprises: an immunohistochemical staining image acquisition unit configured to acquire a high-resolution immunohistochemical staining image of the tumor tissue sample by a full-automatic digital pathology scanning system; an immunohistochemical staining image processing unit configured to perform denoising processing and contrast enhancement on the high-resolution immunohistochemical staining image to obtain a processed immunohistochemical staining image; a binary mask acquisition unit configured to separate a succinylation positive area from a background by a color threshold segmentation algorithm to obtain a binary mask; a staining intensity grade acquisition unit configured to calculate an average staining intensity in the succinylation positive area by a gray level histogram, and grade the average staining intensity according to a standard staining intensity to obtain a staining intensity grade; and a succinylation score calculation unit configured to calculate a succinylation score of the tumor tissue sample according to the staining intensity grade. A positive proportion calculation unit is configured to identify the total cell number of the binarized mask and the positive cell number of the succinylated positive area by using the pre-trained SqueezeNet_Residual model, and obtain a positive proportion of the positive cell number based on the total cell number. A succinylated score calculation unit is configured to multiply the staining intensity grade and the positive proportion to obtain the succinylated score.

4. The esophageal cancer immunization data prediction device according to claim 1, wherein The spatiotemporal alignment module comprises: A time series construction unit is configured to construct the ctDNA sequencing data into a time series X in chronological order, each time point corresponding to a multi-dimensional feature vector, and construct the succinylated score into a time series Y in chronological order, each time point corresponding to a single-dimensional score value. An Euclidean distance matrix calculation unit is configured to calculate an Euclidean distance matrix D of the time series X and the time series Y, wherein D[i, j] is the distance between the ith time point of the time series X and the jth time point of the time series Y. A cumulative distance matrix construction unit is configured to create a cumulative distance matrix C of the same size as the Euclidean distance matrix D and initialize the cumulative distance of the starting point as C[0, 0] = D[0, 0]. A cumulative distance matrix filling unit is configured to fill the cumulative distance matrix C according to the dynamic time warping algorithm, C[i, j] = D[i, j] + min{C[i-1, j], C[i, j-1], C[i-1, j-1]}, wherein i > 0 and j > 0, to obtain a filled cumulative distance matrix. An optimal path backtracking unit is configured to take the lower right corner of the filled cumulative distance matrix as the starting point, backtrack to the upper left corner of the filled cumulative distance matrix with the minimum cumulative distance, and obtain an optimal path. A data linear interpolation unit is configured to perform linear interpolation on the data of the optimal path that moves along a non-diagonal line, so as to align the ctDNA sequencing data and the succinylated score in space and time.

5. The esophageal cancer immunization data prediction device according to claim 1, characterized by, The comprehensive grade score acquisition module comprises: A dual-channel LSTM-Cox hybrid model construction unit is configured to construct a dual-channel LSTM-Cox hybrid model comprising a time series channel and a static channel, wherein the time series channel uses an LSTM network to process time series data, the static channel uses a fully connected network to process static features, and the outputs of the two channels are fused through a concatenation layer and then input into a Cox proportional hazards model. A dual-channel data training unit is configured to input the aligned ctDNA sequencing data into the LSTM network for training by time steps, and input the aligned succinylated score into the fully connected network for training. A comprehensive grade score output unit is configured to concatenate the output features of the trained time series channel and static channel, input them into the Cox proportional hazards model, and jointly train them through a negative log-likelihood loss function, and output the comprehensive grade score through the Cox proportional hazards model.

6. The esophageal cancer immunization data prediction device according to claim 1, wherein The score interval definition module comprises: A comprehensive grade score amplification unit is configured to repeatedly execute the ctDNA sequencing data acquisition module to the comprehensive grade score acquisition module based on a plurality of peripheral blood samples of esophageal cancer patients to obtain a plurality of comprehensive grade scores. The comprehensive grade score clustering unit is configured to cluster all the comprehensive grade scores into three clusters by using a K-means clustering algorithm. The score interval midpoint defining unit is configured to sequentially take the center points of the three clusters as the interval midpoints of the low score interval, the medium score interval and the high score interval. The low score interval obtaining unit is configured to take a zero value as the left boundary of the low score interval, obtain the right boundary of the low score interval by symmetrically using the interval midpoint of the low score interval, and obtain the low score interval. The medium score interval obtaining unit is configured to take the right boundary of the low score interval as the left boundary of the medium score interval, obtain the right boundary of the medium score interval by symmetrically using the interval midpoint of the medium score interval, and obtain the medium score interval. The high score interval obtaining unit is configured to take the right boundary of the medium score interval as the left boundary of the high score interval, obtain the right boundary of the high score interval by symmetrically using the interval midpoint of the high score interval, and obtain the high score interval. The immune effect assigning unit is configured to sequentially assign the low score interval, the medium score interval and the high score interval with labels of good immune effect, general immune effect and poor immune effect.

7. The esophageal cancer immunization data prediction device according to claim 6, characterized by, After the immune effect assigning unit, the esophageal cancer immune data prediction device further comprises: The comprehensive grade score additional amplification unit is configured to repeatedly execute the ctDNA sequencing data obtaining module to the comprehensive grade score obtaining module based on additional peripheral blood samples to obtain additional comprehensive grade scores. The score interval iterative updating unit is configured to repeatedly execute the comprehensive grade score clustering unit to the high score interval obtaining unit with the collection period of the additional peripheral blood samples as a step to iteratively update the low score interval, the medium score interval and the high score interval.

8. An electronic device, comprising: The processor executes the program instructions stored in the memory to realize the operation of the esophageal cancer immune data prediction device according to any one of claims 1 to 7.

9. A storage medium, characterized by The storage medium stores program instructions, and the program instructions are executed by the processor to realize the operation of the esophageal cancer immune data prediction device according to any one of claims 1 to 7.

Images