CN115881296A - A risk-assisted stratification system for papillary thyroid carcinoma (PTC) - Google Patents

Abstract

The invention discloses a Papillary Thyroid Carcinoma (PTC) risk auxiliary layering system, which is characterized by comprising: the system comprises a data acquisition module, a data preprocessing module, a model establishing, practicing and verifying module, a layering module and a database platform. According to the invention, the value of proteomics on PTC danger layering is realized by screening the characteristic protein and the characteristic clinical variable through machine learning and constructing a system prediction risk model, and a decision basis is provided for a PTC patient to make a personalized treatment scheme.

Description

A risk-assisted stratification system for papillary thyroid carcinoma (PTC)

technical field

The present invention relates to an auxiliary stratification system and method for low or medium-high tumor risk, in particular to an artificial intelligence papillary thyroid carcinoma (PTC) risk auxiliary stratification system and method. The invention belongs to the technical field of medical artificial intelligence assisted stratification.

Background technique

Thyroid cancer is one of the most common endocrine malignancies in the world. In recent years, the incidence of thyroid cancer has shown a rapid upward trend. Thyroid cancer includes 4 histological types, among which papillary thyroid cancer (PTC) accounts for 85% of thyroid cancer. Although the overall prognosis of PTC is good, and the 10-year survival rate exceeds 90%, some tumors still exhibit highly aggressive biological behavior in the early stage, and 30-80% of patients have lymph node metastasis at the time of first diagnosis. PTC capsule invasion is found in about 30-50% of cases at the first operation, and the postoperative recurrence rate is still 10-25% even after complete surgical resection. Therefore, early and accurate assessment of PTC lesions and risk levels, customized and individualized surgical plans and specifications are the basis for precise diagnosis and treatment.

At the present stage, doctors mostly rely on their own experience and knowledge in the process of tumor treatment, as well as refer to existing medical guidelines for assessment and treatment. There are problems such as low assessment accuracy, cumbersome work and limited experience. At present, ultrasound, CT, and MRI are the main imaging methods for evaluating PTC lesions and cervical lymph nodes, but previous literatures have not shown that they have high efficacy in evaluating the risk of lesions. Studies exploring PTC risk stratification from genomic changes are relatively limited, and genomics and transcriptomics have revealed somatic mutations in different thyroid cancers, such as BRAFV600E, RAS, TP53, TERT, and RET/PTC gene fusions; among them, BRAFV600E mutation The pathogenesis of PTC has attracted much attention. Relevant data show that 85% of PTC have BRAFV600E mutation, but only a single indicator of BRAFV600E mutation has a weak predictive ability for the prognosis of PTC.

Unlike nucleic acids in genomics, proteins are directly involved in all life processes and determine cellular and organ phenotypes. As the most direct product of life activities, protein can show relatively stable expression in different life stages of cells and different protein cycles. Proteomics has a significant predictive effect on the prognosis of lung cancer, oral squamous cell carcinoma, gastric cancer, esophageal cancer, colorectal cancer and other tumors. Sofiadis (SOFIADISA, DINETS A, ORRE L M, et al. Proteomic study of thyroidtumors reveals frequent up-regulation of the Ca2+ -binding protein S100A6 inpapillarythyroid carcinoma [J]. Thyroid, 2010, 20 (10): 1067-76) and other studies revealed that the frequent upregulation of Ca2+-binding protein S100A6 in PTC can be used as a potential auxiliary marker to distinguish follicular thyroid tumors from PTC. There are also studies (ZHAN S, LI J, WANGT, et al. Quantitative Proteomics Analysis of Sporadic Medullary Thyroid Cancer Reveals FN1 as a Potential Novel Candidate Prognostic Biomarker [J]. Oncologist, 2018, 23(12): 1415-25) on sporadic myeloid cancer Quantitative proteomic analysis of cancer reveals fibronectin 1 as a novel candidate prognostic biomarker.

Traditional proteomic sample preparation techniques such as in-solution digestion or two-dimensional gel electrophoresis (Two-dimensional gel electrophoresis, 2D PAGE) are time-consuming and inefficient. The birth of pressure cycling technology (Pressure cycling technology, PCT) is an important processing technology for constructing proteomics database based on FFPE specimens, using ambient normal pressure (14.7psi) and high pressure ( Up to 45,000 psi) between rapid alternating hydrostatic pressure changes, this method is simple, fast, high-throughput, and effectively promotes the de-crosslinking, extraction and trypsin digestion of proteins in FFPE sections, thus effectively improving the identification of proteins from FFPE samples. protein quantity.

In addition to PCT-assisted sample processing, quantitative and accurate mass spectrometry (MS) analysis is required to identify disease-associated changes in protein expression levels. Data independent acquisition (DIA) is a proteomics technique that performs fragmentation and MS/MS analysis of all ions within a selected mass-to-charge ratio (m/z) range. DIA is an alternative to Data-dependent acquisition (DDA). Compared with DDA, the biggest advantage of DIA is the efficient determination of protein molecules with extremely low abundance in complex samples, which can obtain complete data and achieve protein depth. Coverage and precise quantification greatly improve the reliability of quantitative analysis and have higher quantitative accuracy and repeatability. The combination of DIA-MS with deep neural network-based data processing effectively improves reproducibility, number of identifications, and quantitative accuracy.

With the development of the network and technology, it means that there are a large number of data sets. Artificial intelligence is currently used in many fields. How to use artificial intelligence and mine and process these data to better serve the progress of human life has become a challenge. Now it is the goal that various disciplines and major fields are striving to be the first. When artificial intelligence enters medical care and transforms medical records into medical knowledge, the occurrence and development of diseases can be better understood; biological information such as gene sequences and DNA sequence data can be collected to better understand the differences in the expression of various biological information of diseases. Zhu et al. used machine learning to compare the exosomal metabolome patterns of relapsed and non-relapsed patients and demonstrated that a marker panel consisting of 3′-UMP, palmitoleic acid, palmitaldehyde, and isobutyl decanoate was used to in predicting recurrence of esophageal squamous cell carcinoma. Previous studies have shown that the combination of artificial intelligence and basic disciplines can be effectively applied to clinically difficult problems.

Patent US6005256A provides a detailed description of devices and methods for simultaneously detecting multiple fluorescently-labeled markers in body samples, including devices and methods for the purpose of identifying cancer cells, but does not involve specific applications in cancer assessment or mention Appropriate marker combinations can yield high specificity.

Patent CN110129442A discloses the application of reagents for detecting levels of genes and their expression products in the preparation of products for diagnosing thyroid cancer, characterized in that the genes are selected from LRRN4CL or ZNF883. The products include chips, kits or nucleic acid membrane strips. The gene chip includes an oligonucleotide probe for detecting LRRN4CL or ZNF883 gene transcription level for LRRN4CL or ZNF883 gene, and the protein chip includes a specific binding agent for LRRN4CL or ZNF883 protein; the test kit includes gene detection kit, protein detection kit, the gene detection kit includes reagents for detecting LRRN4CL or ZNF883 gene transcription level, or a chip, and the protein detection kit includes reagents for detecting LRRN4CL or ZNF883 protein expression level, or chip.

The invention patent of patent CN107563383A discloses an auxiliary prediction system for the risk stratification of benign and malignant lesions, including nodule detection, doctor guidance, semantic annotation generation, sample generation, online learning and other steps. The contour of the pulmonary nodule in the area of interest and the prediction probability of each pixel in the contour are calculated, and the labeled samples of pulmonary nodules containing precise semantic contours are generated for learning by the self-learning system.

Currently, the commonly used prognostic assessment systems for thyroid cancer include TNM staging in AJCC/UICC, AMES, AGES, MACIS scoring system, and ATA risk stratification guidelines. AJCC/UICC is a staging system based on pTNM parameters and age, which is applicable to all types of tumors, and is also applicable to thyroid cancer. Prognostic evaluation systems such as AMES, AGES, and MACIS mainly include the presence or absence of lymph node/distant metastasis, patient age, and tumor extent. In addition, the ATA risk stratification guidelines proposed a 3-level stratification system for thyroid cancer to predict the recurrence risk of thyroid cancer. These prognostic evaluation systems are all based on postoperative clinicopathological results, but currently there is no evaluation system that can assess the risk of PTC before operation.

Contents of the invention

In order to solve the deficiencies of the prior art, the purpose of the present invention is to provide an auxiliary risk stratification system for papillary thyroid carcinoma (PTC), improve the confidence of clinicians, and assist clinical management decision-making of papillary thyroid carcinoma.

One aspect of the present invention aims to provide an auxiliary papillary thyroid carcinoma (PTC) risk stratification system, which is characterized in that it includes:

Data collection module: used to collect existing data, including proteomics + clinical indicators + blood immune indicators + BRAFV600E mutation label, after verifying and evaluating data integrity and data quality, perform data desensitization and send to data preprocessing module;

Data preprocessing module: perform missing value interpolation and normalization processing on the collected existing data;

Model building, training and verification module: feature screening based on the whole proteome detection data obtained by the data preprocessing module, and then based on the model of gradient boosting tree - extreme gradient boosting (XGBoost), linking input features with PTC risk stratification, Used to solve supervised learning problems; the discovery set is used as a training set to train and adjust parameters in the model, divide the discovery set into a training sequence and a test sequence, build a model in the training sequence, perform internal verification in the test sequence, select and save The model with the best AUC on the test set was then validated on two independent validation sets and the performance of the model was evaluated by the area under the curve (AUC) of the receiver operating characteristic curve; protein features selected based on the XGBoost algorithm combined with a 10-fold crossover The verification finally determined that 6 characteristic proteins, 5 clinical characteristics, 5 blood immune indicators and BRAFV600E mutation characteristics were used as the characteristics of the model; a risk stratification model for papillary thyroid carcinoma (PTC) was established, and the threshold of the model was determined. Predict and stratify risk lesions at all levels;

Stratification module: analyze the test samples, stratify tumors into low-risk lesions or medium-high-risk lesions, and feed back to the doctor's terminal equipment and database platform;

Database platform: connected with the above modules, used for model building data storage and retrieval, and continuously collects new data to provide online model update function.

The assessment criteria were based on the ATA 2015 management guidelines: samples without tumor invasion and lymph node metastasis were considered low-risk PTC, while samples with tumor invasion or lymph node metastasis were considered intermediate or high-risk PTC.

Preferably, the model threshold is 0.6645.

Preferably, the six characteristic proteins are: DPP7, DPLIM3, COL12A1, CTSL, TUBB2A and ITGB5.

Preferably, the five clinical features are: capsule invasion, extraglandular invasion, tumor diameter, multifocality and age.

Preferably, the five blood immune indexes are: platelet count (Platelet count, plt), neutrophil count (Neutrophil count, N), lymphocyte count (Lymphocyte count, L), monocyte count (Monocyte count , M) and lymphocyte/monocyte ratio (Lymphocyte-to-MonocyteRatio, LMR).

Preferably, the sources of the existing data include: hospital case data.

Preferably, the supervised learning is a process of using a set of samples of known categories to adjust the parameters of the classifier to achieve the required performance.

Preferably, the sample inclusion criteria are as follows: ① initial operation with lymph node dissection; ② no history of chemotherapy and radiotherapy; ③ classic PTC diagnosed by postoperative pathological examination; ④ postoperative pathological diagnosis results containing complete information on patient risk stratification. information. Exclusion criteria: ① History of neck trauma; ② Combination or previous suffering from other cancers; ③ Postoperative pathological diagnosis of other subtypes of PTC or other pathological types; ④ Lack of complete and available postoperative pathology.

Another aspect of the present invention aims to provide a non-diagnostic papillary thyroid cancer (PTC) risk auxiliary stratification system using method, which is characterized in that the above-mentioned auxiliary stratification system is used to obtain stratification results.

The last aspect of the present invention aims to provide the use of the above papillary thyroid carcinoma (PTC) risk auxiliary stratification system in preparing a device for predicting and stratifying the risk of papillary thyroid carcinoma in patients.

The following is an explanation of some terms involved in the present invention:

Abbreviations/Symbol Explanation

English abbreviation English full name Chinese name AJCC American joint committee on cancer American Joint Committee on Cancer ATA The American thyroid association American Thyroid Association AUC Area under curve area under the curve COL12A1 Collagen alpha-1(XII) chain Collagen alpha-1(XII) chains CTSL Procathepsin L Cathepsin L DIA Data-independent acquisition Independent Data Acquisition DPP7 Dipeptidyl peptidase 7 p-dipeptidyl peptidase FFPE formalin-fixed paraffin embedded formalin fixed paraffin embedding FNAB Fine needle aspiration biopsy fine needle aspiration biopsy HT Hashimoto's thyroiditis Hashimoto's thyroiditis ITGB5 Integrin subunit beta 5 Integrin beta-5 LNM Lymph node metastasis Lymph node metastasis PDLIM3 PDZ and LIM domain 3 PDZ and LIM domain protein 3 PTC Papillary thyroid carcinoma papillary thyroid carcinoma ROC Receiver operating characteristic curve receiver operating characteristic curve SHAP SHapley Additive explanations SHapley additional explanation TC Thyroid carcinoma Thyroid cancer TUBB2A Tubulin beta 2A class IIa Tubulin β-2A chain XGBoost Extreme gradient boosting extreme gradient boosting

Dipeptidyl peptidase 7 (DPP7): Regulated by the DPP7 gene, it plays an important role in the degradation of some oligopeptides. It was found in the Bgee gene expression database that DPP7 is expressed in the thyroid lobe. This protein is a marker of poor prognosis in colorectal cancer. However, literature has shown that high expression of DPP7 in breast cancer patients is associated with good prognosis. Previous literature has shown that knockout of DPP7 can up-regulate Bax in HepG2 liver cancer cell line -Bcl2 signaling increases cell apoptosis, so it is also possible to reduce the risk of thyroid cancer through similar pathways, which is consistent with the research results of the present invention. For the amino acid sequence of DPP7, please refer to GenBank accession number NP_037511.2, and for the nucleic acid sequence, please refer to NM_013379.3.

Cathepsin L (CTSL) is a thiol protease that plays an important role in the overall degradation of proteins in lysosomes and is an important protease for maintaining thyroid function cells. It is involved in the limited proteolysis of thyroglobulin in the lumen of thyroid follicles, the dissolution of cross-linked thyroglobulin and the subsequent release of thyroxine T4. For the amino acid sequence of CTSL, please refer to GenBank accession number NP_001244900.1, and for the nucleic acid sequence, please refer to NM_001257971.2.

Integrin beta-5 (Integrin subunit beta 5, ITGB5) is closely related to lymph node metastasis. ITGB5 can induce downstream Src phosphorylation, thereby activating NF-kB signaling pathway and promoting tumor metastasis, which may be related to the tissue heterogeneity of the protein. For the amino acid sequence of ITGB5, please refer to GenBank accession number NP_001341693.1, and for the nucleic acid sequence, please refer to NM_001354764.2.

PDZ and LIM domain 3 (PDZ and LIM domain 3, PDLIM3): Studies have shown that PDLIM3 may play a role in the organization of actin filament arrays in muscle cells. This protein is a marker of poor prognosis in thyroid cancer. It is also reported in literature that PDLIM3 regulates cell proliferation and differentiation through the MAPK signaling pathway, which may lead to the promotion of highly aggressive biological behavior of thyroid cancer, which is also consistent with the research results of the present invention. For the amino acid sequence of PDLIM3, please refer to GenBank accession number NP_001107579.1, and for the nucleic acid sequence, please refer to NM_001114107.5.

Collagen alpha-1(XII) chain (Collagen alpha-1(XII) chain, COL12A1): Studies have shown that in gastric cancer, the COL12A1 system acts through the MAPK pathway to form a positive feedback to promote cell migration, and it is speculated that tumor cell invasion based on this and the high risk of transfer. For the amino acid sequence of COL12A1, please refer to GenBank accession number NP_004361.3, and for the nucleic acid sequence, please refer to NM_004370.6.

Tubulin β-2A chain (Tubulin beta 2A class IIa, TUBB2A) is the main component of microtubules, participates in the cell cycle of mitosis, and is closely related to the growth of tumors. Studies have shown that TUBB2A is a new marker for predicting distant metastasis of breast cancer [48]. In the study of cell lines, when TUBB2A was knocked down, the invasive cells were significantly reduced, thus verifying the distant metastasis potential of TUBB2A . In gastric cancer cells, the expression of TUBB2A also promotes the proliferation, migration and invasion of gastric cancer cell lines. The above studies have shown that the expression of TUBB2A leads to a state of high invasion and high metastasis, which is highly consistent with the results of this study. For the amino acid sequence of TUBB2A, please refer to GenBank accession number NP_001060.1, and for the nucleic acid sequence, please refer to NP_001060.1.

Gradient boosting (GB) algorithm is one of the algorithms of artificial intelligence, which creates more accurate and stronger learners by combining simple and weak decision trees. While the accuracy of the weak tree model reveals a flaw in prediction error, it can be compensated for using the second model. Therefore, combining these successive weak tree models produces a model that is more accurate than the first one. In GB, a weak tree model is fitted to the residuals, and then the forecasts are updated by adding the predicted residuals to the previous forecasts. Extreme gradient boosting (XGBoost) is a model that has recently received much attention in tree-based ensemble learning. Although XGBoost is based on GB, XGBoost can overcome many shortcomings of GB, such as slow execution time and lack of overscaling. Therefore, it can finish training faster than existing GB models.

The benefits of the present invention are as follows: (1) The present invention is based on the unique PCT-MS technology, processing including fresh frozen tissue samples, FFPE samples and fine-needle aspiration (FNA) , and screening features through machine learning Protein and characteristic clinical variables were used to construct a predictive risk model, and the above model was tested through paraffin section samples and biopsy samples. Explore the value of proteomics in risk stratification of PTC, and provide decision-making basis for the development of personalized treatment plans for PTC patients.

(2) At the same time, the present invention constructs a prognosis prediction model derived from preoperative FNA specimen proteins, and finally condenses into a PTC risk assessment system. The prediction efficiency of the model based on this is high, and the AUC of the discovery set reached 0.92, and the paraffin samples and FNA puncture samples in the verification group reached 0.79 and 0.80, respectively. According to the previous literature review, this report is the first study based on proteomics and artificial intelligence to explore the relationship between PTC and risk. It realizes PTC risk stratification management and individualized precise diagnosis and treatment before surgery, providing effective and reliable basic research. , with pioneering and innovative significance.

(3) XGBoost is a model that has recently received much attention in tree-based ensemble learning. XGBoost helps prevent overfitting, thereby improving the generalization ability of the model, which means that it can be more easily used in clinical practice. The SHAP algorithm builds a visual interpretation model and outputs the ranking of important features to decrypt XGBoost (Figure 6). The biggest advantage of the SHAP value is that SHAP can reflect the influence of the characteristics in each sample, and it also shows the positive and negative nature of the influence. The larger the red dot, it represents positive effects, such as capsule invasion, extraglandular invasion, multifocality, and expression of PDLIM3, COL12A1, and TUBB2A. The increase in monocyte count increases the possibility of medium and high risk. High expression of DPP7, CTSL, and ITGB5, as well as increased neutrophil and lymphocyte counts increased the probability of low risk. These models allow doctors to intuitively understand the key characteristics, enabling doctors to effectively apply the models to individual patients, so as to achieve the purpose of hierarchical management and precise treatment.

Description of drawings

Figure 1 is the workflow diagram of PTC sample PCT-DIA;

Fig. 2 is a flow chart of feature selection, building a machine learning model and model verification;

Figure 3 shows the predictive performance of the machine learning model. The machine learning model assists clinicians and only clinicians in distinguishing low-risk and medium-high-risk PTC. Figure 3A is the ROC curve of the retrospective test set, and Figure 3B is the ROC curve of the prospective test set;

Figure 4 is a schematic diagram of proteins with significant expression differences between high-risk group and low-risk group (Figure A), high-risk group and intermediate-risk group (Figure B), and intermediate-risk group and low-risk group (Figure C), each point in the figure Indicates a differential protein, points outside the dotted line are considered to be statistically different. The dots in the upper right quadrant represent up-regulated proteins, and the dots in the upper left quadrant represent down-regulated proteins;

Figure 5 is a summary plot of the SHAP analysis showing the mean absolute SHAP values for the 17 most important features of the model. Feature importance for this model listed from top to bottom. The closer the feature is to the top, the greater the degree of sample differentiation, and the greater the impact of the feature on the output;

Figure 6 is a diagram of the expression levels of 6 proteins screened by machine learning in the high- and high-risk groups and low-risk groups. The upper and lower edges of the box plots are the quartiles of protein expression levels, and the middle line represents the median.

Detailed ways

1. Objects and methods

(1) Research object

1. General Information

The PTC sample set included in this study is divided into a discovery set and an independent validation set. Independent validation sets include retrospective validation sets and forward validation sets. The discovery set retrospectively included PTC paraffin section specimens and clinical information from the Oncology Department of Hangzhou First People's Hospital affiliated to Zhejiang University School of Medicine and Shandong Yuhuangding Hospital. The time span was from June 2013 to November 2020. A total of 283 cases were initially included. Nine cases were excluded due to small or insufficient samples, and finally 274 cases were included (191 training sets/83 testing sets). At the same time, collect 166 cases of PTC paraffin section specimens and clinical information from the above two units from January 2016 to December 2021 as a retrospective verification set, and collect 118 needle biopsy samples from the above two units from January 2020 to December 2021 as a prospective validation set.

Sample inclusion criteria: ①First operation with lymph node dissection; ②No history of chemotherapy and radiotherapy; ③Classic PTC diagnosed by postoperative pathological examination; ④Postoperative pathological diagnosis results included complete information on patient risk stratification. Exclusion criteria: ① History of neck trauma; ② Combination or previous suffering from other cancers; ③ Postoperative pathological diagnosis of other subtypes of PTC or other pathological types; ④ Lack of complete and available postoperative pathology.

2. Clinicopathological information

Extract patient clinical information and tumor characteristics from electronic medical records, including (1) preoperative clinical information: patient gender, age, presence or absence of Hashimoto’s thyroiditis, largest tumor diameter, whether the tumor is multifocal, whether the tumor invades the capsule or extraglandular . (2) Blood immune indicators: platelet count (platelet count, plt), neutrophil count (Neutrophil count, N), lymphocyte count (Lymphocyte count, L), macrophage count (Monocyte count, M), calculation Platelet-to-lymphocyte ratio (Platelet-lymphocyte ratio, PLR), neutrophil-lymphocyte ratio (Neutrophil-lymphocyte ratio, NLR), lymphocyte/macrophage ratio (Lymphocyte-to-monocyte ratio, LMR) and system Systemic immune-inflammatory index (SII). (3) BRAF mutation status.

The tumor staging of the included PTC patients was carried out according to the 8th edition of the TNM staging published by the American Joint Committee on Cancer (AJCC). The risk of PTC tumors refers to the latest 2015 version of ATA for PTC postoperative recurrence risk stratification.

ATA recurrence risk stratification system

Relapse Risk Stratification (Relapse Risk Scale) Meet the criteria Low risk (≤5%) ——PTC meets all of the following conditions without local or distant metastasis; all macroscopic tumors have been completely resected; the tumor has not invaded surrounding tissues; the tumor is of non-invasive histological subtype (such as high cell type, hobnail type, Columnar cell type); if I treatment has been performed, the whole body imaging after the first treatment does not show thyroid iodine-extroverted metastases; no vascular invasion; cN0 or pN1: ≤5 lymph node micrometastases (maximum diameter <0.2 cm) - FV-PTV confined to the thyroid without capsule invasion - well-differentiated FTC confined to the thyroid with capsule invasion, no or only a small amount (<4) of vascular invasion - localized In thyroid, unifocal or multifocal PTMC, with or without BRAF^V600E mutation Medium risk (6%-20%) -Microscopically shows that the tumor invades the soft tissue around the thyroid gland -Iodine-accepting metastases in the neck can be seen on the whole body scan after the first I treatment -Invasive histological subtype (such as high cell type, hobnail type, columnar cell type) - PTC with vascular invasion - cN1 or pN1: >5 lymph node metastases (all <3 cm in maximum diameter) - FV-PTV confined to the thyroid without capsule invasion - multifocal PTMC with extraglandular Invasion and BRAFV600E mutation High risk (>20%) ——The tumor can be seen to invade the soft tissue around the thyroid gland——The tumor cannot be completely resected——Distant metastasis——Postoperative serum TG indicates distant metastasis——pN1: the largest diameter of any metastatic lymph node ≥ 3cm——PTC with extensive blood vessels Infringement (> 4 places)

(2) Tissue sample preparation

Four paraffin thin sections with a thickness of 10 μm were continuously cut from FFPE thyroid cancer tissue blocks on a tissue slicer, and attached to glass slides. Each tissue sample was examined and prepared by two experienced pathologists, compared with hematoxylin-eosin stained pathological diagnostic sections after surgery, and tissue cores were taken after marking the lesion area of 10 μm paraffin thin section under a microscope. The needle biopsy samples came from the preoperative or intraoperative fine-needle aspiration tissue of thyroid nodules, and after being cleared by the pathologist, the needle biopsy samples were stored in a -80°C refrigerator.

(3) Batch design

In the discovery set, in order to minimize the batch effect caused by different experimental batches, 274 paraffin samples were randomly selected as 24 biological replicates and 27 samples as technical replicates, and were randomly assigned to 21 batches. A retrospective test set of 166 samples and 13 technical replicates was divided into 12 batches, and a prospective test set of 118 samples and 8 technical repeats was divided into 9 batches. Each batch includes 15 thyroid samples, one mouse liver sample and pooled thyroid samples with high, medium and low risk as quality control. In the analysis of this discovery set, the histopathological diagnosis of each paraffin section was unambiguous, and thus the data separation model was established.

(4) Dewaxing, rehydration and hydrolysis of FFPE tissue

For each PTC case in the discovery set, a total of 4 FFPE bioreplicate tissue cores were prepared. Samples were deparaffinized in heptane and subsequently hydrated sequentially in 100% ethanol (Sigma), 90% ethanol, and 75% ethanol at room temperature. It was then washed with 100 mM Tris-HCl (pH 10, Sigma) to establish alkaline hydrolysis conditions at 95°C. React at 95°C, 600rmp for 30min, and then quickly cool the sample to 4°C.

(5) Tissue Lysis, Protein Extraction and Protein Digestion

Add 6M urea, 2M thiourea, 10 mM tris(2-carboxyethyl)phosphine hydrochloride (TCEP) and 40 mM iodoacetamide (IAA) to the dewaxed sample, and then use pressure cycling technology (Pressure cycling technology, PCT ), 45000psi for 25s, 10s at 30°C for 90 cycles. Incubate for 30 min with micro-rotation in the dark, then add lysC at a ratio of 40:1 (protein to Lys C (lysc)). PCT-assisted lysC digestion was performed at the following settings: 45 cycles, 20000 psi, 50 s. 10s at 30°C under normal pressure. The final trypsinization was performed with PCT at a ratio of 40:1 (protein to trypsin) with the following settings: 90 cycles, 50 s, 20000 psi. Work under normal pressure at 30°C for 10s. Then add 10% TFA to stop the digestion, adjust the pH to 2-3, then centrifuge at 12000g for 5min. Measure the concentration after reconstitution and adjust to 0.2 ug/ul.

(6) Proteomics data collection and analysis

Clean peptides are separated by nanoLC-MS/MS system, equipped with 15cm*75μm ID column, fixed gradient is 45mins, 3-25% linear gradient (buffer A: 2% acetonitrile, 0.1% formic acid; buffer B: 98% acetonitrile ,0.1% formic acid), the flow rate is 300nL/min. The effluent passes through the QExactiveHF mass spectrometer. Data acquisition was performed in DIA mode. 390-1010 m/z were analyzed in Orbitrap at a resolution of 60,000 (m/z 200) using an AGC target of 3E6 charge and a maximum injection time of 100 ms. After the full MS scan, 24 MS/MS scans were acquired, each with 30,000 resolution (m/z 200), with an AGC target of 1E6 charges, a normalized collision energy of 27%, and a default charge state setting of 2 , the maximum injection time is set to auto. 24 MS/MS scans (center of isolation window) cycled over 3 wide isolation windows (m/z): 410, 430, 450, 470, 490, 510, 530, 550, 570, 590, 610, 630, 650, 670, 690, 710, 730, 770, 790, 820, 860, 910, 970. The entire MS and MS/MS scan acquisition cycle takes approximately 3 seconds and is repeated throughout the LC/MS analysis. The collected data was searched through DIA-NN (1.7.15) with the thyroid peptide spectrum library.

(7) Quality control

(8) Building a prediction model based on XGBoost

The present invention builds a prediction model based on XGBoost, so as to classify any given proteome data sample and clinical features into one of two categories: high-risk and low-risk, so as to achieve the best accuracy. This includes 4 stages: data preprocessing, feature selection, machine learning model building, and model validation.

The following are the detailed steps of the 4 stages (see Figure 3A and Figure 3B)

Phase 1: Data Preprocessing

A total of 3 groups from 2 datasets, which are discovery set, retrospective validation set, and forward validation set, are used to develop DNN models. Of the 274 samples from the discovery set queue, 191 samples are divided into training sets for model building, and 83 samples are divided into test sets for tuning parameters, so that the AUC results of the test set corresponding to the parameters are optimal, and then the trained The model is used to review the validation set, look forward to the data of the validation set queue, and perform external validation to demonstrate the generalization ability of the model of the present invention.

Preprocessing consists of two steps: (1) missing value imputation and (2) normalization. Missing values are inevitably a feature of protein intensity data. Imputation was done by filling all missing values with 0.8 Dmin, considering that most missing values occurred when the protein content was below the detection threshold. Among them, Dmin is the minimum value of all eigenvalues in the discovery set, and in this work, Dmin=13. Therefore, for each feature after the imputation step, estimate the mean and variance of that feature from the discovery set, and normalize the feature for each training sample according to the following.

Phase 2: Feature Selection

There are two reasons for the need for feature selection: (1) Since it is a full proteome detection, most of the detected proteins are not highly relevant to this problem. In addition, too many proteins for machine learning will reduce the generalization ability of the model, resulting in Overfitting, so they should be deleted from the feature matrix in machine learning; (2) Minimize the number of proteins in clinical practice applications, and select the optimal combination to achieve the most effective discrimination effect. It is done in two steps. The first step is feature screening. In the initial protein signature, differentially expressed proteins in the data set stratified by high-medium-low risk and proteins associated with the published literature related to thyroid or thyroid cancer. Of the 274 cases that did not appear in the data set of the present invention, they were excluded. Additionally, this protein was deleted if it was > 45% missing. If the absolute value of the Pearson correlation between a pair of proteins was less than 0.1, they were removed. In the second step, combinatorial optimization is performed to select the best combination of 10 proteins from the screened proteins. Although no algorithm can guarantee a globally efficient optimal solution, a machine learning algorithm is used here to find the best protein combination.

Evolution operations (crossover, mutation, and selection operations) are used to generate new protein feature combinations from existing protein feature combinations. In each iteration, the algorithm eliminates low-fitness combinations and generates new combinations from the remaining high-fitness combinations.

Phase 3: Model Training

The present invention designs a model based on a gradient boosting tree—extreme gradient boosting, which links input features with PTC risk stratification for solving supervised learning problems. Supervised learning is the process of using a set of samples of known categories to adjust the parameters of the classifier to achieve the required performance. In the present invention, a total of 15 possible clinical indicators related to preoperative PTC risk stratification are incorporated into the machine learning model of the present invention. The discovery set is used as a training set to train and adjust the parameters in the model, divide the discovery set into a training sequence and a test sequence, build a model in the training sequence, verify it in the test sequence, select and save the best AUC on the test set The model was then validated on two independent validation sets and the performance of the model was evaluated by the area under the curve (AUC) of the receiver operating characteristic curve. resulting in hierarchical results.

Phase 4: Model Validation

Put unknown samples into the established model. Based on the feature combination selected by the above model, the clinical information and proteomic information of the unknown sample are input, and the possible risk stratification of the patient can be obtained according to the obtained model. The flow chart of feature selection, machine learning model building and model verification is shown in Figure 2.

(9) Statistical analysis

Statistical analysis was performed using R software (version 3.5.1) with heatmap and graphing functions. CV was calculated as the ratio of the standard deviation to the mean. P-values were calculated by one-way ANOVA for expression of protein combination features. Select the volcano map to calculate the differential protein, the screening conditions of the differential protein: 1) Unpaired two side Welch’s t test p<0.05; 2) fold-change>1.2 or fold-change<-1.2. Analyze biological insights through Ingenuity pathway analysis (IPA). Use the Pearson correlation coefficient (Pearson correlationcoefficient) to assess the strength of the repeated data correlation. AUC was used to evaluate the average algorithm performance index. SHAP summary plots to illustrate the importance distribution of each variable.

2. Results

(1) General clinical features

A total of 558 PTC samples were included in the study, with an average age of 45.69 years, 397 female patients and 161 male patients, with a sex ratio of 2.47:1. The average tumor diameter was 13.01mm. There were 244 cases of PTC capsular invasion under ultrasound, accounting for 43.73%, 179 cases of extraglandular invasion under ultrasound, accounting for 32.08%, and 103 cases of multifocal ultrasound under ultrasound, accounting for 18.46%.

The clinicopathological data of 558 PTC patients are shown in Table 1.

Table 1. Clinicopathological data of 558 PTC patients

(2) Construction of proteomics database

In this study, a thyroid proteome database was constructed by PCT-MS to support the identification and quantification of proteins in papillary thyroid carcinoma by DIA-MS. After filtering the polluting proteins and repetitive proteins, the present invention finally identified and quantified 5774, 5025 and 6301 proteins for the three sets of original data of the discovery set, the retrospective verification set and the prospective verification set. Among the three sets of raw data, the constructed thyroid database contained 121,960 peptides and 9,941 proteomes. The resulting DIA dataset validated this library and applied it to the proteomic stratification of high-medium-low risk.

(3) Quality control

In order to ensure the stability and reliability of the subsequent proteomics data included in the model construction, the present invention sets strict quality control.

1. Pool sample analysis: Calculate the summary pool sample variation coefficient (Coefficient variation, CV) between the discovery set paraffin samples, review set paraffin samples and prospective set biopsy samples respectively, the medians are 0.019, 0.014 and 0.011 respectively, the results It suggested a low degree of variation (see Figure 4A for details).

2. Technical repeat analysis: Randomly select 27, 13 and 8 samples from the above three collections, repeat the mass spectrometry analysis once, and compare the mass spectrometry data of the two times. The calculation method is pearson correlation, and the correlation coefficients are all above 0.95. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation is (see Figure 4B).

3. Biological repeat analysis: one sample was randomly selected for each batch in the discovery set, and a total of 22 samples were used as biological repeats for mass spectrometry analysis. The correlation coefficient was above 0.9, suggesting a strong correlation, and there was no obvious batch effect in the discovery set (details See Figure 4C). This means that biases from biology and machines are negligible.

(4) Differential proteins in different risk groups

After the quality control is completed, a complete and stable proteome can be obtained, and proteins with differences can be found by comparing the expression levels among the three groups of high, middle and low risk at all protein levels. In the further analysis of the proteins in the high, middle and low risk groups, as shown in the volcano diagram in Figure 5, Figure A: Compared with high-risk and low-risk samples, the overall protein expression levels are similar, 97 protein expression levels are up-regulated, and 71 protein expression levels are down-regulated. Panel B: The overall protein level of high-risk samples is similar to that of medium-risk samples, with 44 proteins up-regulated and 44 proteins down-regulated. Panel C: The overall protein levels of intermediate-risk samples are similar to those of low-risk samples, with 49 protein expression levels up-regulated and 26 protein expression levels down-regulated.

(5) Machine learning framework prediction model

In order to distinguish PTCs with different risk stratifications by proteomics approach, the present invention established an artificial neural network algorithm based on a feature selection process. The study design and workflow are shown in Figure 2. Modeling includes four parts: data preprocessing, feature selection, deep learning model training and prediction. The machine learning model was built by cross-training and validation based on the discovery set as described previously. Based on the protein features selected by the XGBoost algorithm combined with 10-fold cross-validation, the present invention finally determined 6 feature proteins and 11 clinical features as the features of the model.

The proteomic data and the above detailed clinical and genetic data were used to generate a machine learning model to predict the risk level of PTC samples. Using our training set, Student's t-test and fold change (FC) value calculations were performed for each of the two risk levels and for each protein feature to identify the protein feature that best discriminated PTC samples at different risk levels. Proteins with a P value of 0.05 and |log2(FC)|>0.25 were selected. Proteins with deletion rate ≥ 0.5 were further eliminated. Based on these criteria, we selected 6 proteins. These protein features were normalized between 0 and 1; their missing values were set to 0. We use the same features and perform the same normalization on both test sets. Set the undetectable features in the test dataset to 0 vectors.

Before training the model, the 274 PTC samples from the discovery set were divided into a training set (n = 191) and an internal validation set (n = 83). Then, the training set is used to develop the model, and the internal validation set is used to validate and optimize the model's performance. We design a machine learning architecture including feature selection and risk level classification.

The core part of the algorithm is a cascaded two-step feature selection step that allows selection of protein features and other multi-omic features. First, the parameters and features are optimized using a grid search algorithm; then, given a set of parameters, the XGBoost model is constructed using the protein matrix and the importance of all protein features is ranked. Using this tool, we selected the top 6 protein signatures; protein numbers were selected as average among clinical (7), immunomics (8) and genetic signatures (1). Then, we combine all 22 features to build another XGBoost model with the previous parameters to get feature importances. Finally, the validation set is used to select the top k features with the best area under the curve (AUC) value. This pipeline yielded the following algorithm.

Algorithm 1

Input: protein matrix P; other feature matrix Q; grid space G

Best_AUC = 0

For the grid in G:

Model1 = XGBoost (P, grid)

Importance1 = sort (Model1.importance)

P_selected = P[:,Importance1[:6]]

Multi_omics_features = [Q,P_selected]

Model2 = XGBoost (Multi_omics_features, grid)

Importance2 = sort (Model2.importance)

For num in range (num of Multi_omics_features):

Final_features = Multi_omics_features[:,Importance2[:num]]

Model3 = XGBoost (Final_features, tree_num=20)

Pred_score = Model3(validation_data[:,Final_features])

Temp_AUC = AUC(label,Pred_score)

If Temp_AUC>Best_AUC:

Best_parameter = [grid, num]

Best_features=Final_features

return Best_parameter, Best_features

Output: Model_parameter= Best_parameter, Model_features=Best_features

We generated four models with increasing feature dimensions. Model A used only one dimension of features: preoperative clinical data (n = 7). Model B adds a second dimension to the previous model: immunological indicators (n=8). Model C has added genetic features on the basis of the previous ones. Finally, Model D was combined with the previous proteomic features (n = 6), thus providing the final 4D feature-based model. Finally, we compare the predictive performance of the 4 models.

Clinically, it is more desirable to effectively distinguish low-risk and medium-high risk cases, so the present invention uses this as the outcome index for further analysis. The present invention compares the predictive performance of different feature combinations on the results (see Table 2 for details).

1. In the initial model construction, Model A, which only relied on clinical data (capsular invasion, extraglandular invasion, tumor diameter, multifocality, and age), failed to show good results. In the discovery set, retrospective validation set and The AUCs in the prospective validation set were 0.86, 0.68, 0.71, respectively.

2. Secondly, the present invention incorporates the combination of clinical + immunization model B, which has significantly improved the verification results of each set, and the AUC has reached 0.94, 0.65, and 0.71 respectively.

3. In the model C built with the combination of clinical + immune + BRAF, it was found that the verification effect of the discovery set was significantly improved, and the AUC reached 0.93, 0.68, and 0.73.

4. Finally, the present invention incorporates protein indicators, and the model D of clinical + immune + BRAF + protein is used as the prediction model, and the AUCs in the three sets reach 0.92, 0.79, and 0.80, respectively.

Table 2 AUC of each combined model in the discovery set, retrospective validation set and forward validation set

Inclusion indicators discovery set Review the validation set forward validation set Model A: Clinical 0.86 0.68 0.71 Model B: clinical + immunization 0.94 0.65 0.71 Model C: clinical + immune + BRAF 0.93 0.68 0.73 Model D: clinical + immune + BRAF + protein 0.92 0.79 0.80

(6) Final feature selection and auxiliary risk prediction

Output machine learning to filter the above models to obtain characteristic modeling indicators, characteristic results:

Based on the characteristics of the four dimensions of "clinical + immune + BRAF + protein", the present invention constructs the XGBoost verification model, and the present invention uses "feature selection" to output feature indicators: clinical factors (capsule invasion, extraglandular invasion, tumor diameter, multifocal gender and age), blood immune indicators (preoperative monocyte count (M), preoperative neutrophil count (N), preoperative lymphocyte-to-monocyte ratio (LMR), preoperative platelet count (Plt) and preoperative lymphocyte count (L)), proteomics (DPP7, DPLIM3, COL12A1, CTSL, TUBB2A and ITGB5) and genomics (BRAF gene mutation).

In order to demonstrate the influence of each feature in the above model, the present invention constructs a visual explanation model through the Shapley additive explanations (SHAP) algorithm. For each prediction sample, the model generates a prediction value, and the SHAP value is the value assigned to each feature in the sample (Figure 6), and it is also an array of feature importance. The biggest advantage of the SHAP value is that SHAP can reflect the influence of the characteristics in each sample, and it also shows the positive and negative nature of the influence.

The above XGBoost model outputs a score for each detected patient: if the score is less than 0.6645, the machine learning model considers it low-risk PTC; if the score is greater than 0.6645, the machine learning model considers it medium/high-risk PTC. At the same time, a preliminary independent analysis was performed by 9 clinicians with different qualifications. The analysis results of the clinician are compared with the results of the XGBoost model to assist the doctor to make the final analysis results.

To evaluate the effect of the machine learning model, we compared the effect of the XGBoost model constructed from four-dimensional features and preoperative risk stratification among clinicians of different seniority levels (Fig. 4A-B). Predictive performance results gradually improved with increasing levels of clinician seniority. In the retrospective test set, the accuracy of XGBoost (0.737 [95% CI 0.668-0.802]) was significantly higher than that of only senior clinicians (0.598 [95% CI 0.584-0.606], P<0.0001). In the prospective test set, the accuracy of XGBoost (0.736 [95% CI 0.717-0.740]) was also significantly higher than that of only senior clinicians (0.579 [95% CI 0.573-0.593], P<0.0001) . These results suggest that the XGBoost model has a stronger predictive power than clinicians.

The present invention further explores the use of machine learning to assist the predictive power of clinicians. The results showed that with the aid of the machine learning model, the predictive power of clinicians gradually increased with the level of seniority. In the retrospective test set, the accuracy of the XGBoost model was 0.737 (95% CI 0.668-0.802), while the prediction accuracy of senior clinicians assisted by machine learning improved to 0.844 (95% CI 0.829-0.844) ( P<0.0001; Figure 4A). In the prospective test set, the accuracy of the XGBoost model was 0.736 (95% CI 0.717-0.740), and the prediction accuracy of senior clinicians assisted by machine learning was increased to 0.835 (95% CI 0.822-0.842) (P< 0.0001; Figure 4B). These results suggest that machine learning-assisted clinician predictive performance is superior to machine learning predictive performance.

Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention, those skilled in the art can make the above-mentioned The embodiments are subject to changes, modifications, substitutions and variations.

Claims (9)

1. An auxiliary risk stratification system for papillary thyroid carcinoma (PTC), characterized in that it comprises: Data collection module: used to collect existing data, including proteomics + clinical indicators + blood immune indicators + BRAF ^V600E mutation label, after verifying and evaluating data integrity and data quality, data desensitization is performed and sent to the data pre-processing processing module; Data preprocessing module: perform missing value interpolation and normalization processing on the collected existing data; Model building, training and verification module: feature screening based on the whole proteome detection data obtained by the data preprocessing module, and then based on the gradient boosting tree model-extreme gradient boosting, linking the input features with the PTC risk stratification for solving Supervised learning problem; the discovery set is used as the training set to train and adjust the parameters in the model, the discovery set is divided into training sequence and test sequence, the model is built on the training sequence, internal validation is performed on the test sequence, and the test set is selected and saved The model with the best AUC is then validated on two independent validation sets and the performance of the model is evaluated through the AUC of the receiver operating characteristic curve; the protein features selected based on the XGBoost algorithm combined with 10-fold cross-validation determine that 6 characteristic proteins, 5 clinical characteristics, 5 blood immune indicators and BRAF ^V600E mutation characteristics are used as the characteristics of the model construction; the risk stratification model of papillary thyroid carcinoma (PTC) is established, the threshold value of the model is determined, and risk lesions at all levels can be predicted and stratified according to the value; Stratification module: analyze the test samples, stratify tumors into low-risk lesions or medium-high-risk lesions, and feed back to the doctor's terminal equipment and database platform; Database platform: connected with the above modules, used for model building data storage and retrieval, and continuously collects new data to provide online model update function. 2. The auxiliary stratification system according to claim 1, wherein the six characteristic proteins are: DPP7, DPLIM3, COL12A1, CTSL, TUBB2A and ITGB5. 3. The auxiliary stratification system according to claim 1, wherein the five clinical features are: capsule invasion, extraglandular invasion, tumor diameter, multifocality and age. 4. The auxiliary stratification system according to claim 1, characterized in that: said 5 blood immune indexes are: platelet count (Platelet count, plt), neutrophil count (Neutrophil count, N), lymphocyte Count (Lymphocyte count, L), monocyte count (Monocyte count, M) and lymphocyte/monocyte ratio (Lymphocyte-to-Monocyte Ratio, LMR). 5. The auxiliary stratification system according to claim 1, wherein the sources of the existing data include: hospital case data. 6. The auxiliary hierarchical system according to claim 1, characterized in that the supervised learning is a process of using a set of samples of known categories to adjust the parameters of the classifier to achieve the required performance. 7. The auxiliary stratification system according to claim 1, characterized in that, the sample inclusion criteria are as follows: ① initial operation and lymph node dissection; ② no history of chemotherapy and radiotherapy; ③ classic PTC diagnosed by postoperative pathological examination; ④The results of postoperative pathological diagnosis contain complete information about the risk stratification of patients; exclusion criteria: ① History of neck trauma; ② Combination or previous suffering from other cancers; ③ Postoperative pathological diagnosis of other subtypes of PTC or other pathological types; ④ Lack of complete and available postoperative pathology. 8. A method for using a non-diagnostic papillary thyroid carcinoma (PTC) risk auxiliary stratification system, characterized in that the stratification is obtained based on the auxiliary stratification system according to any one of claims 1-7. result. 9. Use of the auxiliary papillary thyroid carcinoma (PTC) risk stratification system according to any one of claims 1-7 in preparing a device for predicting and stratifying the risk of papillary thyroid carcinoma in patients.

Images