Comprehensive application of AI algorithms with TCR NGS data for Glioma diagnosis

doi:10.21203/rs.3.rs-3596333/v1

Download PDF

Article

Comprehensive application of AI algorithms with TCR NGS data for Glioma diagnosis

https://doi.org/10.21203/rs.3.rs-3596333/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

TCR detection can examine the extent of T-cell immune responses. Therefore, the article analyzed characteristic data of glioma obtained by high-throughput sequencing, to predict the disease with fewer biomarkers and higher accuracy. We downloaded data online and obtained six TCR-related diversity indices to establish a multidimensional classification system. By comparing actual presence of the 602 correlated sequences, we obtained two-dimensional and multidimensional datasets. Multiple classification methods were utilized for both datasets with the classification accuracy of multidimensional data slightly less to two-dimensional datasets. This study reduced the TCR β sequences through dimensionality reduction methods like RFECV. Consequently, using only the presence of these three sequences, the classification AUC value of 96.67% can be achieved. The combination of the three correlated TCR clones obtained at a source data threshold of 0.1 was: CASSLGGNTEAFF_TRBV12_TRBJ1-1, CASSYSDTGELFF_TRBV6_TRBJ2-2, and CASSLTGNTEAFF_TRBV12_TRBJ1-1. At 0.001, the combination was: CASSLGETQYF_TRBV12_TRBJ2-5, CASSLGGNQPQHF_TRBV12_TRBJ1-5, and CASSLSGNTIYF_TRBV12_TRBJ1-3. This method can serve as a potential diagnostic and therapeutic tool, facilitating diagnosis and treatment of glioma and other cancers.

Biological sciences/Cancer/Cns cancer

Biological sciences/Cancer/Cancer genomics

Biological sciences/Cancer/Tumour biomarkers

Health sciences/Diseases/Cancer/Cancer genomics

Health sciences/Diseases/Cancer/Cns cancer

Health sciences/Diseases/Cancer/Tumour biomarkers

glioma

T-cell repertoire(TCR)

prognosis

Artificial Intelligence Techniques

feature selection methods

In recent years, the application of AI algorithms in medicine has become more widespread and intensive, with researchers gradually applying the technology to challenging processes such as cancer diagnosis, treatment, and prognosis. Cancer data encompasses diverse modalities, including medical imaging for visual information and molecular data obtained from biopsies of solid tumors and liquid samples. As a result, numerous studies have focused on analyzing these different modalities using corresponding machine learning algorithms to advance cancer diagnosis, monitoring, and treatment [11].

While imaging studies have greatly impacted current cancer diagnosis and treatment, the rapid progress in sample processing, genome sequencing, and molecular technologies has highlighted the importance of molecular data analysis in cancer research. Molecular data, particularly from solid tumor biopsies and molecular liquid biopsies, contain fine-grained features that are not available in imaging data. This has led to an increasing focus on molecular data analysis in cancer research [11]. For instance, according to [7], the diagnosis of hematologic malignancies can be classified using various blood cell parameters extracted from Cell Population Data (CPD), combined with machine learning classification algorithms. The study found that the artificial neural network (ANN) algorithm achieved the highest AUC of approximately 93.5%.

TCR NGS data is a type of molecular data that holds significance in cancer research. TCR, short for T-cell receptor, is a receptor found on the surface of T lymphocytes that plays a crucial role in recognizing and binding antigens, triggering and regulating the immune response. TCR chains consist of variable and constant regions, with the variable region containing complementarity determining regions (CDRs) responsible for TCR specificity and antigen recognition. The constant region provides structural support and signaling functions. The diversity and assortment of TCRs primarily stem from the highly diverse Complementarity Determining Region 3 (CDR3), which is formed through rearrangement and alteration of V(D)J segments. These V(D)J regions, found within the genes responsible for coding TCRs, contribute to the specificity and heterogeneity of TCRs.

Several studies suggest that TCR diversity plays a crucial role in the diagnosis and treatment of cancer. For instance, a study by Yang-Yang Liu and colleagues observed significant differences in the TCR repertoire between lung cancer patients and healthy individuals, including differences in CDR3 clonotype, diversity, V/J segment utilization, and sequence [9]. Additionally, Huaichao Luo et al. [8] constructed a machine learning classification model for diagnosing pulmonary nodules based on features of peripheral blood T-cell receptor repertoire, such as the Shannon Index, combined with age and corresponding data on benign and malignant pulmonary nodules. The model achieved an AUC value of 0.8, indicating that TCR diversity indices contain relevant information for pulmonary nodules or lung cancer. In this study, we plan to analyze the importance of the T-cell diversity indices to the distinction of glioma patients and healthy individuals.

In addition to TCR diversity indices, several studies have demonstrated the accurate use of TCR sequence distribution and presence data for disease diagnosis and analysis [12, 13]. This forms the focus of our paper.

Gliomas, including glioblastoma, are tumors that originate from precursor or glial cells [1]. According to the CBTRUS Statistical Report, the incidence rate of glioblastoma is reported as high as 3.19 cases per 100,000, but the five-year survival rate for glioblastoma is significantly low at 4.67%. These statistics highlight the prevalence of glioblastoma and theneed for effective diagnostic approaches and emphasize the urgency to improve early detection and accurate diagnosis of glioma[1].

Classification effects on TCR Diversity Indices

To assess the importance of six indicators, we employed four dimensionality reduction methods (Permutation, GBDT, RF, and XGBoost) to rank the importance of this dataset. The corresponding bar graphs are illustrated in Figure 1A to Figure 1D. Figure 1E and Figure 1F represent the sum of importance and sum of importance rank derived from the four dimensionality reduction methods respectively. It is evident that the VJ diversity index (Hvj.Index) holds the highest importance, followed by the proportion of monoclonal T cells (Singleton) and Simpson's diversity index (Invsimpson.Index). Subsequently, Figure 1I to Figure 1K display box plots for these three indicators corresponding to positive-negative samples, confirming their distinct differences.

Furthermore, Figure 1G showcases a correlation coefficient matrix between various indicators and phenotypes. It reveals that the three indicators with higher importance exhibit stronger correlations with phenotypes. Conversely, the indicator with the lowest importance, the Clonality index, demonstrates the weakest correlation with phenotypes.

Figure 1H illustrates the comprehensive ROC curves derived from 11 classification algorithms based on the datasets of TCR Diversity Indices and phenotype of the samples. It can be observed that the Random Forest, XGBoost, and Bagging classifiers all achieve a 100% classification AUC value. The box plots in Figure 1I to Figure 1K provide further evidence supporting the aforementioned statement.

Binary classification algorithms

In Figure 2A, the correlation coefficient plots illustrate the strong correlation between the two-dimensional features and phenotype factors. Specifically, the number of associated clonetypes shows significant relevance to the phenotype, particularly in relation to glioma. This is further confirmed by the box graphs, which reveal a higher number of associated clonetypes in glioma patients compared to healthy individuals. Based on these findings, we construct classifiers using the two-dimensional extracted features from each sample. The data is analyzed using 8 classical classifiers, 4 integrated learning algorithms, and 2 neural network algorithms. The resulting ROC curves are combined in Figure 2D, demonstrating that these algorithms achieve excellent classification performance in the two-dimensional data. Some algorithms even achieve an AUC of 1, indicating perfect classification. This highlights the effectiveness of the two-dimensional feature extraction method for diagnosing glioma. In Figure 2E, F, G, H, I, J, we present the decision boundaries corresponding to several algorithms that exhibit good classification results.

Multidimensional classification algorithms

In Figure 3, Figure 3A presents a Venn diagram depicting the intersection of sequence species between the healthy and glioma patient groups. It is evident that the sequences in the two groups exhibit high divergence, with only a small fraction being common to both.

The linear discriminant method achieved the highest accuracy, with an AUC of 100% (Figure 3B: AUC obtained from 602 sequences with cutoff1 data). In this study, we obtained P-values using Fisher's exact test and selected TCRβ sequences with P-values less than or equal to 10^-2, 10^-3, and 10^-4, resulting in 39, 12, and 1 relevant sequences, respectively.

As depicted in Figure 3C, each column represents binary data (0 or 1) indicating the presence or absence of specific TCRβ sequences in the corresponding samples. We hypothesized that the presence or absence of these sequences is closely associated with glioma status. By analyzing the sequences, we identified a set of TCRβ sequences directly related to glioma and utilized a small number of sequences to accurately predict the disease status.

Next, we applied these sets of sequence data to the 12 classification algorithms mentioned earlier. Notably, when selecting a P-value of 10^-3, several classification algorithms also achieved an AUC of 1. Consequently, we narrowed down the factors to these 12 sequence factors. Figure 3D displays the correlation coefficient graphs, indicating that the phenotypes exhibit a relatively high correlation with these sequences, while the correlation with quality control is generally lower.

By utilizing only these 12 TCRβ sequences, we achieved a high prediction accuracy rate. The specific sequences are as follows: CASSLGGNQPQHF_TRBV12_TRBJ1-5, CASSLRETQYF_TRBV12_TRBJ2-5, CASLSGNTIYF_TRBV12_TRBJ1-3, CASLTGNTEAFF_TRBV12_TRBJ1-1, CASSFSGNTIYF_TRBV12_TRBJ1-3, CASSLGYEQYF_TRBV12_TRBJ2-7, CASFGGNTEAFF_TRBV12_TRBJ1-1, CASQGYEQYF_TRBV3_TRBJ2-7, CASLGETQYF_TRBV12_TRBJ2-5, CASSLQETQYF_TRBV12_TRBJ2-5, CASLGGGYTF_TRBV12_TRBJ1-2, CASLGGNTEAFF_TRBV12_TRBJ1-1. Figure 3E presents a bar chart illustrating the importance corresponding to the P-values of these 12 sequences.

Dimension reduction and Core features’ extraction

Although the two-dimensional features extracted from Fisher's exact test in Figure 2 show promising results in distinguishing healthy individuals from glioma patients, we aim to explore additional approaches to identify disease-associated sequences for easier disease diagnosis and future drug development. In the following analysis, we focus on the screening of glioma-associated sequences.

Figure 4A and Figure 4B illustrate the importance distribution of Lasso with threshold values of 0.1 and 0.001, respectively. We determined that a threshold value of 0.1 yielded an appropriate selection of 19 dimensions for dimensionality reduction, while the thresholds of 0.01 and 0.001 resulted in 11 and 9 dimensions, respectively. There are still too many sequences in relative terms.

While when it comes to the method of RFECV algorithm, as is shown in the Figure 4C, for the 602 sequences with a threshold of 0.1 for feature extraction, the feature extraction algorithm employed is Adaboost, and the classification algorithm used is SVM. This approach yields a feature dimension of 5, achieving an AUC of 0.967 and an accuracy of 0.967. The five selected sequences are: CASSLGGNTEAFF_TRBV12_TRBJ1-1, CASSYSDTGELFF_TRBV6_TRBJ2-2, CASSLTGNTEAFF_TRBV12_TRBJ1-1, CASSQGYEQYF_TRBV3_TRBJ2-7, CASSLSGNTIYF_TRBV12_TRBJ1-3. Among these five sequences, further extraction using the reject-by-exclusion method identifies CASSLSGNTIYF_TRBV12_TRBJ1-3 and CASSQGYEQYF_TRBV3_TRBJ2-7, while the remaining three sequences, CASSLGGNTEAFF_TRBV12_TRBJ1-1, CASSYSDTGELFF_TRBV6_TRBJ2-2, and CASSLTGNTEAFF_TRBV12_TRBJ1-1, contribute to discriminating the prevalence of glioma. With these three core sequences, high AUC and low correlation degree can be obtained in the Figure 4E and 4G.

Using the same method, we analyze the dataset with a threshold of 0.001, which consists of a combination of 12 sequences. As is shown in Figure 4D, the feature extraction algorithm employed is Adaboost, and the classification algorithm used is CV (CalibratedClassifierCV). This analysis also yields five sequences: CASSLGETQYF_TRBV12_TRBJ2-5, CASSLGGNQPQHF_TRBV12_TRBJ1-5, CASSLSGNTIYF_TRBV12_TRBJ1-3, CASSLQETQYF_TRBV12_TRBJ2-5, CASSLRETQYF_TRBV12_TRBJ2-5. With these five sequences, we achieve an AUC of 0.967 and accuracy, with a sensitivity of 0.933 and specificity of 1. Similarly, by screening one by one, we identify a combination of three sequences that also ensure high AUC values: CASSLGETQYF_TRBV12_TRBJ2-5, CASSLGGNQPQHF_TRBV12_TRBJ1-5, and CASSLSGNTIYF_TRBV12_TRBJ1-3. Meanwhile, Figure 4F and Figure 4H demonstrate that the combination of these three less correlated sequences can classify and achieve high classification AUC values. Lastly and not least, Figure 4I is the seqlogo picture of the core three sequences at cutoff 1 data.

As we can see from the classification system constructed based on the diversity indices and the corresponding phenotype of the patients, especially when the classification algorithm is Random Forest, the classification effect is really remarkable, and it can reach 100% differentiation. However, when we consider that the immune system of patients with major diseases is very different from that of healthy people, the AUC value of 1 is meaningless. A large difference in the diversity index can only indicate that the subject may be suffering from a certain disease, but it cannot pinpoint exactly what disease there is. Therefore, this is the significance of the need to target specific TCR sequences associated with gliomas.

By the principle of Fisher's exact test that the lower the P-value the stronger the causality, we can roughly filter out the sequences with strong causality with gliomas. The number of kinds of related sequences at different thresholds reveals a strong correlation with the glioma phenotype in the Fig. 2.

In addition, the multidimensional classification results of Fig. 3 can also enable it to show us that the 12 relevant sequences screened by Fisher's exact test are sufficient for the diagnosis of gliomas and can achieve a high classification accuracy. These 12 sequences are:CASSLGGNQPQHF_TRBV12_TRBJ1-5, CASSLRETQYF_TRBV12_TRBJ2-5, CASLSGNTIYF_TRBV12_TRBJ1-3, CASLTGNTEAFF_TRBV12_TRBJ1-1, CASSFSGNTIYF_TRBV12_TRBJ1-3, CASSLGYEQYF_TRBV12_TRBJ2-7, CASFGGNTEAFF_TRBV12_TRBJ1-1, CASQGYEQYF_TRBV3_TRBJ2-7, CASLGETQYF_TRBV12_TRBJ2-5, CASSLQETQYF_TRBV12_TRBJ2-5, CASLGGGYTF_TRBV12_TRBJ1-2, CASLGGNTEAFF_TRBV12_TRBJ1-1. This result proves the validity of the commonly used Fisher's exact test.

However, because the thresholds to be set by the method are continuous, an infinite number of thresholds can be selected. For example, there are 12 sequences corresponding to a threshold value of 10^-3, while there is only 1 sequence corresponding to a threshold value of 10^-4. The exact number of sequences to which it is optimal to downsize to and the specific threshold between the two thresholds can not be decided right away, and thus the method still has some limitations.

The results indicate that achieving an AUC of 1 is more easily attainable in two-dimensional classification compared to multi-dimensional classification. However, the overall difference between the two approaches is not significant.

The results clearly demonstrate that for glioma diagnosis, higher accuracy can be achieved by utilizing data on the presence of multiple sets of 3D TCR β sequences. Regardless of the specific combination of sequences within each group, all combinations yield an AUC and accuracy of 0.967, with a sensitivity of 0.933 and specificity of 1.

The first combination consists of CASSLGNTEAFF_TRBV12_TRBJ1-1, CASSLTGNTEAFF_TRBV12_TRBJ1-1, and CASSYSDTGELFF_TRBV6_TRBJ2-2. The second combination includes CASSLGETQYF_TRBV12_TRBJ2-5, CASSLGGNQPQHF_TRBV12_TRBJ1-5, and CASSLSGNTIYF_TRBV12_TRBJ1-3.

Our research has demonstrated the practicality of this methodology not only for glioma diagnosis but also for the diagnosis of other viruses such as CMV and Neocon, as well as various types of cancers. This methodology holds great promise for future cancer diagnosis, and we encourage other scientists to apply this approach to the diagnosis of different diseases. Our future work will focus on expanding the application of this methodology to further advance disease diagnosis.

In summary, this study analyzes high-throughput sequencing data from glioma patients and healthy individuals. We explore multidimensional classification and dimensionality reduction of TCR sequence diversity indices, as well as two-dimensional classification and dimensionality reduction analysis of TCR-related sequences. As a result, we identify two sets of core sequences from these analyses. These two sets, each comprising three sequences, are sufficient to achieve a diagnostic accuracy of 96.7% in glioma detection, as evidenced by the AUC values. However, the reproducibility of these two sets of sequences in glioma patient data from different experimental contexts remains to be further investigated and validated.

Data acquisition and Data management

NGS Data

The proposed data can be divided into two groups: 44 training samples and 30 testing samples, with an equal split between positive and negative cases. The data of 22 positive glioma patients in the training set and 15 positive glioma patients in the test set were obtained from 2 researches [2, 3]. Additionally, a total of 37 healthy human negative samples from the training and test samples were obtained from IMMUNOSEQ ANALYZER (https://clients.adaptivebiotech.com/).

The raw data from the training samples underwent high-throughput TCR sequencing analysis, resulting in a total of 1,527,623 TCR β sequences. We obtained information about the presence of each sequence in the samples. However, due to the enormous size of the data, it is challenging to incorporate it directly into the classification model. Therefore, we performed dimensionality reduction using Fisher's exact test as we described previously[13]. We screened 602 TCR β sequences with a p-value less than 0.1, which exhibited a higher correlation with glioma diagnosis, to be used as features for determining glioma.

Machine learning algorithms with different datasets

Classification algorithms on TCR Diversity Indices

For these 74 samples, we analyzed TCR diversity, clonality and singleton frequency. According to the common bioindicator system, we can obtain each common immunome sequencing result index corresponding to positive negative samples, such as Clonality index, Shannon diversity index, inverse Simpson diversity index, VJ diversity index, Singleton ratio, and Simpson diversity index these 6 indices.

Of the above indices, the Clonality index is used to measure the extent of TCR clone amplification and to assess the number and proportion of major clones in tumors, with higher values indicating a higher frequency of amplification of some TCR clones. Clonality takes values between 0 and 1, with a value of 0 indicating that each individual is unique and no clone is amplified; a value of 1 indicating that all individuals are copies of the same clone. The index is calculated as follows, where productive unique represent the total number of TCR clones in the sample and diversity is the result of the Shannon diversity index^[4].

$$\text{C}\text{l}\text{o}\text{n}\text{a}\text{l}\text{i}\text{t}\text{y} = 1-\frac{\text{d}\text{i}\text{v}\text{e}\text{r}\text{s}\text{i}\text{t}\text{y}}{\text{l}\text{n}\left(\text{p}\text{r}\text{o}\text{d}\text{u}\text{c}\text{t}\text{i}\text{v}\text{e} \text{u}\text{n}\text{i}\text{q}\text{u}\text{e}\text{s}\right)}$$

The Hvj diversity index is used to measure the diversity of V-J gene combinations of TCR clones in the sample. The index measures the frequency of the ith V-J gene fragment combination use type in the sample by the ratio of pi, thus reflecting the degree of T cell expansion. This index can also be used to reflect the replicative proliferation capacity following stress generated by T-specific recognition of antigen, with higher values indicating a greater degree of clonal amplification, which is calculated as follows ^[5]:

$${H}^{{\prime }}=- \sum {\text{p}}_{\text{i}}\text{l}\text{n}\left({\text{p}}_{\text{i}}\right)$$

The Shannon Index of diversity is used to measure the diversity of TCR clones in the sample. This index incorporates information regarding species richness (the number of different clone types present) and the evenness of the distribution of individuals among these clone types. A higher Shannon index signifies a greater diversity of TCR clones in the sample. The pi in the formula represents the frequency of the ith specific clone type. Thus, the Shannon index considers the number of TCR clone types and the frequency evenness of each clone ^[6].

$${H}^{{\prime }}=- \sum {\text{p}}_{\text{i}}\text{l}\text{n}\left({\text{p}}_{\text{i}}\right)$$

The Singleton Frequency is used to represent the proportion of Naïve T cells frequency in a sample and is a metric used to assess the degree of backbone of immune system[14]. In TCR studies, several studies have shown that the number of singleton is positively correlated with the body's ability to resist unfamiliar diseases, so this indicator can also reflect the degree of the body's immune response to the disease. The calculation formula is as follows, where n_1 indicates the number of clone types that appear only once in the sample and n indicates the number of all clone types in the sample.

$$\text{S}\text{i}\text{n}\text{g}\text{l}\text{e}\text{t}\text{o}\text{n} = \text{n}\_1/(\text{n}-1)$$

Binary classification algorithms

According to our previous CMV two-dimensional classification paper, it has been demonstrated that extracting two-dimensional features from high-throughput sequencing results can improve diagnostic outcomes [13]. In the context of glioma, the two-dimensional data features extracted from high-throughput sequencing results correspond to a specific threshold of Fisher's exact test p-value. These features include the total clonetypes and associated clonetypes. The total clonetypes represents the total number of distinct TCRβ sequence species observed in each sample, while the associated clonetypes denotes the number of repeated species with associated TCR sequences.

With the two-dimensional data above, we built up the classification system with 16 algorithms including neural network algorithms, integrated learning algorithms and traditional machine learning algorithms.

Multidimensional classification algorithms

The multidimensional dataset of TCR β sequences uses 0–1 format data to represent the presence of each sequence in each sample. If the data equals to 0, the sequence is absent in the subject, else the sequence exists in the sample.

Multidimensional classification algorithms were applied to analyze the tabular data frames of glioma, which contained information on the presence or absence of relevant TCRβs. Several classification algorithms were used, including random forest, support vector machine with linear and polynomial kernel functions, logistic regression, linear discriminant analysis, Bayesian algorithm, GBDT, XGB, and decision tree algorithm.

To identify marker sequences specific to glioma while using fewer markers and maintaining high accuracy, dimensionality reduction techniques were employed. Fisher's exact test, a commonly used method in bioinformatics, was utilized for dimensionality reduction. By applying different cutoffs for Fisher's exact test, groups of sequences with varying degrees of relevance to gliomas were identified. These sequences were then incorporated into the classification algorithm system to achieve higher classification accuracy using a reduced number of sequences. This approach allows for the identification of marker sequences specifically associated with glioma.

Dimension reduction and Core features’ extraction

Due to computational limitations, we did not perform differential gene analysis on the initial millions of sequences. Instead, we focused on 602 sequences corresponding to a Fisher's exact test threshold of 0.1. Based on this subset of sequences, we explored feature extraction methods that outperformed the Fisher's exact test and identified TCRβ sequences that were more significant for glioma diagnosis. Apart from the Fisher's exact test, this study primarily utilized two algorithms for feature extraction: Lasso and RFECV.

Lasso (Least Absolute Shrinkage and Selection Operator) is a regularization method commonly used for linear regression. It can be applied for both feature extraction, which involves generating new features from the original data, and feature selection, which involves identifying important features to retain. Lasso achieves feature selection by introducing an L1 regularization term to the loss function, allowing it to reduce the coefficients of unimportant features to zero. Unlike traditional methods, Lasso considers the collective effect of all features rather than evaluating the importance of each feature individually. In this study, we employed the Lasso algorithm to select features from the multidimensional data obtained through Fisher's exact test using threshold values of 0.1, 0.01, and 0.001.

RFECV (Recursive Feature Elimination with Cross-Validation) is another feature extraction method that incorporates cross-validation. It recursively reduces the number of features and selects the optimal feature subset through cross-validation. Starting with all features, RFECV trains a model at each iteration and eliminates the least important features. Cross-validation is then utilized to evaluate the performance of each feature subset, ultimately selecting the optimal feature subset. RFECV determines the least important features based on model-based feature importance and eliminates them in each iteration. By repeatedly iterating, RFECV identifies the optimal feature subset for feature selection and extraction. One advantage of RFECV is its automatic selection of the number of features and utilization of cross-validation to choose the best feature subset, avoiding the subjectivity and limitations associated with manual feature selection. Additionally, the results obtained from RFECV can enhance the generalization ability and prediction performance of the model. In this study, RFECV was combined with nine classification algorithms for feature extraction, and the effectiveness of feature extraction was evaluated using 12 classification algorithms separately. This approach resulted in improved feature extraction, yielding higher classification accuracy with fewer sequences.

Data availability

The data utilized in this study were obtained from multiple sources, including publicly available datasets published by A vaccine targeting mutant IDH1 in newly diagnosed glioma, TCR Sequencing Can Identify and Track Glioma-Infiltrating T Cells after DC Vaccination and the Adaptive Biotechnologies immuneACCESS, https://doi.org/10.21417/B7001Z. Detailed information on how to access and cite each specific dataset can be found in the respective references provided in this paper.

Code availability

The original data and python code for the experiment are publicly available at https://github.com/Sarakyzhou/Glioma_classify_system/tree/main.

Acknowledgements

This work was supported by the Fundamental Research Funds for the Inner Mongolia University(21100-5187051).

Author information

Authors and Affiliations

Department of Mathematics, School of Mathematical Sciences, Inner Mongolia University, Hohhot, China

Kaiyue Zhou, Zhengliang Xiao, Qi Liu, Xu Wang, Jiaxin Huo, Xiaoqi Wu & Lin Da

Hangzhou ImmuQuad Biotechnologies, Hangzhou, China

Pengfei Xu, Yunyun Deng, Wenwen Xiao & Tao Sun

Institute of Wenzhou, Zhejiang University, Wenzhou, China

Tao Sun

Contributions

LD and TS designed the study and revised the manuscript. KZ performed most of the experiments. KZ, YD, WX and PX processed the primary data. KZ, ZX, QL, XW, JH and XW built the ML algorithms for the study, KZ drafted the manuscript. All authors checked and approved the final manuscript.

Corresponding authors

Correspondence to Tao Sun or Lin Da.

Ethics declarations

Competing interests

The authors declare no competing interests.

Supplementary information

No supplementary information is available for this article.

Ostrom QT, Gittleman H, Farah P, Ondracek A, Chen Y, Wolinsky Y, Stroup NE, Kruchko C, Barnholtz-Sloan JS. CBTRUS statistical report: Primary brain and central nervous system tumors diagnosed in the United States in 2006-2010. Neuro Oncol. 2013 Nov;15 Suppl 2(Suppl 2):ii1-56. doi: 10.1093/neuonc/not151. Erratum in: Neuro Oncol. 2014 May;16(5):760. PMID: 24137015; PMCID: PMC3798196.
Platten, M., Bunse, L., Wick, A. et al. A vaccine targeting mutant IDH1 in newly diagnosed glioma. Nature 592, 463–468 (2021). https://doi.org/10.1038/s41586-021-03363-z
Hsu M, Sedighim S, Wang T, Antonios JP, Everson RG, Tucker AM, Du L, Emerson R, Yusko E, Sanders C, Robins HS, Yong WH, Davidson TB, Li G, Liau LM, Prins RM. TCR Sequencing Can Identify and Track Glioma-Infiltrating T Cells after DC Vaccination. Cancer Immunol Res. 2016 May;4(5):412-418. doi: 10.1158/2326-6066.CIR-15-0240. Epub 2016 Mar 11. PMID: 26968205; PMCID: PMC4873445.
Tumeh PC, et al. PD-1 blockade induces responses by inhibiting adaptive immune resistance. Nature.2014;515:568–571. doi: 10.1038/nature13954.
Maximum Likelihood Estimation of Functionals of Discrete Distributions.2014.
Shannon, C. E. The mathematical theory of communication. 1963.MD Comput. 14, 306–317 (1997).
Syed-Abdul, S., Firdani, RP., Chung, HJ. et al. Artificial Intelligence based Models for Screening of Hematologic Malignancies using Cell Population Data. Sci Rep 10, 4583 (2020). https://doi.org/10.1038/s41598-020-61247-0
Luo, H., Zu, R., Huang, Z. et al. Characteristics and significance of peripheral blood T-cell receptor repertoire features in patients with indeterminate lung nodules. Sig Transduct Target Ther 7, 348 (2022). https://doi.org/10.1038/s41392-022-01169-7
Liu, Y.-Y., Yang, Q.-F., Yang, J.-S., Cao, R.-B., Liang, J.-Y., Liu, Y.-T., Zeng, Y.-L., Chen, S., Xia, X.-F., Zhang, K. and Liu, L. (2019), Characteristics and prognostic significance of profiling the peripheral blood T-cell receptor repertoire in patients with advanced lung cancer. Int. J. Cancer, 145: 1423-1431. https://doi.org/10.1002/ijc.32145
Chen K, Bai J, Reuben A, Zhao H, Kang G, Zhang C, Qi Q, Xu Y, Hubert S, Chang L, Guan Y, Feng L, Zhang K, Zhang K, Yi X, Xia X, Cheng S, Yang F, Zhang J, Wang J. Multiomics Analysis Reveals Distinct Immunogenomic Features of Lung Cancer with Ground-Glass Opacity. Am J Respir Crit Care Med. 2021 Nov 15;204(10):1180-1192. doi: 10.1164/rccm.202101-0119OC. PMID: 34473939; PMCID: PMC8759311.
Swanson K, Wu E, Zhang A, Alizadeh AA, Zou J. From patterns to patients: Advances in clinical machine learning for cancer diagnosis, prognosis, and treatment. Cell. 2023 Apr 13;186(8):1772-1791. doi: 10.1016/j.cell.2023.01.035. Epub 2023 Mar 10. PMID: 36905928.
Emerson, R., DeWitt, W., Vignali, M. et al. Immunosequencing identifies signatures of cytomegalovirus exposure history and HLA-mediated effects on the T cell repertoire. Nat Genet 49, 659–665 (2017). https://doi.org/10.1038/ng.3822
Zhou, K., Huo, J., Gao, C. et al. Applying T-classifier, binary classifiers, upon high-throughput TCR sequencing output to identify cytomegalovirus exposure history. Sci Rep 13, 5024 (2023). https://doi.org/10.1038/s41598-023-31013-z
Sheng J, Wang H, Liu X, Deng Y, Yu Y, Xu P, Shou J, Pan H, Li H, Zhou X, Han W, Sun T, Pan H and Fang Y (2021) Deep Sequencing of T-Cell Receptors for Monitoring Peripheral CD8+ T Cells in Chinese Advanced Non–Small-Cell Lung Cancer Patients Treated With the Anti–PD-L1 Antibody. Front. Mol. Biosci. 8:679130. doi: 10.3389/fmolb.2021.679130

No competing interests reported.

Download PDF

Editorial decision: Revision requested
14 May, 2024
Reviews received at journal
02 May, 2024
Reviewers agreed at journal
18 Apr, 2024
Reviews received at journal
16 Mar, 2024
Reviewers agreed at journal
29 Feb, 2024
Reviews received at journal
15 Dec, 2023
Reviewers agreed at journal
06 Dec, 2023
Reviewers agreed at journal
06 Dec, 2023
Reviewers invited by journal
06 Dec, 2023
Editor assigned by journal
06 Dec, 2023
Editor invited by journal
18 Nov, 2023
Submission checks completed at journal
18 Nov, 2023
First submitted to journal
11 Nov, 2023

You are reading this latest preprint version

Comprehensive application of AI algorithms with TCR NGS data for Glioma diagnosis

Status:

Version 1

Abstract

Figures

Introduction

Results

Classification effects on TCR Diversity Indices

Binary classification algorithms

Multidimensional classification algorithms

Dimension reduction and Core features’ extraction

Discussion

Methods

Declarations

References

Additional Declarations

Status:

Version 1