+1 (502) 904-2126   One Westbrook Corporate Center, Suite 300, Westchester, IL 60154, USA   Site Map
ISSN: 2574 -1241

Impact Factor : 0.548

  Submit Manuscript

Short CommunicationOpen Access

Machine-learning Modeling for Personalized Immunotherapy- An Evaluation Module Volume 47 - Issue 2

Xiaonan Ying1 and Biaoru Li2*

  • 1University of Nebraska Medical Center, Omaha, NE 68131, USA
  • 2Georgia Cancer Center and Department of Pediatrics, Medical College at GA, Augusta, GA 30912, USA

Received: November 05, 2022;   Published: November 18, 2022

*Corresponding author: Biaoru Li, Georgia Cancer Center and Department of Pediatrics, Medical College at GA, Augusta, GA 30912, USA

DOI: 10.26717/BJSTR.2022.47.007462

Abstract PDF


Immune-cell therapy and targeting therapy are in rapid development to treat tumor diseases. However, current immune-cell therapy and targeting immunotherapy often face three challenges (three Ss): safety challenges such as cytokine releasing syndrome (C.R.S.); specificity targeting problems such as low efficacy caused by off-targeting tumor cells; unsatisfying payment are confounded to clinical patients and physicians. We have been studying immunotherapy for more than thirty years, and recently, personalized immunotherapy to treat tumor disease has been proposed. After we discovered quiescent genes from immune cells within the tumor microenvironment, we set up single-cell genomics analysis, studying heterogeneous immune responses from multiple tumor antigens (neo-antigen); here, we further introduce a new generation of immunotherapy module by using a machine-learning model to assess optimal immunotherapy. The machine-learning model combined with single-cell genomic analysis can predict optimal immune-cell (such as T-cells) and other optimal targeting drugs such as PD1 and CTLA4 inhibitors for the patient to use.

Keywords: Machine-Learning; Gene Expression; Pathway Analysis; Tumor-Infiltrating Lymphocytes; Targeting Therapy; Single-Cell Genomic Analysis; Personalized Immunotherapy

Abbrevations: CRS: Cytokine Releasing Syndrome; TCR: T Cell Receptor; CSC: Cancer Stem Cell; TAA: Tumor-Associated Antigens; TSA: Tumor-Specific Antigens; CT: (I) Cancer-testis; AI: Artificial Intelligences; TIME: Tumor Immune Microenvironment; PBMN: Peripheral Blood Mononuclear; CC: Clustering Coefficient; CD: Connectivity Degree


TILs (tumor-infiltrating lymphocytes) are a group of lymphocytes that engage and kill tumor cells in tumor tissues. As with other lymphocytes, the TIL consists of CD3+cell (CD8+T-cell, CD4+Tcell), CD19 (B-cell infiltrating lymphocyte, BIL), CD16+/CD56+cell (NK cells) and other immune cells (macrophage and neutrophil) [1]. T.I.L.s, including CD8+T-cells, demonstrate characteristics of main immune surveillance in the tumor location, which has been specifically recognizing tumor antigens [2]. Moreover, we have also reported that CD8+T-cells from TIL displayed quiescent status and even lower response to tumor-antigen, which are obtained from pre-immunization induction. These cells, during quiescent periods, had no cytotoxicity and low proliferation potential, whereas they could be fully activated when exposed to IL-2 in vitro [3]. In our previous work [4], we employed single-cell genomics analysis to elucidate the quiescent status of CD3+ (such as CD8+T-cells) in tumor microenvironments. By studying the quiescent TIL obtained from a solid tumor, we have identified a group of down-regulated genes, including T cell receptor (TCR), TNF-alpha receptor, TRAIL, and Perforin. More interestingly, we have observed a group of upregulated quiescent genes, including Tob [5], TGF-β, LKLF [6,7], SnoA, Ski, ERF, and REST/NRSF complex. The quiescent genes are discovered in the tumor microenvironment by Western blot and quantitative rtPCR analysis [4]. Furthermore, the observation of actively quiescent T-cells, including CD8+T-cells, is consistent with findings from other laboratories using genomic analysis and animal studies [5,6,8,9].

Since different inhibitors have been found to block immune responses to tumor cells in the tumor environment [8,10], it was further proposed that factors from the quiescent T-cells, such as PD-1 or CTLA-4, are actually one of the cores of immunotherapy [11]. Additionally, the tumor location is a place enriched with specifically activated lymphocytes, such as from an early monoclonal cancer stem cell (C.S.C.) into multiple differentiating tumor antigens within tumor cells, such as tumor-associated antigens (T.A.A.) and tumor-specific antigens (T.S.A.). T.A.A. consists of

(I) Cancer-testis (C.T.) antigens, overexpressed antigens,

(II) Differentiation antigens, and

(III) Oncofetal antigen. T.S.A. contains neo-antigens and oncoviral antigens such as HBV, EBV, CMV, and human papillomavirus (HPV) E6/E7 protein. In addition, neo-antigens have individual expressions for each patient, although shared neo-antigens have been increasingly discovered [12-14]. This background has guided us to set up personalized immunotherapy by singlecell genomics analysis with the machine-learning model.

This manual will use single-cell genomic data to address these questions combined with a machine-learning model. Machine learning is the study of computer algorithms, which can improve automatically through experiments with the use of data. As one of the artificial intelligences (A.I.) algorithms, here, our machine learning algorithms can build a model based on single-cell genomic data from quiescent genes of individual patients. It is called “personal trained data” from the tumor immune microenvironment (TIME) of each patient. The program can make predictions or decisions and then be used in personalized immunotherapy. The treatment module combined three sources:

(I) Immune cell therapy such as T-cell immunotherapy,

(II) Targeting immunotherapy, and

(III) Other different immune therapy methods such as IL2 and TNF and so on. Using the machine-learning analysis, we can address three challenge issues (three Ss) of immunotherapy: Safety: avoiding cytokine releasing syndrome (C.R.S.); Specificity: targeting free off-targeting tumor cell resulting in low efficacy; Satisfying payment avoiding confounded to clinical scientists and physicians. This study will improve our understanding of artificial intelligence, the personal machinelearning application for personalized immunotherapy, and provide a foundation for effective immunotherapy for tumor disease.

Materials and Methods

Immune Cell Isolation

Immune cells were isolated from tumors as described before [15-16]. Briefly, freshly procured tumor tissues were washed in phosphate-buffered saline, cut into small pieces, and digested with 0.25mg/ml of collagenase IV at 4°C for 24 hours. Lymphocytes from tumor tissues and peripheral blood mononuclear (PBMN) cells were centrifuged in Ficoll-Hypaque solution at 500g for 30 minutes and were recovered from the interface of cell suspension. CD3+cell (including CD8+T-cells), non-CD3+cells were isolated from the lymphocytes using magnetic anti-CD3+ microbeads (MACS technology, Miltenyi Biotech, Foster City, CA, U.S.A.) following the manufacturer’s recommendations. The purities of isolated CD3+T-cells were confirmed by fowl cytometry using FITC-labelled CD3+mAb/anti-CD8+mAb. The proliferation potential and cytotoxicity of activated TIL were measured as previously described [4].

R.N.A. Extraction

Purified immune cells were lysed in Trizol reagent (Invitrogen, Carlsbad, CA), and total R.N.A. was extracted using the RNA easy column (Qiagen, Valencia, CA). Briefly, CD8+T-cells were homogenized in Trizol before the phenol-extracted aqueous layer was mixed with chloroform. The aqueous layer from chloroform extraction was precipitated in 70% ethanol and further purified by passing through the RNA easy column. After sequential washing, total R.N.A. was eluted in RNase-free water. Isolated total R.N.A. was quantified, and its integrity was confirmed on an R.N.A. denaturing gel.

cRNA Synthesis

1ug of total R.N.A. was used to prepare biotinylated antisense R.N.A. (cRNA) using Ambion’s Message AmpII-Biotin Enhanced kit (Ambion, Austin, TX) and Affymetrix’s GeneChip sample module kit (Affymetrix, CA). Briefly, total R.N.A. was converted to first-strand cDNA using T7 oligo dT primer. The R.N.A. template was removed from the D.N.A.: R.N.A. hybrid using RNase H before double-stranded cDNA was synthesized. Double-stranded cDNA was purified and used as a template to synthesize biotin-labeled cRNA by in vitro transcription. Purified biotinylated cRNA was fragmented at 94oC and used for gene chip hybridization.

Microarray Hybridization

15ug of fragmented biotinylated cRNA was hybridized to Gene Chip Human Genome U133 Plus 2.0 Array. Following standard conditions described in Affymetrix protocol (Expression Analysis Technical Manual, Affymetrix, Santa Clarita, CA). Briefly, fragmented biotinylated cRNA was incubated with Gene Chip 133 plus 2.0 for 16 hours at 45oC before the hybridized chip was washed and stained in an Affymetrix fluidics station. The processed array was scanned, and the Q.C. was performed using Gene Chip Operating Software (GCOS, Affymetrix, CA). Same specimens performed by RNA-seq will be reported separately (although all data from single-cell RNAseq will be presented by other manuals).

Analysis of Gene Expression

After QC was performed, the expression profile of each sample was generated by Gene Chip Expression Console (Affymetrix, CA) and further processed and analyzed in dChip per the instructions [17]. Briefly, the dChip algorithm was used to summarize the gene expression profile on each sample, where each chip was compared to a reference chip for normalization in overall chip hybridization intensity so as to allow for comparison between different chips. Hierarchical clustering was performed using the same dChip software to examine the similarity of gene expression between different samples. The gene expression profile of each sample was filtered according to the default criteria of dChip. The expression values of genes of interest were listed and compiled. The filtered gene expression profile of each sample was imported into BRB Array Tool [18] for gene enrichment analysis in order to identify statistically significant gene sets that might be involved in the development of immune quiescence in tumors.

Machine-Learning Analysis

Once differential gene expression profiles of TIL cells from tumor tissue were analyzed, a couple of key genes were utilized for quiescent and active analysis by database list. Briefly, interaction data was gathered from a number of different datasets from our work described above. Machine learning was generated by using algorithms that we have published. All datasets with the Regulatory Connectivity and calculated Betweenness of each protein node within these networks were listed and compiled in the Cytoscape platform.

Three indices were used for machine-learning analysis: Betweenness Centrality (B.C.), Connectivity Degree (CD), and Clustering coefficient (CC). In detail,

1) Input a pair of clinical genome profiles from genome diagnosis (including genome profiles CD3+cell (such as CD8+cell), non-CD3+ cells) into Cytoscape platform (plugin Reactome F.I. platform or plugin Mimi platform).

2) Collect a series of genes with a higher degree of BC., lower CD and lower CC, as genes with higher inducing to TIL cells and lower toxicity to T-cells.

3) Enter this genome into the T-cell therapy database (such as CD8+ for TIL, CD16/56+ for N.K., CD19+ for B.I.L., CD11b for TAM1/TAN1 and so on) and the Drug Genome Response Database (DGIdb) for immune targeting therapy or combined chemotherapy.

The personal machine-learning design, performance and clinical information will be reported by another manual from single-cell RNA-seq based Machine learning.


Immune Cells Isolation and Gene Expression Analysis

Table 1.


Figure 1.


Results and performance of immune cells and R.N.A. were isolated from tumors as our previous report. After total R.N.A. extract and microarray performance, Q.C. was performed using the Gene Chip expression console. Each gene chip was examined according to Q.C. standards, including scaling factors and range of percentage of presence. All three pairs of specimens from each patient passed the Q.C. test, while CD3+cell and CD3-cell from TIL with paring T-cell from PBMN as shown in (Figure A). As (Table 1) shows, further immune cells related to biomarker expression to discover the quiescence or activity according to gene expression profiles among the three pairs of samples. In a table list, the CD3+cell from TIL or CD3-cell or CD16/CD56, or CD11b presented quiescence or activity of gene expression. At the same time, peripheral blood of similar patients had quiescence or activity from PBMN of each patient as Table demonstrated that blue color is quiescence and red color is an activity for each immune cell.

The expression values of quiescent-related and activity genes are summarized as a database list if significant changes of quiescent genes were upregulated in CD8+T-cells or CD16/56+ N.K. cell or CD11b (macrophage or neutrophil) isolated from patient tumor samples as compared with those of peripheral blood. For example, T.G.F. beta and REST were upregulated in CD8+cell higher expression section isolated from the tumor as compared with those of PBMN so that the quiescent status was upregulated in CD3+/CD8+cell which we can culture TIL as major cell therapy for the patient because the CD8 quiescent changes can play an important role to treat patient, individually.

Furthermore, the expression profiles of each chip were analyzed in Cluster analysis to investigate the potential enrichment of gene sets. In comparison with those of peripheral blood, quiescent gene profiles were upregulated in CD8+T-cells from TIL; This implies that most of the genes are still in active state tumor while active gene profiles were down-regulated in CD8+T-cells from TIL so that we can further confirm TIL (CD8) as major cell therapy for the patient because inducing the active genes can be significant changes and more function for CD8+T-cells to treat patient.

Machine-Learning Analysis

The genes were entered into the Cytoscape platform for discovering in the network profiles of quiescent CD8+cells, or quiescent CD16/56+cell, or quiescent CD11b, including their seedproteins and their neighbors (which are derived) for a network. All networks are displayed in a configuration in which one seed protein is involved in the network analysis. Three results include Betweenness Centrality (B.C.), Connectivity Degree (CD), and Clustering coefficient (CC). The clinical genome profiles from genome diagnosis (including genome profiles from PBMC and genome profiles from TIL including CD3+ cell such as CD8+cell, non- CD3 cells indicate a higher degree of B.C., lower CD and lower CC, as genes with higher inducing to the immune cells for cell therapy and lower toxicity to the immune cells. The results also can cover other cell therapy such as N.K. cell (CD16/56), and macrophage therapy (CD11b).

Prediction of the Combination of Immune Therapy

Recently, two immune therapies are emerging:

(A) Immune targeting inhibiting therapy such as PD1 or CTLA4 inhibitors and

(B) CAR-T or TCR-T therapy. If we want to find the PD1 or CTLA-4 efficacy, we can put it into Drug Genome Response Database (DGIdb) to discover the immune targeting therapy.

If we want to discover a new generation of cell immune therapy, such as CAR-T or TCR-T therapy, we can use the machinelearning model to discover which cells are optimal. If we want to find a combination, as shown in (Figure-B), therapeutic targeting, including drug targeting, and small molecule targeting, was used for the machine-learning analysis. After the machine-learning was analyzed, (Figure B) finds CAR-T and targeted therapeutic drugs and targeted molecules and links the genome expression profiles with targeted therapeutic drugs and targeted molecules. As the figure demonstrates, CAR-T therapy related to chemotherapy and targeting therapy are discovered to connect as a prediction for the patient individually.


In this manual, after we set up single-cell genomic data to address the quiescent and active status of immune response for each patient, here we combined a machine-learning model to discover an optimal therapy for each patient. Machine learning can improve automatically through experiment data from individual genes so that the machine learning algorithms can be used in personalized immunotherapy. The treatment model also combined three sources:

1) Immune cell therapy such as CD8+T-cells or CD16/56+ N.K. cell or CD11b (macrophage or neutrophil), with totally current thirteen immune cell therapy.

2) Targeting immunotherapy with all clinical trial and F.D.A. approved drugs.

3) All other different treatment such as HDACI or DNMA inhibitors so that we can use the machine-learning analysis to resolve three challenge issues from immunotherapy: safe treatment voiding cytokine releasing syndrome (C.R.S.); specific targeting treatment free off-targeting tumor cell; satisfying payment avoiding confounded to clinical patients for a higher payment. In the near future, we will report using single-cell R.N.A. seq data to mimic machine-learning with the results from clinical patients. Hopefully, it also will support machinelearning to support personalized immunotherapy.


This research was supported by National Cancer Institute IRG- 91-022-09, U.S.A. (B.L.).


  1. Arosa FA (2002) CD8+CD28- T cells: Certainties and uncertainties of a prevalent human T-cell subset. Immunol Cell Biol 80(1): 1-13.
  2. Redmond WL, Sherman LA (2005) Peripheral tolerance of CD8 T lymphocytes. Immunity 22(3): 275-84.
  3. Monsurrò V, Wang E, Yamano Y, Migueles SA, Panelli MC, et al. (2004) The quiescent phenotype of tumor-specific CD8+ T cells following immunization. Blood 104(7): 1970-1978.
  4. Zhang W, Ding J, Qu Y, Hu H, Lin M, et al. (2009) Genomic expression analysis by single-cell mRNA differential display of quiescent CD8 T cells from tumor-infiltrating lymphocytes obtained from in vivo liver tumors. Immunology 127(1): 83-90.
  5. Tzachanis D, Freeman GJ, Hirano N, van Puijenbroek AA, Delfs MW, et al. (2001) Tob is a negative regulator of activation that is expressed in anergic and quiescent T cells. Nat Immunol 2(12): 1174-1182.
  6. Buckley AF, Kuo CT, Leiden JM (2001) Transcription factor LKLF is sufficient to program T cell quiescence via a c-Myc-dependent pathway. Nat Immunol 2(8): 698-704.
  7. Haaland RE, Yu W, Rice AP (2005) Identification of LKLF-regulated genes in quiescent CD4+ T lymphocytes. Mol Immunol 42(5): 627-641.
  8. Crews FT, Bechara R, Brown LA, Guido DM, Mandrekar P, et al. (2006) Cytokines and alcohol. Alcohol Clin Exp Res 30(4): 720-730.
  9. Yusuf I, Fruman DA (2003) Regulation of quiescence in lymphocytes. Trends Immunol 24(7): 380-386.
  10. Mondal TK, Li D, Swami K, Dean JK, Hauer C, et al. (2005) Mercury impairment of mouse thymocyte survival in vitro: Involvement of cellular thiols. J Toxicol Environ Health A 68(7): 535-556.
  11. Eggers V, Pascher A, Althoff H, Thiele S, Mütze J, et al. (2006) Immune reactivity is more suppressed in patients with alcoholic liver disease than in patients with virus-induced cirrhosis after CRH stimulation. Alcohol Clin Exp Res 30(1): 140-149.
  12. Lalor PF, Shields P, Grant A, Adams DH (2002) Recruitment of lymphocytes to the human liver. Immunol Cell Biol 80(1): 52-64.
  13. Braet F, Wisse E (2002) Structural and functional aspects of liver sinusoidal endothelial cell fenestrae: A review. Comp Hepatol 1(1): 1.
  14. Dashek WV, Barker SM, Statkiewicz WR, Shanks ET, Llewellyn GC (1983) Histochemical analysis of liver cells from short-term, aflatoxin-dosed, and nondosed Coturnix coturnix japonica. 1. Aflatoxin-sensitive quail. Poult Sci 62(12): 2347-2359.
  15. Li B, Ding J, Larson A, Song S (1999) Tumor tissue recycling--a new combination treatment for solid tumors: Experimental and preliminary clinical research. In Vivo 13(5): 433-438.
  16. Li B, Chang T, Larson A, Ding J (2000) Identification of mRNAs expressed in tumor-infiltrating lymphocytes by a strategy for rapid and high throughput screening. Gene 255(2): 273-279.
  17. Li C, Wong WH (2001) Model-based analysis of oligonucleotide arrays: expression index computation and outlier detection. Proc Natl Acad Sci USA 98(1): 31-36.
  18. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, et al. (2005) Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proc Natl Acad Sci USA 102(43): 15545-15550.