Identification

of

biomarker

sets

for predicting

the

efficacy

of

sublingual

immunotherapy against pollen-induced allergic rhinitis

Running title: HR/NR patients are distinguishable before SLIT

Minoru Gotoh1, 2, *, Osamu Kaminuma1, *, Akihiro Nakaya1, 3, *, Kazufumi Katayama1, Yuji Motoi1, Nobumasa Watanabe1, Mayumi Saeki1, Tomoe Nishimura1, Noriko Kitamura1, Kazuko Yamaoka1, Kimihiro Okubo1, 2, and Takachika Hiroi1

1

Allergy and Immunology Project, Tokyo Metropolitan Institute of Medical Science, 2-1-6,

Kamikitazawa, Setagaya-ku, Tokyo 156-8506, Japan 2

Department of Otorhinolaryngology, Nippon Medical School, 1-1-5 Sendagi, Bunkyo-ku,

Tokyo 113-8603, Japan 3

Department of Genome Informatics, Graduate School of Medicine, Osaka University, 2-2

Yamadaoka, Suita, Osaka 565-0871, Japan

Correspondence: Dr. Takachika Hiroi Allergy and Immunology Project, Tokyo Metropolitan Institute of Medical Science, 2-1-6, Kamikitazawa, Setagaya-ku, Tokyo 156-8506, Japan. Tel: +81-3-5316-3266, Fax: +81-3-5316-3153, e-mail: [email protected]

* These authors contributed equally to this work.

The total number of pages and figures submitted: 28 pages including 1 table and 6 figures.

© The Japanese Society for Immunology. 2017. All rights reserved. For permissions, please e-mail: [email protected]

2 Keywords: AdaBoost; Cluster analysis; Cytokines; Japanese cedar pollen

3 Abstract Sublingual immunotherapy (SLIT) is effective against allergic rhinitis, although a substantial proportion of individuals is refractory. Herein, we describe a predictive modality to reliably identify SLIT non-responders. We conducted a 2-year clinical study in 193 adult patients with Japanese cedar pollinosis, with biweekly administration of 2000 Japanese allergy units of cedar pollen extract as the maintenance dose. After identifying high-responder (HR) patients with improved severity scores and non-responder (NR) patients with unchanged or exacerbated symptoms, differences in 33 HR and 34 NR patients were evaluated in terms of peripheral blood cellular profiles by flow cytometry and serum factors by ELISA and cytokine bead array, both pre- and post-SLIT. Improved clinical responses were seen in 72% of the treated patients. Pre-therapy IL-12p70 and post-therapy IgG1 serum levels were significantly different between HR and NR patients, although these parameters alone failed to distinguish NR from HR patients. However, the analysis of serum parameters in the pretherapy samples with the Adaptive Boosting (AdaBoost) algorithm distinguished NR patients with high probability within the training data set. Cluster analysis revealed a positive correlation between serum Th1/Th2 cytokines and other cytokines/chemokines in HR patients after SLIT. Thus, processing of pre-therapy serum parameters with AdaBoost and cluster analysis can be reliably used to develop a prediction method for HR/NR patients.

4 Introduction Allergen-specific immunotherapy (AIT) effectively treats allergic diseases (1-3). Recently, the application of sublingual immunotherapy (SLIT), rather than classical subcutaneous immunotherapy (SCIT), has expanded owing to its equivalent efficacy and reduced side effects. SLIT significantly improved severity scores and lowered the use of symptomatic anti-allergy drugs in allergic rhinitis (AR) patients, as demonstrated by several meta-analyses from >60 double-blind, placebo-controlled clinical trials (3,4). Despite its high efficacy, a substantial proportion of patients is refractory to AIT (5,6). Although 50–70% of patients with AR or bronchial asthma respond favorably to SCIT or SLIT, these therapies are ineffective for the remaining ~30% (5,6). Although the mechanisms of AIT in mitigating allergic reactions have been well studied (7-10), the reason for the large proportion of unaffected patients remains unclear. Treating patients who do not respond to AIT is futile and should be avoided. Therefore, it is desirable to identify and exclude nonresponder (NR) patients before initiating AIT; however, no predictive biomarkers or diagnostic tests have been developed (8), at least partially, because the pathophysiological and genetic differences between AIT-responder and NR patients are ambiguous. In addition, as is often the case when searching for biomarkers mediating the development and severity of allergic diseases (11,12), it is difficult to identify a single biomarker that comprehensively predicts the efficacy of SLIT because of the complicated pathogenesis involved. In comparison with the first worldwide research on the prevalence of AR performed in 1998 (13), recent epidemiological studies clearly demonstrate that the number of patients receiving AR has been continuously increasing (14-17). In parallel, an upsurge of patients with Japanese cedar pollen-induced AR (JCPAR) has occurred in Japan (18). JCPAR is a typical seasonal allergy lasting approximately 3 months from February to April, during which time widely dispersed airborne cedar pollen causes AR and/or conjunctivitis in ~25%

5 of the Japanese population (19). SLIT with cedar pollen extract was recently approved to treat JCPAR (20). Although clinical results indicate that SLIT is highly efficacious against JCPAR, NRs also exist (21-23). Therefore, to identify differences between the responder and NR patients, particularly before AIT initiation, we performed a clinical study of patients with JCPAR receiving SLIT. We comprehensively analyzed cellular and serum parameters by using a novel ensemblelearning algorithm to identify biomarkers capable of discerning high-responder (HR) and NR patients before SLIT initiation.

6 Methods Subjects Patients with JCPAR who were residents of Tokyo, Japan; over 20 years old; and attending Nippon Medical School Hospital, Tokyo Metropolitan Hospitals, and clinics in Tokyo from May to June 2006 were included. All patients showed AR symptoms, such as sneezing, rhinorrhea, and nasal congestion, from February to April (pollen season) for at least three consecutive years before undergoing SLIT. They were positive in skin-prick tests and for IgE against cedar pollen allergen. Patients were excluded based on the following criteria: (1) positive for Japanese cypress pollen-specific IgE, (2) complicated symptoms with other nasal diseases, (3) use of drugs affecting SLIT efficacy such as anti-allergic drugs and steroids, (4) history of bronchial asthma or urticaria at age 15 or older, (5) previous treatment with SCIT, (6) pregnancy or potential pregnancy, (7) lactation, and (8) recommended exclusion by attending physicians owing to complications such as liver, kidney, or heart disease, respiratory infection, or other severe disorders.

Allergen administration For the sake of this study, standardized cedar pollen allergen purchased from Torii Pharmaceutical Co., Ltd. (Tokyo, Japan) was self-administered at home, as described previously (22). Briefly, patients held the allergen solution under the tongue for 2 min and then spit it out. The self-administered allergen was gradually increased over 5 weeks from one drop containing two Japanese allergy units (JAU)/ml to up to 20 drops of 2,000 JAU/ml as the maintenance dose by 5 weeks (Table 1), beginning in July (1st year, Fig. 1) and continuing for 2 years. The standard of care treatment and other treatments to control allergic symptoms were not performed, except that a single use of oral and ophthalmic antihistamines was permitted for severely exacerbated patients. These treatments were not

7 expected to significantly affect pre- and post-SLIT parameters because the evaluation occurred after the pollen season ended.

HR and NR patient classification Nasal symptoms recorded by the patients from February 1 to April 30 during three pollen seasons (1st to 4th year) were investigated using the Japanese guidelines for AR as described previously (18,24). Briefly, the numbers of sneezing and nose-blowing episodes, extent of nasal congestion and eye itchiness, symptom duration, and medication compliance were selfrecorded daily using an allergy diary, according to the classification scheme developed by Okuda et al. (25,26). The quality of life of the patients was determined three times at the end of February, mid-March, and mid-April each year using the Japanese Rhinoconjunctivitis Quality of Life Questionnaire No. 1 (JRQLQ1). At the end of each pollen season, SLIT efficacy was evaluated by comparing the patients’ 5-grade severity scores (most severe, severe, moderate, mild, or asymptomatic), as determined by the allergy diary and the JRQLQ1 in each year to that before the start of SLIT as reported previously (25,26). After SLIT, the top 33 HR patients showing >1-step improvement in severity score and the bottom 34 NR patients with unchanged or exacerbated conditions were selected. For all patients matching the above criteria, peripheral blood and serum samples were prepared immediately before (1st year) and after (3rd year) SLIT during the June–July off-season. All samples were randomized before analysis.

Cell preparation Peripheral blood mononuclear cells (PBMCs) were prepared by Ficoll-Paque densitygradient centrifugation, as described previously (27). Then, B cell and macrophage/monocyte fractions were removed using a magnetic cell-sorting system (AutoMACS, Miltenyi Biotec

8 GmbH, Bergisch Gladbach, Germany) with anti-CD14-biotin (eBioscience, San Diego, CA), anti-CD16-biotin (BD Biosciences, San Jose, CA), and anti-CD19-biotin (eBioscience) antibodies (Abs) in conjunction with anti-biotin microbeads (Miltenyi Biotech GmbH). The cells in the negative fraction were stained with anti-CD4-Pacific blue (BD Biosciences), antiCD8-AmCyan (BD Biosciences), anti-CD11c-APC (BD Biosciences), CD123-PE-Cy7 (eBioscience), anti-BDCA-1-FITC (Miltenyi Biotec GmbH), anti-BDCA-4-PE (Miltenyi Biotech GmbH), and anti-HLA-DR-APC-Cy7 (BD Biosciences) Abs. Next, CD4+ T cells, basophils, conventional dendritic cells (cDCs), and plasmacytoid dendritic cells (pDCs) were determined as CD4+HLA-DR-, CD123+CD11clowHLA-DR-, CD11c+BDCA-1+HLA-DR+, and CD123+BDCA-4+HLA-DR+ populations, respectively, using a FACSCanto II flow cytometer (BD Biosciences) (S1 Fig.).

Serum mediator analysis Cedar pollen- and cypress pollen-specific and total IgE were measured using ImmunoCAP Specific IgE and ImmunoCAP Total IgE kits, respectively (Thermo Fischer Scientific Inc., Uppsala, Sweden). Cedar pollen-specific IgG1–4 was measured by ELISA using Cedar Pollen Crude Extract-Cj (Cosmo Bio, Tokyo, Japan) as the capture protein and HRPconjugated mouse anti-human IgG1–4 (Southern Biotech, Birmingham, AL) as the detecting antibody. In total, 50 cytokines of Bio-Plex Pro Human Cytokine Group I (IL-1, IL-1R, IL2, IL-4, IL-5, IL-6, IL-7, IL-9, IL-10, IL-12p70, IL-13, IL-15, IL-17, IFN-, TNF-, IL-8, eotaxin, IP-10, MCP-1, MIP-1, MIP-1, RANTES, FGFbasic, G-CSF, GM-CSF, PDGFBB, and VEGF) and Group II (IL-1, IL-2R, IL-3, IL-12p40, IL-16, IL-18, IFN-2, THF, TRAIL, CTACK, GRO-, ICAM-1, MCP-3, MIF, MIG, VCAM-1, HGF, LIF, M-CSF, NGF, SCF, SCGF-, and SDF1) were measured using assay kits (Bio-Rad Laboratories

9 Inc., Hercules, CA) with a Bio-Plex suspension array system (Bio-Rad Laboratories, Inc.) according to the manufacturer’s instructions. The results for the pre- and post-treatment sample groups were subjected to quantile normalization. The resulting cytokine levels were normalized as a z-score using the equation (z = [v− ]/ ), where v is a subject’s raw data, is the mean value observed in the population, and

is the associated standard deviation in the

population. Using Pearson’s product-moment correlation coefficients for the profiles of cytokine levels in pre-treatment HRs, the data were hierarchically clustered by the unweighted pair-group method with the arithmetic mean (UPGMA). By isolating the dendrogram branches at a correlation coefficient of 0.7, cytokine clusters mutually correlated in HRs were identified.

Adaptive boosting (AdaBoost) analysis AdaBoost is an ensemble-learning algorithm for predicting the attribution of unknown subjects by referring to previously analyzed data (28). To identify factors predicting HR or NR phenotypes, serum cytokine levels were analyzed using AdaBoost. We generated the best additional weak learner (a decision stump) in each incremental step by exhaustively searching for the best explanatory attribute Tt, with its threshold vt and the directions of inequalities. AdaBoost determines the weight for the majority decision, wt. Letting ft(x) and wt respectively denote the result of the t-th weak learner for subject x and the weight for the majority decision, a strong learner is constructed using the weighted linear summation of all results from weak learners, F(x) = Σ(wt × ft(x)). When the resulting F(x) is >0 or 1-step improvement in severity score increased from 52% in the second year to 72% in the third year (Fig. 2B). SLIT effectiveness lasted ≥1 year after ceasing allergen administration. Most importantly, and consistent with previous investigations (5,6), ~30% of patients failed to show improvement after 2 years. The top 33 HRs (aged 39.3 ± 11.0 years, male:female = 8:17) and the bottom 34 NRs (aged 47.8 ± 12.1 years, male:female = 11:14) were selected (Fig. 2C). No significant pre-treatment differences in clinical scores were observed between the HRs (3.94 ± 0.98) and NRs (3.12 ± 0.98). In both groups, > 95% patients received SLIT by following the exact allergen administration schedule. Severe exacerbation was not observed in HR patients. As determined by Durham’s method (29), the airborne cedar-pollen count in Chiyoda-ku, Tokyo, Japan during the first 2 years was lower than the average of the last 10 years, whereas that during the last 2 years was comparable to this average (http://www.tokyo-eiken.go.jp/kj_kankyo/kafun/data_spring/; S2 Fig.).

13 Cell population analyses CD4+ T cells and cDCs, which are involved in antigen recognition and presentation, respectively, are implicated in AIT pathogenesis (7-10,30). Basophil and pDC involvement in allergic diseases was also shown recently (31,32). We analyzed these cell types to investigate the mechanisms of SLIT (S1 Fig.). In addition to equivalent pre-treatment levels of CD4+ T cells, basophils, cDCs, and pDCs in CD14/CD16/CD19 cell-depleted PBMCs in HR and NR patients, these populations were unaffected by SLIT (Fig. 3), indicating that SLIT efficacy was not mediated by differences in the frequency of CD4, basophil, cDC, and pDC immune subsets.

Serum immunoglobulin and cytokine analysis Serum immunoglobulin levels in the HR and NR groups were essentially equal before the study (Fig. 4A and B). SLIT significantly increased allergen-specific IgE and the allergenspecific/total IgE ratio in both groups. Total IgE levels and allergen-specific IgG subtypes were elevated only in the NR group (Fig. 4A and B). Serum levels of several cytokines were significantly changed following SLIT (S3 Fig.); only IL-12p70 and IgG1 levels were significantly different between HR and NR patients before and after SLIT, respectively (Fig. 4B and C).

Serum factor analysis with AdaBoost Although the differences in IL-12p70 and IgG1 levels were significant, the HR and NR groups could not be distinguished based on these parameters alone. Therefore, all cytokine data measured in pre-treatment sera were processed with AdaBoost (28). From the first weak learner, IL-12p70, with an accuracy of 0.57, the accuracy improved as weak learners were added, finally reaching 0.97 using 24/50 categories of cytokines measured (Fig. 5A). Of 94

14 total mid-range patients, 51 not selected as HR or NR were subsequently categorized as NR by the AdaBoost learners generated using the HR plus NR training data set. Therefore, an additional AdaBoost analysis was performed for the resulting NR plus mid-range patients. From the first weak learner, TRAIL, NR, and mid-range patients were divided with an accuracy of 0.70, which ultimately reached 0.95 when using 25 cytokine categories (Fig 5B). These findings suggest the usefulness of this strategy for constructing strong learners that enable identification of NR patients before SLIT with high probability by collecting serum cytokine data from many patients.

Clustering analysis of serum factors To obtain the basal cytokine profile of HR patients, serum cytokine data were hierarchically clustered according to their profiles in the pre-treatment HR group. We found that the correlations in six groups of serum factors, namely TRAIL, SDF-1, M-CSF, -NGF, IL2R and MIF; IL-12p40, IL-1, TNF- and LIF; FGFbasic and IL-8; IFN-, IL-4, G-CSF, IL-10, IL-5, TNF-, IL-13, IL-7, MCP-1, IL-15, IL-6, IL-2, IL-1R, IL-1 and eotaxin; IFN-2, IL-16, MCP-3, and IL-3; and GM-CSF and IL-12p70 were clearly different between the HR and NR groups (Fig. 6A). Particularly, a strong correlation was observed among typical Th1 cytokines such as IFN-, TNF- and IL-1 as well as Th2 cytokines such as IL4, IL-5, and IL-13 in the HR group (Figs. 6A and E). The average correlation coefficients in this

IFN-/IL-4/G-CSF/IL-10/IL-5/TNF-/IL-13/IL-7/MCP-1/IL-15/IL-6/IL-2/IL-1R/IL-

1/eotaxin cluster were significantly different between HR and NR groups both before and after SLIT. The significant difference between the HR and NR groups was also seen in the relationships in the TRAIL/SDF-1/M-CSF/-NGF/IL-2R/MIF cluster (Fig. 6B). The level of correlation in the former cluster in HR patients was decreased in response to SLIT, whereas that in the later cluster in NR patients was increased. The relationships in the

15 FGFbasic/IL-8 (Fig. 6D) and GM-CSF/IL-12p70 (Fig. 6G) clusters were much weaker in the NR group.

16 Discussion The data presented herein demonstrate that HR and NR JCPAR patients can be distinguished before the initiation of SLIT based on serum cytokine levels, suggesting a possible predictive means for SLIT efficacy. It is obviously useful to identify NR patients to be excluded from SLIT, which must be continued for several years once it is begun, although few studies have addressed the issue of pre-therapy identification of potential non-responders. Di Lorenzo et al. demonstrated a significant correlation between the serum allergen-specific/total IgE ratio and clinical responses to SLIT and SCIT (5); higher AIT efficacy was achieved for rhinitis patients showing higher serum allergen-specific/total IgE ratios, and HR and NR patients could be predicted from this ratio with 77.8–100% accuracy (5), as observed in another SLIT trial (23). These findings are impressive because relatively high predictive values were obtained using a single parameter (5). However, this pattern was not reproducible here. Although cedar pollen-specific/total IgE ratios were significantly increased by SLIT, these increases were largely equivalent in the HR and NR groups of our study cohort. The reason for the discrepancy is unclear, and further studies involving detailed time courses and longer followup are required to validate the usefulness of such parameters in predicting AIT efficacy unequivocally. Only pre-treatment IL-12p70 and post-treatment IgG1 levels were significantly different between HR and NR in serum analyses. Increased serum IgG1 level is a characteristic of AIT-treated patients (8) and has been associated with IgG-mediated serum inhibitory activity for allergen-IgE binding to B cells (33), indicating that IgG1 is involved in AIT pathophysiology. However, this was contradicted in the present study, in which a significant increase in IgG1 was seen only in post-treatment NR patients. Further investigation will be required to confirm whether IgG1 facilitates SLIT resistance.

17 In addition, the significant increase of IL-12p70 levels in NR patients appears consistent with previous reports demonstrating the elevation of nasal IL-12 as well as other Th1 cytokines following AIT (34,35). However, the serum IL-12p70 level in HR patients was already high before treatment and was dramatically reduced by SLIT. As such, Th1 and Th2 cytokines might be co-regulated in HR patients but not in NR patients. This Th1/Th2 theory is widely recognized in the field of allergy (36,37), and the shift of Th2 CD4+ T cells to Th1 cells by AIT is known (7,8). Nevertheless, our findings on serum immunoglobulin levels were inconsistent with the Th1/Th2 theory. For instance, pre-treatment IL-12p70 was higher in HR, whereas Th1-related IgG2 and IgG4 were equivalent in HR and NR. In contrast, Th2-related IgG1 was higher in post-treatment HR, although the difference in IgE levels was not clear. While the precise reason for these discrepancies remain unclear, the timescale of cytokine and immunoglobulin production was different: cytokine synthesis is regulated within hours or days, whereas immunoglobulins are produced over the course of weeks or months. Thus, the differences in cytokine levels—such as pre-treatment IL12p70—might reflect immunoglobulin levels at a later time point. Since IgE switching occurs directly or via IgG (38), higher post-treatment IgG1 in HR might induce the elevation of IgE after the study end-point. As such, a longer and more frequent evaluation of serum parameters is required to clarify the detailed relationship between serum cytokines and immunoglobulins in patients receiving SLIT. Furthermore, the complex contribution of T cells to allergic disease pathogenesis is being increasingly recognized. Cytokines synthesized by several new T-cell subsets, such as Th9, Th17, Th22, and Treg cells, can play different roles (39,40). More severe and chronic disease conditions is associated with the presence of a wider variety of T-cell subsets (41,42). The suppressive function of Treg cells against Th cells shows some selectivity (43), and some Th cells promote Treg cell function (44). These complications in allergic inflammation

18 might affect the responsiveness of patients receiving T-cell-targeted therapies, including AIT. Collectively, the poles separating HR and NR patients in our clustering analysis indicated that the high Th1/Th2 correlation in HR patients might be pivotal in their responses to SLIT. Since we did not observe differences in the frequency of CD4+ T cells, basophils, cDCs, and pDCs between HR and NR patients, it may be necessary to evaluate the individual cell subsets and/or their activation status. Wachholz et al. demonstrated that the effect of AIT was related to Th1/Th2 cytokine levels in nasal tissues, but not in serum (34). Therefore, further investigation of nasal parameters might provide additional information useful in predicting SLIT effectiveness. Intriguingly, SLIT-naïve HR and NR patients were distinguishable by serum cytokine profiles in our training data set. AdaBoost differs from a bootstrap aggregating algorithm that randomly generates equally weighted weak learners independently (45). Because AdaBoost was confirmed to be a robust and reliable method for decision making (46), it has been applied for many classification tasks, for example in image processing related to various diseases (47-49). Because AdaBoost incrementally and iteratively generates variously weighted weak learners, an additional weak learner in each round of iteration enables focused examination of subjects that have been falsely predicted by the previously generated weak learners. The threshold values in the new weak learner were determined by exhaustive searching. The resulting strong learner is optimized to predict values of the objective attribution in the training dataset (28). High prediction accuracy was achieved by this optimization strategy, which was originally implemented for this study. Substantial predictability was obtained by using only the training data set herein, and two-round selection steps were needed to extract NR from HR plus mid-range patients. Therefore, additional analyses of larger cohorts are required to unequivocally establish and validate the robustness of our approach in identifying patients most likely to benefit from SLIT.

19 In conclusion, we demonstrated a predictive modality for SLIT responsiveness in AR patients by employing AdaBoost. The clustering analysis revealed that differences in the relationships among cytokines may be involved in the mechanisms underlying SLIT efficacy. Further examination of larger cohorts with additional analytical parameters—such as gene expression and polymorphism—would expand our understanding of the immunological mechanisms of SLIT in AR.

20 Funding This study design and data collection aspects of this work were funded by the Tokyo Metropolitan Government. This work was also supported by funding from the Kurozumi Medical Foundation and the Japan Allergy Foundation.

Acknowledgments We thank the physicians and nurse practitioners of Nippon Medical School Hospital, Tokyo Metropolitan Komagome Hospital, Tokyo Metropolitan Hiroo Hospital, Tokyo Metropolitan Fuchu Hospital, Tokyo Metropolitan Otsuka Hospital, Tokyo Metropolitan Health and Medical Treatment Corporation Ebara Hospital, Endo ENT Clinic, and Hirooka Clinic who assisted with the collection of clinical samples. We also thank Dr. Satoshi Yuzawa for his technical assistance. We are grateful to Dr. Liyanage P. Perera for reviewing this manuscript. The manuscript was edited by the professional English language editing service Editage.

Conflicts of interest statement: the authors declared no conflicts of interest.

21 Table 1. Allergen administration schedule

Day 1 2 3 4 5 6 7

Weeks Dose (JAU)

1 2

2 20

3 200

4 2000

5 2000

6 2000

≥7 2000

1 2 3 4 6 8 10

1 2 3 4 6 8 10

1 2 3 4 6 8 10

1 2 4 8 12 18 20

20

20

20*

20

The number of drops of allergen solution administered to subjects at each time point is indicated. *Biweekly.

22

References 1

Bousquet, J., Lockey, R. and Malling, H. J. 1998. Allergen immunotherapy: therapeutic vaccines for allergic diseases. A WHO position paper. J. Allergy Clin. Immunol. 102:558.

2

Frati, F., Incorvaia, C., Lombardi, C., et al. 2012. Allergen immunotherapy: 100 years, but it does not look like. Eur. Ann. Allergy Clin. Immunol. 44:99.

3

Passalacqua, G., Garelli, V., Sclifo, F., et al. 2013. Sublingual immunotherapy for allergic rhinitis and conjunctivitis. Immunotherapy 5:257.

4

Incorvaia, C., Di Rienzo, A., Celani, C., et al. 2012. Treating allergic rhinitis by sublingual immunotherapy: a review. Ann. Ist. Super. Sanita 48:172.

5

Di Lorenzo, G., Mansueto, P., Pacor, M. L., et al. 2009. Evaluation of serum sIgE/total IgE ratio in predicting clinical response to allergen-specific immunotherapy. J. Allergy Clin. Immunol. 123:1103.

6

Manzotti, G. and Lombardi, C. 2013. Allergen immunotherapy as a drug: the new deal of grass allergen tablets from clinical trials to current practice. Eur. Ann. Allergy Clin. Immunol. 45:34.

7

Moingeon, P. and Mascarell, L. 2012. Induction of tolerance via the sublingual route: mechanisms and applications. Clin. Dev. Immunol.:1.

8

Cappella, A. and Durham, S. R. 2012. Allergen immunotherapy for allergic respiratory diseases. Hum. Vaccin. Immunother. 8:1499.

9

Jutel, M., Van de Veen, W., Agache, I., et al. 2013. Mechanisms of allergen-specific immunotherapy and novel ways for vaccine development. Allergol Int 62:425.

10

Akdis, C. A. and Akdis, M. 2015. Advances in allergen immunotherapy: Aiming for complete tolerance to allergens. Sci. Transl. Med. 7:280ps6.

11

Kaminuma, O., Suko, M. and Mori, A. 2006. Genetic factors in the treatment of bronchial asthma. Expert Rev. Clin. Immunol. 2:727.

12

Ricciardolo, F. L., Sorbello, V. and Ciprandi, G. 2015. A pathophysiological approach for FeNO: A biomarker for asthma. Allergol. Immunopathol.

13

ISAAC. 1998. Worldwide variation in prevalence of symptoms of asthma, allergic rhinoconjunctivitis, and atopic eczema. Lancet 351:1225.

23 14

Yamazaki, S., Shima, M., Nakadate, T., et al. 2015. Patterns of Sensitization to Inhalant Allergens in Japanese Lower-Grade Schoolchildren and Related Factors. Int. Arch. Allergy. Immunol. 167:253.

15

Schwindt, C. D. and Settipane, R. 2012. Allergic rhinitis (AR) is now estimated to affect some 1.4 billion people globally and continues to be on the rise. Editorial. Am. J. Rhinol. Allergy 26 Suppl 1:S1.

16

Wheatley, L. M. and Togias, A. 2015. Clinical practice. Allergic rhinitis. N Engl J Med 372:456.

17

Meltzer, E. O., Blaiss, M. S., Naclerio, R. M., et al. 2012. Burden of allergic rhinitis: allergies in America, Latin America, and Asia-Pacific adult surveys. Allergy Asthma Proc. 33 Suppl 1:S113.

18

Fujieda, S., Kurono, Y., Okubo, K., et al. 2012. Examination, diagnosis and classification for Japanese allergic rhinitis: Japanese guideline. Auris Nasus Larynx 39:553.

19

Yamada, T., Saito, H. and Fujieda, S. 2014. Present state of Japanese cedar pollinosis: The national affliction. J. Allergy Clin. Immunol. 133:632.

20

Mitobe, Y., Yokomoto, Y. and Ohashi-Doi, K. 2015. Safety evaluation of standardized allergen extract of Japanese cedar pollen for sublingual immunotherapy. Regul. Toxicol. Pharmacol. 71:529.

21

Horiguchi, S., Okamoto, Y., Yonekura, S., et al. 2008. A randomized controlled trial of sublingual immunotherapy for Japanese cedar pollinosis. Int. Arch. Allergy Immunol. 146:76.

22

Okubo, K., Gotoh, M., Fujieda, S., et al. 2008. A randomized double-blind comparative study of sublingual immunotherapy for cedar pollinosis. Allergol. Int. 57:265.

23

Fujimura, T., Yonekura, S., Horiguchi, S., et al. 2010. Increase of regulatory T cells and the ratio of specific IgE to total IgE are candidates for response monitoring or prognostic biomarkers in 2-year sublingual immunotherapy (SLIT) for Japanese cedar pollinosis. Clin. Immunol. 139:65.

24

Okubo, K., Kurono, Y., Fujieda, S., et al. 2011. Japanese guideline for allergic rhinitis. Allergol. Int. 60:171.

25

Okuda, M. 2001. Grading the severity of allergic rhinitis for treatment strategy and drug study purposes. Curr. Allergy Asthma Rep. 1:235.

24 26

Okuda, M., Ohkubo, K., Goto, M., et al. 2005. Comparative study of two Japanese rhinoconjunctivitis quality-of-life questionnaires. Acta Otolaryngol. 125:736.

27

Kaminuma, O., Kitamura, F., Miyatake, S., et al. 2009. T-box 21 transcription factor is responsible for distorted T(H)2 differentiation in human peripheral CD4+ T cells. J. Allergy Clin. Immunol. 123:813.

28

Freund, Y. and Schapire, R. E. 1997. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci. 55:119.

29

Durham, O. C. 1946. The volumetric incidence of atmospheric allergens; a proposed standard method of gravity sampling, counting, and volumetric interpolation of results. J. Allergy 17:79.

30

Allam, J. P., Bieber, T. and Novak, N. 2009. Dendritic cells as potential targets for mucosal immunotherapy. Curr. Opin. Allergy Clin. Immunol. 9:554.

31

Falcone, F. H., Zillikens, D. and Gibbs, B. F. 2006. The 21st century renaissance of the basophil? Current insights into its role in allergic responses and innate immunity. Exp. Dermatol. 15:855.

32

Jahnsen, F. L., Farkas, L., Lund-Johansen, F., et al. 2002. Involvement of plasmacytoid dendritic cells in human diseases. Hum. Immunol. 63:1201.

33

Wachholz, P. A., Soni, N. K., Till, S. J., et al. 2003. Inhibition of allergen-IgE binding to B cells by IgG antibodies after grass pollen immunotherapy. J. Allergy Clin. Immunol. 112:915.

34

Wachholz, P. A., Nouri-Aria, K. T., Wilson, D. R., et al. 2002. Grass pollen immunotherapy for hayfever is associated with increases in local nasal but not peripheral Th1:Th2 cytokine ratios. Immunology 105:56.

35

Hamid, Q. A., Schotman, E., Jacobson, M. R., et al. 1997. Increases in IL-12 messenger RNA+ cells accompany inhibition of allergen-induced late skin responses after successful grass pollen immunotherapy. J. Allergy Clin. Immunol. 99:254.

36

Umetsu, D. T. and DeKruyff, R. H. 1997. TH1 and TH2 CD4+ cells in human allergic diseases. J. Allergy Clin. Immunol. 100:1.

37

Romagnani, S. 1998. The Th1/Th2 paradigm and allergic disorders. Allergy 53:12.

38

Eckl-Dorna, J. and Niederberger, V. 2013. What is the source of serum allergenspecific IgE? Curr. Allergy Asthma Rep. 13:281.

39

Wambre, E., James, E. A. and Kwok, W. W. 2012. Characterization of CD4+ T cell subsets in allergy. Curr. Opin. Immunol. 24:700.

25 40

Lloyd, C. M. and Saglani, S. 2013. T cells in asthma: influences of genetics, environment, and T-cell plasticity. J. Allergy Clin. Immunol. 131:1267.

41

Wilke, C. M., Bishop, K., Fox, D., et al. 2011. Deciphering the role of Th17 cells in human disease. Trends Immunol. 32:603.

42

Souwer, Y., Szegedi, K., Kapsenberg, M. L., et al. 2010. IL-17 and IL-22 in atopic allergic disease. Curr. Opin. Immunol. 22:821.

43

Stewart, C. A., Metheny, H., Iida, N., et al. 2013. Interferon-dependent IL-10 production by Tregs limits tumor Th17 inflammation. J. Clin. Invest. 123:4859.

44

Jabeen, R. and Kaplan, M. H. 2012. The symphony of the ninth: the development and function of Th9 cells. Curr. Opin. Immunol. 24:303.

45

Breiman, L. 1996. Bagging Predictors. Mach. Learn. 24:123.

46

Lu, T., Hu, Y. H., Tsai, C. F., et al. 2016. Applying machine learning techniques to the identification of late-onset hypogonadism in elderly men. SpringerPlus 5:729.

47

Mita, T. 2007. AdaBoost and its application to face detection. CVIM tutorial series. Comput. Vis. Im. Med. 42:265.

48

Zhu, W., Chen, H., Zhao, H., et al. 2016. Automatic Three-dimensional Detection of Photoreceptor Ellipsoid Zone Disruption Caused by Trauma in the OCT. Sci. Rep. 6:25433.

49

Zhao, Y., MacCormick, I. J., Parry, D. G., et al. 2015. Automated Detection of Vessel Abnormalities on Fluorescein Angiogram in Malarial Retinopathy. Sci. Rep. 5:11154.

26 Figure legends

Figure 1. Experimental protocol. SLIT was performed for 2 years beginning in June 2006. Blood sampling was performed immediately before and after SLIT. CD4+ T cells, basophils, cDCs, and pDCs were recovered from PBMCs. Serum levels of IgE, IgG, and cytokines were measured as described in the Methods section. Gray bars indicate the pollen seasons.

Figure 2. Changes in severity scores following SLIT. Severity scores were evaluated based on patient allergy diaries as well as JRQLQ1 during the pollen season in each year (A; n = 142–193). The efficacy of SLIT was determined each year in comparison with the severity score in 2006 (B; n = 142–173). Changes in severity scores in HR and NR patients from the 1st to 3rd year were calculated and expressed as the mean ± SEM with individual data (C). ##P < 0.01, ###P < 0.001 (n = 33–34, Mann–Whitney’s U-test).

Figure 3. Effects of SLIT on cellular populations. Populations of CD4+ T cells, basophils, cDCs, and pDCs in PBMCs of HR and NR patients before and after SLIT were analyzed as described in the Methods section and Supporting Information Fig. 2 and expressed as the mean ± SEM with individual data (n = 33–34).

Figure 4. Effects of SLIT on serum factors. Total and pollen-specific IgE, specific/total IgE ratio (A), pollen-specific IgG subclasses (B), and IL-12p70 (C) in the sera of HR and NR patients were measured as described in the Methods section and expressed as the mean ± SEM (n = 25–34). *P < 0.05, **P < 0.01, ***P < 0.001 (Wilcoxon signed-rank test); #P < 0.05, ###P < 0.001 (Mann–Whitney’s U-test).

Figure 5. The pre-therapy serum cytokine levels shown in Fig. 4 and Supporting Information

27 Fig. 3 along with those of mid-range patients were processed with AdaBoost. Each row in the left-hand panel shows a weak learner represented by a combination of the name of an explanatory attribution; Tt, its threshold; vt, the direction of inequality, and the weight for the majority decision; wt. Each column in the right-hand panel corresponds to a subject. The top rows of the columns represent the true classification of the subjects, where the red, blue, and green cells represent NR, HR, and mid-range patients, respectively. Each row in those columns shows the prediction results, F(x), for the strong learner generated in each iteration round by AdaBoost. In the t-th row, the strong learner was constructed using the first t of weak learners, and the red and blue cells respectively show whether the value of F(x) is positive (predicted as NR) and negative (predicted as HR) (A). The resulting positive NR and mid-range patients predicted as NR by the first-round AdaBoost were further processed with a second-round AdaBoost (B). The red and green cells indicate whether the value of F(x) was positive (predicted as NR) or negative (predicted as mid-range), respectively. The numbers in the right column show the cumulative accuracy, P1~t, attained using the first to the t-th weak learner.

Figure 6. Clustering and correlation analysis of serum factors. Clustering of pre-treatment (upper) and post-treatment (lower) serum cytokine levels was performed (A, left panels). Each row and line indicates an individual patient and the indicated cytokine, respectively. The dendrogram attached to the left of the pre-treatment profiles shows the clustering process. With this order of cytokines in both the vertical and horizontal axes, a correlation matrix among cytokines was constructed (A, right panels). The HR and NR data are shown in the lower left and upper right panels, respectively. Clusters in which the correlation coefficients of the HR data were greater than 0.7 are indicated by bold squares. Differences in the averages of correlation coefficients in the TRAIL, SDF-1α, M-CSF, β-NGF, IL-2Rα

28 and MIF (B); IL-12p40, IL-1α, TNF-β and LIF (C); FGFbasic and IL-8 (D); IFN-, IL-4, GCSF, IL-10, IL-5, TNF-α, IL-13, IL-7, MCP-1, IL-15, IL-6, IL-2, IL-1Rα, IL-1β and eotaxin (E); IFN-α2, IL-16, MCP-3, and IL-3 (F); and GM-CSF and IL-12p70 (G) clustering groups in HR (dark bars) and NR (light bars) patients were determined using samples obtained before (Pre) and after (Post) SLIT and expressed as the mean (D, G; n = 2) or mean ± SEM (B, C, E, F; n = 4–15). ***p < 0.001 (Wilcoxon signed-rank test), #p < 0.05, ##p < 0.01, ###p < 0.001 (Mann–Whitney’s U-test).

29 Figure 1.

30 Figure 2

31 Figure 3

32 Figure 4

33 Figure 5

34 Figure 6