Glycomics meets artificial intelligence – Potential of glycan analysis for identification of seropositive and seronegative rheumatoid arthritis patients revealed
Introduction
Rheumatoid arthritis (RA) is a chronic autoimmune disorder characterized by joint inflammation, swelling and synovium destruction [1,2]. In RA, the immune complexes are present in serum including autoantibodies antiCCPs (anti-cyclic citrullinated peptide antibodies) and RFs (rheumatoid factors) for seropositive RA patients (S+ RA). In seronegative RA patients (S− RA, about one-third of RA patients) the serological level of these markers is very low or can be absent. Moreover, the disease of S− RA patients is more severe comparing to S+ RA patients, when taking into account disease activity and joint damage [3]. Treatment of S− RA patients can be successful if identified at an early stage and thus, novel and reliable methods applicable for diagnostic and prognostic purposes of S− RA disease should be available [3]. Additionally, seropositive and seronegative RA disorder although similar in the phenotype might represent two different diseases with different pathogenesis, disease development and prognosis [2].
Glycomics is an emerging scientific field with a huge impact on biomedicine [4]. Glycoconjugates are in vivo involved in a diverse range of cellular events and changes in their structure are associated with different disease development and progression. RA associated changes of a biantennary complex type N-glycan on Asn297 in the Cy2 domain on each Fc constant fragment of an antibody involve degalactosylation and desialylation. As a result the underlying N-acetylglucosamine (GlcNAc) residues are more accessible to mannose-binding protein, activating the complement via an alternative lectin pathway [5]. Moreover, GlcNAc inhibits the binding of C1q (part of a complex activating the complement via a classical pathway) [6]. AntiCCPs were shown to acquire a pro-inflammatory glycoprofile (removal of galactose residues) three months prior to RA diagnosis. Changes in fucosylation prior to the onset of RA were also observed [7]. The glycans also control the thermodynamic stability and maintain the quaternary structure of the Fc domain. Fcγ receptors, expressed on leukocytes (including macrophages, NK cells, lymphocytes, eosino- and neutrophils) require IgG Fc glycans for optimal binding to all three classes of their receptors (RI-III, e.g. CD64, CD32 and CD16) [8]. For a reliable glycan analysis during RA development and progression, expensive and time-consuming methods are employed including capillary electrophoresis or MALDI-TOF-MSn and/or their combination [9].
An alternative to instrumental-based analytical approach is to apply a panel of lectins as natural glycocode decipherers for glycan analysis [10]. Recently we published a series of papers focused on an alternative IgG glycoprofiling of RA patients serum samples [11] and isolated IgGs [12] using electrochemical biosensors integrated with lectins. The aim of this study was to investigate if glycan analysis with or without standard immunoassays can be applied to successfully discriminate between non-RA and S− RA patients. Additionally the same approach was also tested to discriminate between non-RA and S+ RA patients. Data mining was performed using artificial neural network (ANN) and L1 regularized logistic regression model. Glycan analysis of serum samples or antibodies isolated from serum samples was performed using lectin microarray (MA) or enzyme-linked lectin-binding assay (ELLBA).
Recently, machine learning algorithms (MLA) and cloud computing gained an attention in healthcare services for automated and more precise diagnostics with a potential to decrease the financial burden in the healthcare sector. ANN, applicable to pattern recognition tasks, diagnostics and data classification using a learning process [13], was successfully used for RA diagnostics using a panel of inflammatory cytokines expressed in serum samples [14]. It was also already shown that ANN-based algorithms could be used to quantify joint space width for RA assessment [15]. A novel method for predicting the clinical response to a chimeric anti-TNF monoclonal antibody drug infliximab based on machine-learning algorithm was also recently developed [16], together with hybrid clustering-classification ANN for solving diagnostic tasks of RA [17]. The main aim of the study was to combine machine learning algorithms with glycomic studies applied for analysis of human sera of RA patients. The ANN results are summarized in a form of 2 × 2 confusion matrix with all parameters defined in Fig. S1.
Section snippets
Human serum samples
Serum samples of 31 seropositive patients (5 males, 26 females, mean age 60.9 ± 12.8 years) and 16 seronegative patients (3 males, 13 females, mean age 59.9 ± 12.6 years) were included in the study, as well as 53 gender and age-matched non-RA controls. The control group (6 males, 47 females, mean age 57.6 ± 11.3 years) co-morbidities were of non-autoimmune origin, consisting mainly of hypertension and osteoporosis. RA patients were treated with various immune suppressive drugs (non-steroidal
Results
Five lectins including RCA, SNA, GSL-II, LEL and LOL were considered for glycan analysis. GSL-II, LEL and LOL lectins were excluded from the study, because of the lack of sensitivity of analysis. For example, LOL lectin was not able to distinguish between S− RA or S+ RA patients and controls at all (sensitivity of detection of 0%). Lectin LEL offered the sensitivity of detection of 3.2% (non-RA vs. S+ RA patients) and sensitivity of detection of 0% (non-RA vs. S− RA patients). GSL-II lectin
Non-RA vs. S+ RA patients
Computer-aided diagnostics is of great interest with an extensive effort put into its development. The greatest significance is to lower the mortality caused by misdiagnosis of highly severe, e.g. oncologic disorders, such as breast cancer [21]. The learning process of ANN can be applied to digital images (MRI, ultrasound, etc.), results from laboratory blood tests and also historical data. Here, we can conclude that ANN applied on data acquired from standard immunoassays correctly identified
Conclusions
The results provided in this study strongly indicate the importance of using machine-learning algorithms, capable of processing large density of clinically relevant data. We can make the following conclusions: 1) It is better to use whole serum samples compared to isolated IgG for glycan analysis. 2) The preferred choice for glycan analysis is ELLBA format compared to lectin MA. 3) Glycan analysis performed with protein A adsorbed on the ELLBA plates provides better results compared to ELBBA
Acknowledgements
Financial support received from the Slovak Scientific Grant Agency VEGA 2/0137/18 and Slovak Research and Development Agency APVV 14-0753 is acknowledged. The research received funding from the European Research Council (No. 311532). This publication was made possible by NPRP grant no. 6-381-1-078 from the Qatar National Research Fund. This publication is the result of the project implementation: Centre for materials, layers and systems for applications and chemical processes under extreme
References (40)
Terminal sugars of Fc glycans influence antibody effector functions of IgGs
Curr. Opin. Immunol.
(2008)- et al.
Cardiac ScoreCard: a diagnostic multivariate index assay system for predicting a spectrum of cardiovascular disease
Expert Syst. Appl.
(2016) - et al.
Machine learning techniques for breast cancer computer aided diagnosis using different image modalities: a systematic review
Comput. Methods Prog. Biomed.
(2018) - et al.
A method to identify trace sulfated IgG N-glycans as biomarkers for rheumatoid arthritis
Nat. Commun.
(2017) - et al.
A machine learning model for improving healthcare services on cloud computing environment
Measurement
(2018) - et al.
Automated diagnosis of focal liver lesions using bidirectional empirical mode decomposition features
Comput. Biol. Med.
(2018) - et al.
Breast cancer data analysis for survivability studies and prediction
Comput. Methods Prog. Biomed.
(2018) - et al.
Classification of cancer cells using computational analysis of dynamic morphology
Comput. Methods Prog. Biomed.
(2018) - et al.
Discovering disease-associated genes in weighted protein–protein interaction networks
Physica A
(2018) - et al.
Machine learning techniques for medical diagnosis of diabetes using iris images
Comput. Methods Prog. Biomed.
(2018)
Association of rheumatoid arthritis and primary osteoarthritis with changes in the glycosylation pattern of total serum IgG
Nature
Pentraxin 3 in serum and synovial fluid of patients with rheumatoid arthritis with and without autoantibodies
Scand. J. Rheumatol.
Rheumatoid arthritis: seronegative and seropositive RA: alike but different?
Nat. Rev. Rheumatol.
Nanotechnology in glycomics: applications in diagnostics, therapy, imaging, and separation processes
Med. Res. Rev.
Glycosylation changes of IgG associated with rheumatoid arthritis can activate complement via the mannose-binding protein
Nat. Med.
Anti-citrullinated protein antibodies acquire a pro-inflammatory Fc glycosylation phenotype prior to the onset of rheumatoid arthritis
Ann. Rheum. Dis.
The impact of glycosylation on the biological function and structure of human immunoglobulins
Annu. Rev. Immunol.
Characterization of IgG glycosylation in rheumatoid arthritis patients by MALDI-TOF-MSn and capillary electrophoresis
Anal. Bioanal. Chem.
Altered glycan accessibility on native immunoglobulin G complexes in early rheumatoid arthritis and its changes during therapy
Clin. Exp. Immunol.
Ultrasensitive impedimetric lectin biosensors with efficient antifouling properties applied in glycoprofiling of human serum samples
Anal. Chem.
Cited by (24)
Artificial intelligence in the analysis of glycosylation data
2022, Biotechnology AdvancesCitation Excerpt :Burgeoning glycomics datasets (e.g., Ranzinger et al., 2010) have been generated with these advanced technologies. These data, both new datasets and those catalogued in databases, provide opportunities for the use of Artificial intelligence (AI) for analyzing experimental glycan data (Chocholova et al., 2018). Another major roadblock to exploring glycomes with AI stems from challenges in database coordination.
Glycans in blood as biomarkers for forensic applications
2020, TrAC - Trends in Analytical ChemistryCitation Excerpt :Since the use of machine learning algorithm has helped to improve the diagnostic capability of biomarker for diseases [63], the development of prediction models to account for these coexisting relationships, as well as relationships with pathophysiological manifestations can be further explored to fully realise the potential of using glycan analysis in blood and other bodily fluids for profiling multiple physiological/lifestyle parameters.
Precision medicine in the care of rheumatoid arthritis: Focus on prediction and prevention of future clinically-apparent disease
2020, Autoimmunity ReviewsCitation Excerpt :It has been known for more than two decades that alteration in the glycosylation status of antibodies is associated with autoimmunity [40]. Several studies have demonstrated that the Fc portion of IgG show different glycosylation patterns in patients with autoimmune diseases when compared to healthy individuals (HI) [41–44]. In HI, the glycosylation site in the constant region is typically fully glycosylated, while in certain autoimmune conditions, IgG becomes increasingly deglycosylated.
Big Data and artificial intelligence: Will they change our practice?
2020, Revue du Rhumatisme (Edition Francaise)Big Data and artificial intelligence: Will they change our practice?
2020, Joint Bone Spine