Deep Learning Unveils Characteristics of Immune B Cell Aging in >17,000 Samples and Identifies Changes in the Immunoglobulin Repertoire Distribution and the Development of Monoclonal Gammopathy

Immune cell aging is characterized by increased dysfunction of both innate and adaptive immune responses, response to pathogens, and greater morbimortality. Most studies have focused on the T cell repertoire, but few have examined the effect of aging on the B cell repertoire in a large cohort. Here, we leveraged mass spectrometry (MS) and supervised deep learning to identify changes in immunoglobulin diversity and characterize B cell diversity in association with age and race in a large cohort of healthy individuals from several centers in the US and South Africa.

We used raw quantitative matrix-assisted laser desorption ionization-time of flight data from 17,151 serum samples obtained with EXENT technology (Binding Site, part of Thermo Fisher Scientific). Participants were enrolled in one of five multicenter cohort studies between 1994 and February 2024: 4760 samples from the PROMISE screening study; 7274 from the Mass General Brigham Biobank (MGBB); 1985 from the PROMISE South Africa screening study; 2225 samples from the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer screening study; and 907 from the PCROWD study. 5386 (31%) were from individuals who self-identified as Black/African American, 9684 (56%) as White/European, 2081 (12%) as other races or did not disclose; 9872 (58%) as female and 6793 (40%) as male (486 (3%) did not disclose); median age was 57 (IQR 46.4-65.0). We identified healthy individuals with no monoclonal gammopathy (MG) (no peak > 0.015 g/L and normal FLC ratio (0.26 to 1.65)) and individuals with quantifiable peaks (MG of Indeterminate Potential, MGIP group: 0.015-0.2 g/L; MG of Undetermined Significance, MGUS group > 0.2 g/L). We trained deep neural networks to predict the age, race, and sex of healthy individuals based on immunoglobulin distribution assessed by MS. We then studied how race, sex, and MG influenced those predictions and used feature explanation methods to explore how they were predicted.

Of the 17,151 serum samples, we identified 9159 samples in the healthy group. 8243 were used for training models and 916 as test samples. Deep learning models were able to predict a humoral age (h-Age, r = 0.37) and race (ROC-AUC = 0.74) but not sex (ROC-AUC = 0.56). Feature exploration revealed that h-Age was explained by the presence of subtle peaks below the limit of detection, but also a skewing in the proportion of polyclonal heavy mass kappa light chain-associated IgM and IgA. Previous studies had found that this kappa fraction reflects B cells with certain germline alleles, predominantly encountered in patients with chronic lymphocytic leukemia and MM (Barnidge et al. J. Proteome Res. 2015). Race prediction was mainly explained by the kappa/lambda light chain ratio of IgM and IgA isotypes. Patients with MG (MGIP, n = 3440; MGUS, n = 3947) had a significantly higher h-Age compared to healthy of the same chronological age (p < 0.0001), suggesting accelerated h-Age: young (<50 years) patients with MGUS were predicted to have the same h-Age as older (>65 years) healthy. h-Age did not change when masking peaks from the model’s vision of the immunoglobulin distribution (r = 0.89-0.94), confirming an overlap of features explaining h-Age. Race prediction showed similar accuracy in patients with MG, whether peaks were masked from the models’ vision (ROC-AUC 0.70-0.81). For 63 patients with MGUS, we have matched MS and single-cell B cell receptor (BCR) sequencing from bone marrow plasma cells. These multiomic data showed that BCR diversity (Chao1 index) decreased with h-Age (r = -0.52, p < 0.05), consistent with our MS findings.

This comprehensive analysis is the first effort, to our knowledge, to leverage MS coupled with deep learning on a large cohort of more than 17,000 samples to explore how immunoglobulin repertoire, age, and race correlate in healthy individuals. Deep learning was able to detect differences correlated with race and age by observing exclusively healthy individuals. Those differences could still be detected with similar accuracy in patients with MG, and humoral age was accelerated in patients with MG and correlated with restricted clonal diversity even when masking monoclonal peaks, supporting a close correlation between immune aging, polyclonal repertoire, and MG development, even at its earliest stages. This approach could offer an easy, fast, and affordable approach to assess immunosenescence and changes in B cell diversity across other B cell disorders.