A Machine Learning Framework for Screening Plasma Cell-Associated Feature Genes to Estimate Osteoporosis Risk and Treatment Vulnerability

Osteoporosis, in which bones become fragile owing to low bone density and impaired bone mass, is a global public health concern. Bone mineral density (BMD) has been extensively evaluated for the diagnosis of low bone mass and osteoporosis. Circulating monocytes play an indispensable role in bone destruction and remodeling. This work proposed a machine learning-based framework to investigate the impact of circulating monocyte-associated genes on bone loss in osteoporosis patients. Females with discordant BMD levels were included in the GSE56815, GSE7158, GSE7429, and GSE62402 datasets. Circulating monocyte types were quantified via CIBERSORT, with subsequent selection of plasma cell-associated DEGs. Generalized linear models, random forests, extreme gradient boosting (XGB), and support vector machines were adopted for feature selection. Artificial neural networks and nomograms were subsequently constructed for osteoporosis diagnosis, and the molecular machinery underlying the identified genes was explored. SVM outperformed the other tuned models; thus, the expression of several genes (DEFA4, HLA-DPB1, LCN2, HP, and GAS7) associated with osteoporosis were determined. ANNs and nomograms were proposed to robustly distinguish low and high BMDs and estimate the risk of osteoporosis. Clozapine, aspirin, pyridoxine, etc. were identified as possible treatment agents. The expression of these genes is extensively posttranscriptionally regulated by miRNAs and mA modifications. Additionally, they participate in modulating key signaling pathways, e.g., autophagy. The machine learning framework based on plasma cell-associated feature genes has the potential for estimating personalized risk stratification and treatment vulnerability in osteoporosis patients.