An exploratory exome-wide machine learning analysis identifies candidate host gene signatures associated with Long...

Genetic factors have been suggested as modifiers of vulnerability to postCOVID-19 sequelae, referred to as Long COVID (LC). We hypothesize that LC may involve central nervous system (CNS)-related mechanisms, influenced by neuroinflammatory, autoimmune, viral mechanisms, and genetic factors. In this work, we used whole-exome sequencing in conjunction with a machine learning-based prioritization framework to investigate the connection between LC and host genomic variation.

Our patient group included 312 individuals previously infected with SARS-CoV-2 enrolled in two public hospitals of Vitoria city, Brazil, between November 2020 and July 2023. After rigorous quality control in accordance with reference guidelines, the exome data revealed 651,652 variants in our cohort. To rank candidate variants, a supervised machine learning framework combining Recursive Feature Elimination (RFE) and XGBoost was implemented. Five variants were found to be statistically significant after Benjamini-Hochberg false discovery rate (FDR) correction in subsequent logistic regression analyses that were adjusted for age, sex, and principal components of ancestry under an additive genetic model.

The variant at LERFS - rs200443822 was associated with increased odds of LC (OR = 6.21, 95% CI 2.98-12.91, FDR-adjusted p < 0.01). Similarly, the variant at PNKD - rs1870125 (OR = 2.33, 95% CI 1.52-3.57, FDR-adjusted p < 0.01) and LIPA - rs1051338 (OR = 2.87, 95% CI 1.77-4.65, FDR-adjusted p < 0.01) showed increased odds. In contrast, variants at chromosome 10 (rs7912524 - HK1 and LAMB4 - rs1735499) were associated with reduced odds of LC, with ORs ranging from 0.38 to 0.50 (all FDR-adjusted p < 0.01).

Our findings do not support single-gene causal effects; rather, they are consistent with the notion that common variants may collaboratively influence inter-individual variations in LC manifestations within a more extensive polygenic framework. Clinical features such as fatigue, pain, anosmia, dysautonomia, and cognitive impairment may reflect interactions between host genomic background and clinical or demographic factors. Overall, this study provides a hypothesis-generating integrative framework for investigating host genetic contributions to LC in an underrepresented admixed population. Targeted functional studies are important to ascertain the biological significance and translational applicability of these findings.