A HEK293T-derived explainable CatBoost signature for estimating HCoV-OC43 viral burden from host transcriptomes

초록

While HCoV-OC43 provides a tractable experimental model, it should be recognized as a limited model for studying generalized betacoronavirus host programs rather than a direct surrogate for SARS-CoV-2. We developed a CatBoost regression model using single-cell RNA sequencing data from 12,980 HEK293T cells to predict viral burden at single-cell resolution (full model 5-fold cross-validated R-2 = 0.870). By leveraging CatBoost's native feature importance metrics, we identified 10 core host genes whose expression was most predictive of infection intensity, with TPI1 ranking as the single most important predictor. SHAP (SHapley Additive exPlanations) values were subsequently utilized to interpret the directional impact of these core genes on individual cellular predictions. A compact model trained exclusively on these 10 genes retained strong predictive performance on a held-out test set (R-2 = 0.860, n = 2, 598). When applied to independent bulk RNA-seq data, the compact model showed strong concordance with experimental viral production in OC43-infected MRC-5 lung fibroblasts (Pearson r = 0.93, P = 0.022, 95% bootstrap CI [0.88, 1.00], Spearman rho = 1.00; n = 5 time points). When applied to a non-viral stress dataset of K562 cells treated with tunicamycin, the model produced predicted viral scores that were elevated relative to the healthy baseline-consistent with severe stress-induced transcriptional perturbation-but remained distinguishable from OC43-high-infection cells, while showing partial overlap with low-infection cells. Effect size analysis using Cliff's Delta confirmed that tunicamycin-stressed cells were moderately separated from both low-infection (d = - 0.44, P < 0.001) and high-infection (d = - 1.00, P < 0.001) groups, demonstrating that model distinguishes generic stress from high-burden viral states, while partially overlapping with low-burden infection. Ultimately, this pipeline offers a systematic approach to identifying interpretable transcriptional signatures of viral replication across heterogeneous cellular states.

키워드