TensorPLS: an R package for exploring molecular and temporal patterns in longitudinal omics data using PLS-DA
Longitudinal omics datasets combine high-dimensional repeated measurements with missing observations, complicating supervised modelling and temporal interpretation. We present TensorPLS, an R package integrating missing-data handling, matrix-based partial least squares discriminant analysis (PLS-DA), structured variable importance, subject-level nested cross-validation and downstream functional enrichment. Data are represented as subjects x features x time tensors. Missing values are handled by Tucker3 reconstruction or feature-time mean imputation, after which tensors are unfolded along the subject mode for PLS-DA. Cumulative variable importance in projection (VIP) scores are mapped to feature-time coordinates and summarised across time and cumulative PLS models. During nested cross-validation, imputation, component tuning, VIP-based feature selection and refitting are performed within training partitions. Full-cohort models are fitted separately for descriptive interpretation. We evaluated TensorPLS on two public longitudinal transcriptomic datasets using five-by-five nested cross-validation. In GSE68310 (109 subjects), Tucker3 reconstruction yielded Q2Y = 0.670, AUROC 0.971 and balanced accuracy 0.931. In GSE182616 (66 subjects), feature-time mean imputation yielded Q2Y = 0.294, AUROC 0.827 and balanced accuracy 0.756. Predictive metrics were calculated from pooled outer held-out predictions. TensorPLS supports predictive assessment, functional enrichment of selected genes and temporal interpretation of model-assigned feature importance.