scTREND: An annotation-free single-cell time-resolved and condition-dependent hazard model

Prognosis in cancer and other complex diseases is shaped by heterogeneous cell states and clinical or genetic contexts whose effects change over time. Yet most survival analyses assume predefined cell types and time-invariant covariates, and cohorts pairing single-cell transcriptomes with outcomes are limited, obscuring within-type heterogeneity, time-varying risk, and condition-specific effects. We developed scTREND, an annotation-free, time-resolved, condition-dependent hazard model that learns variational single-cell embeddings, deconvolves bulk or spatial samples without labels, and fits a conditional piecewise-constant hazard model to estimate cell-level hazard coefficients across discrete time bins and conditions. scTREND enables cell-level risk attribution without prior cell-type annotations, dynamic risk modeling over time, and mutation- or treatment-specific risk assessment. In simulations, scTREND recovered ground-truth temporal and condition-specific coefficients with strong concordance and improved prediction over an existing method. In melanoma, scTREND identified subpopulations with risk amplified in BRAF-mutant tumors; in COVID-19, cytotoxic T-cell subpopulations with early-to-late risk reversal and pathways associated with prognosis under remdesivir; and in spatial transcriptomics of clear cell renal carcinoma, spatial regions whose prognostic relevance shifts over time. Across diseases and modalities (bulk RNA-seq and spatial transcriptomics), scTREND provides a time-resolved, condition-aware link between single-cell states and clinical outcomes and is implemented in Python (GitHub:https://github.com/R301Carbine/scTREND).