Learning interpretable compositional features associated with continuous covariates

Abstract

Compositional variation as a function of continuous covariates, such as time, presents challenges for learning interpretable features and accounting for their construction in subsequent inference. Motivated by temporal change in isoform usage in genomics, we introduce IsoVita, a statistical framework integrating covariate-informed balance learning designed for compositional data with complementary tests for abrupt and gradual compositional variation. IsoVita estimates a dominant covariate-dependent direction in the log-ratio coordinates and approximates it as a log ratio of their geometric mean proportions between two learned component groups, linking statistical evidence to explicit component membership. Change point and spline tests evaluate this interpretable feature, while a selection-aware permutation procedure repeats feature construction and testing under each permutation. In transcript-level simulations, IsoVita maintained low empirical false discovery proportions with power to detect differences and change point accuracy improving as temporal signals strengthened. However, the performance depended on trajectory shape and study design. Applications to two human brain bulk RNA-sequencing datasets identified widespread temporal variation in isoform composition while retaining transcript-level interpretability. IsoVita integrates interpretable compositional feature learning with inference along continuous covariates, with potential applications beyond transcriptomics. IsoVita is available as a PyPI package (https://pypi.org/project/isovita).