Learning interpretable compositional features associated with continuous covariates

IsoVita learns an interpretable covariate-dependent log-ratio direction for compositional data, approximated by a log ratio of geometric-mean proportions between two component groups.

• Why it matters: Continuous covariates like time make it hard to learn interpretable compositional features and correctly account for how those features are constructed during inference.
• What they did: The framework estimates a dominant direction in log-ratio coordinates, tests abrupt versus gradual change using change point and spline methods, and uses a selection-aware permutation procedure that repeats feature construction.
• The result: In transcript-level simulations, it maintained low empirical false discovery proportions with power to detect differences and improved change point accuracy as temporal signals strengthened, and it found widespread temporal isoform composition shifts in two human brain bulk RNA-seq datasets.

AI summary of the abstract · may contain errors · not medical advice · read the paper.