Deconformer: efficient cell-type composition profiling of cfRNA across diverse cell types

Resolving cell-type contributions in cell-free RNA (cfRNA) profiles is important for liquid-biopsy research, but profiling many potential cellular sources remains computationally challenging. Here, we present Deconformer, a biologically informed deep-learning framework that integrates pathway knowledge into a Transformer architecture to estimate relative cfRNA contributions across 60 major human cell types. In systematic benchmarks using cross-platform simulations selected for similarity to real plasma cfRNA, Deconformer showed improved robustness from pathway-informed modeling and achieved competitive accuracy with substantially lower computational cost than established deconvolution methods. Applied to real cfRNA cohorts, it recapitulated known gestational trajectories and detected disease-associated composition shifts, supporting its use as a practical framework for large-scale cfRNA profiling.