RNA-SpecFlow: RNA-Conditioned Spectral Flow for Histology Synthesis

Abstract

RNA-conditioned histology synthesis models the range of tissue appearances compatible with a transcriptomic representation. We present RNA-SpecFlow, a continuous-latent FlowAR architecture that applies molecular conditioning at adjacent-scale feature transitions. A Biological Latent Adapter maps RNA representations into the generator's condition space. A shared spectral fusion module separates aligned features into complementary low- and high-frequency bands, modulates each band independently with RNA, and combines them through residual fusion. The generator uses five latent grids of sizes 1, 2, 4, 8, and 16, with per-scale velocity flow matching and an auxiliary objective for retaining the condition in backbone features. We accompany this architectural specification with an analysis of archived outputs from five TCGA cancer cohorts, including 500 saved real/generated tile pairs. Nuclear analysis yields a concrete measurement finding: increased detection counts can coincide with reduced segmentation agreement. On 12 public images with 303 annotated nuclei, twofold enlargement increases HoVer-Net predictions from 287 to 415 while mean panoptic quality changes from 0.673 to 0.458. Switching analysis pipelines also reverses the direction of real/generated count differences in two cohorts. RNA-SpecFlow provides a design for repeated frequency-specific molecular conditioning, and the accompanying experiments establish the sensitivity of morphology-based evaluation to detector inputs and analysis choices.