Hoike: A Joint-Embedding Predictive Architecture for Transcriptome Data Generation with Diffusion Models
In biomarker discovery, access to sufficient quantities of condition-specific transcriptomic data is often limited by cohort size, privacy concerns, and domain shift between normal and condition populations. Generative modeling can augment scarce cohorts and probe distributional transitions. Furthermore, synthetic transcriptome generation can support differential expression analyses, machine…
Biomarker discovery often faces challenges due to limited availability of condition-specific transcriptomic data. Issues such as small cohort sizes, privacy concerns, and differences between normal and condition populations hinder progress in this field. Generative modeling techniques can help overcome these limitations by creating synthetic transcriptomes.
These synthetic datasets support various applications, including differential expression analysis, machine learning, privacy-preserving data sharing, benchmarking, and hypothesis generation.
Introducing Hoike, a novel framework that integrates a Cross-Domain Joint-Embedding Predictive Architecture (JEPA) with a latent diffusion model. Hoike is designed to generate disease-specific bulk transcriptomes from normal reference contexts. In this architecture, normal tissue profiles serve as continuous conditioning signals, while the model learns disease-associated shifts in latent space. The reconstructed gene-level expression is expressed in log2(TPM+1) space.
Hoike's implementation allows for paired normal-condition training, tissue-aligned conditioning, and constrained non-negative decoding to ensure biologically valid outputs. The architecture, objective design, and evaluation protocol were thoroughly described and tested across various GTEx-derived normal references and multiple TCGA condition cohorts. This serves as the technical foundation for the Hoike framework and its reproducible analysis workflow.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
