Molecular Systems Biology Information Systems

xDecoder unlocks the potential of genomic foundation models for few-shot personal gene expression prediction

Li, Shumin, Luo, Ruibang, Huang, Yuanhua
Springer Nature · 2026-11-02 · DOI 10.1038/s44320-026-00238-1

Abstract

Few-shot personal gene expression prediction from DNA sequence was improved by xDecoder, while zero-shot analyses revealed a cross-locus transfer bottleneck linked to missing regulatory-state information. Few-shot personal gene expression prediction from DNA sequence was improved by xDecoder, while zero-shot analyses revealed a cross-locus transfer bottleneck linked to missing regulatory-state information. Cross-individual prediction for seen genes was improved by xDecoder trained on personal DNA-RNA pairs using gLM and S2F embeddings. Sequence-based modeling enabled potential discovery of candidate cis-regulatory variants at loci not applicable to gene-specific statistical models. Zero-shot transfer to unseen loci was generally unreliable, with a weak signal observed in more cis-predictable genes. Increasing the number of training individuals improved seen-gene performance, whereas zero-shot performance at unseen loci remained near zero. Matched ATAC-seq improved unseen-locus prediction, while ATAC-only controls indicated that measured chromatin accessibility largely explains the gain. Large-scale genomic language models (gLMs) hold promise for modeling gene regulation, yet their ability to capture personal gene expression variations remains unresolved. We developed xDecoder, a unified decoding framework that utilizes gLMs and sequence-to-function (S2F) embeddings to learn how personal genetic variation shapes gene expression from paired genome-transcriptome data. Compared to the pretrained genomic models, xDecoder with personalized DNA–RNA training makes cross-individual prediction tractable for seen genes in a few-shot setting. However, zero-shot prediction at unseen loci remains unreliable and gene-dependent, revealing a cross-locus transfer bottleneck of current sequence models. Experiments incorporating individual-level chromatin accessibility suggested that regulatory-state information important for unseen-locus prediction is not fully captured by current DNA-only models. Overall, these results highlight the potential utility of the few-shot setting, the limitations of DNA-only models, and point toward multi-omic, variant-aware frameworks as a promising direction for building personalized regulatory models.

Read at publisher →