Visualizing viral replication organelles: conventional electron microscopy reveals distinct vesicle formation and membrane dynamics in Flaviviruses and Coronaviruses
Dr. Le Zheng (Post-doctoral Fellow)
[Supervisor: Professor Tao Ni]
The formation of membrane-bound replication organelles (ROs) is crucial for positive-sense RNA viruses. Comparing their ultrastructure reveals key differences in viral replication strategies. Here, we employed conventional room-temperature electron microscopy (RT-EM) to compare RO biogenesis in cells infected with the Zika virus (Flaviviruses) or SARS-CoV-2 (Coronaviruses). While cryo-electron microscopy offers high resolution, RT-EM provides superior membrane contrast, a wider field of view, and greater practicality for statistically robust comparative cytopathology. To achieve optimal structural preservation, infected cells were processed using high-pressure freezing and freeze-substitution, followed by epoxy resin embedding, ultramicrotomy, and heavy-metal staining for high-contrast imaging. Our results show distinct membrane-remodelling pathways: Zika virus replication was associated with the endoplasmic reticulum and formation of single-membrane vesicles (SMV) within its lumen. In contrast, SARS-CoV-2 replicon cells induced the formation of double-membrane vesicles (DMV), predominantly derived from the ER-Golgi intermediate compartment. This study demonstrates that advanced sample preparation combined with conventional RT-EM is a powerful approach for comparative ultrastructural analysis. It effectively visualizes the divergent host membrane-hijacking mechanisms employed by Flaviviruses and Coronaviruses, providing a complementary structural framework that deepens understanding of viral replication organelle biology.

xDecoder unlocks the potential of genomic foundation models for few-shot personal gene expression prediction
Dr. Shumin Li (Post-doctoral Fellow)
[Supervisor: Professor Yuanhua Huang & Professor Ruibang Luo]
Large-scale genomic language models (gLMs) hold promises 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-tofunction (S2F) embeddings to learn how personal genetic variation shapes gene expression from paired genome–transcriptome data. Compared to the pre-trained genomic models, we first demonstrate that xDecoder with personalized DNA–RNA training makes it tractable for seen genes, as a few-shot prediction. Moreover, while all current models fail in the zeroshot setting of unseen loci—revealing a fundamental limitation of sequence models, we show that this bottleneck can be partially overcome by integrating individual-level chromatin accessibility. 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.

All are welcome.