Multimodal zero-shot learning of previously unseen epitranscriptomes from RNA-seq data.

Precise identification of condition-specific epitranscriptomes is of critical importance for investigating the dynamics and versatile functions of RNA modification under various biological contexts. Existing approaches for predicting condition-specific RNA modification are usually trained on epitranscriptome data obtained from the same condition, which limited their usage, as such data are available only for a small number of conditions due to the technical difficulties and high expenses of epitranscriptome profiling technologies. We present ExpressRM, a multimodal zero-shot learning framework for predicting condition-specific RNA modification sites in previously unseen contexts from genome and RNA-seq data. Different from existing in-condition learning approaches, this method does not rely on matched epitranscriptome data for training, which greatly expands its applicability. On a benchmark dataset comprising epitranscriptomes and matched transcriptomes of 37 human tissues, we demonstrate that ExpressRM can accurately predict epitranscriptomes of previously unseen conditions from their transcriptomes only, and the performance is comparable to existing in-condition learning algorithms that require epitranscriptome data from the same condition. Additionally, the method has the capability of differentiating highly dynamic RNA methylation sites from more static (or house-keeping) ones. With a case study, we show that ExpressRM can uncover N6-methyladenosine RNA methylation sites in glioblastoma using only its RNA-seq data, and unveils novel and previously validated pathological insights. Together, these results suggest that the proposed multimodal zero-shot learning framework can effectively leverage transcriptome knowledge to explore the dynamic roles of RNA modifications in previously unseen experimental setups, providing valuable insights into vast biological contexts where RNA-seq is routinely used but epitranscriptome profiling has not yet been covered.