TriTower-m6Am: a triple-tower heterogeneous deep learning architecture integrating semantic, sequential, and structural information for mRNA N6,2'-O-dimethyladenosine site prediction

N6,2 -O-dimethyladenosine (m6Am) is a cap-proximal modification found in eukaryotic mRNA, and predicting its location can provide insight into how cells regulate mRNA stability and translation. However, existing computational predictors often rely on a single sequence representation, failing to capture the semantic, sequential, and structural signals present in RNA sequences.

Researchers have now developed TriTower-m6Am, a triple-tower deep learning architecture that integrates three distinct representations to improve m6Am site prediction. These representations include semantic (RNA-FM with BellPooling), sequential (One-Hot BiLSTM), and structural (RGCN with three typed edges). When tested on a set of 640 independent sequences, TriTower-m6Am achieved an AUC of 0.776, a Matthews Correlation Coefficient (MCC) of 0.440, and an SN (Sensitivity) of 0.888.

This represents an 8.8 percentage-point improvement in sensitivity compared to the current state-of-the-art method, DTC-m6Ams. For every 100 real m6Am sites, the new model correctly identifies approximately 9 additional sites that were missed by the previous best approach. The researchers found that the RGCN tower alone provided the strongest single signal, and by using an AUC-weighted ensemble of the three towers, they were able to raise sensitivity from the 0.55-0.76 range of individual towers to an impressive 0.89.

Further analysis revealed that backbone connectivity played a crucial role in capturing the structural signal, accounting for most of the RGCN's contribution. The model's design allows for a clear understanding of how each tower contributes to the final prediction, making it more transparent than black-box models. This design pattern could be applied to other RNA modification site prediction tasks, potentially improving our understanding of RNA modifications and their roles in cellular processes.