TriTower-m6Am: A triple-tower heterogeneous deep learning architecture integrating semantic, sequential, and structural information for mRNA N6,2′-O-dimethyladenosine site prediction
- 1School of Information Engineering, Jingdezhen Ceramic University, Jingdezhen, China
Abstract
Background: N6,2′-O-dimethyladenosine (m6Am) is a cap-proximal mRNA modification deposited by PCIF1 (phosphorylated C-terminal domain (CTD)-interacting factor 1) at the first transcribed nucleotide of eukaryotic mRNAs. Knowing where m6Am sites sit would help explain how cells tune mRNA stability and translation, but current predictors typically depend on a single representation and do not jointly model multiple views of the same RNA sequence. Results: We present TriTower-m6Am, combining semantic (RNA-FM, a pretrained RNA foundation model, with BellPooling), sequential (One-Hot BiLSTM, bidirectional long short-term memory), and structural (RGCN, a relational graph convolutional network, over the 41-nt window) representations of each RNA window. On an independent test set of 640 sequences, TriTower-m6Am reaches an area under the ROC curve (AUC) of 0.776 and a Matthews correlation coefficient (MCC) of 0.425, against DTC-m6Am’s AUC = 0.765 and MCC = 0.411, using 4.4 times fewer trainable parameters; at every matched specificity between 0.40 and 0.70 it recovers 1–7 more true sites out of 320 than DTC-m6Am. The AUC-weighted voting keeps every tower’s contribution readable for any individual prediction. Analyses of the graph tower delimit its benefit to the relational aggregation of local backbone adjacency over sequence-derived node features: the edge-type ablation, the node-feature ablation, and a structure-permutation control agree that the specific RNAfold-predicted fold carries no attributable signal, and a probing analysis shows that per-base base-pairing status is largely recoverable from the sequence-derived RNA-FM embeddings (AUC = 0.72 vs. 0.65 from the raw sequence), so the fold annotation is substantially redundant with the semantic representation the ensemble already receives. Conclusions: Combining semantic, sequential, and structural views of the same RNA sequence improves m6Am prediction beyond any single representation, and the ensemble’s voting weights make each prediction auditable without a post-hoc explainer. Whether the triple-tower pattern transfers to other RNA modification prediction tasks remains to be tested in future work. Author summary: Predicting where m6Am modifications occur on messenger RNA is important for understanding how cells regulate transcript stability and translation. Existing computational methods typically encode the RNA sequence in a single way, such as k-mer counts or a one-hot code, and treat the model as a black box that emits a prediction without explaining which features drove it. We built TriTower-m6Am to address both limitations. Our model combines three independent encoders—a pretrained RNA language model for semantic patterns, a bidirectional LSTM for local nucleotide order, and a relational graph convolutional network for the structural fold—and fuses their outputs by AUC-weighted voting, so the contribution of each tower to a given prediction is a readable number rather than a hidden parameter. On an independent benchmark the ensemble achieves ranking quality comparable to the prior best method with 4.4 times fewer parameters, and its voting weights make each tower’s contribution auditable for any individual prediction. A controlled structural ablation shows that linear backbone adjacency, rather than long-range base-pairing, carries whatever signal the graph tower uses, and that randomizing the predicted secondary structures does not degrade performance. This controlled analysis delineates what the graph tower contributes: its benefit lies in the relational encoding of local adjacency over sequence-derived features, which the control retains, rather than in the specific predicted fold. Whether the same triple-tower pattern transfers to other RNA modification site prediction tasks is a question for future work.
Related
- Architecture-Guided Sequential Editing of Distinct Disulfides in Unprotected Peptides.
- DynML-Net: a porcine enteric virus identification network based on protein language models and a dynamic heterogeneous multi-branch architecture
- Heterogeneous primary nucleation determines Aβ42 fibrilstructure.
- Mechanism of the Sequential Trajectory of Allosteric Ring Closure in Human Chaperonin TRiC
- Sequential TPRX1 and HERVL activation delineates a progression toward the totipotent-like state
BibTeX
@article{xiong2026tritowerm6am,
title = {{TriTower-m6Am: A triple-tower heterogeneous deep learning architecture integrating semantic, sequential, and structural information for mRNA N6,2′-O-dimethyladenosine site prediction}},
author = {Kaifeng Xiong and Jianhua Jia},
journal = {PLOS Computational Biology},
year = {2026},
month = nov,
doi = {10.1371/journal.pcbi.1014939},
url = {https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014939},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - TriTower-m6Am: A triple-tower heterogeneous deep learning architecture integrating semantic, sequential, and structural information for mRNA N6,2′-O-dimethyladenosine site prediction
AU - Kaifeng Xiong
AU - Jianhua Jia
PY - 2026
DA - 2026/11/05
JO - PLOS Computational Biology
DO - 10.1371/journal.pcbi.1014939
UR - https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1014939
AB - Background: N6,2′-O-dimethyladenosine (m6Am) is a cap-proximal mRNA modification deposited by PCIF1 (phosphorylated C-terminal domain (CTD)-interacting factor 1) at the first transcribed nucleotide of eukaryotic mRNAs. Knowing where m6Am sites sit would help explain how cells tune mRNA stability and translation, but current predictors typically depend on a single representation and do not jointly model multiple views of the same RNA sequence. Results: We present TriTower-m6Am, combining semantic (RNA-FM, a pretrained RNA foundation model, with BellPooling), sequential (One-Hot BiLSTM, bidirectional long short-term memory), and structural (RGCN, a relational graph convolutional network, over the 41-nt window) representations of each RNA window. On an independent test set of 640 sequences, TriTower-m6Am reaches an area under the ROC curve (AUC) of 0.776 and a Matthews correlation coefficient (MCC) of 0.425, against DTC-m6Am’s AUC = 0.765 and MCC = 0.411, using 4.4 times fewer trainable parameters; at every matched specificity between 0.40 and 0.70 it recovers 1–7 more true sites out of 320 than DTC-m6Am. The AUC-weighted voting keeps every tower’s contribution readable for any individual prediction. Analyses of the graph tower delimit its benefit to the relational aggregation of local backbone adjacency over sequence-derived node features: the edge-type ablation, the node-feature ablation, and a structure-permutation control agree that the specific RNAfold-predicted fold carries no attributable signal, and a probing analysis shows that per-base base-pairing status is largely recoverable from the sequence-derived RNA-FM embeddings (AUC = 0.72 vs. 0.65 from the raw sequence), so the fold annotation is substantially redundant with the semantic representation the ensemble already receives. Conclusions: Combining semantic, sequential, and structural views of the same RNA sequence improves m6Am prediction beyond any single representation, and the ensemble’s voting weights make each prediction auditable without a post-hoc explainer. Whether the triple-tower pattern transfers to other RNA modification prediction tasks remains to be tested in future work. Author summary: Predicting where m6Am modifications occur on messenger RNA is important for understanding how cells regulate transcript stability and translation. Existing computational methods typically encode the RNA sequence in a single way, such as k-mer counts or a one-hot code, and treat the model as a black box that emits a prediction without explaining which features drove it. We built TriTower-m6Am to address both limitations. Our model combines three independent encoders—a pretrained RNA language model for semantic patterns, a bidirectional LSTM for local nucleotide order, and a relational graph convolutional network for the structural fold—and fuses their outputs by AUC-weighted voting, so the contribution of each tower to a given prediction is a readable number rather than a hidden parameter. On an independent benchmark the ensemble achieves ranking quality comparable to the prior best method with 4.4 times fewer parameters, and its voting weights make each tower’s contribution auditable for any individual prediction. A controlled structural ablation shows that linear backbone adjacency, rather than long-range base-pairing, carries whatever signal the graph tower uses, and that randomizing the predicted secondary structures does not degrade performance. This controlled analysis delineates what the graph tower contributes: its benefit lies in the relational encoding of local adjacency over sequence-derived features, which the control retains, rather than in the specific predicted fold. Whether the same triple-tower pattern transfers to other RNA modification site prediction tasks is a question for future work.
SN - 1553-7358
ER -