MCAMEF-BERT: an efficient deep learning method for RNA N7-methylguanosine site prediction via multi-branch feature integration

Abstract

Accurate identification of N7-methylguanosine (m7G) modification sites plays a critical role in uncovering the regulatory mechanisms of various biological processes, including human development, tumor initiation, and progression. However, existing prediction methods still suffer from limited representational power, redundant feature fusion, insufficient utilization of biological prior knowledge, and poor interpretability. In this study, we propose a novel deep learning model named MCAMEF-BERT. This model adopts a parallel architecture that integrates both a DNABERT-2-based pretrained model branch and multiple traditional feature encoding branches, enabling comprehensive multi-perspective sequence feature extraction. To address the redundancy issue in feature fusion, we introduce a multi-channel attention module. Our model demonstrates superior accuracy and effectiveness on datasets from m7GHub, outperforming other state-of-the-art classifiers. Furthermore, we validate the interpretability of MCAMEF-BERT through in silico saturation mutagenesis experiments, and confirm its robustness in motif recognition. Moreover, its generalization capability is validated across diverse RNA modification site prediction tasks.

Keywords: in silico saturation mutagenesis experiments; DNABERT-2 pretrained model; N7-methylguanosine modification; multi-channel attention; multi-encoding fusion.

Figure 1

The architectures of MCAMEF-BERT. (A) The development process of MCAMEF-BERT. (B) Encoding module. Transforming the input sequences. (C) fine-tuned DNABERT-2 module. Fine-tuning the DNABERT-2 pre-trained model. (D) MEF module. Extracting feature vectors of various encoding methods. (E) Multi-Channel attention block. Removing redundancy and achieving effective feature fusion of embedded vectors. (F) Representation alignment module. Reconciling the differences in feature representations between the two branches. (G) Classification. Predicting the RNA modifications.

Figure 2

Comparison results of different input sequence lengths. Six input length (51 bp, 101 bp, 201 bp, 301 bp, 401 bp, and 501 bp) were utilized.

Figure 3

Comparison results of different classifiers. iRNA-m7G was evaluated with 41 bp and 201 bp inputs, m7GPredictor with 101 bp and 201 bp, moss-m7G with 201 bp and 501 bp, and MCAMEF-BERT with 201 bp input.

Figure 4

Comparison results of different model variants and UMAP visualization of different model variants. (A) The performances results of different model variants. (B-H) The UMAP visualization of different model variants (MCAMEF-BERT, W/o pretrain, W/o mca, W/o fuse, W/o LSTM, W/o transformer, W/o CNN).

Figure 5

Base-wise sensitivity and importance scores derived from the ISM experiment, along with representative sequence motifs identified by the MEME Suite tool. (A) Base sensitivity of the model at each position derived from the ISM experiment. Dashed boxes indicate regions overlapping with motifs identified in (C), while solid boxes highlight motifs to which the model exhibits high sensitivity. (B) Position-wise importance scores in model predictions derived from the ISM experiment. (C) Motifs widely identified in the dataset by STREME from the MEME Suite.