Background <p>RNA methylation (RM) regulates gene expression regulation, RNA stability, and protein translation. Accurate prediction of RM modification sites is essential for understanding their biological functions. However, existing wet-lab detection techniques face challenges including operational complexity and high costs. Deep learning (DL) methods have been applied to this task. However, existing methods show performance degradation with smaller training datasets. For instance, the Bidirectional Gated Recurrent Unit (BGRU) demonstrates substantial performance degradation. Contrastive Learning Network (CNN) can extract local pattern features but learns overly specific patterns with sample-limited data, resulting in poor feature generalization. Bidirectional Long Short-Term Memory (BiLSTM) excels at modeling long-range dependencies but cannot sufficiently learn gating mechanism parameters to capture effective sequence representations with limited samples. Transformer processes sequences in parallel and captures global dependencies through self-attention, but its quadratic computational complexity and large parameter count make it prone to overfitting on small datasets. Current DL methods show reduced performance when training data is limited.</p> Results <p>This study proposes a Multi-view Contrastive Learning with CNN-BiLSTM-Attention (MCLCBA) framework for RM modification site prediction. The multi-view approach comprises a primary view and auxiliary view, where the primary view utilizes DNA Bidirectional Encoder Representations from Transformers (DNABERT) to extract sequence contextual features, and the auxiliary view employs Chaos Game Representation (CGR) to extract structural features. Feature extraction includes four components: data augmentation, multi-view encoders, projection heads, and contrastive loss functions. By implementing dual differential data augmentation strategies and constructing multi-view network architectures for feature processing and fusion, the model learns discriminative feature representations invariant to data augmentation through maximizing positive sample similarity while minimizing negative sample similarity. This effectively addresses sample-limited feature learning scenarios. Experimental results on the sample-limited m⁷G dataset demonstrate that MCLCBA achieves AUROC and AUPRC of 85.64% and 86.94%, respectively, improving upon existing methods by 5–6% in both metrics.</p> Conclusions <p>Through multi-view contrastive learning, MCLCBA provides an approach for RM sites under sample-limited scenarios.</p>