m6Am-DLcat: A Deep Learning Model to Detect m6Am Sites on Different Amalgamation of mRNA Sequences

N6,2’-O-dimethyladenosine (m6Am) sites are characteristic Ribonucleic Acid (RNA) modifications in which adenosine undergoes 2’-O ribose methylation after being methylated at the N6 position. These m6Am alterations are linked to several illnesses, including cancer, and are normally located next to the m7G cap at the 5’ cap of messenger RNAs (mRNAs), as a result finding these locations is essential to furthering medical research. However, existing computational approaches for m6Am detection often rely on limited feature representations, shallow Machine Learning (ML) classifiers, or lack mechanisms to capture long-range dependencies within RNA sequences, which restricts their generalization ability and biological interpretability. To address these limitations, this paper introduces m6Am-DLcat, a novel Deep Learning (DL) framework based on a CNN-Attention architecture for identifying m6Am locations in mRNA sequences. Five complementary feature representations, Nucleotide Chemical Property (NCP), Adaptive Skip Dinucleotide Composition (ASDC), Pseudo Electron-Ion Interaction Potential (PseEIIP), Dinucleotide Binary Encoding (DBE), and K Monomeric Units (Kmer), are extracted and integrated, followed by a secondary feature selection stage using Light Gradient-Boosting Machine (LGBM), eXtreme Gradient Boosting (XGB), and Chi-square (X2 ) approaches. We evaluate our approach using independent traintest splits and 10-fold cross-validation with both DL and ensemble ML baselines. Experimental results demonstrate that m6Am-DLcat consistently outperforms state-of-the-art predictors, by achieving an accuracy of 87.76% and an Area Under the Curve (AUC) of 0.95 using XGB-based feature selection. Feature importance analysis further reveals strong associations between physicochemical properties and m6Am signals, providing biological insight into m6Ammediated gene regulation. These findings highlight the effectiveness of attention-driven deep learning models for robust and interpretable m6Am site prediction.