(A) An outline of the EANMD pipeline. AS events are identified from long- or short-read data, followed by transcript-level NMD prediction. The XGBoost machine learning model is applied to classify NMD flags and assign NMD scores. (B) Experimental datasets used for AS-NMD validation, including (a) Neuro2a cells with Upf1 KD (n = 2 biological replicates), (b) primary mouse neurons treated with CHX (n = 2), and (c) Upf2 conditional knockout in embryonic mouse neocortex (n = 2) (Lin et al., ref. 20). (C) Violin plots showing ΔPSI (left) and log₂FC in gene expression (right) for NMD versus non-NMD SE events upon Upf1 KD. Events triggering NMD exhibit increased transcript abundance after NMD inhibition, serving as a proxy for NMD efficiency. NMD_in (n = 43), NMD_ex (n = 63), ORF preserving (n = 90), and ORF changing (n = 4) events (1-way ANOVA followed by Tukey’s multiple-comparison test). (D) SHAP values of the minimum stop codon position fraction (Min stop Pos F) and the maximum stop codon distance to the last exon-exon junction (Max DJ), highlighting their interaction effect on NMD efficiency. (E) SHAP analysis of 3′-UTR length. NMD efficiency increased when 3′ UTRs were either longer than 1,150 nt or shorter than 180 nt. MFE, minimum free energy. (F) Receiver operating characteristic (ROC) curves for NMD classification. Using the 50 nt rule, the ROC-AUC is 0.680 (gray line). Incorporating the machine learning–predicted NMD score improves classification performance (AUC = 0.752, total SE events, navy blue line). ML, machine learning. (G) Filtering events based on NMD scores enhances AS-NMD detection.