Abstract
Accurate SUMOylation site prediction is crucial for deciphering gene regulation and disease mechanisms. However, distinguishing SUMO1 and SUMO2 modifications remains a major challenge due to their structural similarities. Conventional prediction models often struggle to differentiate between these paralogues, limiting their applicability in biological research. To address this, we introduce SUMO-LMNet, a deep learning-based framework for the precise prediction of SUMO1 and SUMO2 sites. Unlike previous models, SUMO-LMNet integrates a lossless mapping strategy and deep learning architectures to enhance both prediction accuracy and interpretability. Our model extracts high-dimensional features from sequences and transforms them into two-dimensional feature maps, enabling convolutional neural networks (CNNs) to effectively capture both local and global dependencies within the data. By leveraging a Lossless Mapping Network (LM-Net), this approach preserves the original feature space, ensuring that feature integrity is retained without loss of spatial information. While Grad-CAM highlights key features in individual predictions, it lacks consistency across samples and does not provide a dataset-wide evaluation of feature importance. To address this, we introduce Combined Heatmap Feature Analysis (CHFA), which systematically aggregates feature importance across multiple samples, providing a more reliable and interpretable dataset-wide assessment. Experimental results reveal distinct feature dependencies between SUMO1 and SUMO2, underscoring the necessity of paralogue-specific predictive models. Through a systematic comparison of multiple neural network architectures, we demonstrate that our model achieves over 80 % accuracy in distinguishing SUMO1 and SUMO2 modification sites. By prioritizing candidate sites for further study, our model aids experimental design and accelerates the discovery of biologically relevant SUMOylation targets. SUMO-LMNet is publicly available at https://predictor.isu.edu.tw/sumo-lmnet.