Objective Cysteine S-carboxyethylation is a recently identified post-translational modification with potential biological relevance, but its experimental identification is costly and time-consuming. Existing computational approaches mainly rely on sequence information and may not fully exploit positional differences around the central cysteine or the complementary representation capabilities of heterogeneous neural networks. This study aimed to develop an effective sequence-based model for prioritizing potential cysteine S-carboxyethylation sites.Methods We developed Cys-DHE, a dual-branch heterogeneous deep ensemble model for cysteine S-carboxyethylation site prediction. The benchmark dataset contained 1 204 training samples, including 602 positive and 602 negative samples, and an independent test set containing 913 samples, including 83 positive and 830 negative samples. Each sample was represented by a 41-aa sequence window centered on the target Cys. Binary-weight encoding (BWE) was used to combine amino-acid identity with a distance-dependent weighting mechanism, thereby emphasizing residues closer to the central site. The resulting 41×20 feature matrix was simultaneously input into two complementary branches: CNN-BiLSTM-Attention for extracting local patterns and bidirectional contextual dependencies, and CNN-Transformer for modeling cross-position interactions and long-range dependencies. The prediction probabilities of the two branches were integrated using fixed weights of 0.75 and 0.25, respectively, determined according to five-fold cross-validation results on the training set. Model performance was evaluated using sensitivity (Sn), specificity (Sp), accuracy (Acc), Matthews correlation coefficient (MCC), AUROC, and AUPR, and was further compared with individual branches and conventional machine-learning methods.Results In five-fold cross-validation, Cys-DHE achieved Sn, Sp, Acc, MCC, AUROC, and AUPR values of 0.680 3, 0.842 7, 0.751 9, 0.526 2, 0.825 7, and 0.856 0, respectively. Compared with the individual branches, the fixed-weight ensemble improved Sn, Acc, AUROC, and AUPR, indicating complementary information between the two network architectures. Cys-DHE also outperformed SVM, RF, XGBoost, and LightGBM across the evaluated metrics under the same cross-validation setting. On the independent test set, Cys-DHE achieved an Acc of 89.60%, Sn of 69.88%, Sp of 91.57%, MCC of 0.508 8, AUROC of 0.865 4, and AUPR of 0.485 2. Compared with DLBWE-Cys, Cys-DHE showed higher Sp, Acc, and MCC, although Sn, AUROC, and AUPR were slightly lower. Additional local Cys analysis showed that positive sites were more frequently accompanied by additional Cys residues in their local sequence neighborhoods. Among samples containing an additional Cys within the 41-aa window, the median distance from the central Cys to the nearest additional Cys was approximately 5 residues in positive samples and 10 residues in negative samples, with a significant difference (P=1.35×10??).Conclusion Cys-DHE effectively integrates position-aware sequence encoding with complementary heterogeneous deep-learning architectures and provides balanced performance for cysteine S-carboxyethylation site prediction. The model is intended as a sequence-based candidate prioritization tool for narrowing potential modification sites before subsequent structural analysis and experimental validation. Sequence proximity should not be interpreted as direct evidence of three-dimensional spatial proximity or disulfide-bond formation. Future work will incorporate structural descriptors, solvent accessibility, spatial neighborhood information, and reliable disulfide-bond annotations, while expanding experimentally validated datasets to further improve model robustness, interpretability, and generalization across proteins. An online prediction server is available athttp://www.lzzzlab.top/Cys-DHE/.