Molecular recognition features (MoRFs) play important roles in various biological processes and are associated with numerous diseases. However, existing methods are limited in multisource feature modeling, contextual dependency capture, and robust feature fusion. To address these issues, we propose a novel framework, DeepDPM. We selected Prot-T5 and ESM-2 as feature extractors and designed four novel modules after evaluating several state-of-the-art models. BioWaveKAN introduces a dual-path activation mechanism integrating wavelet functions with linear transformations to capture multifrequency information in protein sequences. DCAttention leverages dynamic parameters and convolution operations to enhance contextual modeling. BiLCrossAttention employs weighted dual-channel cross-attention to achieve precise and efficient integration of sequential and spatial information. MultiScaleSmoothing applies multiscale convolution and filtering strategies to improve feature smoothness and robustness, strengthening model stability. In addition, we refined Focal Loss by incorporating label smoothing and temperature scaling to mitigate class imbalance. On the benchmark data set Test1, DeepDPM achieved an MCC of 0.7739 and an AUC of 0.9760, corresponding to relative improvements of 8.68 and 4.75% over the best baseline. These results demonstrate the accuracy and robustness of DeepDPM, underscoring its broad potential in protein function analysis and disease-related studies. The associated data and source code are available at https://github.com/SunHuaiYangCCZU/DeepDPM.
Sun et al. (Wed,) studied this question.