Prediction of cyclin-dependent proteins using pre-trained protein language models and effective neural network architectures.

Cyclin-dependent proteins (CDPs), including cyclins, cyclin-dependent kinases (CDKs), and cyclin-dependent kinase inhibitors (CKIs), are pivotal regulators of the eukaryotic cell cycle. Dysregulation of these proteins can lead to disrupted cell cycle control and contribute to cancer progression. Traditional methods for identifying CDPs, especially homology-based approaches, face limitations when dealing with non-homologous data. Additionally, conventional machine learning models relying on manually extracted features often fail to capture critical functional characteristics of proteins. To address these challenges, we proposed a novel computational method that integrates pre-trained protein language models (PLMs) with deep learning techniques for CDP prediction. Specifically, the ESM model was employed to generate high-quality protein embeddings, thereby eliminating the need for manual feature extraction. In addition, a customized convolutional neural network (CNN) architecture was designed to automatically learn discriminative features relevant to CDP identification. The proposed computational method was further evaluated on two independent test sets, consistently outperforming existing state-of-the-art approaches. The experimental results demonstrated that the method exhibits strong generalization and robustness across different datasets, highlighting its effectiveness for accurate identification of CDPs.