Protein-protein interactions (PPIs) govern essential biological processes, relying on specific binding sites for molecular machinery in cells. Identifying these binding sites is crucial, with computational methods emerging as efficient alternatives to labor-intensive experimental approaches. While various techniques leverage sequential and structural information of amino acids, the limited availability of protein structural data in databases makes sequential-based models more practical. The proposed model, named TranP-B-site, employs a convolutional neural network on the transformer model’s embeddings of the sequential information of the amino acids to predict the binding sites of PPIs. First, two types of features are extracted for each amino acid in a protein sequence: one-hot encoding representing the low-level features and transformer model-based embeddings, which contain information about the entire protein sequence. These one-hot encodings and amino acid embeddings are concatenated to form two matrices. Then, two local feature sets are created by employing a windowing technique across the acquired matrices. The amino acid–based local feature set is fed into a CNN architecture, while the one-hot encoding-based local features are fed into a neural network. Finally, classification is performed on the concatenated output of the CNN and neural network using a sub-neural network. The proposed model demonstrates an improvement of 3% in MCC and 7% in accuracy compared to the previous state-of-the-art sequence-based model for independent dataset. Additionally, a new test dataset was curated from recently published protein sequences in the PDB database, and the proposed model outperformed other state-of-the-art models.