Deep learning has greatly improved protein function prediction (PFP); however, they often struggle to efficiently process biological sequences. While current models are primarily focused on natural language processing (NLP) tasks, consequently, “protein-centric” deep learning models are needed for biological sequences. This paper presents an efficient Lite-SeqCNN + multi-attention-based PFP framework, with a focus on developing protein-centric models. Lite-SeqCNN is a simple dilated-CNN model with the ability to capture short-and-long dependencies, whereas multi-attention aids in information refinement by emphasizing focus on key words through a set of parallel attention layers. The experimental results analysis, employing multiple datasets, suggests the potential of the proposed solution as well.