Accurate prediction of protein structural properties is critical for understand-ing biological function and advancing biomedical applications. Existing methods often overlook task interdependencies and rely on limited feature representations, hindering generalization. To address these issues, we pro-pose ProteinMM, a unified multi-task learning framework for residue-level protein property prediction. ProteinMM introduces three key innovations: (i) a multi-view representation that integrates contextual embeddings from pre-trained protein language models with evolutionary features; (ii) a positive-transfer-driven task grouping strategy that adaptively clusters tasks to enhance feature sharing and feature extraction; and (iii) a task-specific evolutionary deep fusion network that dynamically selects and integrates group-relevant features without manual tuning. Extensive experiments on bench-mark datasets demonstrate that ProteinMM consistently outperforms state-of-the-art methods across multiple structural prediction tasks, offering a robust and scalable solution for protein modeling.