In this paper, we propose a novel learning-aided sphere decoding (SD) scheme\nfor large multiple-input--multiple-output systems, namely, deep path\nprediction-based sphere decoding (DPP-SD). In this scheme, we employ a neural\nnetwork (NN) to predict the minimum metrics of the ``deep'' paths in sub-trees\nbefore commencing the tree search in SD. To reduce the complexity of the NN, we\nemploy the input vector with a reduced dimension rather than using the original\nreceived signals and full channel matrix. The outputs of the NN, i.e., the\npredicted minimum path metrics, are exploited to determine the search order\nbetween the sub-trees, as well as to optimize the initial search radius, which\nmay reduce the computational complexity of SD. For further complexity\nreduction, an early termination scheme based on the predicted minimum path\nmetrics is also proposed. Our simulation results show that the proposed DPP-SD\nscheme provides a significant reduction in computational complexity compared\nwith the conventional SD algorithm, despite achieving near-optimal performance.\n