Antenna selection in reconfigurable intelligent surface (RIS)-assisted multiple-input multiple-output (MIMO) systems presents has significant computational challenges.  The contribution of each transmit antenna is determined by the combined effects of direct and RIS-reflected channels. To address the complexity of combinatorial search a graph neural network (GNN)-based antenna selection framework is proposed. In this framework transmit antennas are represented as graph nodes with channel-correlation information forming. The graph edges and magnitude-phase channel statistics serve as node features. A three-layer feedforward GNN is trained using greedy-selection labels generated from 1000 channel realizations and evaluated on 200 independent test realizations. For a 16×8 MIMO system assisted by a 64-element RIS at 28 GHz the proposed method achieves a spectral efficiency of 63.41 bits/s/Hz and corresponding to 94.5% of the greedy baseline performance of 67.09 bits/s/Hz. While reducing the average selection time from 0.441 ms to 0.025 ms. These results determine that graph-structured learning enables near-greedy antenna selection with considerably lower inference complexity. The current study is limited to simulated settings with fixed system dimensions and idealized channel assumptions; future work will address broader channel models and larger-scale configurations.