<p>The assessment of seismic response to buildings on a regional scale under mainshock-aftershock (MS-AS) sequences is crucial for post-earthquake emergency response and disaster recovery. Existing methods predominantly focus on individual buildings, limiting their ability to evaluate building clusters with diverse structural types. Moreover, these methods often lack computational efficiency for large-scale analyses. To overcome these challenges, this research introduces a deep learning framework for efficiently assessing regional seismic response to buildings subjected to MS-AS sequences. A representative building inventory, featuring a wide range of structural characteristics, was compiled and combined with city-scale nonlinear time-history analyses (THA) and recorded seismic events to generate a response database for building clusters. Leveraging intensity measures (IMs) from both mainshocks and aftershocks, together with relevant building attributes, the model predicts the seismic responses of structures exposed to MS-AS sequences. A deep neural network with an attention mechanism was designed to map these input-output relationships. Results show that the model achieves R² values exceeding 0.999 on the test set, with an average prediction error of less than 12% compared to observed seismic responses. Additionally, the model predicts maximum inter-story drift ratios with an R² value of 0.91 compared to existing models. With a computation time of less than 0.01&#xa0;s per 1000 buildings, the method improves computational efficiency by a factor of 1,533, meeting the requirements for real-time and accurate seismic response prediction of building clusters under MS-AS sequences.</p>