The dual metamodeling approach is usually adopted to tackle the stochastic nature of earthquakes in seismic reliability analysis relying on the lognormal response assumption. Alternatively, a direct response approximation approach where separate metamodels are constructed for each earthquake is attempted here avoiding prior distribution assumption. Further, an adaptive support vector regression-based metamodeling is proposed that selects new training samples near the failure boundary with due consideration to accuracy and efficiency. The effectiveness of the approach is elucidated by comparing it with the results obtained by the direct Monte Carlo simulation technique and a state-of-the-art active learning-based Kriging approach.