Testing self-driving cars (SDCs) requires extensive simulation-based testing, making efficient test case selection important. This paper, submitted to SBFT 2025, presents a transformer-based approach for test selection in SDC testing that uses attention mechanisms and search-based techniques to learn complex relationships between road geometry features. The selector combines geometric analysis with deep learning, using a transformer encoder that processes comprehensive road features including curvature profiles, segment relationships, and historical failure patterns. The approach implements a feature engineering pipeline that extracts meaningful geometric properties and uses self-attention to capture long-range dependencies in road sequences. Our search strategy uses a sophisticated scoring mechanism that balances predicted failure probabilities with test diversity measures. Evaluated in the BeamNG.tech simulation environment, our experimental results show strong performance with an initialization time of 5.09s and selection time of 0.48s. The selector achieves a time-to-fault ratio of 217.95 and a fault-to-selection ratio of 0.19, while maintaining a diversity score of 0.037. These results demonstrate the effectiveness of combining transformer-based architectures with search-based optimization techniques in capturing complex road geometry patterns and predicting challenging test scenarios for SDCs.