This paper proposes a novel retrosynthetic model, RetroChimera, to overcome the limitations of AI-based synthetic planning models. RetroChimera is based on two newly developed components with complementary inductive biases, combined using a novel framework that integrates predictions from multiple sources via a learning-based ensemble strategy. Experimental results demonstrate that RetroChimera significantly outperforms existing leading models, exhibits robust performance outside the training data, and demonstrates, for the first time, the ability to learn even with a very small number of examples for each reaction class. Furthermore, industrial organic chemists prefer RetroChimera's predictions to reactions from the training data, demonstrating a high degree of consistency. Finally, zero-shot transfer on an internal dataset from a major pharmaceutical company demonstrates robust generalization under distributional shift.
