Assessing the Extrapolation Capability of Template-free Retrosynthesis Models

Abstract

Template-free retrosynthesis models offer the potential to extrapolate beyond established chemical reaction spaces, addressing inherent limitations of template-based approaches. However, it remains unclear whether these models can reliably predict accurate, novel, and chemically feasible pathways outside their training distribution. In this study, we rigorously assess the extrapolation ability of state-of-the-art template-free models using carefully constructed out-of-distribution (OOD) benchmarks derived from USPTO datasets. While these models can generate novel synthetic routes, their exact-match accuracy on OOD reactions is remarkably low (typically <1%). Moreover, round-trip performance (≈5–30%) is influenced by the performance of the forward model and may not fully capture some chemically reasonable predictions. Complementary manual inspection mitigates this limitation by revealing that the surrogate forward model produces false negatives, where chemically feasible reactions are incorrectly predicted as infeasible, and vice versa for false positives. These results underscore a critical challenge: current models may exhibit little creative extrapolation yet lack mechanisms to ensure chemical feasibility. Addressing this gap is essential for developing retrosynthesis models that are not only innovative, but also reliable for real-world synthesis planning.