Can Machine Intelligence Solve Retrosynthesis of Complex Natural Products Respecting Stereochemistry and Synthetic Practicality?

Download PDF

Machine intelligence has transformed retrosynthesis from rule retrieval into data-driven reaction prediction and search, yet structurally complex natural products expose a harder question than whether a model can recover plausible precursors. Their synthesis often depends on stereochemical information, chemoselective control, protecting-group logic, long-range strategic choices, convergent assembly, specialized building blocks, and practical execution under conditions that are only incompletely represented in reaction corpora. This critical review evaluates whether contemporary machine-learning retrosynthesis can meet those demands. We distinguish single-step prediction from route planning, formal precursor validity from experimentally meaningful selectivity, and route search from synthetic strategy. Across current approaches, template-based models retain chemically explicit transformation priors but inherit coverage limits, whereas template-free architectures broaden prediction flexibility while remaining sensitive to representation, dataset bias, and local rather than route-level objectives. Stereochemistry is a particularly revealing stress test: improved encodings and editing strategies can reduce errors, but correct atom-level configuration does not by itself guarantee stereoselective feasibility across a multistep route. We therefore propose that natural-product retrosynthesis should be evaluated as a constrained decision problem in which structural correctness, selectivity, strategic coherence, material availability, and practical risk remain separable dimensions. The literature supports substantial progress in reaction prediction and search, but evidence for autonomous solution of complex natural-product synthesis remains limited by benchmark composition, sparse experimental validation, incomplete modeling of protecting groups and conditions, and weak route-level practicality criteria. Progress will require benchmarks built around expert syntheses, prospective laboratory testing, and models that represent uncertainty and strategy explicitly.