The complexity cliff in natural product discovery: why full automation remains elusive and the case for collaborative...

Covering: 1992 to 2025Natural products contribute to roughly half of FDA-approved small-molecule drugs, yet their discovery has remained surprisingly resistant to the automation that has reshaped synthetic medicinal chemistry. In this review, we examine why. We propose that natural product workflows encounter a "complexity cliff", a threshold beyond which automation success rates fall sharply rather than degrading gradually. We define the complexity cliff in operational terms as a discontinuity in the performance of standardized, scaled pipelines, which is triggered when chemical, biological or ecological complexity exceeds the assumptions encoded in the workflow. The cliff manifests across three coupled axes: a chemical axis, where unprecedented scaffolds and stereochemistry overwhelm pattern-based dereplication and structure elucidation; a biological axis, where unculturable organisms, silent biosynthetic gene clusters and context-dependent metabolite production frustrate standardized cultivation; and an operational axis, where matrix-specific extraction, scale-up and sustainable sourcing resist consistency-driven design. We distinguish challenges that are intrinsic to natural product research from those that are unique to automation, and we argue that several often-cited "automation barriers" (such as scale-dependent yields) are, in fact, universal pre-automation problems for which automation may help rather than hinder. Emerging technologies, including AI- and machine-learning-guided genome mining, CRISPR-enabled pathway engineering and ecosystem-scale digital twins, address parts of the cliff but not its serendipitous core. We close by arguing that the most productive path forward is not full autonomy but a collaborative intelligence framework: machine throughput and pattern recognition coupled with human curiosity, biological intuition, and willingness to follow anomalies. This framing reframes the central question from "how do we automate natural product discovery?" to "how do we partition natural product discovery between humans and machines so that each side does what it does best?"