Jordan P. Liles, Nicholas Casetti, Thijs Stuyver, Connor W. Coley (2026)
Highlighted by Jan Jensen
One thing I tend to go on and on about the is the dearth of prospective experimental verification of computational predictions, so this study is a welcome and interesting addition.
The authors combine a robust TS optimizer with a reactive MLIP that can efficiently compute barriers for sigmatropic rearrangement and use the methodology to screen for feasible reactions.
For example, they targeted natural products from the COCONUT database that lack reported synthetic preparations. Beginning with the full database, they applied a set of retrosynthetic rearrangement templates to enumerate precursors capable of forming natural products through a single hypothetical rearrangement. They used ASKCOS to identify precursors accessible in three or fewer synthetic steps. Structural complexity was assessed using the synthetic accessibility score, and kinetic feasibility was defined by a barrier-height threshold of 40 kcal/mol. This workflow yielded 23 promising candidates and their corresponding rearrangements. Among these candidates, they identified (±)-norbrachycoumarin. Although it was first isolated in 1986 it had never been chemically synthesized. They subsequently confirmed the predicted rearrangement experimentally!
More generally, prospective, physics-based evaluation of reaction mechanisms could accelerate synthetic chemistry. Analyzing kinetically relevant elementary steps can inform the feasibility and selectivity of key transformations while revealing useful retrosynthetic disconnections that might otherwise be overlooked. However, accurate reaction modeling remains challenging because of factors such as solvation, transition-state conformational diversity, catalysis, open-shell species, and metal-containing systems. Mechanistic uncertainty presents an additional difficulty, as reliable predictions require competing reactions and off-cycle pathways to be considered.
The general strategy of applying physics-driven mechanism evaluation prospectively holds great promise for accelerating synthetic chemistry. Probing kinetically relevant elementary steps can help evaluate the feasibility and selectivity of key reactions while planning synthetic routes as well as help reveal strategic disconnections that may go overlooked in initial route designs. However, several challenges in reaction modeling remain, including solvation, conformational ensembles of transition states, catalysis, open-shell systems, and heavier elements such as metals. Equally complicating can be mechanistic uncertainty, wherein accounting for competing side reactions and/or off-cycle pathways is crucial for oracle accuracy. Nonetheless, our results demonstrate that current methods are, at present, capable of delivering actionable predictions at a meaningful scale for well-behaved reaction classes.
This work is licensed under a Creative Commons Attribution 4.0 International License.
