TS-Free Prediction of Activation Strain Components of Diels–Alder Reactions with Graph Neural Networks. – ( Daiann Sosa Carrizo / LPCNO / Seminar). – 16/04/2026, 14H

Séminaire du LCPQ

Daiann Sosa Carrizo, LPCNO

Seminar LCPQ, 16/04/2026, 14H

Summary :
Machine learning is transforming computational chemistry. From predicting molecular properties — solubility, spectroscopic, and drug-likeness — to guiding retrosynthetic planning and catalyst design, ML models have progressively replaced expensive quantum-chemical calculations with fast, learned approximations. Yet some class of properties has remained largely out of reach: those that encode transition-state geometry.[1]
The Activation Strain Model (ASM) decomposes activation barriers into fragment deformation energies (ΔE‡_strain) and interfragment interaction energy (ΔE‡_int), explaining why a reaction is fast or slow rather than merely predicting trends. Computing these quantities requires explicit DFT transition-state optimisation — costly and incompatible with high-throughput screening.[2-3]
In this seminar, I will discuss the application of the directed message-passing neural network (Dual D-MPNN) to predict ASM components from reactant SMILES, without TS optimisation or hand-crafted descriptors. The model was trained on 802 endo Diels–Alder reactions computed at the M06-2X/6-311+G(d,p) level of theory.[4]
Error analysis reveals a chemically interpretable result: diene conformational rigidity, not electronic complexity, is the primary determinant of model accuracy. Rigid heteroaromatic dienes such as furans are predicted near chemical accuracy, while flexible hetero-acyclic dienes bearing conjugated carbonyl groups remain the principal failure mode. This is not an architectural limitation — it reflects a fundamental constraint of 2D molecular graphs, which cannot encode the conformational ambiguity that determines TS geometry in flexible systems.

As a prospective application, this model will be used to screen virtual diene libraries and identify reactive candidates for Diels–Alder reactions targeting terpenoid and alkaloid natural products. In these systems, the ASM decomposition provides design principles — distinguishing strain-controlled from interaction-controlled reactivity — that barrier heights alone cannot offer.

• [1]. Chen, X. et al. Nat. Synth. 2025, 4, 877–887.
• [2] Bickelhaupt, F. M. et al. Angew. Chem. Int. Ed. 2017, 56, 10070–10086
• [3] Brunard, E. et al. J. Am. Chem. Soc. 2024, 146, 5843–5854.
• [4] Vargas, S. et al. J. Chem. Theory Comput. 2021, 17, 6098–6110.