Speaker: Federico Lazzari, Scuola Superiore Meridionale

Accurate electronic energies for large molecules are now routine, yet underlying geometries are often unreliable downstream. Conformational workflows thus separate exhaustive exploration from costly high-level refinement, a split hard to reconcile.

We introduce a descriptor-guided pipeline in which the same molecular-perception features (“synthons”), first devised for Δ-ML geometry correction,[1,2] steer both steps.

Hard degrees of freedom are tackled with Pharma-LEGO: SMILES from the 300-structure LCB25 set[3] are fragmented into rings, recombined into chemically valid scaffolds, and ranked by synthon similarity before optimisation.

Soft degrees of freedom (torsions) are searched with a hydrogen-bond-aware island-model genetic algorithm.[4,5] A fitness penalty discourages poses that collapse into known minima, boosting diversity. When geometry and energy correlate strongly, an active Bayesian loop can replace the genetic stage, exchanging breadth for query efficiency.

Selected conformers enter exploitation and are re-optimised by two complementary schemes:

• BDPCS3[6]: revDSD double-hybrid plus a-posteriori core–valence correlation; bond-length errors <0.001 Å.

• BHPCS2[7]: a Δ-ML surrogate trained on B3LYP geometries; similar accuracy at far lower cost.

Because both methods share the synthon feature space, exploration and refinement remain internally consistent.

The modular protocol reunites exploration and exploitation. Ring perception, torsional search, and high-level optimisation are decoupled yet seamless, enabling automated, chemically meaningful sampling of macrocycles and drug-like molecules, sidestepping transition-state barriers and providing tunable cost-accuracy trade-offs for high-resolution spectroscopy, conformational mapping, and reactive-intermediate studies.

References:

[1] Lazzari et al., J. Chem. Phys. 2025, 162, 114310.
[2] Lazzari et al., J. Phys. Chem. A 2024, 128, 1385.
[3] Di Grande et al., J. Chem. Theory Comput. 2024, 20, 9243.
[4] Mancini et al., J. Chem. Phys. 2020, 153, 124110.
[5] Barone et al., J. Chem. Theory Comput. 2023, 19, 1243.
[6] Barone & Lazzari, J. Phys. Chem. A 2023, 127, 10517.
[7] Lazzari et al., J. Chem. Theory Comput. 2025, in press.