Taming T-REX: A Canonical Language for Geometry-Aware Generative Design of Transition-Metal Complexes

Abstract

Canonical string representations have transformed organic cheminformatics, yet transition-metal complexes (TMCs) lack an equivalent that captures coordination geometry, stereochemistry, and donor topology. We introduce Trans-pair Relations EXpression (T-REX), a canonical line notation encoding geometry, topology, and metal-centered chirality (@/@@, Δ/Λ) via trans-pair maps. Applied to 63,375 DFT-optimized structures from the tmQMg dataset, T-REX identifies five distinct isomer classes (coordination, enantiomeric, linkage, hemilabile, and geometric) and reveals that fewer than 1.2% of complexes capable of stereoisomerism are resolved as such in crystallographic data. Combinatorial enumeration expands these parent structures into 149,228 unique topological variants; modular ligand substitution generates millions of additional candidates. Across one bond-only baseline and four geometry-aware architectures, encoding the T-REX coordination map consistently improves prediction of HOMO, LUMO, gap, and dipole moment. Dipole moment shows the largest gains (R² = 0.845 vs. 0.715 for the baseline), and three architecturally distinct models with a direct coordination-sphere readout achieve equivalent performance, confirming that T-REX topology, not architecture choice, drives the improvement. Geometry-aware models reach equivalent accuracy with roughly four times less training data, positioning T-REX as both an interoperable data format and an ML-ready representation for transition-metal chemistry.