Feature-Driven Prediction of HOMO-LUMO Gaps in Transition Metal Complexes Using the SLEET Model: A SMILES-Based Transformer Framework

A feature-driven model - SLEET built upon the early reported SchNet-bs-RAN framework, that combines the approaches of SchNet and the bondstep representation weighted by the reduced atom number, is reported for evaluating the molecular electronic structure properties of transition metal complexes (TMCs). Ligands were derived by segmenting purely two-dimensional SMILES representations, and metal-ligand interactions were modeled using a Transformer-like architecture to construct a property prediction framework that aligns closely with chemical knowledge. This approach effectively captures the characteristics of ligand field within TMCs. Consequently, the SLEET model delivers precise HOMO-LUMO gap predictions comparable to those achieved by three-dimensional information-based models, while also demonstrates strong performance in predicting the molecular-weight-independent electronic properties.