Classical substructure search versus LLM-assisted keyword retrieval in SciFinder: A case study on GABAB1 ligand discovery using a phenibut scaffold.

The effectiveness of classical substructure search versus LLM-assisted keyword-based retrieval in the SciFinder database was compared for identifying potential GABAB1 receptor ligands using a phenibut-like scaffold. Substructure search based on the reference compound C1 yielded 53 retrieved structures, including a promising candidate (s34, CAS 97125-82-1) that showed most favorable docking affinity (-8.9 kcal/mol in AutoDock Vina) and MM/GBSA binding free energy (-52.80 kcal/mol) compared to phenibut. Molecular dynamics (MD) simulations (500 ns) suggested a dynamic but persistent binding mode of s34 within the binding site of GABAB1. In contrast, LLM-assisted keyword search returned a set of heterogeneous records, including formulations, mixtures, and compounds lacking directly dockable structures, resulting in a significantly reduced number of molecules suitable for structure-based evaluation. These results highlight differences in the chemical utility of retrieval strategies for structure-based screening. Toxicity predictions indicated that both s34 and phenibut fall within ProTox Class 4, which suggests moderate predicted acute toxicity. This study demonstrates how search methodology can influence the chemical space of candidates.