When Grounding Fails: Input Verification for Agentic Virtual Screening

Retrieving a scientific record correctly does not establish its suitability for computation. We investigate this distinction in a virtual screening agentic workflow that connects medicinal-plant records to molecular screening and traceable candidate reports. The workflow combines catalogue retrieval, receptor checks, docking, property prediction and rule-based assessment. A retrospective audit identified incorrect protein assignments in 10 of 33 structure records; nine of 18 target entries met the evaluated docking-eligibility rules. Paired experiments with eight compounds on each of two targets showed that replacing inherited search boxes with structure-derived boxes changed mean docking scores by −3.95 and −5.00 kcal/mol. A subsequent campaign recorded 210 target–compound evaluations, with 78 passing predefined computational filters. These findings document input inconsistencies and their consequences for screening outputs. They motivate explicit verification between retrieval and execution, while leaving verification accuracy, unnecessary withholding of valid requests and biological predictive performance to independent evaluation.