Natural-language prompting builds and operates an open-source software workflow for virtual drug-discovery

Virtual docking and free-energy calculations can accelerate drug discovery, but practical use requires structural chemistry expertise, complex parameter selection and software integration. Here we show that iterative dialogue with an artificial intelligence (AI) coding assistant, OpenAI Codex, can guide a non-specialist scientist to build and operate a scripted, auditable computational workflow without direct coding. The assistant provided structural chemistry guidance and assembled an integrated pipeline for structure retrieval and preparation, binding-site definition, ligand preparation, virtual docking, interaction analysis, molecular dynamics (MD) and alchemical relative binding free-energy calculations. It also built a script generator for launching, monitoring and interpreting workflow runs. We tested the workflow on public cyclin-dependent kinase 2 (CDK2) benchmarks. After primary docking and multi-seed hit refinement, the active-ligand fraction increased from 1.73% in the input library to 80% in the ten-candidate handoff set (46.4-fold over input); a short MD stage acted as a structural quality-control gate on the leading candidates. Docking and hit refinement also generalized to human immunodeficiency virus reverse transcriptase and kinesin family member 11 benchmarks. In a CDK2 free-energy benchmark, the workflow achieved root-mean-square error = 0.66 kcal/mol and Pearson r = 0.90. Prompt-guided AI can orchestrate computational drug-discovery workflows for non-specialists while the underlying calculations remain the work of established external software and every run remains auditable.