Agentic AI-Orchestrated Computational Drug Discovery: A Case-Study-Demonstrated Open-Source Workflow Blueprint

Bringing a new drug to market can exceed $1 billion, and integrated commercial computational-chemistry platforms cost $200,000–$5,000,000 per year, placing them beyond most resource-constrained research groups. We demonstrate that open-source tools with large language model (LLM) orchestration can democratize early-stage New Chemical Entity (NCE) prioritisation. Three computational case studies (Case A: proprietary target, class withheld, public experimental co-crystal structure; Case B: kinase with co-crystal structure; Case C: protein–protein-interaction interface, apo) each ran Tier 1 (target-enriched ZINC20 library) and Tier 2 (GNINA v1.1 docking with CNN rescoring, plus SwissADME and ADMETlab 2.0 profiling). Tiers 3–4 are proposed future work. A tiered Claude Opus 4/Sonnet 4 agent orchestrated the workflow. Each library (~154,000–400,000 compounds) was reduced to 200 or fewer docked and 10 prioritised candidates per case, with Case C treated as an exploratory PPI stress test. LLM citation accuracy on independent PubMed verification was 88%, 78%, and 90% (pooled 85%, 23/27). No experimental assays were performed. Incurred LLM API cost was $1,450–$1,800; projected discovery-phase cost $3,450–$10,800 (lower bound). A retrospective public-target benchmark (acetylcholinesterase, PDB 4EY7) reproduced the crystallographic pose to 0.31 A but showed that the docking score alone did not enrich known actives (ROC-AUC 0.23); mechanism-based interaction filtering restored above-random early recognition (BEDROC 0.02 to 0.44), supporting the multi-evidence, human-gated design rather than docking-based ranking alone. This provides a reproducible, low-cost blueprint for computational NCE prioritisation in resource-constrained settings, not a validated drug-discovery platform; prioritised compounds are computational predictions requiring subsequent Tier 3 experimental confirmation.