PMxAgent: An Agentic Platform for Pharmacometrics

Abstract

AI agents are transforming computational workflows, yet pharmacometric analyses remain manual, relying on significant data wrangling, custom scripts, and specialized software. We present PMxAgent, an open-source agentic platform for developing and deploying specialized agent-callable pharmacometric tools. The platform uses Docker to orchestrate an R-based API server alongside a Python-based model context protocol (MCP) server, which automatically generates agent-callable tools from OpenAPI specifications, making pharmacometric functions discoverable and usable by AI agents. To exemplify pharmacometric applications, five tools were developed: non-compartmental analysis (NCA) using PKNCA, exposure-response (ER) modeling, pharmacokinetic (PK) simulation using mrgsolve, data standardization (DATA), and generation of population PK datasets from the nlmixr2lib model library (LIBRARY). As a proof-of-concept case study, PMxAgent orchestrated a multistep pharmacometric workflow, consisting of a PK simulation of 60 subjects across three dose groups, NCA to derive individual exposure metrics, and ER analysis. To evaluate analytical accuracy and reproducibility, the agentic NCA workflow was benchmarked against four frontier AI agents across 182 drugs and 1820 simulated subjects, using PKanalix as reference. PMxAgent's NCA accuracy (98.3%) matched or exceeded that of all frontier agents evaluated. PMxAgent produced deterministic, reproducible results in contrast to GPT and Claude agents that performed NCA by generating new code each run. PMxAgent was demonstrated using two MCP-compatible AI agents (Cursor and Claude Code) and is designed to integrate with any AI agent supporting the MCP protocol. PMxAgent provides an extensible foundation for integrating pharmacometric tools into human-supervised AI-driven workflows while ensuring reproducibility, transparency, and detailed documentation required for model-informed drug development.