Abstract
Artificial intelligence (AI) is making a significant impact across various industries, including healthcare, where it is driving innovation and increasing efficiency. In the fields of Quantitative Clinical Pharmacology (QCP) and Translational Sciences (TS), AI offers the potential to transform traditional practices through the use of agentic workflows—systems with different levels of autonomy where specialized AI agents work together to perform complex tasks, while keeping “human in the loop.” These workflows can simplify processes, such as data collection, analysis, modeling, and simulation, leading to greater efficiency and consistency. This review explores how these AI-powered agentic workflows can help in addressing some of the current challenges in QCP and TS by streamlining pharmacokinetic and pharmacodynamic analyses, optimizing clinical trial designs, and advancing precision medicine. By integrating domain-specific tools while maintaining data privacy and regulatory standards, well-designed agentic workflows empower scientists to automate routine tasks and make more informed decisions. Herein, we showcase practical examples of AI agents in existing platforms that support QCP and biomedical research and offer recommendations for overcoming potential challenges involved in implementing these innovative workflows. Looking ahead, fostering collaborative efforts, embracing open-source initiatives, and establishing robust regulatory frameworks will be key to unlocking the full potential of agentic workflows in advancing QCP and TS. These efforts hold the promise of speeding up research outcomes and improving the efficiency of drug development and patient care.
6 Case Studies and Practical Applications
6.1 Example #1: InsightRX Apollo-AI
InsightRX Apollo-AI is a practical example of how agentic workflows can support quantitative clinical pharmacologists. Currently under development, Apollo-AI aims to enhance the analytical capabilities of QCP and TS experts by offering tools for PK and PD analyses. The system addresses several limitations of traditional LLM-based tools, such as the risk of hallucinations—where models generate incorrect or nonsensical information—and challenges associated with user interface and workflow design.
To address these challenges, the design of the agent-based analysis system was guided by several key principles: clearly defining agent roles and responsibilities, ensuring that each agent's tasks were narrowly focused, and maintaining clear human-agent interaction throughout the analysis process. The application was developed with a customized user interface (UI) and backend infrastructure. A well-designed UI/UX is essential not only for enhancing the platform's overall usability and ensuring reliable code output but also for understanding human intent throughout the analysis process. A pure chat-based UI like ChatGPT is likely to be suboptimal for PKPD analysis. For example, user workflows for QCP/TCS analysis will require a user interface that can accommodate multiple analysis tasks such as data visualization, user collaboration, analysis management, and code editing. For modeling tasks, users should be able to develop and diagnose models in an iterative manner as well as submit multiple jobs simultaneously. While low-level APIs to LLMs are available to develop a robust analysis system, they often present similar workflow challenges and are generally beyond the technical expertise of most users. Additionally, the underlying software infrastructure was customized to ensure robust data security and compliance throughout the analysis process.
The Apollo-AI system employs a variety of specialized AI agents, each with distinct roles (Figure 4), as defined below, that contribute to a cohesive and efficient analytical workflow. Central to this architecture is the Agent–Computer Interface (ACI), which enhances the functionality and efficiency of these agents.
Figure 4: Overview of the Apollo-AI system agentic workflow.
6.1.1 Conversational Agent
The Conversational Agent acts as the primary interface between the user and the system. It is specific role is to process user input, such as natural language queries, and translates them into tasks for other agents to execute. By leveraging example queries, analysis plans, and code snippets, the conversational agent ensures that the user's requests are accurately interpreted and carried out effectively.
For example, if a clinical pharmacologist wants to model a patient's drug concentration levels, the Conversational Agent will first confirm that the request pertains to population PK modeling with some preliminary analysis requirements before passing it on to a planning agent. This step is crucial for capturing user intent accurately and serves as a safeguard against downstream errors or hallucinations.
6.1.2 Planning Agent
The Planning Agent organizes the steps necessary to fulfill a user's analysis request, ensuring that everything is aligned with the user's objectives and any predefined study requirements. Before a plan is executed, the user has a chance to review, modify, and approve the plan developed by the agent, which serves as an important quality check. Similar to the conversational agent, the planning agent keeps the human in the loop by understanding user intent and making the underlying analysis process transparent. For example, an analysis plan for a basic NCA could outline the data variables to be used, provide a step-by-step guide for the analysis process (including which PK parameters to include and the method for calculating terminal half-life), and specify how to handle data below the limit of quantification (BLQs), among other considerations. The user will then have the ability to directly modify the plan before proceeding with the analysis.
6.1.3 Task Agents
Task Agents are tools designed to perform tasks throughout the analysis process, like finding data outliers, excluding data, running analysis (e.g., exploratory analysis, pop-PK, NCA), making aesthetic modifications to plots/tables, and managing the analysis workflow. Task agents will follow the plan created by the Planning Agent and use the system's resources through the ACI to complete their tasks. For example, a Task Agent might flag and remove unusual data points that could throw off the results, helping to keep the analysis accurate. For an analysis task like NCA, the task agent could invoke specific R libraries or other computational packages to fulfill the analysis request. With access to example code, outputs, and the ACI, Task Agents are able to do their tasks reliably.
6.1.4 Global Agent
The Global Agent monitors and coordinates the activities of all individual agents, is aware of the end user's interactions, and has access to the knowledge/data within the computer. Its primary objective is to offer timely recommendations and orchestrate agent actions to achieve optimal outcomes. For example, during model development, the Global Agent will track all prior modeling runs, remain aware of the study context and data constraints, and offer suggestions to the end user throughout their workflow. These recommendations may include changes to the model structure, covariance matrix, or error model.
6.1.5 Agent–Computer Interface
The ACI is a crucial component of the Apollo-AI system designed to enhance overall system performance by providing agents with an environment similar to the tools used by software engineers. This interface enables agents to navigate code repositories, access data, edit files, and execute tests. The ACI enables the retrieval of accurate and relevant knowledge to supplement an agent's response to help prevent downstream hallucinations. Specifically tailored to the operational characteristics of LLMs, the ACI mimics the interactive features of integrated development environments (IDEs) used by developers. Both Task and Planning Agents within Apollo-AI leverage the ACI to search files, write code, view and edit data, as well as run analysis code.
6.1.6 Computational Infrastructure
Referred to as the “Computer,” the underlying computational infrastructure contains all the necessary data, files, PK/PD software, and code and output examples. It interacts with the agents through the ACI, supplying the necessary resources for analysis and code execution. This part of the system acts as a repository for the AI agents, designed to have all of the necessary components required to perform clinical pharmacology analysis.
While still in development, the Apollo-AI system exemplifies how a well-coordinated agentic workflow could be built for QCP. By giving each agent a specific role and making sure they work smoothly with the available technology, the system aims to address many of the limitations associated with traditional workflows.
Shahin, M.H., Goswami, S., Lobentanzer, S. and Corrigan, B.W. (2025), Agents for Change: Artificial Intelligent Workflows for Quantitative Clinical Pharmacology and Translational Sciences. Clin Transl Sci, 18: e70188.
