Ali Farnoud1, Anuraag Saini1
1Boehringer Ingelheim Pharma GmbH & Co. KG
Introduction/Objectives: Quantitative Systems Pharmacology (QSP) combines systems biology, mechanistic modelling, and pharmacology to simulate disease progression and therapeutic effects [1]. QSP model development requires significant manual effort, including literature review, building biologically relevant mathematical framework, hypothesis generation, model calibration, and validation [2]. The process is often time-consuming, iterative requiring close collaboration between biologists, pharmacologists, and mathematical modelers [3]. To improve these aspects, we introduce QSP-Copilot, a cloud-based, AI-powered multi-agent system designed to streamline QSP workflows. QSP-Copilot integrates multiple large language models (LLMs) [4], retrieval-augmented generation (RAG) [5], and AI-driven multi-agent orchestration [6] to accelerate model building, calibration, and validation, while maintaining human oversight for final decision making, compliance and interpretability [6]. Implementation of QSP-Copilot reduces the end-to-end QSP project execution time by 30% to 40%. Methods: QSP-Copilot is built on a modular, multi-agent AI framework where specialized agents collaborate on various tasks of the model development. While the AI assists in automation, retrieval, and reasoning, users remain in control, ensuring that all modelling decisions align with scientific best practices. The system includes: 1. Multi-Agent System for QSP Modelling: QSP-Copilot consists of multiple AI agents, each performing distinct tasks: • Literature Review Agent: provides a summary and potential mechanism of actions for a disease area. This agent identifies biological mechanisms, components, interaction types, reaction types and relevant scientific references. Furthermore, this agent summarizes the mechanism of actions and allows the user to review it and once the user has confirmed it, interaction diagram for a specific disease area will be created. • Model Building Agent: Based on interaction diagram, it builds the set of ordinary differential equations (ODEs) describing the system interactions. It supports model simulations, parameter estimation and performs model calibration, sensitivity analysis, and virtual cohort generation. • Orchestration: A framework is designed to orchestrate role-playing of AI agents. Unlike traditional single-agent approaches, this approach enables the creation of multiple specialized agents that can work collaboratively for QSP model development. 2. Dynamic LLM Selection for Optimized Performance: Unlike static AI implementations, QSP-Copilot allows users to dynamically select from multiple LLMs, tailoring AI responses based on model complexity, reasoning depth, and computational constraints. It supports faster response (Claude 3 Haiku, GPT-4o-mini) for rapid information retrieval and high reasoning capability (GPT-4o, Claude 3 Sonnet) for complex QSP tasks such as multi-scale model generation. 3. Human-in-the-Loop (HITL) for Decision Making: QSP-Copilot is designed as a human-in-the-loop (HITL) system, ensuring that all AI-generated outputs are reviewed and validated by user before implementation. Ensuring interpretability & transparency users can inspect retrieved literature information, model assumptions, and suggested parameters before integrating them into the QSP model. In addition, users stay in control of the execution and final decision-making while AI accelerates the modelling process. Results QSP-Copilot reduces QSP model development time up to 40% compared to current manual workflows. Moreover, the AI powered literature review agent significantly speeds up the extraction of mechanistic insights, pharmacokinetic data, and biological parameters, reducing the time needed for manual literature review. 1. Enhanced Interdisciplinary Collaboration: QSP-Copilot improved the efficiency of cross-disciplinary communication. 2. Flexible AI Integration: The ability to select from multiple LLMs keeps the cost low. 3. Transparency: The human-in-the-loop approach ensures interpretability, transparency, and compliance, making QSP-Copilot a practical AI assistant for QSP modelling. Conclusions: QSP-Copilot represents a significant advancement in AI-assisted QSP modelling, offering an intelligent, multi-agent framework that enhances efficiency while maintaining user oversight. By automating key processes such as literature review and model development workflow, it enables modellers to accelerate QSP model development. The combination of multi-agent AI architecture, RAG, and dynamic LLM selection makes QSP-Copilot a modular and scalable AI tool. Future developments will focus on: 1. Expanding task-specific agents to assist in interactive structural modification like adding biological pathways. 2. Integration with industry-standard modelling platforms such as R, MATLAB, Julia and Python based QSP frameworks. 3. Enhancing user experience with interactive dashboards and real-time model evaluation tools. By leveraging GenAI, multi-agent AI coordination, and a human-in-the-loop design, QSP-Copilot provides a powerful, efficient, and scientifically rigorous solution for modern QSP model development and optimization.
1. Friedrich, C. (2024). What Is QSP and Why Does It Exist?: A Brief History. 2. Gadkar, K., Friedrich, C., Hurez, V., Ruiz, M. L., Dickmann, L., Kumar Jolly, M., … & Ramanujan, S. (2022). Quantitative systems pharmacology model-based investigation of adverse gastrointestinal events associated with prolonged treatment with PI3-kinase inhibitors. CPT: Pharmacometrics & Systems Pharmacology, 11(5), 616-627. 3. Helmlinger, G., Sokolov, V., Peskov, K., Hallow, K. M., Kosinsky, Y., & Voronova, V. (2019). Quantitative systems pharmacology: an exemplar model-building workflow with applications in cardiovascular, metabolic, and oncology drug development. CPT: Pharmacometrics & Systems Pharmacology, 8(6), 380–395. 4. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., … & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30. 5. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33, 9459-9474. 6. Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., … & Wang, C. (2023). Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155.
Reference: PAGE 33 (2025) Abstr 11656 [www.page-meeting.org/?abstract=11656]
Poster: Methodology – AI/Machine Learning