The UK QSP Training Workshop brings together a range of researchers working on QSP in academia and industry. It provides an opportunity to learn from expert speakers, poster presentations, and interactive training sessions.

Can’t Miss Presentation!

Title: Actual intelligence, thoughtfully integrated: Computation-in-the-loop as a paradigm for agentic QSP
Speaker: Priyata Kalra, Product Manager, Simulations Plus
Date/Time: Friday, July 10th, 1:15 PM – 2:15 PM

Mechanistic models in quantitative systems pharmacology (QSP) provide a principled framework for integrating biological knowledge with data, yet their development remains largely bespoke, manual, and time-        intensive. Translating a scientific hypothesis into validated, decision-ready insight typically demands months of iterative model execution, scenario testing, parameter sourcing, and documentation, the majority of which is translational work rather than mechanism development: mapping clinical datasets to model inputs, reconciling study-design details, extracting and unit-converting literature parameters with traceable provenance, editing configurations, and assembling results into reporting-grade outputs.

This session introduces and demonstrates a Computation-in-the-Loop framework that couples large language models (LLMs) with a deterministic mechanistic QSP engine to enable natural-language–driven 19 execution, interrogation, and interpretation of QSP models; while leaving the biology, the equations, and the numerics firmly under human control. The central architectural commitment, and the one most relevant to a methodologically rigorous audience, is the inversion of the usual trust relationship. The mechanistic engine remains the sole source of numerical truth; the LLM acts only as an orchestration and reasoning layer within a multi-agent system. Stochastic AI reasoning is thereby anchored in validated mechanistic computation: intent expressed in natural language is routed only to deterministic, validated procedures and never to improvised numerics.

This anchoring is what mitigates hallucination risk and preserves auditability; every action is logged, and every result traces back to the request and the source prompt and data that produced it. Working through a representative disease case study, participants will engage with three coordinating agents that interact through standardized Model Context Protocols (MCP). A Data agent performs reproducible preparation – ingesting raw trial deliverables and analysis plans, surfacing missing values and design assumptions rather than hiding them, and enforcing formatting rules as reusable, version-controlled tests so that bad data fails before it reaches the engine. A Modeling agent drives deterministic execution and structured scenario exploration – composing combination and population scenarios with validity enforced before evaluation, running virtual cohorts, and returning outputs with appropriate uncertainty and reference comparators.

A Reporting agent captures the exploratory session as a transparent, auditable record and re-expresses models into standard formats (with explicit acknowledgement that solver and hardware differences require human intervention). The session is framed honestly around its limits. We will distinguish where agents add genuine value, schema-validated extraction, free-text-to-parameter mapping grounded in literature-informed decisions, model construction and reasoning from where human judgement remains non-negotiable: mechanistic structure, governance of ambiguous parameter assignments, and scientific interpretation. By the end, participants will be able to run a complete curation-to-report cycle through an agentic workflow, critically evaluate a trust-by-construction architecture and learn concepts useful against the fast moving AI for QSP definition, and articulate for their own work the boundary between what should be automated and what must stay human. Familiarity with QSP or population modeling is assumed; no prior agentic-AI experience is required.