AstraZeneca researchers present a five-agent AI system that retains a brief history of drug-effect calculations across runs
AstraZeneca researchers present a five-agent AI system that retains a brief history of drug-effect calculations across runs
Mathematical models help researchers estimate how a drug moves through the body and how its concentration relates to its effect. Ensemble QSP divides this work among five AI agents and carries forward only the information needed to continue the calculations in the next run. In tests using synthetic data, the system selected the correct model form in all 20 tasks.
A pharmacokinetic model describes how a drug enters the bloodstream, spreads through tissues, and is eliminated from the body. A pharmacodynamic model relates its concentration to its biological effect. Researchers use these calculations in drug development, including the development of candidates for treating age-related diseases.
A researcher returns to the same model many times to change an equation, correct the data, and recalculate the parameters. Each new run needs to know what has already been done and why. A complete conversation quickly fills the language model’s context window, while a short summary may omit a decision that the calculation depends on.
In the Ensemble QSP preprint, five agents search the literature, construct equations, fit parameters, write reports, and check code. A central coordinating program assigns the tasks, while agents equipped with pharmacological rules check the calculations against the model’s physical constraints. The system carries a compact map of the work into the next run, including the latest requests, modified files, unfinished tasks, and the outcome of the previous session. A separate log records the reasons behind the decisions.
Across 104 runs, the median length of this map was 301 tokens. A token is a small unit of text that a language model uses to measure the size of its context window. This compact record leaves room for new data without erasing the reasoning from earlier work.
The authors tested the system on 20 synthetic tasks for which the correct forms of the pharmacokinetic and pharmacodynamic models were known in advance. Ensemble QSP selected the correct equation structure in all 20 cases. A single agent based on GitHub Copilot did so in 14 cases and required nine debugging cycles involving a human. This test measured equation selection on synthetic data and did not use clinical data.
The preprint still lists the repository address as “[repository URL]”, so an external researcher cannot reproduce the calculations using the same code. Once the code is publicly available and the system has been tested on data not created by its authors, it will be possible to determine whether this approach maintains the quality of its calculations beyond the authors’ own tests.