Building Biology’s Blueprints: An AI for Systems Pharmacology

Author: Denis Avetisyan

Researchers have developed a new artificial intelligence system that automates the creation of complex models of biological systems, accelerating drug discovery and personalized medicine.

The system fosters a collaborative ecosystem where agents iteratively decipher and reconstruct quantitative systems pharmacology (QSP) code into equivalent MATLAB implementations-a process validated through continuous feedback-and subsequently refine these models based on natural language directives, autonomously debugging and versioning outputs to ensure traceable, evolving functionality.

GRASP leverages graph neural networks and multi-agent systems to construct Quantitative Systems Pharmacology models while preserving biological constraints and incorporating human expertise.

Quantitative Systems Pharmacology (QSP) modeling is critical for drug development yet remains a time-consuming bottleneck for researchers. To address this, we present ‘GRASP: Graph Reasoning Agents for Systems Pharmacology with Human-in-the-Loop’, a novel multi-agent framework that encodes QSP models as biological knowledge graphs and translates them into executable code while upholding critical constraints. GRASP leverages graph-based reasoning and natural language interaction to automate model construction, achieving superior biological plausibility and mathematical correctness compared to existing methods. Could this approach unlock a new era of accessible and rigorous QSP modeling, accelerating the pace of biomedical discovery?

The Inevitable Complexity of Biological Modeling

Quantitative Systems Pharmacology (QSP) modeling, while powerful, historically presents a significant bottleneck in drug discovery due to the substantial time and manual effort required for model creation and refinement. Building a predictive model often necessitates painstaking curation of biological data, followed by iterative cycles of model construction, parameter estimation, and validation – a process that can span months or even years. This reliance on manual intervention limits the speed at which hypotheses can be tested and explored, particularly in the face of evolving experimental data or the need to investigate numerous potential drug candidates. The laborious nature of traditional QSP approaches therefore restricts the scalability of systems pharmacology, hindering its broader adoption and impacting the efficiency of pharmaceutical research and development.

Biological systems, from single cells to entire organisms, present an unparalleled level of intricacy, demanding innovative approaches to transform raw experimental observations into functional, predictive models. Traditional methods often falter when faced with the sheer volume of interacting components and feedback loops characteristic of living processes. Consequently, researchers are actively developing computational frameworks that can efficiently assimilate diverse datasets – including genomic, proteomic, and metabolomic information – and translate them into mathematical representations of biological phenomena. These emerging techniques prioritize not only accuracy in replicating observed behaviors, but also scalability, allowing for the modeling of increasingly complex systems and the prediction of responses to novel stimuli or perturbations. The ultimate goal is to move beyond descriptive models towards mechanistic understandings, enabling precise predictions and ultimately, informed interventions in biological processes.

Current quantitative systems pharmacology (QSP) methodologies often falter when adapting to the inherent intricacies of biological systems, specifically regarding the preservation of biological realism during model refinement. The iterative process of model building frequently involves adjustments to parameters and structural components to achieve a desired fit with experimental data; however, these modifications can inadvertently introduce implausible relationships or distort established biological knowledge. Capturing the nuanced interplay of complex interactions – such as feedback loops, allosteric regulation, and protein-protein interactions – proves particularly challenging. A model that accurately reflects initial conditions may drift from biological validity as successive alterations are made, potentially leading to predictions that, while mathematically consistent, lack biological meaning. This disconnect necessitates careful validation and often requires substantial manual curation to ensure the final model remains grounded in established biological principles and accurately represents the system’s underlying mechanisms.

The final model successfully implements cooperative trimer formation, demonstrating GRASP’s ability to handle complex biological scenarios with sequential binding, cooperative kinetics, and stoichiometry while upholding existing constraints.

Automating the Inevitable: A Graph-Reasoning System

GRASP utilizes a multi-agent system architecture to automate the development of Qualitative Spatial Programs (QSPs). This framework decomposes QSP model construction into discrete tasks assigned to specialized agents, facilitating parallel and distributed reasoning. A core component is the integration of Graph Neural Networks (GNNs) which operate directly on the knowledge graph representing the system’s components and their spatial relationships. These GNNs learn to predict relationships and constraints, informing the agent’s decision-making process and ultimately generating the QSP model. The system’s novelty lies in its ability to autonomously translate high-level spatial queries into executable QSP code by leveraging the combined strengths of agent-based task decomposition and graph-based relational reasoning.

GRASP utilizes a Knowledge Graph (KG) to represent biological entities – such as genes, proteins, and metabolites – and their known interactions. This KG serves as the foundational data structure for automated model construction, with nodes representing biological components and edges defining relationships like protein-protein interactions, metabolic reactions, or regulatory links. The graph-based representation allows GRASP to perform structured reasoning through graph traversal and pattern matching, identifying relevant components and relationships for Quantitative Systems Pharmacology (QSP) model building. This approach facilitates efficient model construction by predefining potential interactions and reducing the search space for model parameters, ultimately accelerating the development of predictive biological models.

GRASP incorporates Human-in-the-Loop (HITL) approaches to leverage domain expertise during QSP model development. This involves presenting intermediate model states and reasoning paths to human experts for validation and refinement. Specifically, experts can review proposed relationships within the Knowledge Graph, correct inaccuracies in component assignments, and guide the selection of appropriate model parameters. This iterative process of automated generation and human curation addresses the limitations of purely automated systems, which may struggle with complex biological nuances, and ensures higher model accuracy and reliability compared to models built solely through automated methods. The HITL framework also facilitates knowledge transfer, allowing domain expertise to be embedded within the automated system over time.

GRASP effectively translates natural language prompts into increasingly complex pharmacokinetic models, progressing from a basic two-compartment system to one incorporating receptor binding, dual-target modulation, and ultimately, cooperative trimer formation with preserved mass balance.

Preserving the Illusion of Biological Validity

GRASP utilizes constraint preservation mechanisms to ensure the biological realism of generated models by enforcing both mass balance and stoichiometric consistency. Mass balance is maintained by tracking the conservation of all chemical species within the modeled system, preventing the artificial creation or destruction of matter. Stoichiometry is enforced by verifying that all reactions adhere to correct molar ratios of reactants and products, as defined by the underlying biochemical pathways. These constraints are applied throughout the model generation process, ensuring that the resulting network represents a physically and chemically plausible biological system. Failure to meet these criteria results in model rejection or iterative refinement until constraints are satisfied.

GRASP utilizes iterative validation procedures to enhance model reliability by systematically assessing both network topology and syntactical correctness. Topology checks verify the logical connectivity and structure of the generated biological network, identifying inconsistencies or invalid connections between nodes. Simultaneously, syntax checks confirm that all model components adhere to predefined rules governing reaction definitions and parameter assignments. These checks are performed iteratively during model generation and refinement, allowing for the identification and correction of errors before they propagate through the simulation. This process improves model robustness and reduces the likelihood of biologically implausible results.

Parameter alignment within the GRASP system utilizes a Breadth-First Search algorithm to systematically explore possible parameter configurations during model modification. This approach prioritizes consistency and biological relevance by evaluating parameter sets in a layered fashion, ensuring that changes adhere to established biological principles. Quantitative assessment demonstrates a high biological plausibility score of 9/10 using this method, representing a measurable improvement over Subject Matter Expert (SME)-guided methods which achieve a score of 7/10. This indicates the algorithm’s effectiveness in generating biologically sound models with a reduced reliance on manual curation.

The model successfully incorporates target-mediated drug disposition mechanisms-such as receptor binding, internalization, and degradation-without disrupting existing pharmacokinetic characteristics.

Orchestrating the Inevitable: A System of Systems

LangGraph functions as the central nervous system of the GRASP framework, providing the architecture necessary to coordinate a network of interacting agents. This orchestration is not simply sequential task management; instead, LangGraph enables dynamic routing of information and responsibilities between specialized agents – each contributing to the overall goal of quantitative systems pharmacology (QSP) model development. The framework manages the complex workflow by defining clear communication channels and decision-making processes, allowing agents to collaboratively generate, evaluate, and refine model components. This agent-based approach, facilitated by LangGraph, moves beyond traditional monolithic modeling by distributing the workload and fostering a more adaptable and robust system for tackling complex biological problems, ultimately streamlining the entire QSP modeling lifecycle.

A critical component of the GRASP system is the utilization of Large Language Models (LLMs) for rigorous code evaluation and comparative modeling analysis. These LLMs don’t simply execute generated code; they actively assess its quality, identifying potential errors, inefficiencies, and adherence to established programming standards. Furthermore, the system employs LLMs to compare multiple modeling approaches-such as different sets of ordinary differential equations or alternative parameterizations-by analyzing their outputs, evaluating their biological plausibility, and predicting their performance against specified criteria. This automated, LLM-driven evaluation process ensures a high degree of quality control and efficiency, allowing the system to rapidly iterate through diverse modeling strategies and converge on optimal solutions without extensive manual intervention. The objective comparison facilitated by the LLMs enables the selection of the most robust and accurate models for quantitative systems pharmacology investigations.

The practical utility of GRASP stems from its deliberate design for interoperability with existing quantitative systems pharmacology (QSP) tools, notably MATLAB and SimBiology. This integration isn’t merely about compatibility; it allows researchers to leverage established models, data, and workflows without requiring a complete overhaul of their current practices. By seamlessly connecting to these platforms, GRASP can efficiently utilize existing resources, accelerating the model generation and refinement process. This approach was rigorously validated through pairwise preference testing, where GRASP consistently outperformed baseline methods with a statistically significant 71% win rate, demonstrating a substantial improvement in both efficiency and the quality of generated models within standard QSP pipelines.

Using SME-guided prompts, GRASP outperforms Chain-of-Thought and Tree-of-Thoughts reasoning when evaluated by a large language model.

The Expanding Horizon of Automated Modeling

The GRASP system demonstrates a particular strength in modeling intricate biological processes, notably those governed by Target-Mediated Drug Disposition (TMDD). These systems, characterized by complex interactions between drugs, targets, and downstream effects, often pose significant challenges for traditional quantitative systems pharmacology (QSP) approaches. GRASP’s architecture is specifically designed to dissect and represent these relationships, effectively capturing the nuances of drug distribution and elimination influenced by target engagement. This capability is crucial for predicting drug behavior in vivo and optimizing dosage regimens, ultimately accelerating the development of more effective therapeutics by providing a robust framework for simulating complex pharmacological scenarios.

Quantitative systems pharmacology (QSP) model creation is traditionally a labor-intensive process, demanding significant time and expertise; however, the GRASP system demonstrably streamlines this workflow through automation. Studies reveal a substantial efficiency gain, as GRASP not only accelerates model building but also exhibits superior performance in identifying missing parameters crucial for accurate simulations. Specifically, GRASP achieves an impressive F1 score of 0.95 in detecting these parameters, a considerable improvement over the 0.68 score attained through manual model extension – highlighting its potential to significantly reduce development timelines and enhance the reliability of drug discovery efforts. This automated approach promises to unlock faster, more accurate predictions of drug behavior, ultimately facilitating the advancement of novel therapeutics.

Continued development centers on enhancing GRASP’s foundational knowledge base through an expanded knowledge graph, encompassing a wider range of biological mechanisms and drug interactions. Simultaneously, efforts are directed towards refining the reasoning capabilities of the agent-based system, allowing for more nuanced and accurate predictions regarding complex pharmacological behaviors. Crucially, integration with diverse data sources – including clinical trial results, genomic data, and real-world evidence – will be pivotal in validating model predictions and improving the overall robustness and translational potential of the platform, ultimately accelerating the pace of pharmaceutical innovation.

The model accurately integrates a dual-target TMDD system with R2 receptor dynamics, demonstrating consistent mathematical handling of complex multi-target pharmacology including competitive drug binding and parallel pathways.

The pursuit of automated model construction, as exemplified by GRASP, inevitably reveals the inherent limitations of any formalized system. It attempts to capture the fluidity of biological processes within the rigid confines of graph-based reasoning, a compromise frozen in time. As Barbara Liskov observed, “It’s one of the most difficult things to do well – to design something that is flexible enough to meet all the requirements.” GRASP, while promising in its ability to integrate biological knowledge and constraint preservation, merely delays the inevitable entropy. Dependencies will shift, understanding will evolve, and the model, however meticulously crafted, will become a snapshot of a past reality. The ecosystem always finds a way.

Where Do We Go From Here?

The pursuit of automated model construction, as exemplified by GRASP, feels less like engineering and more like tending a garden. One prunes, encourages, and hopes for robustness, knowing full well that the most carefully cultivated system will eventually succumb to unforeseen pressures. Scalability is, after all, just the word used to justify complexity. The elegance of constraint preservation is a temporary reprieve from the inevitable drift of real-world data, a delaying action against the entropy inherent in biological systems.

The natural language interface, while promising, highlights a deeper issue: the translation of human intuition into formal logic is fundamentally lossy. Each successful interaction is not a step towards true understanding, but a carefully constructed illusion of it. The system doesn’t reason with biology; it mirrors the reasoning of those who built it, amplifying their biases and limitations.

Perhaps the true challenge isn’t automating model construction, but accepting the inherent imperfection of all models. Everything optimized will someday lose flexibility. The perfect architecture is a myth to keep sane, and the next step isn’t a more sophisticated algorithm, but a more humble approach-one that embraces uncertainty, prioritizes interpretability, and recognizes that the map will never be the territory.

Original article: https://arxiv.org/pdf/2512.05502.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/