AI Agents Chart a Course for Faster Drug Discovery

Author: Denis Avetisyan

A new framework combines the power of large language models with structured workflows to accelerate and improve the reliability of identifying potential new medicines.

Mozi establishes a hierarchical agent system-comprising a central Coordinator and specialized Research and Computation workers-that executes scientific workflows through dynamic task instantiation, reflection-driven monitoring, and structured report synthesis, all integrated via the MCP Platform to standardize access to computational biology tools and biomedical databases for autonomous discovery.

Mozi enables governed autonomy for drug discovery by integrating LLM reasoning with workflow graphs, ensuring auditability and reproducibility.

While large language models show promise in scientific discovery, their deployment in high-stakes fields like drug discovery is hindered by issues of reliability and reproducibility. To address this, we introduce Mozi: Governed Autonomy for Drug Discovery LLM Agents, a novel framework that integrates the reasoning capabilities of LLMs with a dual-layer architecture of governed workflows and skill graphs. This approach enforces robust tool usage, safeguards scientific validity through strategic human-in-the-loop checkpoints, and mitigates error accumulation across complex pipelines. By transforming LLMs from conversationalists into reliable co-scientists, can Mozi unlock a new era of accelerated and trustworthy drug discovery?

The Inherent Flaws of Conventional Therapeutic Discovery

The pharmaceutical industry faces a growing crisis in research and development, characterized by protracted timelines and ballooning expenses, especially when tackling complex diseases like Alzheimer’s or many cancers. Historically, bringing a single new drug to market required, on average, over a decade and upwards of two billion dollars; recent analyses suggest these figures are worsening. This inefficiency stems from a reliance on sequential experimentation – a linear process of hypothesis, testing, and analysis – which proves inadequate for diseases governed by intricate biological networks. The sheer number of potential drug candidates, combined with the high failure rate in clinical trials-often due to unforeseen side effects or lack of efficacy-contributes significantly to these escalating costs. Consequently, innovation is stifled, and the development of therapies for widespread, debilitating conditions is unduly delayed, creating a critical need for fundamentally new approaches to therapeutic discovery.

The escalating intricacy of biological systems presents a fundamental challenge to therapeutic discovery, necessitating a shift beyond traditional knowledge representation. Simply accumulating data is insufficient; effective strategies require methods capable of structuring biological information in a manner that mirrors the interconnectedness of life itself. Researchers are exploring approaches like knowledge graphs and computational ontologies to move past linear understandings of disease, instead modeling complex relationships between genes, proteins, pathways, and environmental factors. These systems allow for reasoning about potential drug targets, predicting off-target effects, and identifying novel therapeutic strategies that would remain hidden using conventional methods. This move toward sophisticated knowledge representation isn’t merely about storing more facts; it’s about building a computational framework capable of mimicking the nuanced, dynamic logic of biological systems and enabling truly predictive and personalized medicine.

Contemporary therapeutic development often falters when addressing diseases requiring extended, sequential interventions – termed ‘long-horizon tasks’. Traditional drug discovery pipelines excel at identifying targets and initial compounds, but struggle with predicting efficacy through multiple biological stages and anticipating adaptive responses within a dynamic system. This limitation arises because most current methods prioritize short-term predictions based on static data, failing to adequately model the cascading effects of interventions over extended periods. Effectively navigating these complex scenarios demands computational approaches capable of simulating biological processes with sufficient fidelity to anticipate unforeseen consequences and iteratively refine therapeutic strategies, a significant hurdle given the inherent uncertainties within biological systems and the vastness of the relevant data space.

Modern therapeutic discovery is increasingly defined by the need to synthesize information from disparate and voluminous datasets – genomic sequences, proteomic profiles, clinical trial results, and even patient lifestyle data – creating a complex analytical challenge. This integration isn’t simply a matter of data aggregation; it demands sophisticated computational methods capable of identifying subtle patterns and causal relationships within noisy and incomplete information. The scientific landscapes surrounding complex diseases are inherently ambiguous, filled with conflicting hypotheses and incomplete mechanistic understanding. Consequently, successful therapeutic strategies require navigating this uncertainty, leveraging probabilistic reasoning and machine learning to prioritize promising avenues of research and mitigate the risks associated with pursuing poorly understood biological pathways. Ultimately, the ability to seamlessly connect data and confidently interpret ambiguity will be crucial for accelerating the development of effective treatments for challenging diseases.

Our Mozi platform generated novel [latex]LRRK2[/latex] inhibitors with comparable or improved predicted binding affinity (measured by ipTM) to existing clinical candidates and those from alternative biomedical agent platforms, as demonstrated by comparative analysis of molecular structures and predicted interface TM-scores.

Mozi: An Agentic Framework for Rigorous R&D

The Mozi framework addresses limitations inherent in both traditional research and development workflows and those employing Large Language Models (LLMs). Traditional methods often lack the flexibility to rapidly adapt to new data or explore diverse hypotheses. LLM-based approaches, while offering increased automation, can suffer from issues of hallucination, lack of reproducibility, and difficulty adhering to established scientific protocols. Mozi distinguishes itself through an architecture that combines the reasoning capabilities of LLMs with a structured system for workflow management and control, aiming to mitigate these deficiencies and enable more reliable and scalable scientific discovery.

Mozi’s dual-layer architecture distinguishes between a governance layer and an execution layer to improve control and adaptability in research and development workflows. The governance layer, comprised of high-level objectives and constraints, directs the overall process without directly implementing steps. Conversely, the execution layer handles the specific tasks and calculations necessary to achieve those objectives. This separation allows for dynamic adjustment of the execution layer – including the LLM and associated tools – without altering the overarching research goals, and facilitates intervention and oversight to ensure adherence to scientific rigor and safety protocols. The architecture enables rapid iteration and refinement of experimental procedures while maintaining a consistent framework for reproducibility and validation.

The Mozi Control Plane utilizes prompt engineering to direct the Large Language Model (LLM) toward scientifically valid reasoning and outputs. This is achieved through carefully constructed prompts that define the task, specify constraints, and request outputs in a structured format. Prompt engineering within Mozi includes techniques like providing relevant background information, decomposing complex problems into sub-steps, and requesting justifications for each step. By explicitly guiding the LLM’s reasoning process, the Control Plane mitigates the risk of hallucinations or logically unsound conclusions, ensuring that generated hypotheses and experimental designs align with established scientific principles and methodologies. The system also allows for iterative refinement of prompts based on performance feedback, enhancing the LLM’s ability to consistently produce reliable and reproducible results.

Mozi utilizes skill graphs as a structured representation of standardized scientific workflows, facilitating both reproducibility and scalability in research and development processes. These graphs define tasks as nodes and their dependencies as edges, allowing for the decomposition of complex experiments into discrete, manageable units. By explicitly defining these workflows, Mozi ensures consistent execution and facilitates the tracking of experimental parameters and results. The skill graph approach enables efficient reuse of existing workflows, adaptation to new research questions, and automated execution of multi-step experiments, thereby improving the speed and reliability of scientific discovery. Furthermore, the standardized format allows for easy sharing and collaboration among researchers, promoting the verification and extension of existing findings.

Layer B's workflow plane dynamically routes tasks through discovery stages or a complete pipeline ([1-3]) utilizing a stateful execution mechanism with enforced data contracts and Human-in-the-Loop validation, enabling iterative refinement, expert oversight, and controlled termination to prevent error propagation. — Layer B’s workflow plane dynamically routes tasks through discovery stages or a complete pipeline ([1-3]) utilizing a stateful execution mechanism with enforced data contracts and Human-in-the-Loop validation, enabling iterative refinement, expert oversight, and controlled termination to prevent error propagation.

Workflow Integrity and Data Provenance within Mozi

The Mozi ‘Workflow Plane’ functions as an intermediary layer translating high-level experimental protocols into executable skill graphs. These graphs represent a series of interconnected computational ‘skills’ – discrete operations such as molecule generation, property prediction, or statistical analysis – and define the precise sequence and dependencies for automated execution. This materialization allows for the automation of complex, multi-stage experiments, removing the need for manual intervention between steps and facilitating rapid iteration and analysis. The skill graphs are dynamically constructed based on the specified protocol, enabling flexible experimental design and adaptation to varying research objectives. Each skill within the graph is modular and reusable, promoting efficiency and reducing redundancy in the experimental process.

The Mozi framework incorporates the principal stages of therapeutic discovery – target identification, hit identification, and lead optimization – into a unified, automated pipeline. Target identification is facilitated through integrated genomic and proteomic data analysis, enabling the selection of relevant biological targets. Hit identification leverages virtual screening and in silico modeling to pinpoint potential compounds exhibiting activity against the chosen target. Subsequently, lead optimization employs iterative design, prediction, and evaluation of molecular properties to refine identified hits into drug-like lead candidates. This seamless integration allows for rapid progression through each phase, reducing the time and resources traditionally required for drug discovery efforts.

Mozi consistently completes all steps within three established drug discovery pipelines – Crohn’s disease, Parkinson’s disease, and sepsis – with a 100% success rate. This performance was observed across multiple iterations of each pipeline, indicating a high degree of reliability in executing complex, multi-stage experimental workflows. The consistent completion rate suggests the system effectively manages dependencies, handles potential errors, and maintains operational stability throughout the entire process, validating its robustness for automated therapeutic discovery.

Lead candidates identified through the Mozi framework demonstrate favorable predicted pharmaceutical properties. Specifically, optimized compounds consistently achieve Quantitative Estimated Drug-likeness (QED) scores ranging from 0.933 to 0.944, indicating a high probability of possessing characteristics suitable for oral bioavailability and desirable pharmacokinetic profiles. Furthermore, these compounds attain a Synthetic Accessibility Score of 1.0, signifying that their chemical structures are readily synthesizable using established laboratory procedures and available reagents, thereby facilitating rapid progression into physical synthesis and subsequent biological validation.

State management within Mozi utilizes a persistent and versioned representation of all experimental parameters, intermediate results, and system configurations. This ensures that each computational step operates on a well-defined and consistent dataset, preventing data corruption or inconsistencies that could arise from mutable state. The system tracks the lineage of all data objects, enabling rollback to previous states and reproducibility of experiments. This is achieved through a combination of immutable data structures and a centralized state store, allowing for deterministic execution of workflows even across distributed computing environments. Furthermore, the state management system integrates with the data provenance tracking, providing a complete audit trail of all changes to the experimental context.

Data provenance within Mozi is maintained through a comprehensive tracking system that records the origin, processing steps, and dependencies of all generated data. This includes precise logging of input parameters, software versions, execution timestamps, and the specific algorithms applied at each stage of a workflow. By meticulously documenting this information, Mozi ensures complete reproducibility of results and facilitates thorough auditing for validation and error tracing. This level of detail allows users to confidently assess the reliability of findings and trace the lineage of any particular data point back to its original source, supporting robust scientific investigation and decision-making.

The target identification workflow involves a multi-stage process to accurately locate and characterize desired objects.

The Future of Rational Pharmaceutical Innovation

The Mozi framework represents a significant leap forward in pharmaceutical innovation through its agentic approach, enabling a system capable of independently formulating and testing scientific hypotheses. Unlike traditional AI models that require explicit instructions, Mozi operates with a degree of autonomy, leveraging reinforcement learning to navigate the complex landscape of biological data and propose potential drug candidates. This isn’t merely pattern recognition; the system actively designs experiments – in silico, initially – to validate its own predictions, iteratively refining its understanding and pinpointing promising compounds. By automating the crucial early stages of drug discovery – hypothesis creation and initial validation – Mozi dramatically accelerates the research timeline and reduces reliance on manual, time-consuming processes, paving the way for faster development of novel therapeutics.

The Mozi framework distinguishes itself through an architecture designed for seamless data integration and experimental flexibility. Its modularity allows researchers to readily incorporate a wide spectrum of information – from genomic datasets and protein structures to chemical libraries and patient records – without requiring extensive code modification. Crucially, this isn’t limited to specific data types; the framework accommodates diverse experimental modalities, including high-throughput screening, computational simulations, and even data gleaned from clinical trials. This scalability ensures that as new data sources and analytical techniques emerge, they can be efficiently layered into the existing system, fostering a continuously evolving and increasingly powerful drug discovery pipeline. The result is a system capable of handling the complexity of modern biomedical research and adapting to future innovations with relative ease.

The Mozi framework demonstrates a remarkable capacity for in silico molecular design, consistently generating novel compounds exhibiting strong binding affinities – a critical metric for drug candidates – ranging from -8.8 to -9.2 kcal/mol. This performance isn’t limited to a single disease area; the system has successfully designed potential therapeutics across multiple, diverse therapeutic areas, indicating a broad applicability and robustness. Such consistently high predicted binding affinities suggest a significant probability of real-world efficacy, as compounds within this range are considered highly promising leads for further investigation. The ability to reliably generate molecules with these characteristics represents a substantial leap forward in AI-driven drug discovery, offering the potential to drastically accelerate the identification of effective treatments for a wide spectrum of diseases.

The escalating costs and protracted timelines inherent in traditional pharmaceutical research create significant barriers to developing treatments for rare diseases and addressing health disparities globally. Mozi, an AI-driven framework for drug discovery, addresses this critical challenge by substantially lowering both the financial investment and time required to identify promising therapeutic candidates. This reduction in resource demands isn’t merely incremental; it envisions a future where drug development becomes accessible to a broader range of institutions and researchers, including those in resource-limited settings. Consequently, Mozi’s potential extends beyond simply accelerating innovation – it promises to democratize access to life-saving therapies, fostering a more equitable landscape for global health and potentially unlocking treatments for conditions currently neglected due to economic constraints.

The advent of frameworks like Mozi signals a paradigm shift in biomedical research, moving beyond AI as a mere tool and towards a collaborative partnership between artificial intelligence and human scientists. This isn’t simply about automating existing processes; it’s about establishing a system where AI proactively generates hypotheses, designs experiments, and analyzes data, freeing researchers to focus on interpretation, validation, and the nuanced aspects of scientific inquiry. Such a synergy promises to dramatically accelerate the pace of discovery, potentially compressing years of traditional research into months or even weeks, and enabling exploration of previously inaccessible biological complexities. The resulting speed and efficiency aren’t just incremental improvements, but rather a catalyst for tackling previously intractable diseases and ushering in a new era of precision medicine and therapeutic innovation.

The development of Mozi, as detailed in the article, echoes a fundamental principle of computational elegance. The framework’s emphasis on structured workflow graphs, ensuring reliability and auditability in drug discovery, aligns with a commitment to provable solutions. As John McCarthy famously stated, “The best way to predict the future is to invent it.” Mozi doesn’t simply attempt to apply large language models to a complex scientific problem; it actively constructs a system where their reasoning is governed and verifiable. This isn’t about achieving a functional outcome, but about building a predictable and mathematically sound approach to accelerating drug discovery-a testament to the power of defined boundaries and consistent logic.

Future Directions

The framework presented here, while a step toward demonstrable reliability in LLM-driven scientific workflows, merely shifts the locus of potential error. The inherent stochasticity of large language models remains, and governing autonomy – however structured – does not eliminate uncertainty, but rather attempts to constrain its propagation. Future effort must focus not on elaborate workflow choreography, but on formal verification of the underlying reasoning engines themselves. A workflow graph is only as sound as the computations within each node.

The current reliance on LLMs as black boxes is particularly troubling. The temptation to treat these models as oracles, merely because they occasionally yield correct answers, is a profound intellectual failing. True progress demands the ability to prove the correctness of an algorithm, not merely demonstrate its efficacy on a limited dataset. The field needs to move beyond empirical validation and embrace formal methods, even if it necessitates sacrificing some degree of flexibility.

Ultimately, the most significant challenge lies in minimizing redundancy. Each layer of abstraction, each intervening component in the workflow, introduces potential for error. The pursuit of elegance dictates a relentless drive toward minimalism – a direct mapping of problem to solution, expressed in the most concise and provable form. Only then can one hope to build truly robust and trustworthy scientific AI.

Original article: https://arxiv.org/pdf/2603.03655.pdf

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

The Inherent Flaws of Conventional Therapeutic Discovery

Mozi: An Agentic Framework for Rigorous R&D

Workflow Integrity and Data Provenance within Mozi

The Future of Rational Pharmaceutical Innovation

Future Directions

See also: