The landscape of computational drug discovery is undergoing a profound transformation with AI Scientists. As artificial intelligence systems become increasingly capable, we’re witnessing the emergence of autonomous systems that can design, execute, and interpret complex scientific workflows. At OpenBioSim, we’ve been at the forefront of this evolution, and we’d like to share our journey and reflections on where this field is heading.
The OpenBioSim Foundation: Bridging the Interoperability Gap
OpenBioSim was founded to address a fundamental challenge in computational chemistry: the interoperability of open-source research software. The academic community has produced an impressive array of specialized tools, many hosted on GitHub, each excelling at specific tasks—from molecular dynamics simulations to free energy calculations to protein-ligand docking. However, these tools were developed independently, often with different data formats, incompatible dependencies, and varying levels of documentation.
This fragmentation created a significant barrier. While individual tools could perform remarkable feats, assembling them into higher-level workflows capable of solving industry-scale problems proved extraordinarily difficult. Scientists found themselves spending more time wrestling with software integration than actually doing science. The promise of combining cutting-edge academic methods with production workflows remained largely unfulfilled, not because the science was lacking, but because the software ecosystem wasn’t designed for interoperability.
We recognised that bridging this gap required more than just technical solutions—it demanded a sustainable community-driven approach to maintaining and evolving interconnected research software. This realisation became the founding principle of OpenBioSim Community Interest Company.
Building Production-Ready Workflows
Almost four years ago, we created a business around providing services to scientists who wanted to access the latest academic research in computer-aided drug design while working with robust software systems that could be reliably deployed in production environments. Our approach centered on creating libraries of interoperable workflow components that could be assembled into modular scientific workflows.
This strategy has proven successful. Several organisations have adopted our tools, integrating them into their drug discovery pipelines. By focusing on modularity and interoperability, we’ve enabled researchers to leverage cutting-edge computational methods without sacrificing the reliability and reproducibility demanded by production environments.
The AI Scientists Imperative
As maintainers of open-source community projects, we witnessed firsthand how the rapid advances in large language models were changing the way users learned to work with our toolkits. Code assistance, documentation queries, and troubleshooting were increasingly mediated through AI systems. But earlier this year, we recognized something more fundamental: Agentic AI was going to transform not just how scientists learn software, but how they consume it.
We expect that state-of-the-art AI models will remain the province of a few leading organisations. However, we believe there will be significant utility in different implementations of specialised AI scientist systems tailored to specific domains. Just as with traditional software systems, these frameworks will require ongoing maintenance, evolution, and domain expertise to remain effective. This prompted us to begin building our own AI Scientist system to explore autonomous experiments using the OpenBioSim stack and other functionality commonly used in computer-aided drug design.
Technical Insights from the Journey
Currently, the effectiveness of agentic systems is bounded by context window limitations. Many architectures attempt to mitigate this constraint through a user-facing orchestrator LLM that maintains a task list, delegating subtasks to specialized subagents executing specific tools. In our work, we’ve found that restricting the toolkit available for a given scientific problem helps decrease stochasticity between repeat sessions and reduces token overhead.
We found Claude’s Agent Skills framework easier to work with than alternative approaches like MCPs (Model Context Protocol). Skills provide a convenient abstraction for documenting command-line interfaces and scripts while making them accessible to agents. We’ve structured our skills such that tools return only the key results needed to support decision-making at the next reasoning phase. Additional tool outputs are stored in the filesystem and read only when necessary, with hierarchical storage structures enabling context-efficient lookups.
A particular challenge for drug design systems is the vast variation in time-to-result across different computational chemistry tools—ranging from near-instantaneous to hours or even days. Asynchronous tool polling has proven essential, allowing the system to progress multiple tasks in parallel until decisions can be made.
With these design principles in mind, it becomes relatively straightforward to assemble an agentic AI system that can be prompted to carry out specialised CADD tasks. The prompt below illustrates a request that triggers execution of docking, scoring and chemoinformatics tools to profile a set of candidate ligands for a protein. The system then performs a nuanced evaluation of the outputs to ultimately prioritise a compound for progression.
‘’I have uploaded a set of SDF ligand files for the protein target Tyk2 in the folder “inputs/ligands”. The protein can be found in “inputs/protein” in both PDB and AMBER format files. Using the skills in this repository predict the water solubility and binding affinity of each ligand. When done, create a comprehensive report in a formatted pdf document that summarises, the methodology, the results, conclusions and recommendations for prioritising a compound for experimental testing.’’
The system generates a human readable report to support the decision-making process. A section of the report is reproduced below.
The Future of CADD Software: Challenges and Opportunities
We’re currently witnessing frantic efforts to wrap as many existing CADD functionalities as possible into agent-accessible tools, with the reasoning that systems with access to larger toolsets will be more powerful. However, this overlooks a fundamental challenge: essentially all molecular modeling methods fail to make reliable predictions on at least some problem classes.
Experienced CADD scientists interpret model results with considerable nuance. The default position is often healthy skepticism unless provided with evidence that a model has been validated for the specific problem at hand. Factors like the chemical space of the dataset, the quality of the training data, the appropriateness of the physical approximations, and the limitations of sampling methods all inform expert judgment. Embodying this capability in an LLM-based system poses significant challenges that must be addressed to progress beyond simply running benchmark pipelines.
The true vision of an AI Scientist isn’t a system that can execute a predefined sequence of CADD tools—it’s a system capable of autonomously discovering experimentally validated ligands for novel drug targets. Achieving this will require not just tool access, but the kind of contextual understanding and critical evaluation that currently exists only in experienced practitioners.
Commoditisation and Its Implications
We believe that the commoditisation of AI Scientists will have intriguing implications for the CADD software ecosystem. Commoditised agentic AI systems will lower adoption barriers for diverse software tools, including open-source solutions that may have been historically neglected due to integration costs. When an AI agent can learn to use a new tool from documentation and examples, the traditional barriers of complex installation procedures, unfamiliar interfaces, and steep learning curves diminish significantly.
We envision a future where greater value is placed on teams of scientists who can produce high-quality specialized tools, while middleware and integration tooling becomes less prominent. The competitive advantage will shift from who has the most polished user interface to who has the most reliable and innovative underlying methods. This democratization of access could accelerate the pace of methodological innovation while making cutting-edge computational chemistry more accessible to a broader community of researchers.
The journey toward truly autonomous AI scientists in drug discovery is just beginning, but the implications for how we develop, maintain, and deploy computational chemistry software are already becoming clear. At OpenBioSim, we’re excited to continue building the interoperable, open-source foundation that will enable this future.
Written by Julien Michel and proofed by Claude
