Taking AI from Answering Questions to Running Simulations: Schrödinger Launches Molecular Discovery Assistant Bunsen

The bottleneck in drug design is often not simply a lack of computational tools, but how to break down a biological question into appropriate computational steps and then decide on the next experiment from a large volume of results. Schrödinger has launched an early-access version of Bunsen, seeking to help artificial intelligence cross this threshold: it not only answers researchers’ questions but can also plan and execute molecular simulation workflows.

According to Schrödinger, researchers can first describe their scientific objectives, project background, and constraints in natural language. Bunsen then selects appropriate computational methods, organizes multistep workflows, allocates computing resources, and ultimately analyzes and visualizes the output before proposing directions for further research. Potential applications include evaluating how compounds bind to proteins, comparing candidate molecules, or narrowing the search space before laboratory synthesis and testing.

The difference between this type of “agentic AI” and conventional literature question-answering tools is its ability to call specialized software and initiate computations. Bunsen is built on Schrödinger’s physics-based molecular modeling platform. The company says its internal drug and materials research teams have already used the system across different projects, with the aim of enabling computational chemists to advance more work in parallel while allowing drug researchers unfamiliar with modeling operations to use more complex methods.

The computing infrastructure comes from Schrödinger’s collaborations with NVIDIA and Google Cloud. Bunsen can integrate with the NVIDIA BioNeMo Agent Toolkit and use elastic computing provided by Google Cloud as well as NVIDIA Blackwell hardware. Information separately released by NVIDIA indicates that BioNeMo allows AI agents to call tools for molecular generation, docking, property prediction, protein design, and genomic analysis, enabling them to run computations, interpret outputs, and then select the next step. Schrödinger is one of the platform providers integrating these capabilities into scientific applications.

However, the evidence disclosed to date mainly describes functionality and commercial deployment. It does not provide public benchmark tests, the evaluation datasets used, or comparisons with manually planned workflows, nor does it disclose whether Bunsen improves experimental hit rates, reduces the number of syntheses, or shortens the time required for candidate drugs to enter preclinical research. The company’s statement that the modeling methods have been validated does not mean that the AI can consistently select the right tools and correctly interpret results across different targets and chemical spaces; computational rankings still require experimental confirmation.

Bunsen is currently available only to selected customers for early access, and Schrödinger expects a full commercial launch by the end of 2026. Before formal adoption, research institutions will also need to examine data confidentiality, model and software version tracking, error handling, and workflow reproducibility. Whether it can become a reliable research collaborator will depend not on how many automated steps it completes, but on whether each step can be audited and whether its final predictions are borne out in real-world experiments.