In a landmark shift for the pharmaceutical industry, Schrödinger, the pioneer of physics-based computational drug discovery, has announced a strategic collaboration and software agreement with Bristol Myers Squibb (BMS). At the heart of this partnership is "Bunsen," Schrödinger’s newly unveiled "AI co-scientist." This agentic engine is designed to act as an orchestrator for complex computational chemistry, transforming the way researchers navigate the staggering complexity of molecular design.
As pharmaceutical giants race to integrate artificial intelligence into their research and development pipelines, the definition of "scale" is being rewritten. By marrying Schrödinger’s established physics-based modeling with the iterative, reasoning-heavy capabilities of agentic AI, the companies aim to transition from analyzing billions of molecules to exploring trillions, all while enhancing the precision of their predictive filters.
A Chronology of Computational Evolution
The roots of this collaboration run deep. BMS has been a long-standing partner of Schrödinger, relying on the company’s suite of platforms to streamline the discovery of small-molecule therapeutics. This new agreement marks a maturation of that relationship, focusing on the deployment of Bunsen and the integration of RetroSynth, Schrödinger’s AI-driven synthesis planning platform.
The evolution toward agentic workflows did not happen overnight. It is the culmination of decades of advancement:
- The Physics-Based Foundation (1990–2010s): Schrödinger built its reputation on high-accuracy physics simulations, such as Free Energy Perturbation (FEP+). These methods allowed researchers to predict binding affinity with high confidence, though they were computationally expensive.
- The GPU Revolution (2010s): The advent of powerful GPU computing allowed Schrödinger to scale these physics-based methods, enabling larger screening campaigns than previously possible.
- The Rise of Large-Scale Exploration (2020–2024): With the launch of platforms like LiveDesign and the 2022 introduction of AutoDesigner, Schrödinger began automating the generation and progressive filtering of massive chemical spaces. By early 2024, the company successfully coupled generative AI with Active Learning FEP+, creating a feedback loop between molecular design and rigorous evaluation.
- The Agentic Inflection Point (2025–Present): The release of advanced coding agents like Claude Code provided a proof-of-concept for how autonomous systems could handle multi-step, iterative tasks. Schrödinger recognized that its own "expert" predictive tools—which act as compilers and unit tests for chemistry—required an intelligent "agent" to oversee the process. Bunsen was born from this realization.
Supporting Data: The Expanding Funnel
The necessity for Bunsen is driven by the sheer magnitude of modern drug discovery. According to Pat Lorton, Schrödinger’s CTO and COO, the scale of molecular exploration has undergone an exponential shift. Five years ago, a typical weekly cycle involved enumerating approximately 100 billion compounds. Today, that number has reached the trillions.
This expansion is not just about quantity; it is about quality. As researchers demand predictions across a wider array of biological and pharmacological properties, the "funnel" must be widened at the top to ensure a robust selection of high-quality candidates at the bottom.
Lorton draws a compelling analogy to the history of aircraft design. In the mid-20th century, engineers relied on physical wind tunnels and balsa wood models to test drag and aerodynamics. As computational fluid dynamics and supercomputing matured, the industry shifted toward simulation, drastically reducing the need for empirical physical testing. Similarly, drug discovery is moving away from a reliance on purely experimental "trial and error" toward a simulation-first approach, powered by high-performance computing.
The Role of Supercomputing in Pharma
The partnership comes as the pharmaceutical industry enters an era of "Big AI." Major players are investing heavily in massive on-premises supercomputing clusters to power their internal models.
- Lilly: Launched "LillyPod" in February 2026, utilizing over 1,000 NVIDIA Blackwell GPUs.
- Roche: Deployed an AI factory spanning more than 3,500 Blackwell GPUs.
- Bristol Myers Squibb: Recently announced the deployment of one of the industry’s most powerful AI supercomputers, built on eight NVIDIA DGX Vera Rubin NVL72 systems.
Schrödinger has designed Bunsen to be agnostic regarding the underlying LLM architecture. Whether a company chooses to run its workflows through cloud-based APIs like Claude or host its own open-weight models on a private DGX SuperPOD, Bunsen is engineered for total flexibility. This ensures that as hardware capabilities improve and proprietary foundational models become more sophisticated, the Bunsen engine remains a plug-and-play solution for the user’s preferred compute environment.
Official Perspectives and Strategic Implications
Stephen Johnson, Vice President of Computational Sciences at BMS, noted that Bunsen empowers scientists to rethink the application of physics-based tools. "It allows our scientists to think differently about how physics-based tools can be used to navigate molecular design space and accelerate the discovery of innovative medicines for patients," Johnson stated.
Erin Davis, BMS Vice President of Research Business Insights & Technology, added that the company is actively building its own foundational models, a task that demands significant computational firepower. By integrating Bunsen into this infrastructure, BMS intends to maintain its "Predict First" strategy, ensuring that AI-generated insights are a standard component of both small- and large-molecule programs.
Implications: Solving the "Human Bottleneck"
Perhaps the most profound implication of the Bunsen collaboration is its potential to alleviate the acute shortage of computational chemistry talent.
Currently, the scarcity of experts is a significant blocker. At many pharmaceutical companies, a single computational modeler might be split across multiple programs, leaving them with little time for the manual, arduous tasks of configuring software, managing data flows, and setting up complex simulations.
"For them, the arduous task of just putting these things together is a limiter," Lorton explained. "It limits their impact dramatically."
Bunsen acts as a force multiplier. By providing a natural language interface, it allows a scientist to act as an "architect" of a study rather than a "technician" of a software suite. A researcher can define a high-level goal—such as improving hydrogen bonding in a specific pocket—and Bunsen can suggest the necessary workflow, execute the calculations, and report the findings.
A Controlled Partnership
Importantly, the system is designed with safety rails. Bunsen is not a "black box" that operates entirely without oversight. Scientists have the option to:
- Direct the Agent: Use their expertise to guide the agent toward specific regions of the chemical space.
- Review the Plan: Have the agent present a "thesis" of the proposed calculation strategy for human approval before execution.
- Iterate: Correct the agent when it makes mistakes, ensuring that human intuition remains the final arbiter of scientific direction.
This collaborative model acknowledges a fundamental truth in drug discovery: while an AI can process millions of data points, it cannot replace the years of experience a medicinal chemist brings to the table. By encoding that expertise into the way scientists direct the agent, the industry can leverage the best of both worlds.
Looking Ahead
The collaboration between Schrödinger and BMS signals the end of the early, fragmented phase of AI in pharma. As workflows become longer and more complex, the ability to chain together predictive tools without losing accuracy—or time—will distinguish the industry leaders.
By treating the "AI co-scientist" as a partner that can handle the heavy lifting of execution, data retrieval, and iterative testing, Schrödinger and BMS are setting a new standard. The shift toward trillion-scale exploration, once considered a distant ambition, is now the operational reality for organizations ready to embrace the agentic era. As these tools continue to evolve, the bottleneck of drug discovery may no longer be the lack of computational power, but the speed at which human experts can dream up the next great therapeutic challenge.
About the Author
