In a landmark development for the pharmaceutical industry, Schrödinger, the pioneer of physics-based computational drug discovery, has announced a strategic collaboration and software agreement with long-term partner Bristol Myers Squibb (BMS). The centerpiece of this alliance is "Bunsen," Schrödinger’s newly unveiled agentic AI co-scientist. This partnership signifies more than just a software license; it represents a fundamental shift in how large-scale pharmaceutical research organizations leverage artificial intelligence to navigate the trillion-scale landscape of chemical design.
As drug discovery moves toward increasingly complex molecular spaces, the need for automated, intelligent, and scalable execution engines has become paramount. Bunsen is designed to serve as an autonomous bridge between high-level human scientific intuition and the rigorous, physics-based simulations that have long defined Schrödinger’s platform.
The Evolution of Scale: From Balsa Wood to Trillions
The trajectory of drug discovery is being compared to the evolution of aeronautical engineering. In the mid-20th century, engineers relied on physical wind tunnels and manual testing—a slow, expensive, and empirical process. Today, aerospace relies on sophisticated computational fluid dynamics and supercomputing to model performance before a single piece of metal is cut.
Schrödinger’s Pat Lorton, Chief Technology Officer and Chief Operating Officer, argues that pharmacology is undergoing a similar metamorphosis. Five years ago, a typical "large-scale" computational screen might involve enumerating roughly 100 billion compounds around a lead molecule. Today, that number has ballooned into the trillions.
This expansion is not merely about volume; it is about the "quality" of the funnel. With the ability to predict a wider array of molecular properties—solubility, metabolic stability, potency, and selectivity—simultaneously, the computational filter has become significantly more selective. "The 50 candidates coming out at the end of the process are becoming higher and higher quality," Lorton explained. "Because we are optimizing for more properties, we must widen the top of the funnel to ensure we still end up with a robust set of candidates."
A Chronology of Computational Innovation
Schrödinger’s path to the "agentic" era was not an overnight pivot. It is the culmination of decades of incremental technical milestones:
- 1990s–2000s: The foundational era, characterized by the development of robust, physics-based molecular simulation tools.
- 2010s: The GPU revolution. The advent of high-performance graphics processing units allowed Schrödinger to scale physics-based methods like Free Energy Perturbation (FEP) to levels previously thought impossible.
- 2020: The IPO filing. Schrödinger codified its vision for "Large-Scale Molecule Exploration" (computational enumeration) and "Large-Scale Molecule Evaluation" (applying machine learning and physics to massive idea sets).
- 2022: The publication of AutoDesigner, a tool capable of generating and iteratively filtering large chemical spaces.
- 2024: The integration of AutoDesigner with Active Learning FEP+, bridging generative design with precise physics-based evaluation.
- 2025–2026: The Agentic Shift. Inspired by the capabilities of coding agents like Claude Code—which can write, compile, test, and debug its own code—Schrödinger recognized the potential to apply similar autonomous loops to scientific workflows.
The Agentic Workflow: Claude Code as a Catalyst
The emergence of "agentic" AI—systems that do not just provide information but act on it—has provided the final piece of the puzzle. According to Lorton, the "agent" concept is not new, but the ability to create reliable, self-correcting loops is.
In a software engineering context, agents like Claude Code act as a partner: they write code, inspect errors, unit-test results, and modify their approach based on the feedback. Bunsen applies this same architecture to chemistry. Because Schrödinger’s predictive tools (like FEP+) provide quantitative, concrete data, they serve as the perfect "unit tests" for an AI agent.
"We have these predictive tools that are highly accurate," Lorton noted. "They are expert tools that require something smart to interpret the output and decide what to do next. That is what Bunsen does. It understands the quantitative answers from our simulations and acts as a researcher that can plan the next step."
Big Pharma’s Supercomputing Arms Race
The collaboration arrives at a time when the pharmaceutical industry is aggressively deploying massive, on-premises AI infrastructure. The cost of entry for large-scale discovery is rising, with companies investing in NVIDIA DGX SuperPODs to power their proprietary models.
- Lilly: Launched "LillyPod" in February 2026, featuring 1,016 Blackwell GPUs.
- Roche: Debuted an AI factory in March 2026, spanning over 3,500 Blackwell GPUs.
- BMS: Announced in July 2026 that it is deploying one of the most powerful AI supercomputers in the industry, built on eight DGX Vera Rubin NVL72 systems.
Schrödinger has designed Bunsen to be hardware- and model-agnostic. While it integrates seamlessly with leading LLMs like Claude, it is built to run on a client’s local infrastructure. "If you want to run your own open-weight LLM locally on your own SuperPod, we’ve engineered the system to accommodate that," Lorton said. This flexibility allows firms like BMS to integrate Bunsen into their "Predict First" strategy—a mandate to bring AI-generated predictions into every single small-molecule program and the majority of large-molecule programs.
Implications: Addressing the Talent Bottleneck
One of the most profound implications of the Bunsen collaboration is its potential to solve the industry’s most persistent bottleneck: the shortage of expert computational modelers.
Currently, the distribution of expert talent is uneven. While Schrödinger may assign five experts to a single internal discovery program, many industry programs have only a fraction of a modeler’s time available. These experts often find themselves bogged down in "the arduous task" of manual workflow assembly—locating data, setting up complex environments, and waiting for manual feedback.
"It is a limiter," Lorton said. "When they see a plain-text interface that acts as a postdoc-level partner—someone who can fill in the details and ask the right questions—their impact increases dramatically."
Human-in-the-Loop: The Future of Collaborative Science
Despite the autonomy of the Bunsen platform, Schrödinger is clear that the human element remains non-negotiable. The platform is designed with a "human-in-the-loop" philosophy:
- Guided Autonomy: Scientists can choose to have Bunsen propose a plan and pause, waiting for human approval before launching expensive, large-scale calculations.
- Review Mechanisms: The agent is designed to return to the human scientist when it encounters ambiguity, acknowledging that even the most advanced AI can make errors.
- Strategic Steering: The most effective use of the tool, according to Lorton, is when the scientist directs the agent, leveraging their years of experience to point the AI toward specific chemical pockets or binding hypotheses that the agent might overlook on its own.
As Stephen Johnson, Vice President of Computational Sciences at BMS, noted in the official announcement, the technology "allows our scientists to think differently about how physics-based tools can be used to navigate molecular design space."
Conclusion
The partnership between Schrödinger and Bristol Myers Squibb represents a shift from "tools as software" to "tools as collaborators." By combining the raw power of NVIDIA-driven supercomputing with the iterative, self-correcting logic of agentic AI, the industry is entering an era where the speed of drug discovery is no longer limited by the number of hours in a day, but by the creativity of the scientists guiding the machines.
As Bunsen begins its deployment within BMS, the industry will be watching closely to see how this "AI co-scientist" model impacts the velocity and success rates of the drug discovery pipeline. If successful, the era of the trillion-scale search may soon become the standard, fundamentally shortening the time it takes to deliver life-saving medicines to patients worldwide.
About the Author
