Schrödinger says its Bunsen agent now handles 80% of the company’s computational drug discovery tasks

Schrödinger says its Bunsen agent now handles 80% of the company’s computational drug discovery tasks

In an October 10 interview with BiopharmaTrend, Schrödinger’s chief scientific officer Robert Abel said that Bunsen now handles about 80% of the computational jobs on the company’s cluster, performing calculations previously handled by computational scientists. Schrödinger plans a full commercial launch by the end of 2026.

For 35 years, Schrödinger has sold pharmaceutical companies software that uses physics to model molecules at the atomic level and calculate how tightly a drug will bind to its target protein. In 2016, Nimbus Therapeutics, co-founded by Schrödinger, sold Gilead Sciences a molecule designed this way in a deal worth up to $1,2 billion. Those calculations had produced an actual drug candidate.

In late July, Schrödinger released Bunsen in early access. The pharmaceutical company Bristol Myers Squibb was among its first customers. In August, BMS agreed to develop Bunsen for molecular calculations and integrate RetroSynth, a program that plans the synthesis of selected molecules. A scientist describes a task in everyday language, and Bunsen prepares a computational plan for approval. Once approved, it runs the entire campaign without continuous human supervision. Abel puts a number on its use:

“We recently found that about 80% of the computational jobs on our cluster are now being handled by Bunsen... Bunsen is doing serious work.”

Within Bunsen, a language model selects the appropriate validated method and the sequence of operations. The method itself produces the numerical result, such as a binding energy. “The numerical results come from these thoroughly validated computational methods, rather than from the language model itself,” Abel says. This is how Bunsen avoids the main risk posed by AI agents in science: a number that looks plausible but is physically wrong.

Two years ago, in November 2024, Schrödinger CEO Ramy Farid called AI in drug discovery “nonsense,” arguing that physics was more accurate than a neural network. At Bunsen’s launch in July, he acknowledged:

“I was extremely skeptical, to be honest, and I could not be happier to have been wrong. This changes everything.”

The company began developing agentic systems in late 2023, immediately after GPT-4 was released. For a year and a half, they made mistakes on simple tasks. Abel draws a comparison with MYCIN, a Stanford expert system developed in the 1970s to diagnose infections. In a blinded evaluation in 1979, physicians judged its antibiotic recommendations acceptable in 65% of cases, higher than the scores for staff infectious disease physicians (42,5–62,5%). It never entered clinical practice. Questions of legal liability for a computer-generated diagnosis and cumbersome data entry prevented adoption of the more accurate system. Bunsen has checks at every step, and the simulation itself indicates when a calculation fails to converge on a physically consistent result. “MYCIN had nothing like that beyond waiting for the patient’s outcome,” Abel says.

Asked whether Bunsen improves decisions as well as making them faster, Abel gives a direct answer: comparing drug discovery programs with and without the agent would be too expensive an experiment. He instead points to two measures: the accuracy of the method and its coverage, meaning how many compounds it brings into consideration as priorities that would otherwise have gone unnoticed.

In July, Farid described Bunsen’s licensing model as charging according to the volume of campaigns run. In October, Abel clarified that formal terms had yet to be announced. He said the main constraint for pharmaceutical companies was computational scientists’ time. They typically work on three or four projects at once and cannot complete all the modeling needed. Bunsen frees up that time. To accommodate the increased workload expected at full launch, Schrödinger is deploying Bunsen on Nvidia and Google Cloud infrastructure.

This infrastructure accelerates drug discovery across therapeutic areas. The agent handles routine computational work, while the scientist defines the question and performs the final review.