Schrödinger and Bristol Myers Squibb Deepen Their AI Drug-Discovery Pact

Schrödinger and Bristol Myers Squibb are expanding a long-running collaboration to bring more artificial intelligence into how BMS discovers new drugs.

Announced August 6, 2026, the expanded deal gives BMS access to two newer tools: Bunsen, an agentic AI co-scientist that helps run molecular-discovery workflows, and RetroSynth, an AI system that plans and rapidly evaluates chemical-synthesis routes. These sit alongside Schrödinger’s established physics-based simulation. Financial terms were not disclosed.

What physics-based simulation does

Schrödinger’s core technology models how candidate molecules behave using the physics of atomic interaction rather than pattern-matching against known compounds.

The central calculation is binding free energy: given a target protein structure and a proposed molecule, simulate the atoms and estimate how tightly the two will associate. Doing it properly means accounting for the water molecules displaced when a drug enters a binding pocket, the flexibility of both molecules, and the entropy cost of locking a floppy molecule into one conformation.

The appeal is that it works on molecules nobody has made. A statistical model trained on known compounds extrapolates poorly beyond its training distribution; a physical calculation applies equally to anything.

The cost is computation. These simulations are expensive, which limits how many candidates can be evaluated.

Why synthesis planning is the real bottleneck

RetroSynth addresses a problem that has undermined computational design repeatedly.

Generative models are good at proposing molecules with desirable predicted properties, and a substantial fraction of what they propose cannot practically be made. There is no known route, or the route requires many steps with poor yields, or it needs reagents that are hazardous or unavailable.

Retrosynthetic analysis works backwards from a target molecule to available starting materials, decomposing it into simpler precursors step by step. Chemists have done this by hand for decades; automating it lets a designed molecule be screened for feasibility before anyone commits laboratory time.

Adding that filter to generative design is what makes the output usable rather than merely plausible — which is why it is packaged with the design tools rather than sold separately.

What an agentic co-scientist means

Bunsen is described as an agentic AI that helps run molecular-discovery workflows, and the term is worth unpacking.

A conventional computational tool answers one question — how tightly does this molecule bind, how might this be synthesised. Discovery involves chaining many such questions, deciding what to ask next based on the last answer, and managing the resulting data.

An agentic system operates across that chain: selecting which calculations to run, sequencing them, and interpreting results well enough to decide the next step. The value claimed is not better individual predictions but less human time spent orchestrating them.

BMS’s Stephen Johnson said Bunsen “allows our scientists to think differently about how physics-based tools can be used to navigate molecular design space.”

Why the combination is the bet

Pairing physics-based modelling with generative AI is a widely watched approach, and the two are genuinely complementary rather than redundant.

Generative models produce candidates quickly and cheaply but predict properties unreliably outside familiar chemistry. Physics-based simulation predicts more reliably but too slowly to survey a large space.

Using generation to propose and simulation to verify plays to both. It is also why a company with deep physics capability is well positioned as AI methods spread: the generative half is increasingly commoditised, while accurate simulation remains hard.

Why early-stage discovery attracts this investment

“BMS is a long-standing customer and collaborator,” said Schrödinger chief scientific officer Robert Abel.

The deal is one more sign that big pharma is embedding AI deeper into the earliest, most uncertain stage of development — and that stage is targeted for a specific reason.

Late-stage failure is expensive but caused by biology: the target turns out not to drive the disease, or the drug has effects in humans that no model predicted. Early-stage work is a chemistry and optimisation problem, which is far more tractable computationally.

Schrödinger’s two-sided business

The company occupies an unusual commercial position that shapes deals like this one.

It sells software and services to pharmaceutical companies, which produces predictable revenue and gives it visibility across many discovery programmes. It also runs its own internal drug pipeline, using the same tools on targets it selected.

The tension is evident: a company whose customers are large pharmaceutical firms is also, in its pipeline business, a potential competitor to them. Partnerships accordingly tend to be structured as tool access rather than as joint programmes, keeping the customer relationship separate from the pipeline.

The advantage is credibility. A software vendor whose tools have never produced a clinical candidate faces reasonable scepticism about whether the predictions hold up; running an internal pipeline is how that gets tested in public. It also means the tools are refined against real programme requirements rather than benchmark datasets, which is a meaningfully different feedback loop.

The honest framing is that these tools make the tractable part faster and cheaper. They do not address why most drugs fail. Business news, not investment advice.