Boltz-2 Open-Source Model Wins Pfizer as IsoDDE Stays Closed

Two paths for AI drug design

AI-driven drug discovery is splitting into two camps. One is building powerful proprietary engines and showing only results. The other is releasing models that anyone can download, inspect, and retrain. A recent pairing of announcements shows the split clearly. Pfizer has signed on with an open-source lab, while Google DeepMind's drug-discovery spin-off is promoting a closed system it says beats the open alternative.

In January 2026, Pfizer announced a collaboration with Boltz PBC, an AI research lab, to apply its open-source models, Boltz-2 and BoltzGen, to designing both small molecules and biologics 3. Around the same time, Isomorphic Labs unveiled IsoDDE, a tool for predicting how proteins interact with drug candidates 4. Isomorphic says IsoDDE outperforms Boltz-2 and physics-based methods at estimating binding affinity 4. Scientists quoted in coverage were impressed enough to compare it to an "AlphaFold 4" 4.

What Boltz-2 actually does

Boltz-2 was developed by researchers at MIT working with Recursion 12. Its main advance is predicting two things at once: the three-dimensional structure of a molecular complex and how strongly its parts bind 12. Binding affinity is important because it is a direct proxy for whether a molecule will have a biological effect 1. Measuring it in the lab remains a major time and cost bottleneck in early-stage discovery 2.

Until now, the best computational option has been free-energy perturbation (FEP). FEP uses long, atom-level physics simulations that are accurate but too slow and expensive to use widely in small-molecule work 2. Recursion presents Boltz-2 as the first deep learning model to approach FEP-level accuracy on affinity 2. One industry analysis says it runs roughly 1,000 times faster than FEP 3. The same analysis reports a Pearson correlation of about 0.62 on affinity benchmarks, comparable to open-source FEP pipelines 3. According to Recursion, the model can support virtual screening for hit discovery and hit-to-lead work, and it improves on the prior state of the art on unseen targets without fine-tuning 2.

These figures come mainly from the model's developers and from commentators who favor it. They should be read as promising benchmark results, not proof of real-world drug discovery performance.

Openness as the selling point

All the sources agree on one point: Boltz-2's defining trait is openness. The model code, trained weights, and full training pipeline are publicly available 1. That openness is concrete enough that independent developers have published walkthroughs of building simplified versions of the architecture from scratch 1. Analysts describe Boltz as part of a broader move toward open-source biomolecular foundation models, which are larger and more general than older tools such as QSAR regressors and molecular docking 3.

IsoDDE sits at the other end. Researchers have praised its capabilities, but they also point out that it is proprietary. The technical paper released alongside it gives little guidance on how others could reproduce its results 4.

Why Pfizer's choice matters

IsoDDE's reported edge over Boltz-2 makes Pfizer's move easy to misread. It looks odd for a major drugmaker to partner with an open project while a closed rival claims better accuracy. But accuracy on a benchmark is only one factor for a large pharmaceutical company.

An open model gives Pfizer things a closed one cannot:

• Fine-tuning: it can train the model on its own internal assay data.
• Private deployment: it can run the model inside its own infrastructure.
• Auditing: it can inspect how the model reaches its predictions.
• No lock-in: it does not depend on a competitor's roadmap.

Isomorphic Labs is also a drug developer, not only a tool vendor. That likely makes some pharmaceutical companies cautious about relying on it for core research.

The open approach also benefits from the wider research community. Every academic group that tests, extends, or criticizes Boltz-2 in public adds evidence about where it works and where it fails. A proprietary model with a thin technical paper cannot build that kind of shared evidence base.

The likely outcome

The most plausible reading is that this is not a contest one side will win outright. IsoDDE's results suggest closed systems with deep resources may stay ahead on raw performance, at least for now 4. Pfizer's deal suggests that for many industrial users, the ability to own, adapt, and verify a model matters as much as a small accuracy gap 3.

If open models keep closing in on physics-based accuracy at a fraction of the cost, as Boltz-2's benchmarks indicate 23, the field may follow a pattern seen elsewhere in AI. Closed labs set the frontier, and open projects become the default infrastructure that most real work runs on. Pfizer appears to be betting on the second.