Pfizer Is Betting That Drug Discovery No Longer Needs Evolutionary History

For years, some of the biggest advances in computational biology came from a simple idea: if researchers could compare a protein to enough related proteins, they could learn how it was likely to behave.

That principle helped power modern protein structure prediction systems, including AlphaFold, which uses evolutionary relationships captured through Multiple Sequence Alignments (MSAs) as part of its approach to predicting protein structures.

The challenge appears when researchers encounter proteins with few known relatives. Rare disease targets, highly mutated proteins, and previously uncharacterized biological pathways often lack the evolutionary information that has traditionally helped computational systems make accurate predictions.

That problem sits at the center of Pfizer's new agreement with Chai Discovery. Under the deal, Pfizer will gain access to Chai-3, the company's latest biological foundation model, and integrate it into its drug discovery infrastructure. Chai will also build a custom version of the model using Pfizer's proprietary data and workflows.

The agreement reflects a growing effort across the pharmaceutical industry to develop AI systems for biological targets that offer limited historical data.

Drug Discovery's Dependence on Historical Data

Modern computational biology has benefited enormously from evolutionary history.

MSAs work by comparing proteins across organisms and identifying patterns that have been preserved over time. Those patterns help researchers determine which parts of a protein are likely to be important and how the protein may fold.

The approach has delivered major advances because evolution effectively generated a vast biological dataset that AI systems could learn from.

Researchers have also identified limitations. Proteins with few known relatives can be harder to analyze because there is less comparative information available. Studies examining low-homology and orphan proteins have highlighted the challenge that sparse evolutionary data can create for computational prediction systems.

That limitation has become increasingly important as drugmakers pursue more difficult disease targets.

Many of the proteins involved in cancer, rare diseases, and emerging pathogens do not fit neatly into well-understood biological families. The less historical information available, the harder it becomes to use conventional computational approaches.

Researchers are now developing new methods designed to perform when evolutionary information is limited. Recent work has focused specifically on reducing dependence on MSA-based approaches for low-homology proteins and other difficult targets.

The industry is now testing whether AI can design effective molecules for biological targets that have limited historical data.

From Searching for Molecules to Generating Them

That question is at the heart of Chai Discovery's technology.

Earlier this year, the company published results for Chai-2, a multimodal generative model for antibody design. In a preprint, the researchers described the system as capable of "fully de novo antibody design," meaning it generates entirely new candidate molecules instead of searching existing biological libraries.

The distinction is important.

Traditional drug discovery often begins with large collections of molecules that researchers screen and optimize over time. Generative biological models create candidate molecules directly from information about a target.

Chai reported a 16% hit rate for Chai-2 and described improvements that exceeded prior computational approaches by more than 100-fold. The company says Chai-3 doubles the success rate achieved by its previous model in antibody design tasks.

Those results remain subject to the realities of pharmaceutical development. Drug discovery is still a long process that requires laboratory validation, clinical testing, and regulatory review.

The agreement is notable because of how Pfizer plans to deploy the technology.

Pfizer is placing a frontier biological model directly inside its discovery engine and pairing it with decades of internal experimental results and discovery workflows. Chai says the objective is to help researchers generate biomolecules from scratch and reduce discovery cycles from months or years into much shorter development periods.

Why Pharmaceutical Companies Are Building Biological AI Infrastructure

The Pfizer deal reflects a broader shift taking place across AI-driven drug discovery.

Companies including Isomorphic Labs, Recursion Pharmaceuticals, Generate Biomedicines, Absci, and EvolutionaryScale are developing foundation-model approaches for biology. While their technical methods differ, they share a common objective: using AI to design and understand biological systems at a scale that was previously impossible.

The structure of the Pfizer agreement also highlights how pharmaceutical companies are changing their AI strategies.

Earlier partnerships often centered on external prediction tools. Technology companies generated insights and pharmaceutical companies evaluated the outputs.

The Chai agreement is closer to infrastructure deployment.

Pfizer is integrating the model into its own research environment and training a custom version around proprietary data assets accumulated across decades of drug development.

Proprietary experimental data, validation systems, and research workflows are becoming increasingly important assets as biological foundation models mature. Pfizer's agreement with Chai places those assets alongside a new generation of AI systems designed for difficult biological targets, offering a view into how large pharmaceutical companies are approaching the next phase of computational drug discovery.