b12 Labs Built a GPS for Chemistry and Published the Proof in Nature

The Macro: Drug Discovery Chemistry Is Still Guesswork

I want to explain why the first stage of drug discovery is so wasteful, because it frames why b12 Labs matters. When a pharmaceutical company identifies a promising molecule, the next step is figuring out how to actually make it. That process, called synthetic chemistry, involves choosing reagents, solvents, temperatures, catalysts, and reaction times from an enormous space of possibilities. A typical medicinal chemistry campaign tests hundreds of reaction conditions before finding one that works reliably.

The numbers are staggering. The pharmaceutical industry spends over $80 billion annually on R&D. A meaningful chunk of that goes to early-stage chemistry optimization, the repetitive cycle of designing an experiment, running it, analyzing the results, adjusting parameters, and trying again. The average drug takes 10 to 15 years to develop. The chemistry phase alone can consume 2 to 4 years of that timeline. Every month of delay costs millions in lost patent life.

The industry knows this is inefficient. That is why lab automation has been a growth area for a decade. Companies like Chemspeed, Opentrons, and Unchained Labs make robotic platforms that can execute experiments faster than humans. The problem is that these robots are fast at running experiments but they do not know which experiments to run. A robotic platform can execute 96 reactions in parallel, but if those 96 reactions are poorly designed, you get 96 failures very quickly.

This is the gap. The robotics exist. The chemistry knowledge exists, scattered across millions of papers and patents. What is missing is the connective tissue between domain knowledge and automated execution. Benchling handles lab data management. Schr0dinger does computational chemistry. CDD Vault tracks compound libraries. None of them close the loop between “what experiment should I run?” and “run it on the robot.”

The Micro: Two Founders, a Nature Paper, and Full Conversion in One Shot

b12 Labs is an AI copilot for pharmaceutical chemistry that automates experiment planning and translates plans directly into robotic lab protocols. You describe what you want to synthesize in natural language, and the system designs the experiment, selects the conditions, and generates executable code for your robotic platform. It integrates with Chemspeed, Opentrons, Unchained, and other lab automation systems.

Andres Bran is a founder. Zlatko Joncev is a founder. The team is two people, based in San Francisco, part of Y Combinator’s Summer 2025 batch. Two people is tiny for a company tackling pharmaceutical R&D, but the technical credibility is unusually strong for a seed-stage startup. Their work has been published in Nature Machine Intelligence, which is one of the most selective journals in the field. That is not a preprint on arXiv. That is peer-reviewed science in Nature’s flagship AI journal.

The headline result from their pharma pilots is that the AI achieved full conversion in a single attempt, eliminating the usual 8 to 12 iteration cycle. If you are not a chemist, let me translate. Full conversion means the reaction worked perfectly. Every molecule of starting material turned into the desired product. Doing that on the first try, when human chemists typically need 8 to 12 attempts to optimize conditions, is a significant result.

The product connects multiple AI agents in a unified platform. One agent handles experiment design. Another translates the design into robotic protocols. Another integrates with chemistry literature and expert tools for context. The result is a system that goes from natural language description to robotic execution without requiring the chemist to write code or manually program the robot.

What I find compelling about b12 Labs is that it is solving both sides of the lab automation problem simultaneously. The robots are there. The knowledge is there. b12 connects them. A chemist no longer needs to be an expert in both synthetic chemistry and robotic programming. They describe the goal and the system handles the translation.

The company name is a reference to vitamin B12, which has one of the most complex molecular structures in biochemistry. Its total synthesis by Robert Burns Woodward in the 1970s is considered one of the greatest achievements in organic chemistry. Naming the company after that molecule signals ambition about the complexity of chemistry problems they intend to solve.

The Verdict

b12 Labs has something that most early-stage AI companies lack: published, peer-reviewed evidence that the technology works. Nature Machine Intelligence is not a pay-to-play journal. Getting published there means the methodology and results survived rigorous scrutiny from domain experts. For a two-person startup, that is an extraordinary credential.

The risk is the sales cycle. Pharmaceutical companies are among the slowest technology adopters in any industry. Validation studies take months. Procurement takes months. Integration with existing lab infrastructure takes months. A two-person team selling to pharma needs either a very patient investor or a very fast path to initial revenue.

The other risk is competition from well-funded players. Recursion Pharmaceuticals is spending heavily on AI-driven chemistry. Insilico Medicine has raised hundreds of millions. Isomorphic Labs, the Alphabet spinout, has deep pockets. But these companies are building vertically integrated drug discovery platforms. b12 Labs is building a tool that works with existing lab infrastructure, which is a fundamentally different go-to-market. A pharma company that has already invested millions in Chemspeed robots does not want to replace them. They want to make them smarter. That is what b12 offers.

In thirty days, I want to know how many pharma companies are running pilots with b12. Sixty days, the question is repeatability. The single-attempt full conversion result is impressive, but is it consistent across different reaction types and molecular targets? Ninety days, I want to see whether the platform is expanding beyond synthetic chemistry into other areas of lab automation, like formulation or analytical method development. The science is real. The problem is real. The market is enormous. Two founders against the entire pharmaceutical industry is a dramatic mismatch in resources, but the Nature paper gives them credibility that money cannot buy. If they can translate that credibility into enterprise pilots before a larger competitor copies the approach, they have something special.