I first met Gabriele Corso, then a PhD student at MIT, in November of 2023. The year before, he’d published DiffDockDiffDock: Diffusion Steps, Twists, and Turns for Molecular DockingGabriele Corso, Hannes Stärk, Bowen Jing, Regina Barzilay, Tommi JaakkolaPredicting the binding structure of a small molecule ligand to a protein -- a task known as molecular docking -- is critical to drug design. Recent deep learning methods that treat docking as a regression problem have decreased runtime compared to traditional search-based methods but have yet to offer substantial improvements in accuracy. We instead frame molecular docking as a generative modeling problem and develop DiffDock, a diffusion generative model over the non-Euclidean manifold of ligand poses. To do so, we map this manifold to the product space of the degrees of freedom (translational, rotational, and torsional) involved in docking and develop an efficient diffusion process on this space. Empirically, DiffDock obtains a 38% top-1 success rate (RMSD<2A) on PDBBind, significantly outperforming the previous state-of-the-art of traditional docking (23%) and deep learning (20%) methods. Moreover, while previous methods are not able to dock on computationally folded structures (maximum accuracy 10.4%), DiffDock maintains significantly higher precision (21.7%). Finally, DiffDock has fast inference times and provides confidence estimates with high selective accuracy.arXiv:2210.01776v2View paper, which was a big deal. It was the first application of modern generative diffusion models to the problem of molecular docking: predicting how a small molecule will bind to a protein. Suddenly, pharma companies and biotechs around the world wanted to know where Gabriele, a 27-year-old who’d made his first foray into the life sciences during graduate school, would take his research next. DiffDock became a core part of NVIDIA’s digital biology offering. All of this commercial interest made Gabriele wonder if he should pursue a startup. But he decided to stay to finish his thesis.

This decision proved to be impactful. In 2024, a schism formed between industry and academia. Molecular machine learning, Gabriele’s field of research, was white-hot after the AlphaFold 1 and AlphaFold 2 breakthroughs delivered by Google DeepMind. Both models were made open-source at the time of release for academic use. But AlphaFold 3 wasn’t. Suddenly, brilliant academic researchers, like Gabriele and his labmates, had to test their ideas by carefully choosing how to use the ten free predictions they were allotted each day. He experienced firsthand what it was like to not have access to the tools he needed to do his best work.

This lit a fuse. Together with Jeremy Wohlwend and Saro Passaro, he launched the Boltz project. The vision was to make an open-source structure prediction model rivaling the performance of the leading closed-source models. Doing this would put frontier tools back in the hands of researchers around the world. They succeeded massively. With a shoe-string academic budget that allowed for only one large-scale training run, they delivered Boltz-1, which rivaled AlphaFold 3’s performance but had no restrictions for use.

Boltz was named after the Boltzmann distribution, which is the fundamental physical function that AI structure prediction models are approximating. But the name took on a double meaning, because the project caught lightning in a bottle. Upon release, Boltz-1 was adopted by over 100,000 researchers, thousands of companies, and every top 20 pharma company. The release of Boltz-2 and BoltzGen scaled this even further.

The Boltz team then formed a new hypothesis. They reasoned that it was possible to build a thriving business that was exclusively focused on democratizing access to frontier molecular machine learning models. The future didn’t need to have a stark dividing line between large drug discovery companies with proprietary models within their walled gardens and small labs and companies who lacked access. If these AI models are the new experimental tools of the 21st century, they could be the next Illumina, giving everybody access to the best picks and shovels. Everybody could win together.

They set out to raise Seed capital to run an experiment. The goal was to demonstrate that they could translate their widespread open-source adoption into early signs of a viable and differentiated business. We at Amplify developed deep conviction in this team and thesis, and had the privilege of leading their Seed round alongside Zetta and a16z.

The results of this experiment have been absolutely extraordinary. Since launching Boltz Lab in January this year, thousands of scientists have signed up to use it. A wide range of companies, including six of the top twenty pharma, are on the platform. Generation of molecular tokens, the central unit of work, has increased fifty-fold since June. And they have achieved this rapid growth all while maintaining their commitment to pursuing frontier research and making their tools accessible to every scientist.

Now, they are advancing their mission to the next stage that can be encapsulated in one word: scale. Boltz is scaling their research, commercial, and product efforts to deliver the best possible AI models for molecular design in the hands of as many scientists, labs, and companies as possible. If they succeed, the gap between a scientist’s idea and a new molecule that can change the world will only be inference.

Bessemer is leading a $110M Series A to make this happen. We are excited to be doubling down in our support of this team in this new round. Gabriele, Jeremy, and Saro are exactly the type of visionary, mission-driven technical founders that Amplify Bio was built to serve. We’re proud to be able to continue supporting them in this next stage of their journey.