🔬Beyond AlphaFold: How Boltz is Open-Sourcing the Future of Drug Discovery - Latent Space Recap
Podcast: Latent Space
Published: 2026-02-12
Duration: 1 hr 21 min
Guests: Gabriele Corso, Jeremy Wohlwend
Summary
Boltz is revolutionizing protein structure prediction by open-sourcing tools to democratize access, moving beyond AlphaFold's capabilities. Their models focus on generative protein design and complex interactions, promising advancements in drug discovery.
What Happened
Gabriele Corso and Jeremy Wohlwend, co-founders of Boltz, are pioneering the democratization of protein structure prediction with their open-source models, Boltz-1 and Boltz-2. These models were developed as alternatives to AlphaFold 3, which was not open-sourced, thereby limiting accessibility to advanced protein modeling technology. Boltz-1 was trained using shared resources from the Department of Energy, highlighting the challenges of scaling such technologies without proprietary infrastructure.
The podcast discusses the evolution of structural biology from AlphaFold's initial breakthroughs in single-chain protein prediction to more complex modeling tasks. AlphaFold 2 achieved significant progress, but AlphaFold 3 introduced a generative modeling approach that Boltz seeks to replicate and expand upon through open-source means. The focus is now on modeling protein-ligand and protein-protein interactions, as well as generative protein design, fields that hold great promise for drug discovery.
Boltz has developed a community-driven approach, leveraging thousands of contributors in their Slack community to refine and improve their models. This collaborative platform has enabled Boltz to rapidly iterate and validate their models across multiple targets, achieving high accuracy in predicting nanomolar binders for therapeutic purposes.
The Boltz Lab platform is a key product offering, designed to support scientists in protein and small molecule design. It provides a range of tools and agents optimized for large-scale computations, allowing for faster and more efficient screenings compared to traditional open-source methods. The platform also supports human-in-the-loop processes to aid medicinal chemists in the therapeutic development pipeline.
Real-world validation of Boltz's models has shown promising results, with successful designs on targets with zero known interactions in the training data. This demonstrates the models' ability to generalize and innovate beyond existing data sets, a critical step for advancing drug discovery.
The episode concludes with a discussion on the future directions of Boltz, including developing the 'Virtual Cell' concept and engaging with skeptical medicinal chemists to showcase the practical applications of their models. Boltz continues to build partnerships with academic and industry labs, emphasizing the importance of open-source contributions and collaborations in pushing the boundaries of structural biology.
Key Insights
- Boltz-1 was trained with shared resources from the Department of Energy. It's like building a rocket ship using public libraries, showing that you don't need a Tesla-sized budget to revolutionize protein modeling.
- Boltz is turning thousands of Slack users into a protein-prediction powerhouse. By crowdsourcing improvements, they're rapidly achieving feats like predicting nanomolar binders, proving that sometimes the best lab is a virtual chatroom.
- Designing drugs for diseases with no known treatment targets - Boltz's models are doing just that, successfully predicting interactions with targets they weren't even trained on. It's like solving a puzzle without looking at the picture on the box.
- While some pharmaceutical giants keep their secrets close, Boltz is open-sourcing the blueprint for the future of drug discovery. Their 'Virtual Cell' concept might just make skeptics rethink the power of collaboration in science.