Podcast Summary: Latent Space – Beyond AlphaFold: How Boltz is Open-Sourcing the Future of Drug Discovery

Date: February 12, 2026
Guests: Gabriela Corso & Jeremy Volvind (Co-founders, Boltz)
Hosts: Latent Space Team

Episode Overview

This episode dives deep into the evolution of protein structure prediction, tracing the impact of AlphaFold 2 and 3, and exploring how Boltz is open-sourcing advanced models at the cutting edge of drug discovery. Boltz's mission is to democratize biology’s next wave, making foundational models and tooling accessible for researchers, industry, and academics alike. The conversation covers the state of the field, technical innovations, community building, validation, and the launch of BoltzLab, their product platform.

Key Discussion Points & Insights

1. Retrospective: The AlphaFold Revolution (03:33 – 13:07)

• Scientific Milestone: AlphaFold 2’s breakthrough made structure prediction of single-chain proteins dramatically more accurate, shifting computational biology’s landscape.

  • “AlphaFold was a real breakthrough in this problem of protein folding, which is trying to understand the structure of a single protein.” — Jeremy Volvind (01:33)
  • Motivated many ML researchers (including today’s guests) to pivot toward structural biology.

• CASP Competitions: Provided key testbeds for benchmarking progress and challenging models on unseen biological structures.

  • “The goal remains to really challenge the models, like how well do these models generalize?” — Gabriela Corso (05:05)

• Limits of ‘Solving’ Protein Folding: Progress has been made in static structure prediction, but protein folding pathways and dynamic conformations remain complex, open problems.

  • “We’ll steer away from the term ‘solved’... but the problem that a lot of progress was made on was the ability to predict the structure of single chain proteins.” — Gabriela Corso (06:49)
  • Static models don’t capture the physical process or the multiple states a protein may occupy.

2. Why Structure Matters (09:55 – 11:30)

• Protein structures underpin all cellular functions, disease mechanisms, and therapeutics; understanding misfolding is crucial in diseases.
  • “Proteins are kind of the machines of our body.” — Jeremy Volvind (09:55)

3. Technical Deep Dive: How AlphaFold Works (13:40 – 21:42)

• Evolutionary Data and Coevolution: Clarified the role of multiple sequence alignments (MSAs) in revealing interaction hints between amino acids.

  • “If I have some amino acid that mutates, it’s going to impact everything around it... the protein, through random mutations in evolution, ends up figuring out that this other amino acid needs to change as well...” — Gabriela Corso (13:40)

• Architecture Innovation: Dual-phase modeling—coarse global relationship estimation and fine-grained optimization—with an evolutionary hint guiding the search.

  • “The interesting thing about AlphaFold is it’s got this very peculiar architecture that operates on this pairwise context between amino acids...” — Jeremy Volvind (15:31)

4. Progression from AlphaFold 2 → 3 (16:53 – 23:41)

• AlphaFold 3’s Leap: Unified architecture for proteins, small molecules, nucleic acids—modeling biological interactions beyond individual chains.

  • “AlphaFold 3 was a significant advancement on the problem of modeling interactions... proteins with small molecules... with RNA and DNA...” — Jeremy Volvind (17:16)

• Technical Advancements: Shift from regression to generative modeling; improved handling of structural uncertainty; architectural simplification towards (yet distinct from) transformers.

  • “Moving from modeling structure prediction as a regression problem... to a generative modeling problem...” — Jeremy Volvind (19:49)
  • Model size remains modest in parameter count but requires heavy computational effort due to cubic scaling.

• “Bitter Lesson” and Specialization: Despite interest in transformer-based architectures, specialized equivariant models still outperform generic large models in this domain.

5. Open Source Crisis and Boltz’s Genesis (26:17 – 28:36)

• AlphaFold 3 Not Open-Sourced: With DeepMind’s shift toward commercialization (Isomorphic Lab), access to state-of-the-art models was restricted.
• Boltz’s Response: Boltz 1 developed as a fully open-source, high-accuracy alternative; rapid build (under heavy compute constraints).
  • “We only trained the big model once. That’s how much compute we had.” — Gabriela Corso (29:58)
  • Community and academic effort, later supported by industry partners.

6. Benchmarking and Evaluation (32:03 – 36:15)

• Public Datasets and the PDB: Boltz and AlphaFold compared using time-split datasets; focus on generalization to unseen structures.
• Importance of Open Benchmarks: Iterative improvement depends on community feedback from open-use and transparent benchmarking.
  • “It’s always critical... to set clear benchmarks. And as you start doing progress of certain benchmarks, then you need to improve the benchmarks and make them harder and harder.” — Jeremy Volvind (33:50)

7. Boltz’s Community and Open Science Ethos (41:27 – 44:46)

• Community Growth: Slack and GitHub community grew rapidly; design focused on usability and accessibility.

  • “We have a Slack community that has thousands of people on it and it’s actually self-sustaining now...” — Gabriela Corso (41:45)

• Surprising Community Contributions: GPU kernels, creative hacks for new use cases, novel inference search strategies.

  • “One individual... wrote a complex GPU kernel for part of the architecture...” — Gabriela Corso (45:01)
  • “He basically gave random hints to the model... and then looked at the confidence of the model in each of those cases and took the top...” — Jeremy Volvind (45:40)

8. Boltz Suite and the Rise of Generative Biology (48:34 – 54:40)

• Boltz Model Stack:

  • Boltz 1: Open-source AlphaFold 3-level prediction.
  • Boltz 2: Added affinity prediction—how strongly proteins/small molecules interact.
  • Boltzgen: Protein design foundation model—simultaneously predicts structure & sequence (scalable, atomic-level encoding).

• Pipeline Workflow: Diffusion models generate sequence/structure jointly; ranked and filtered using learned scoring—consistency and affinity scoring are key.

• Protein Design Validation: Wide-scale cross-institutional collaborations for real-world (wet lab) validation on diverse tasks: peptides, nanobodies, small-molecule binders.

  • “He basically put together, I think it was something like 25 different academic and industry labs that committed to testing some of the designs from the model...” — Jeremy Volvind (55:18)
  • Highlight: Achieved nanomolar binders on 2/3 of targets in blind generalization tests.

9. Launch of BoltzLab: Platform and Product (62:21 – 67:48)

• BOLT Agents: Modular “recipes” that coordinate complex multi-step modeling pipelines for protein and small molecule design. Not LLM wrappers, but specialized, orchestrated ML workflows.
• Infrastructure: Purpose-built GPU fleet enables cost-effective, massively parallel screens (amortizing compute for users).
• User Experience: API and web interface for both computational experts (companies) and broad audiences (biologists, chemists); collaborative tools for analysis and selection.
• Access: Academics receive free credit; startups collaborations encouraged; custom enterprise deployment possible.

10. Validation and the Importance of Real-World Testing (68:52 – 73:07)

• Lab Testing Is King: No amount of computational metrics can replace empirical test—hit rates and binding strength in wet lab experiments drive true progress.

  • “As beautiful as the platform can be, as nice as the molecules might look that the model predicted, I think what really convinces people is hits.” — Gabriela Corso (79:40)

• Agentic System Validation: Diverse, representative targets; partnerships with CROs for scale and repeatability.

• Division of Responsibility: Boltz builds the tools; actual therapeutic development remains with research and pharma partners.

11. Future Directions & Skepticism (73:07 – End)

• Boltz’s “Layer”: Intends to remain a platform/tool provider, not a therapeutics company. Next breakthroughs: expand to more developability traits, include cellular-level modeling iteratively, integrate feedback from real-world outcomes.
• Convincing Skeptical Medicinal Chemists: Trust is earned by showing surprising, valuable results, not just beautiful interfaces.
  • “For people to be convinced, you have to show them something that they didn’t think was possible. And until you have that ‘aha’ moment, I think the skepticism will remain.” — Gabriela Corso (79:09)

Noteworthy Quotes & Moments

• “AlphaFold was a real breakthrough... seeing the structures coming out of these models, where you see this beautiful creation of life, is something that was very inspiring to me.” — Jeremy Volvind (01:33)
• “It was really clear... there was so much that we were not really doing... there was a lot to catch up on. But... we had some experience working with the data and... with these type of models. And that put us already in a good place to produce [Boltz 1] quickly.” — Gabriela Corso (28:36)
• “Actually, we only trained the big model once. That’s how much compute we had... we were finding bugs left and right... we just kept training it with the bug fixes along the way...” — Gabriela Corso (29:58)
• “Putting a model on GitHub is definitely not enough to get chemists and biologists... to use your model in their therapeutic programs. So a lot of what we think about boldly beyond just the models is... all the layers that come on top of the models to get... something that can really enable scientists.” — Jeremy Volvind (37:25)
• “You have to show them something that they didn’t think was possible... then, it’s like ‘oh, wow, okay, I can do something with this.’” — Gabriela Corso (79:09)

Timestamps by Topic

• 00:00 – 04:48 — AlphaFold’s Impact and Introduction to Structural Biology
• 04:48 – 06:27 — CASP Competitions and AlphaFold’s Breakthrough
• 09:10 – 12:10 — Why Protein Structures and Foldings Matter
• 13:07 – 16:44 — AlphaFold's Algorithms and Model Insights
• 16:53 – 21:42 — AlphaFold 3 and Advances in Modeling Interactions
• 21:42 – 26:17 — Model Architecture, Compute, and Open Source Limitations
• 26:17 – 32:03 — Boltz’s Creation in Response to AlphaFold3 Closed Model
• 32:03 – 36:15 — Benchmarking, Community Validation, and Field Progress
• 41:27 – 44:46 — Community, Contributions, and Open Science Culture
• 48:34 – 54:40 — Boltz’s Product Stack & Protein Design Foundation Models
• 55:18 – 62:21 — Lab Validation and Experimental Collaboration
• 62:21 – 67:48 — BoltzLab: Infrastructure, Access, and Products
• 68:52 – 73:07 — Validation in Practice & Staying Platform-Focused
• 76:25 – 79:54 — Skepticism Among Experts and Winning Trust
• 80:05 – End — Boltz’s Future and Call for Team Expansion

Conclusion

Boltz exemplifies the shift towards open science, community-driven innovation, and productization in generative biology. By openly building advanced models and platforms—backed by deep technical rigor, broad validation, and a collaborative community—they’re powering the next wave of discovery in drug design. Their commitment to democratization, product excellence, and iterative experimental feedback sets a new path for scalable, accessible science in the AI-for-biology era.