🔬Generating Molecules, Not Just Models

Описание: This episode traces the remarkable journey from AlphaFold2’s landmark achievement in protein structure prediction to the broader landscape of molecular interaction modeling and protein design. The problem AlphaFold2 addressed—predicting the structure of single-chain proteins—was long considered intractable due to its perceived NP-hard nature. The breakthrough came not only from advances in machine learning but also from leveraging evolutionary data to infer co-evolution of amino acids, providing powerful hints about spatial proximity in protein structures. Yet, as the guests explain, the field quickly moved beyond this milestone toward more complex questions, like how proteins interact, how they fold dynamically, and how to model these interactions with small molecules, RNA, and DNA.

AlphaFold3 marks a critical shift in this evolution, moving from static structure prediction to modeling heterogeneous molecular interactions. Rather than treating these interactions as isolated problems, AlphaFold3 unifies them within a single model trained across modalities. This progress also reflects a broader trend in machine learning: the shift from regression-style prediction to generative models capable of expressing uncertainty and capturing system dynamics. By sampling from a distribution of plausible structures and interactions, these models allow researchers to better understand the flexibility and variability of biological systems. However, such models also introduce new challenges, particularly around validation and ranking of generated outputs.

Enter Boltz and its suite of tools, which aim to democratize access to these cutting-edge capabilities. Boltz builds on open-source principles and a strong community foundation to deliver models that are both state-of-the-art and accessible, with a focus on usability, extensibility, and real-world validation. Boltz2 and BoltzGen combine structure prediction, affinity estimation, and generative design in one pipeline, enabling users to design new proteins and small molecules with high confidence. Notably, Boltz emphasizes the importance of experimental validation, collaborating with partners across academia and industry to test new designs in the lab. This feedback loop is essential to the iterative improvement of models and benchmarks.

BoltzLab, the newly launched platform, encapsulates this vision by providing a cloud-based interface for running large-scale protein and molecule design campaigns. With support for both computational and experimental scientists, BoltzLab offers APIs, collaboration tools, and automated agentic workflows to make advanced molecular modeling accessible to users with varying levels of computational expertise. It embodies the shift from abstract model development to practical deployment, where infrastructure, cost-efficient compute, and user-friendly interfaces make a meaningful difference. As the guests emphasize, the real progress lies in enabling scientists to use these tools creatively and collaboratively to accelerate discovery in biology and medicine.

Timestamps

00:00 Introduction to Benchmarking and the “Solved” Protein Problem
06:48 Evolutionary Hints and Co-evolution in Structure Prediction
10:00 The Importance of Protein Function and Disease States
15:31 Transitioning from AlphaFold 2 to AlphaFold 3 Capabilities
19:48 Generative Modeling vs. Regression in Structural Biology
25:00 The “Bitter Lesson” and Specialized AI Architectures
29:14 Development Anecdotes: Training Boltz-1 on a Budget
32:00 Validation Strategies and the Protein Data Bank (PDB)
37:26 The Mission of Boltz: Democratizing Access and Open Source
41:43 Building a Self-Sustaining Research Community
44:40 Boltz-2 Advancements: Affinity Prediction and Design
51:03 BoltzGen: Merging Structure and Sequence Prediction
55:18 Large-Scale Wet Lab Validation Results
01:02:44 Boltz Lab Product Launch: Agents and Infrastructure
01:13:06 Future Directions: Developpability and the “Virtual Cell”
01:17:35 Interacting with Skeptical Medicinal Chemists