AlphaFold 2 was a breakthrough for predicting single protein structures. However, this success highlighted the much larger, unsolved challenges of modeling protein interactions, their dynamic movements, and the actual folding process, which are critical for understanding disease and drug discovery.
Modern protein models use a generative approach (diffusion) instead of regression. Instead of predicting one "correct" structure, they model a distribution of possibilities. This better handles molecular dynamism and avoids averaging between multiple valid states, which is a flaw of regression models.
Lacking massive compute resources, the Boltz team could only afford one training run for their model. They discovered and fixed bugs mid-training by stopping the run, patching the code, and resuming. This created a powerful but technically irreproducible model born from necessity.
Contrary to trends in other AI fields, structural biology problems are not yet dominated by simple, scaled-up transformers. Specialized architectures that bake in physical priors, like equivariance, still yield vastly superior performance, as the domain's complexity requires strong inductive biases.
Unlike LLMs, parameter count is a misleading metric for AI models in structural biology. These models have fewer than a billion parameters but are more computationally expensive to run due to cubic operations that model pairwise interactions, making inference cost the key bottleneck.
To get scientists to adopt AI tools, simply open-sourcing a model is not enough. A real product must provide a full-stack solution, including managed infrastructure to run expensive models, optimized workflows, and a UI. This abstracts away the complexity of MLOps, allowing scientists to focus on research.
To convince skeptical medicinal chemists of AI's value, you must deliver a result that surpasses their intuition. It's not about the user interface, but about the model generating a genuinely surprising and effective molecule. This "aha" moment, validated by lab results, is the ultimate way to build trust.
Models like AlphaFold don't solve protein folding from physics alone. They heavily rely on co-evolutionary data, where correlated mutations across species provide strong hints about which amino acids are physically close. This dramatically constrains the search space for the final structure.
A key strategy for improving results from generative protein models is "inference-time scaling." This involves generating a vast number of potential structures and then using a separate, fine-tuned scoring model to rank them. This search-and-rank process uncovers high-quality solutions the model might otherwise miss.
By open-sourcing its model, Boltz created a feedback loop where the community discovered novel use-cases, like a crude but effective "inference-time search" for antibody prediction. This demonstrates how open access allows external users to find creative applications the original developers hadn't considered.
To avoid overfitting and prove true generalization, Bolts validates its protein design models by testing them across a wide array of targets from over 25 external academic and industry labs. This diverse, real-world testing is the ultimate benchmark of a model's utility in drug discovery.
Bolts Gen's protein design model simplifies its task by predicting only the final 3D atomic structure. Because different amino acids have unique atomic compositions, the model's placement of atoms implicitly determines the protein's sequence, elegantly merging two traditionally separate prediction tasks.