Protein binders, such as antibodies and nanobodies, are powerful tools used in biochemistry, structural biology, imaging, therapeutics and other areas of biological sciences. Yet, for many proteins of interest, no binders exist. Designing them efficiently de novo has long been a dream of the protein design community, but for years this was limited to a small number of labs with the right combination of design expertise and screening capacity. Not to mention the sheer complexity of accurately designing protein-protein interactions; something Bruno has worked on for many years.
Most researchers who would benefit from binders are not computational designers, but biologists who simply require the right tool to facilitate their wetlab experiments. The state of the art for binder discovery until recently relied heavily on yeast display screening due to low design success rates, which was slow, resource-intensive, and frustratingly unreliable. In our own experience, reproducibility was a constant issue: hits from one round would often fail when expressed outside of the yeast context. Moreover, yeast display imposed a significant time and feedback bottleneck: moving from the choice of a target to having a validated binder often required multiple rounds of library sorting and deep sequencing. We wanted to break away from this, and to develop a method where only a handful of candidate binders needed to be designed, ordered as genes, and directly validated; minimizing false positives, false negatives, and wasted time.
The motivation was also personal. Martin had been trying for years to design binders against Cas9, with every method failing. Around the same time, there was more and more evidence for AlphaFold2’s power in selecting promising binder designs through self-consistency tests (Bennett et al. 2023 Nature communications). From earlier work in the group, we already knew that AlphaFold2 “backpropagation” could also yield highly soluble, well-expressed proteins (Goverde et al. 2024 Nature). That sparked the obvious question: if AlphaFold2 is what makes binder prediction more accurate, why not use it directly for design?
Sergey's open source ColabDesign codebase allowed us to relatively quickly assemble a pipeline capable of generating binders to arbitrary protein targets. Our first test case was PD-L1, our lab’s benchmark protein. When we did the very first in vitro pulldown experiment, 7 out of 9 designs bound. None of us believed it at first; this was an enormous leap compared to our previous method of choice, MaSIF (Gainza et al. 2020 Nature). Lennart then confirmed the results with SPR. That was the moment we realized something had fundamentally changed.
From there, we pushed into increasingly difficult targets. Cas9 and Ago were natural choices, given Martin’s background. The AAV part orchestrated by Christian and Lennart followed, directly screening for biological function in the actual context of the application, rather than previously screening for binding. And finally, allergens — motivated by Martin’s and Lennart’s own allergies. Each case brought new challenges but also reinforced the same conclusion: BindCraft could reliably generate binders across a wide range of proteins.
We are incredibly grateful to see the community embracing BindCraft, providing both valuable feedback and inspiring motivation. Our hope is that BindCraft continues to empower researchers, especially those outside the traditional protein design field, to make protein design a common and accessible approach. For us, it has already played a central role in a wide range of projects, and we are excited to see the diverse challenges that other scientists and labs will tackle using it. Lastly, we also hope that the wide adoption of the pipeline effectively demonstrates the strengths of open source and user friendly software and will motivate future, more efficient models to adopt a similar philosophy.