Designer Spotlight: Proteins, small molecules, DNA - making binders for all with BoltzDesign2 - Yehlin Cho

TL;DR

• BoltzDesign is a protein design tool that inverts structure prediction models to create new binders or potentially any type of protein.

• It speeds up computation by focusing on key modules: the Pairformer and Confidence modules.

• BoltzDesign works for proteins, small molecules, metals, and nucleic acids.

• BoltzDesign2, currently in development, can use structural templates instead of MSAs.

Aim: 2k word count

Color palette for figures: blue #56A6D4 and purple #8B90DD are the main ones

#8CD2F4, #3E6175, #56A6D4, #8B90DD, #F5A43E, #FFB3BA

Fonts: Inter or Roboto, bold titles

Introduction

In the past few years, deep learning has completely transformed protein research making it possible to predict structures with remarkable accuracy, explore vast molecular spaces, and accelerate experimental breakthroughs. Tools like RoseTTAFold-AllAtom and AlphaFold3 have pushed these capabilities even further, expanding prediction beyond proteins to include small molecules, nucleic acids, ions, and even covalent modifications.

But what if we could take these models and run them in reverse not just predicting how molecules look, but actually designing new proteins to bind specific targets? That’s exactly what we set out to do with BoltzDesign1.

Problem statement

Built on Boltz-1, an open-source reproduction of AlphaFold3, BoltzDesign1 can utilize a structure prediction model and invert it to design proteins. However, unlike AF2, AF3 and Boltz models incorporate diffusion modules for structure prediction after processing through the MSA and Pairformer modules. Backpropagating through the 200 steps of the diffusion model is computationally expensive, requires substantial memory, is prone to vanishing-gradient issues, and produces only a single sample from the probability distribution of the pair features.

Solution: BoltzDesign1

BoltzDesign1 uses a streamlined approach leveraging only the Pairformer and Confidence modules to cut down computation while maintaining high success rates. Instead of tweaking a single predicted structure, we directly optimize the probability distribution of atomic distances, guiding designs toward sequences that naturally form stable, well-defined complexes.

In silico and experimental results

As an initial demonstration, we benchmarked against small molecules tested in RfDiffusionAA and designed binders. BoltzDesign1-generated binders showed higher in silico success rates and greater structural diversity compared to RfDiffusionAA. Experimental validation is currently in progress.

The main difference between BoltzDesign1 and BoltzDesign2 is that, for protein–protein binder design, we can use templates from natural proteins instead of relying on MSA. As shown in BindCraft, using templates can reduce bias from evolutionary information and also decrease computational time for both forward and backward prediction processes.

Thankfully, AdaptyvBio has also contributed to designing experimentally validated protein–protein binders. We tested our newly updated BoltzDesign2 pipeline, which adopts Boltz2 as its framework, to design protein binders without MSA information relying solely on template information. Using BoltzDesign2, we designed binders for the IL-7Rα and BHRF1 targets.

We further demonstrate that our approach can also be applied to the design of metal binders, nucleic acid binders, and post-translationally modified proteins.

Limitations and what’s next

Method that can generate diverse, high-quality protein–ligand complexes with flexible conformations opening new possibilities for biosensors, enzyme engineering, therapeutics, and beyond.

BoltzDesign is a method that can generate diverse, high-quality protein–ligand complexes with flexible conformations, opening new possibilities for biosensors, enzyme engineering, therapeutics, and beyond.

Users can also specify their own loss functions depending on their goals for example, maximizing β-sheet secondary structure, increasing contacts between the ligand and protein, and more.

However, a current limitation is that the design success rate is somewhat correlated with the structure prediction model’s ability to accurately predict the target structure. For harder targets, such as DNA or RNA, the success rate decreases, and we cannot fully guarantee the confidence module’s output because the model is undertrained on those specific biomolecules.

Therefore, this approach should be developed in parallel with advancements in structure prediction models. The strength of BoltzDesign, however, is that it can theoretically be integrated into any type of model to enable design capabilities.

Resources and links

• Have some novel proteins you want to test in the lab? Come talk to us — we’d like to run many more of those protein designer spotlights, so if you have a cool new hypothesis or model to test we’d love to hear from you!