BindCraft Background

BindCraft methodology overview

Pacesa et al, Nature, 2025 - One-shot design of functional protein binders with BindCraft

Figure 1a from Pacesa et al 2025 (https://doi.org/10.1038/s41586-025-09429-6) - a high level overview of the BindCraft design pipeline

Hallucination with the Alphafold2 model as implemented by ColabDesign.

The trajectory is split into four stages, where in each stage the probability distribution of the sequence at each position predicted by the Alphafold2 model is made increasingly discrete.

ie, we go from a ‘fuzzy’ sequence with many possible residues at a position, to fixing on the highest probability residue.

From Martin Pacesa and Lennart Nickel’s presentation https://www.youtube.com/watch?v=u5yijcBsonw

The final ‘semi-greedy’ stage tests the impact of randomly mutating a small fraction of positions (based on a PSSM derived from the sequence logits)

The BindCraft loss function

It’s instructive to look at what goes into the BindCraft loss function that guides the trajectory and generates the initial design. We have weighted terms that drive the confidence of the binder and confidence of the interface, as well as encouraging more non-helical secondary structure and a more compact radius of gyration.

\[\begin{align} ℒ &= 0.1⋅(1−pLDDT) \nonumber \\ &\quad + 0.05⋅ipTM \nonumber \\ &\quad + 0.4⋅pAE_{binder} \nonumber \\ &\quad + 0.1⋅ipAE_{interchain} \nonumber \\ &\quad + 1.0⋅hotspot\_contact_{loss} \nonumber \\ &\quad + 0.3⋅radius_{gyr} \nonumber \\ &\quad - 0.3⋅helicity_{loss} \end{align}\]