Binder Design
Leading protein designers, including the teams at MIT and EPFL who authored BindCraft and ColabDesign release Boltzdesign1. The protocol enables AI/structure prediction to design binders to small molecules, nucleic acids, metal ions, and PTMs.
Much like previous approves from the team, Boltzdesign1 iteratively generates new sequences, predicts the structure of that sequence against the target of interest. Then the scores generated through the structure prediction are used to inform the next sequence to test. Boltzdesign1 uses Boltz-1 (also replaceable with Chai-1 or AF3 for academics), which is a open-source/commercially available reproduction of AlphaFold3.
Try Boltzdesign1 out on Tamarind now: https://app.tamarind.bio/tools/boltzdesign
Results
Below are the promising in silico benchmarks reported by the authors
Benchmark | Metrics used | Comments |
|---|---|---|
Small-molecule binders (IAI, FAD, SAM, OQO) | AF3 pLDDT > 0.7 & ipAE < 10; cross-model RMSD; self-consistency; Gnina CNN-VS; diversity (pairwise TM-score) | 87-96 % designs pass strict AF3 filter; mean TM-score ≈ 0.36 (high diversity); 50-63 % pass cross-model & self-consistency < 2 Å; 4-9 % designs beat native ligand in docking |
Protein-protein set (212 BindCraft binders) | Precision@K of contact maps | 76 % designs P@K > 0.5 with Pairformer alone; recycling > 3 gives no benefit |
Metals, DNA & PTM sites | AF3 confidence; AllMetal3D classification; qualitative interface inspection | Octahedral Zn and tetrahedral Fe coordination recovered; B-DNA binder shows shape complementarity & electrostatic match; binders cradle phospho-Tyr211 (PCNA), phospho-Ser201/203 (Smad2), and GlycNAc-Asp101 (CD45) |
Ablation tests | Recycling vs. none; fixed vs. re-designed interface | Highest success with **no recycling and interface fixed**, implying adversarial sequences arise during deep recycling |
The Future
Boltzdesign1 is still in active development, and there is still no experimental validation to the binders generated. Notably, the in silico metrics discussed above were shown to correlate with experimental success in precursors like BindCraft, although it may not necessarily be the case that AF3 metrics match AF2 for e.g. small molecule-protein interactions vs PPIs for AF2. The authors note that there are limitations especially for nucleic acid targets.
Here are most immediate next steps we've identified from the authors:
Stabler, target-aware optimisation – develop loss schedules that are tuned to a ligand or interface class and couple them with joint sequence-and-structure optimisation.
Tackle harder targets: • highly flexible DNA/RNA • proteins bearing multiple covalent modifications • binders that must discriminate among PTM states.
Introduce template conditioning for uncertain folds.
Integrate nucleic-acid MSAs to boost protein–DNA/RNA design accuracy.
Separate design and scoring models to curb over-fitting.
Read the original work:
Image credit: Figure 5 from Boltzdesign1 preprint