Binder Design Tool Selection
Which tool wins
No single tool is best for every target. Hit-rate is strongly target-dependent, so
choose by target type, what you want to control, and available compute.
The clearest signal comes from head-to-head competitions where many methods design
against the same target. On the Adaptyv Nipah de novo target, the public results show:
| Method | Tested | Binders | Hit-rate |
|---|
| Mosaic (gradient, multi-model) | 9 | 8 | 89% |
| ProteinMPNN hybrid | 28 | 7 | 25% |
| RFdiffusion | 60 | 13 | 22% |
| BindCraft | 98 | 7 | 7% |
| BoltzGen | 182 | 6 | 3% |
Mosaic had the highest hit-rate here, but on a small, expert-tuned sample. The ranking
shifts on other targets, and that target-dependence is true of every method (BoltzGen,
Boltz, BindCraft, Mosaic). You cannot know a priori which will win on a new target, so
this is not a fixed leaderboard.
Because of that, choose a starting point by cost and effort to a binder, not by
assuming a method has the best hit-rate. BoltzGen is the suggested default because it is
turnkey and all-atom, so it gets you testable designs fastest with the least setup.
Mosaic is the high-ceiling option when you can invest time tuning the objective. On a
hard or important target, running more than one method in parallel is reasonable.
De novo binder design?
│
├─ Lowest cost/effort to testable designs → BoltzGen (default)
├─ Hard/important target, can invest tuning → Mosaic (gradient, multi-model)
├─ Ligand / small-molecule binding → BoltzGen (all-atom)
├─ Diversity / exploration → RFdiffusion + ProteinMPNN
├─ End-to-end with built-in validation → BindCraft
└─ Antibody / nanobody (VHH) → germinal skill (also mber, iggm in biomodals)
Tool comparison
| Tool | Strengths | Weaknesses | Best for |
|---|
| BoltzGen | All-atom, single-step, turnkey | One model in the loop; mid-range cost per design | Lowest-effort default, ligand binding |
| Mosaic | Composable multi-model objective, won hard head-to-heads | Needs tuning, local JAX only | Hard or important targets, expert use |
| BindCraft | End-to-end, built-in AF2 validation | Less diverse | Production campaigns |
| RFdiffusion | High diversity | Requires ProteinMPNN; not in biomodals | Exploration, diversity |
| Germinal | Antibody and nanobody formats | Finicky | scFv / VHH design |
Compute cost per design
Adaptyv's own tests of these models showed the following compute cost per accepted
design, averaged across 7 targets (it varies several-fold by target):
| Method | Cost per design |
|---|
| RSO | ~$0.15 |
| RFdiffusion | ~$0.25 |
| Mosaic | ~$0.55 |
| ESMFold2 inversion | ~$0.85 |
| mBER | ~$1.40 |
| Germinal | ~$1.60 |
| BoltzGen | ~$1.80 |
| BindCraft | ~$2.90 |
Per-design compute cost is not the same as cost to a binder, which also depends on the
hit-rate on your target. The gradient methods (RSO, Mosaic) are cheap per design but
need setup and tuning; BoltzGen and BindCraft cost more per design but are turnkey, so
their advantage is low human effort rather than lowest compute cost.
Compute vs effort tradeoff
- Lowest human effort: BoltzGen needs no tuning and runs through biomodals. Good
first pass and good for ligand binding.
- Highest ceiling on a hard target: Mosaic, given time to design and tune the
objective. It runs locally on a JAX GPU rather than through biomodals, and is cheap
per design.
- Whatever the generator, validate with
boltz or chai and rank with ipsae.
Other biomodals-backed options: modal_rso.py (Rejection Sampling Optimization, an
AlphaFold-based gradient method) for minibinders, and modal_mber.py for VHH
nanobodies.
Example pipeline: BoltzGen → Chai → QC
BoltzGen provides all-atom design with built-in side-chain packing. This is one
turnkey path; swap in Mosaic, RFdiffusion, or BindCraft depending on the target.
Target → BoltzGen → Validate → Filter
(pdb) (all-atom) (chai) (qc)
1. Target preparation
# Fetch structure from PDB
# Use pdb skill for guidance
- Trim to binding region + 10A buffer
- Remove waters and ligands
- Renumber chains if needed
2. Hotspot selection
- Choose 3-6 exposed residues
- Prefer charged/aromatic residues
- Cluster spatially (within 10-15A)
3. Design with BoltzGen
First, create a YAML config file (e.g., binder.yaml):
entities:
- protein:
id: B
sequence: 70..100
- file:
path: target.cif
include:
- chain:
id: A
binding_types:
- chain:
id: A
binding: 45,67,89
Then run:
modal run modal_boltzgen.py \
--input-yaml binder.yaml \
--protocol protein-anything \
--num-designs 50
Why BoltzGen?
- All-atom output (no separate ProteinMPNN step needed)
- Better for ligand/small molecule binding
- Single-step design (backbone + sequence + side chains)
4. Alternative: RFdiffusion Pipeline
For maximum diversity or when backbone-only is preferred:
# Step 1: Backbone generation (RFdiffusion, run from the official repo)
python run_inference.py \
inference.input_pdb=target.pdb \
contigmap.contigs=[A1-150/0 70-100] \
ppi.hotspot_res=[A45,A67,A89] \
inference.num_designs=500
# Step 2: Sequence design
modal run modal_ligandmpnn.py \
--input-pdb backbone.pdb \
--params-str "--number_of_batches 16 --temperature 0.1"
5. Validation
modal run modal_chai1.py \
--input-faa sequences.fasta \
--out-dir predictions/
6. Filtering
Apply standard thresholds:
- pLDDT > 0.80
- ipTM > 0.50
- PAE_interface < 10
- scRMSD < 2.0 A
See protein-qc skill for details.
Number of designs
| Stage | Count | Purpose |
|---|
| Backbone generation | 500-1000 | Diversity |
| Sequences per backbone | 8-16 | Sequence space |
| AF2 predictions | All | Validation |
| After filtering | 50-200 | Candidates |
| Experimental testing | 10-50 | Final selection |
Common mistakes
Wrong hotspots
- Using buried residues
- Too many hotspots (over-constrain)
- Wrong chain/residue numbers
Insufficient diversity
- Too few designs generated
- Low temperature in ProteinMPNN
- Not exploring multiple backbones
Poor target preparation
- Including full protein instead of binding region
- Missing important structural features
- Wrong protonation states
Timeline guide
| Step | Compute Time |
|---|
| RFdiffusion (500 designs) | 2-4 hours |
| ProteinMPNN (8000 sequences) | 1-2 hours |
| AF2 prediction (8000 sequences) | 12-24 hours |
| Filtering and analysis | 1-2 hours |
Total: 1-2 days of compute
