BindCraft2 protein-binder campaigns
BindCraft2 designs protein binders using AlphaFold and ProteinMPNN/HyperMPNN. Status: planned. The integration awaits clean installation and bounded campaign verification on real Linux/NVIDIA hardware. Normal installation stays blocked until that gate passes.
Requirements and license
| Package | bindcraft 1.0.3 from pinned source |
| Python | Managed 3.12 |
| Platform | Linux x64 + NVIDIA; CUDA 12 / JAX 0.11.2 |
| Disk / RAM | At least 25 GB disk and 24 GB host RAM; VRAM depends on complex size |
| Input | JSON campaign referencing PDB/mmCIF/FASTA targets and optional custom scaffold |
| Results | Trajectory/refolded/ranked CSVs, mmCIF structures, campaign metadata and explicit accepted-design count |
BindCraft2 Source-Available License (Hosting-Restricted): local/internal commercial use is permitted. Exposing substantial functionality as a hosted service requires a separate commercial license. BindCraft2 is not OSI open source. MoleculeDesk runs it on your own workstation or rented GPU server.
Bounded campaigns
{
"targets": [{ "name": "my_target", "target_path": "target.pdb", "chains": "A" }],
"modality": "binder",
"binder_lengths": [50, 60],
"number_of_final_designs": 1,
"max_trajectories": 1,
"resume": false
}Both design count and trajectory budget must be positive integers. Target and scaffold files stay within the campaign directory tree and are copied into the run's immutable input area. Each invocation creates a fresh independent run.
moldesk run bindcraft2 campaign.json --param max_trajectories=2Supported modalities: binder, peptide, cyclic_peptide, VHH, ARP.
Use --param number_of_final_designs=1 or paired min_length/max_length
overrides. VHH/ARP use shipped scaffolds. Resume, automatic workers, arbitrary
preset files and modality combinations are not supported yet. CUDA is required;
CPU fallback is refused.
Results
A campaign can exhaust its budget with zero accepted designs. That is a
successful bounded computation, recorded explicitly in moldesk-summary.json,
with trajectory/refolded records retained. Accepted structures appear only when
designs pass upstream filters. All collected outputs are checksummed in run.json.
Predicted candidates still require experimental validation.
BindCraft2 benchmarks: speed, VRAM and cost per design campaign
On an NVIDIA GeForce RTX 3090, BindCraft2 takes about 3.8 min per design campaign once warm (2.5 min on the first run), roughly 15 design campaigns per GPU-hour, or about $0.032 per design campaign.
| Metric | NVIDIA GeForce RTX 3090 |
|---|---|
| Time per run, first (cold) | 2.5 min |
| Time per run, warm | 3.8 min |
| Design campaigns per GPU-hour | 15 |
| Cost per run, first (cold) | $0.021 |
| Cost per run, warm | $0.032 |
| Peak GPU memory | 4.5 GiB |
| Peak GPU utilization | 100 % |
| Peak GPU power | 303 W |
| Install time | — |
| Disk per install | 13 GiB |
Workload: Bounded campaign (examples/bindcraft2/campaign.json): 50-residue binder against a short peptide target, 1 trajectory, 1 final design, seed 101. Median of 3 runs.
Machine: RunPod GPU pod, Linux x64, 125 GiB RAM; AMD EPYC 7H12 (32 vCPU allocated); driver 580.126.20.
Cost: at $0.50/GPU-hour (RunPod Secure Cloud, EU-CZ-1, compute only; storage adds about $0.02/hr); excludes storage and idle time.
Runs took 152 s, 232 s and 229 s: design search is stochastic, so time varies run to run and there is no cold-start penalty. The 'warm' figure is the mean of runs 2 and 3. Real campaigns with more trajectories scale roughly linearly with trajectory count.
Verification status
Linux dependency resolution, a real installation and bounded campaigns on CUDA
are verified on an NVIDIA RTX 3090 (see the benchmarks above); cancellation and
zero-design behavior still need GPU testing. The repository's
models/bindcraft2/README.md contains exact private test-registry commands,
source/asset provenance and the live verification checklist. No macOS, MPS,
CPU, AMD or multi-GPU support is advertised for this integration.