New job — BINDER

Clustering & novelty H

Groups by embedding when available, otherwise by sequence composition. A way to pull a representative per cluster instead of keeping only the top score.

ESM-2 embeddings G/H GPU

Protein language model embeddings — input for novelty and clustering. The first run downloads about 2.5GB of model weights.

ESM-2 model download (prewarm) -

Downloads the weights ahead of time, so a network failure shows up as a prewarm failure instead of a mysteriously failed embedding job.

Test job (noop) -

A harness for verifying the queue, progress, log, cancel, and orphan recovery without a GPU.

ProteinMPNN F GPU

Fixes the target chain and designs only the binder. Use SolubleMPNN weights to lower surface hydrophobicity. Pure PyTorch, so it runs as-is on pt_gpu.

RFdiffusion E GPU

Generates backbones by diffusion. No side chains, so sequences must be added with ProteinMPNN. The default weights favor helical binders.

Sequence properties batch F/H

Recomputes MW, pI, net charge, hydrophobicity, extinction coefficient, and developability warnings.

AF2 initial guess G GPU

Re-predicts the complex to produce pae_interaction / plddt_binder.

BINDER_AF2IG_PATH not set — result import only — import results instead

BindCraft E–G GPU

Hallucinates binders by backpropagating through AF2-multimer. Needs JAX, ColabDesign, AF2 weights, and PyRosetta.

BINDER_BINDCRAFT_PATH not set — result import only — import results instead

Boltz G/M GPU

For orthogonal validation — confidence rises when two models agree on the same pose.

BINDER_BOLTZ_PATH not set — result import only — import results instead

Greyed-out engines are not installed on this machine. They stay listed because this list is the roadmap — and results from them can be imported right now.