RFdiffusion2 Remote Tool#

Validation and deployment status (2026-10-06)#

RFdiffusion2 (commit d365cbf, official bakerlab_rf_diffusion_aa.sif image and RFD_173.pt weights) ran on one NVIDIA RTX 5000 Ada (32 GB) under rootless Apptainer 1.5.4. Through a direct loopback MCP connection, ToolUniverse discovered rfdiffusion2_design and completed the upstream active_site_unindexed_atomic demo (lactate dehydrogenase 1LDM with NAD and oxamate) in 188 seconds. The returned 180-residue backbone kept both ligands and all four active-site residues, and its motif atoms superimpose on the input at 0.09 A RMSD. Reduced-step, two-design, and ligand-only runs also returned valid designs. Platform relay sharing (tu remote share), non-loopback binding, multi-GPU or concurrent load, and the LigandMPNN and Chai-1 pipeline stages were not tested; keep this deployment private.

  • Operation: rfdiffusion2_design

  • Start: python -m tooluniverse.remote.rfdiffusion2.rfdiffusion2_mcp_server

  • Endpoint: http://127.0.0.1:8033/mcp

  • Provider configuration: RFDIFFUSION2_COMMAND, RFDIFFUSION2_WORKDIR, and RFDIFFUSION2_INPUT_ROOT are required. Keep the container image, weights, and job scratch space outside Git. The server runs one design at a time; a call that cannot start within its own timeout gets a busy error.

  • End-to-end check: python scripts/remote_validation/rfdiffusion2_live_validation.py --out rfdiffusion2_validation --reference-pdb "$RFDIFFUSION2_WORKDIR/rf_diffusion/benchmark/input/mcsa_41/M0584_1ldm.pdb" (the reference PDB enables the motif RMSD check)

Non-loopback binding requires TOOLUNIVERSE_API_TOKEN; otherwise keep the server on loopback.

This implementation is not in the tu remote share catalog or the shared setup-skill preflight, because neither could be exercised without a Platform service key.

See the setup and usage guide for installation, parameters, and output format.