Skip to content

Biotin binder

Objective and inputs

Generate a protein pocket around the bundled linker-bearing biotin analogue, sequence it, and evaluate the holo and apo structures separately. The input files/biotin_all.pdb and files/biotin.json retain the original ligand atom names and hydrogen-bond, burial, and linker-exposure conditioning.

Pipeline

RFD3 → contact/RASA filters → LigandMPNN → soft ligand DDG triage
     → AF3 holo → ligand remapping → contact/H-bond/RASA filters
     → AF3 apo → backbone RMSD/pLDDT filters

RFD3 defaults to two backbones; MPNN makes two sequences per backbone. The original stringent gates may eliminate every candidate in a smoke run. An empty results table is a valid scientific outcome, but does not verify execution of the downstream stages that were skipped.

Run

Configure the Foundry, LigandMPNN, and AlphaFold 3 containers and their model weights. AlphaFold 3 requires separately licensed parameters. The ligand_ddg_soft block also requires PyRosetta in its worker environment. Install HBPLUS and set HBPLUS_COPY to its executable for hydrogen-bond fixing during MPNN and the final hydrogen-bond metric. Set MOLFILE_TO_PARAMS_SCRIPT to Rosetta’s molfile_to_params.py (with its rosetta_py helpers) for on-demand ligand parameter generation.

python examples/biotin-binder/pipeline.py --test --rundir=/absolute/path/biotin-smoke

Use --local --use_local_gpus inside a GPU allocation. Configure site resource mapping rather than embedding a particular cluster's partition in the recipe.

Gates and interpretation

The source recipe requires zero ligand clashes, non-loop fraction ≥ 0.2, minimum ligand distance ≥ 2 Å, and aggregate core contact score ≥ 110. At least one linker atom C14/C15/C16 must have RASA ≥ 0.1. Core burial is an RFD3 conditioning request; the aggregate contact score does not prove that every core atom is buried. DDG triage retains the lowest 10% per parent.

Holo filtering requires interaction PAE ≤ 1.2, at least three reported hydrogen bonds, and contact score ≥ 120. Apo filtering requires backbone RMSD ≤ 2 Å and pLDDT ≥ 85. These computational criteria do not establish binding or affinity. The DDG value is a model-dependent score, not a measured binding free energy.

Outputs and validation

results.csv, engine manifests, intermediate structures, and per-block metrics are written beneath --rundir. Ligand remapping makes predicted ligand names consistent before atom-specific metrics run. CPU configuration tests cover the ported recipe; bounded GPU execution is described below.

Adapted from the original pipelines examples/biotin-binder at revision 95ef0b6fd87eb27e1e73d2cfaa1482a39a0a132f.

MPNN returns side-chain-packed structures, with one packing sample per designed sequence; refolding consumes those sequences and structure context.

GPU execution check

A bounded CoreWeave H200 diagnostic completed RFD3, one side-chain-packed LigandMPNN sequence, holo AF3, ligand remapping, hydrogen-bond/contact/RASA/RMSD metrics, and apo AF3/RMSD. The same packed design also completed five PyRosetta DDG replicates on CPU, with a finite trimmed-mean score. Generated and predicted coordinates were finite. This used native AtomWorks 2 RFD3, the public MPNN container, and an existing AF3 image with Chaperone's public inference runner; it was not a fresh container build test.

The diagnostic deliberately evaluated later stages even when a quality gate would stop the normal recipe. This candidate failed the holo gates: two hydrogen bonds (required three) and minimum interaction PAE 1.69 (maximum 1.2). Execution success therefore does not imply a successful biotin binder. Holo backbone RMSD was 1.09 Å; apo-versus-holo backbone RMSD was 0.40 Å.

Packed-output checks also verified that the sampled FASTA matches the packed protein sequence and all 21 biotin heavy atoms retain their identities and coordinates. After restoring the five fixed protein residues during compilation, all 39 observed fixed heavy-atom coordinates match the staged input exactly.