De novo protein binder design via motif-scaffolded diffusion + sequence design + structural validation.
BinderDiffuser is an end-to-end pipeline for designing protein binders against a chosen target. It composes three battle-tested components into one reproducible workflow:
- Motif-scaffolded RFdiffusion generates novel backbones conditioned on a fixed functional motif (the binding hot-spot of the target).
- ProteinMPNN designs sequences for those backbones, with target chain and motif residues held fixed.
- AlphaFold2 (via ColabFold) re-folds each designed sequence and the pipeline computes self-consistency metrics (scRMSD, scTM, pLDDT, ipTM, pAE_interface) to filter and rank candidates.
The reference example targets PD-L1 (programmed death-ligand 1, gene
CD274) — the canonical immune checkpoint shared by pembrolizumab,
atezolizumab, and durvalumab.
Each stage writes intermediates under output_dir/ so a crashed run can be
resumed by re-invoking from the failed stage; the orchestration code lives in
src/binderdiffuser/pipeline.py.
Motif scaffolding > unconditional generation. Naive backbone diffusion
samples folds without regard to where they should bind on the target. By
fixing a motif (e.g. PD-L1 residues 56-65) and letting RFdiffusion grow
scaffold around it, the generated cohort concentrates at the productive
binding face from sample one. The contig grammar exposed by RFdiffusion is a
small DSL; this repo includes a typed builder
(motif_spec.py) that
generates valid contigs from a high-level MotifSpec.
Self-consistency > MPNN scoring alone. A ProteinMPNN log-likelihood score reflects how plausible a sequence is given a backbone, but it does not check whether the sequence will actually fold to that backbone. The self-consistency loop solves that: re-fold the designed sequence with AF2, then check whether AF agrees with the diffuser. Designs where the two disagree (high scRMSD, low scTM) are filtered out before any wet-lab work.
Backend swap-in for AlphaFold3. AlphaFold3 has higher accuracy and a
public CLI is on the way. The
AlphaFoldRunner
abstracts over the backend so an AF3 swap is a one-line config change once
the public CLI lands.
After a run, binderdiffuser report <run_dir> regenerates the canonical
diagnostic plots from designs.csv:
The lower-left/upper-right quadrant (low scRMSD, high pLDDT, high ipTM) is where the productive designs cluster. The dashed lines mark the default filter thresholds (scRMSD = 2 Å, pLDDT = 70).
The violin plots make it easy to see whether a run is producing a meaningful right tail (top-K marked in orange) or a uniformly mediocre cohort.
Note: the figures above are rendered from a synthetic cohort (
figures/sample_designs.csv) so the README is interactive without requiring the reader to run GPU inference. Replace the CSV with a real run viabinderdiffuser report path/to/your/run/.
git clone https://github.com/deepmind11/BinderDiffuser.git
cd BinderDiffuser
pip install -e ".[dev,notebooks]"The CPU-friendly stages (motif extraction, contig builder, metric helpers, filter/rank, viz) work out of the box. RFdiffusion and ColabFold are external GPU-bound binaries; install them via the Dockerfile or follow upstream install guides:
- RFdiffusion: https://github.com/RosettaCommons/RFdiffusion
- ProteinMPNN: https://github.com/dauparas/ProteinMPNN
- ColabFold: https://github.com/sokrypton/ColabFold
Sanity-check the contig setup without GPU:
binderdiffuser show-contigs examples/pdl1_binder/config.yaml -n 5Sample output:
[00] seed=348282721 B1-65/0 12,B56-65,15
[01] seed=121478392 B1-65/0 8,B56-65,22
[02] seed=940482910 B1-65/0 19,B56-65,11
...
binderdiffuser run examples/pdl1_binder/config.yamlThis drives RFdiffusion → ProteinMPNN → ColabFold and writes:
examples/pdl1_binder/runs/v1/
├── backbones/ # RFdiffusion outputs
├── sequences/ # ProteinMPNN outputs (per-backbone FASTAs + bundle)
├── alphafold/ # ColabFold outputs (per-sequence PDBs + score JSONs)
├── designs.csv # one row per (backbone, sequence) pair with all metrics
└── figures/ # regeneratable plots
src/binderdiffuser/
├── config.py # Pydantic schemas; YAML round-trip
├── targets.py # PDB parsing, motif residue extraction
├── diffusion/
│ ├── motif_spec.py # RFdiffusion contig string builder
│ └── rfdiff_wrapper.py # subprocess wrapper for run_inference.py
├── sequence/
│ └── mpnn_runner.py # ProteinMPNN runner + FASTA parser
├── validation/
│ ├── alphafold_runner.py # ColabFold / AF2/3 wrapper
│ ├── metrics.py # scRMSD, scTM, pLDDT, ipTM, pAE_iface
│ └── filters.py # threshold filtering + composite ranking
├── pipeline.py # end-to-end orchestration (run_pipeline)
├── viz.py # scatter / violin / PyMOL script helpers
└── cli.py # `binderdiffuser run | show-contigs | report`
examples/pdl1_binder/ # PD-L1 worked example
notebooks/ # walkthroughs (5 sections, no GPU required)
figures/ # README-facing artifacts
docker/ # GPU image (CUDA 12.1 + PyTorch 2.2 + ColabFold)
tests/ # 51 unit tests
.github/workflows/ci.yml # py3.10/3.11/3.12 matrix; ruff + pytest
pytest --cov=binderdiffuser tests/51 unit tests covering:
| Module | Tests |
|---|---|
targets |
PDB loading, chain selection, residue extraction, motif segmentation (contiguous, disjoint, unsorted, duplicates, error paths) |
motif_spec |
Validation, contig format, length envelope honoured, reproducibility under seed |
metrics |
Kabsch (identity, translation, rotation, known offset, shape mismatch), TM-score (perfect alignment, monotonic with noise, d0 clamp), pLDDT, ipTM (multiple JSON field names, missing/non-numeric), pAE_interface |
filters |
Threshold pass/fail, composite ordering, top-K, missing optional metrics, DataFrame round-trip |
- Wire AlphaFold3 backend once the public CLI is released
- Add a Boltz-1 backend alongside AF2 for cross-tool consensus filtering
- Add Rosetta InterfaceAnalyzer scoring as a post-filter signal
- Add a
binderdiffuser sweepsubcommand for hyperparameter sweeps - Add a Streamlit dashboard for run inspection
- Watson J. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089-1100 (2023). https://www.nature.com/articles/s41586-023-06415-8
- Dauparas J. et al. Robust deep learning-based protein sequence design using ProteinMPNN. Science 378, 49-56 (2022). https://www.science.org/doi/10.1126/science.add2187
- Jumper J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583-589 (2021). https://www.nature.com/articles/s41586-021-03819-2
- Cao L. et al. Design of protein-binding proteins from the target structure alone. Nature 605, 551-560 (2022). https://www.nature.com/articles/s41586-022-04654-9
- Mirdita M. et al. ColabFold: making protein folding accessible to all. Nature Methods 19, 679-682 (2022). https://www.nature.com/articles/s41592-022-01488-1
MIT — see LICENSE.