DNA binder

Objective and inputs

Design a protein against the bundled full-atom double-stranded DNA duplex CGCGAATTCGCG, based on PDB 1BNA. The target PDB uses chains A and B, each numbered from 1. The recipe validates sequence, numbering, and hydrogen-bond conditioning atoms before generating a run-local RFD3 input JSON.

Pipeline

RFD3 → geometry filter → LigandMPNN → AF3 → backbone RMSD/filter
     → LigandMPNN → AF3 → tighter backbone RMSD/filter

The protein has 120–150 residues. Both DNA strands are fixed context; the sense strand supplies major-groove donor/acceptor conditioning. The origin is the fixed duplex center of mass. The predecessor marked the whole duplex as hotspots and inferred an outward direction from those same atoms; this gives a zero-length direction and nonfinite coordinates in both the original Foundry implementation and PR #397. Using the supported COM origin repairs this geometry error while keeping the hotspot and hydrogen-bond selections. RFD3 generates two backbones and each MPNN round makes two sequences by default.

Run

Install the public RFD3 checkpoint and configure Foundry, LigandMPNN and AlphaFold 3 containers/weights. The public RFD3 model supports nucleic-acid binder generation; a private DNA checkpoint is not required. AlphaFold 3 parameters require their own license/access.

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

Optional --dna_checkpoint=/absolute/path/model.ckpt overrides the installed RFD3 checkpoint. To change the target, supply both --dna_target and --dna_sequence; this script validates existing coordinates and does not construct a new duplex. Use --local --use_local_gpus inside a GPU allocation.

Gates and interpretation

The first filter requires zero chain breaks and non-loop fraction ≥ 0.2. Round one requires ipTM ≥ 0.6, minimum interaction PAE ≤ 2.5 and backbone RMSD ≤ 8 Å. Round two tightens these to 0.7, 1.25 and 3 Å. These are inherited computational gates, not evidence of sequence-specific DNA binding. Specificity against competing DNA sequences is not evaluated by this example. min_pae_interaction is the minimum over all chain pairs, so it can describe the DNA–DNA interface rather than the protein–DNA interface. These inherited gates therefore do not establish binder-specific interface confidence; inspect the chain-pair confidence matrices before promoting a design.

Outputs and validation

results.csv and the generated inputs/dna_binder.json are under the run directory, together with engine manifests and intermediate structures. Empty survivor tables are allowed; a smoke run that stops at a gate does not verify later inference stages. CPU tests check target preparation and block settings; Bounded GPU execution is described below.

The public Foundry models/rfd3/docs/examples/na_binder_design.md documents using rfd3_latest.ckpt for dsDNA binder generation. This port adapts original pipelines examples/dna-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

After the COM-origin repair, a CoreWeave H200 diagnostic completed one RFD3 backbone (123 protein residues and two 12-nucleotide DNA strands), packed LigandMPNN, AF3/RMSD, and a second packed MPNN → AF3/RMSD round. DNA context was preserved through packing, and generated/predicted coordinates and RMSDs were finite. This used native AtomWorks 2 RFD3, the public MPNN image, and an existing AF3 image with Chaperone's public runner; fresh image builds were not tested.

Quality filtering was bypassed only to exercise both rounds. The candidate failed the first-round ipTM gate (0.09 versus 0.6 required); the second diagnostic round also failed its confidence gates. Backbone RMSDs were 0.59 Å and 0.84 Å. These results verify execution, not DNA binding or sequence specificity.

Both packed MPNN rounds matched their sampled protein FASTA sequences and preserved all 24 fixed DNA residues, including atom identities and coordinates. The checks covered 486 nonprotein heavy atoms in round one and 494 after AF3 preparation in round two; displacement from each round's own input was zero.