For RIFdock, follow instructions in 1_rifgen_rifdock. The pipeline will require installation of Rifdock: https://github.com/rifdock/rifdock. You will require a DNA structure with your target sequence, which can be generated with x3dna (https://x3dna.org) or other means. All PDBs of HTH scaffolds are provided at https://files.ipd.uw.edu/pub/2023_dna_binders/pdbs.zip.

LigandMPNN and Rosetta steps can easily be demoed using a protein-DNA complex PDB, such as those in the corresponding data. Runtime for a single PDB on a standard computer ranges from 10s (LigandMPNN, no Rosetta relax) to 2 min (Rosetta design or LigandMPNN + Rosetta Relax)

We recommend the following end-to-end pipeline, which are detailed in the companion notebooks 1_rifgen_rifdock and 2_dna_interface_design.ipynb.

1. RIFgen/RIFdock
2. LigandMPNN Prefilter Calibration
3. LigandMPNN Stage 1 + Filtering
4. Loop inpainting diversification (or partial RFdiffusion)
5. LigandMPNN Stage 2 + Filtering
6. LigandMPNN Stage 3 (Ligandmpnn-Rosetta FastRelax recycling) (optional)
7. AlphaFold2 Filtering

Running MPNN requires installation of several programs, including LigandMPNN, PyRosetta, PSI_pred, and PSI_blast. The paths to these programs must be encoded in the 'external_paths.txt' folder. The mpnn_design.py script automates the process of MPNN design and has several options including design of structures in silent files or a folder of pdbs. Additional options include the ability to generate many MPNN sequences and run FastDesign on the resulting outputs constrained by a PSSM from the MPNN designs.

An apptainer containing the dependencies necessary for running mpnn_design.py may be built from mpnn_design.spec: apptainer build mpnn_design.sif mpnn_design.spec

Options are also available to freeze resis labeled as hotspots in the pdb labels, freeze specific resis provided in a comma separated list, and skip designs that lack mainchain-phosphate hydrogen bonds or residues labeled with MOTIF or RIFRES in the recognition helix.

Design can be run in prefilter mode, which first checks if a design passes specified prefilters that are calibrated for their likelihood to pass final metric cutoffs. To run a design in prefilter calibration mode, use the following example:

DEMO: the following scripts can be run on a protein-DNA complex PDB, such as those included in the supplementary data files. Expected run time is approximately 10s - 2 min depending on settings (using Rosetta relax will take ~ 2 min per PDB).

python  {full_path_to_mpnn_design.py} -silent {silent} \
                                -seq_per_struct 1 \  ### this option specifies the number of MPNN sequences to generate per input structure
                                -n_per_silent 1 \ ### In prefilter calibration, you should use this to randomly sample designs from rifdock silent file
                                -temperature 0.2 \ ### This specificies the temperature used in LigandMPNN design
                                -redesign_base_hbonds 1 \ ### In some cases, you want to turn off redesign of residues that form hydrogen bonds to bases, but not here.
                                -hbond_energy_cut -0.5 \ ### This specificies the energy cutoff for what counts as a hydrogen bond in filtering steps
                                -run_predictor 1 \ ### This specifies whether you should run a prefilter before full fast relax. If set to 0, the script will proceed straight to full relax after MPNN sequence design
                                -run_relax 1\n ### Setting this to 1 specificies that a full relax should be performed on the final structures. If set to 0, the script will save a structure with the MPNN sequence threaded onto the input design model.

Once you have generated 1000-10000 designs in prefilter calibration mode, you can follow the notebook 'prefilter_calibration.ipynb' to generate files that specify prefilter cutoffs.

Next, you we run design on the entire set of rifdock outputs using prefiltering to avoid full relax on bad docks. When paths to prefilter_eq_file and prefilter_mle_cut_file are specified, bad docks will be pre-empted from full relax.

python {full_path_to_mpnn_design.py} -silent {silent} \
                                -seq_per_struct 1 \
                                -temperature 0.2 \
                                -redesign_base_hbonds 1 \
                                -hbond_energy_cut -0.5 \
                                -prefilter_eq_file {prefilter_eq_file} \ ### specify the path to the generated prefilter_eq_file
                                -prefilter_mle_cut_file {prefilter_mle_cut_file} \ ### specificy the path to the generated prefilter_mle_cut_file
                                -run_predictor 1 \
                                -run_relax 1\n

You may iterate through MPNN design multiple steps to improve design metrics using a command similar to below. Evidence shows that this iteration between MPNN and Fast Relax can improve sequence design quality. Typically, one can do this with a lower MPNN temp.

python {full_path_to_mpnn_design.py} -silent {silent} \
                                -seq_per_struct 1 \
                                -temperature 0.1 \
                                -redesign_base_hbonds 1 \
                                -hbond_energy_cut -0.5 \
                                -run_relax 1\n`