binder_design

Binder design

Binder design pipelines. Mostly still work in process..

Round 1 - binder scaffold docking

Idea is to get get initial binder sequence, that binds to the target protein. If you dont care about where binders bind or want to explore prefered binding spots, start with 01a_binder_random_dock.ipynb. If you want to target specific residues, start with start with 01b_binder_dock.ipynb for docking scaffolds (usually 3HB) or 01c_binder_dock.ipynb for docking random scaffolds.

1a Random binder location docking

In the 01a_binder_random_location_dock.ipynb notebook, main thing you need to do is to change parameters, especailly path to target_pdb, target_chain and setup number of iterations.

Final command will look somewhat like this:

sbatch --output=/dev/null --array=0-{scaf_num-1}%{array_lim} helper_scripts/predict_one_binder_random_loc.sh {target_pdb} {target_chain} {scaf_dir} {out_dir} {iterations} {designs_per_iteration} {proteinmpnn_per_input} {recycles} {sampling_temp}

• target_pdb - path to target pdb
• target_chain - name of target chain (can be multiple chains)
• scaf_dir - path to directory with binder scaffold pdbs
• out_dir - output directory
• iterations - number of iterations to perform when optimizing one binder (usually 10-15 is enough)
• designs_per_iteration - total number of (best) designs to go into next iteration as inputs (default: 10)
• proteinmpnn_per_input - total number of sequences to sample per input per iteration (default: 5)
• recycles - number of AF2 recycles (default: 12)
• sampling_temp - ProteinMPNN sampling temperature (default: 0.15)

Note: Continue with 015b_binder_random_dock_analysis.ipynb notebook, to quickly get best pdbs for visual inspection

1b RFdiffusion selected scaffold docking

In the 01b_binder_scaffold_dock.ipynb notebook, you need to setup target_pdb path and target proteinhotspots, specify scaffold_dir (with/without backbone threading before ProteinMPNN) and a number of dockings (num_of_diffusions) to perform for each scaffold.

Final command:

sbatch --output=/dev/null --array=1-{array_number}%{array_limit} helper_scripts/binder_dock.sh {num_of_diffusions} {total_mpnn} {mpnn_per_design} {num_recycles} {sampling_temp} {scaffold_dir} {prefix} {target_pdb} {hotspots}

• num_of_diffusions - number of dockings to perform
• total_mpnn - total number of sequences for per design
• mpnn_per_design - total number of sequences to sample per design (filter from total_mpnn)
• num_recycles - number of AF2 recycles (default: 12)
• sampling_temp - ProteinMPNN sampling temperature (default: 0.15)
• scaffold_dir - path to directory with binder scaffolds ()
• target_pdb - path to target pdb
• hotspots - hotspot residues to target with binder (usually around 5)

Note: If the Scaffold directory contains multiple subdirectories, the script will run each directory separately.

1c RFdiffusion random binder scaffold docking

In the 01c_random_binder_scaffold_dock.ipynb notebook, you need to setup target_pdb path and target proteinhotspots, number of dockings (num_of_diffusions) to perform for each scaffold, along with other MPNN/AF2 parameters. Make sure to setup config file for binder prediction, specifying desired size.

Final command:

sbatch --output=/dev/null --array=0-{array_number-1}%{array_limit} helper_scripts/binder_dock_random_scaff.sh {num_of_diffusions} {total_mpnn} {mpnn_per_design} {num_recycles} {sampling_temp} {prefix} {target_pdb} {hotspots} {contigs} {thread_binders}

• num_of_diffusions - number of dockings to perform
• total_mpnn - total number of sequences for per design
• mpnn_per_design - total number of sequences to sample per design (filter from total_mpnn)
• num_recycles - number of AF2 recycles (default: 12)
• sampling_temp - ProteinMPNN sampling temperature (default: 0.15)
• prefix - input name
• target_pdb - path to target pdb
• hotspots - hotspot residues to target with binder (usually around 5)
• contigs - RFdiffusion contigs for binder generation
• thread_binders - select if you want to thread binder scaffold with sequence and FastRelax before designing

Round 2 - binder optimization

The idea of 02_binder_optimization.ipynb notebook to get primary binders and optimize them with iterations of AF2 and ProteinMPNN, sorting by i_pae or some custom scoring function. Filter the binders and for each binder run an optimization script. Make sure to specify number of iterations and a way to sort.

Final command:

sbatch --output=/dev/null --array=0-{array_number-1}%{array_limit} helper_scripts/binder_opt.sh {binders_input} {output_dir} {iterations} {starting_mpnns} {designs_per_iteration} {mpnn_per_design} {sort} {af2_recycles} {mpnn_sampling_temp}

• binders_input - path to directory with input binder pdbs
• output_dir - output directory
• iterations - number of iterations per binder (10-15 is usually enough)
• starting_mpnns - number of sampled sequences from input binder
• designs_per_iteration - total number of (best) designs to go into next iteration as inputs (default: 10)
• mpnn_per_design - total number of sequences to sample per input per iteration (default: 5)
• sort - sort metrics between iterations (default: "i_pae")
• af2_recycles - number of AF2 recycles (default: 12)
• mpnn_sampling_temp - ProteinMPNN sampling temperature (default: 0.15)

Round 3 - binder analysis

The idea of 03_binder_analysis.ipynb notebook is to filter optimized sequences and calculate additional metrics, that would help use discriminate good binders from the bad ones. The notebook divides datasets into smaller chunks of 1000 and subsequently executes them individually on a cluster.

Final command:

sbatch --output=/dev/null --array=0-{array_number-1}%{array_limit} helper_scripts/binder_analysis.sh {input_file} {target_chain} {binder_chain} {metric_path}

• input_file - list of proteins to calculate metrics for (1000 per array job)
• target_chain - target chain
• binder_chain - binder chain
• metric_path - file for writing results (prepared with notebook)
• xml_file - rosetta xml file

Round 4 - binder filtering and sequence clustering

The idea of 04_binder_filter.ipynb notebook is to filter binders by different metrics, predict them again with AF2 and RF2 (and ESMfold), cluster similar sequences together and group final results.

Workflow:

• filter sequences based on the metrics from 03_binder_analysis.notebook (i_pae, ddg_dsasa_100, charge, hyd_contacts, shape_comp, ...)
• validate sequences again with different prediciton methods (AlphaFold2, RoseTTAfold2, ESMfold)

sbatch --output=/dev/null --array=0-{array_number-1}%{array_limit} helper_scripts/predict_binders.sh {msa_folder} {output_path} {binder_analysis} {prediction_tools}

msa_folder - folder with input msa files
output_path - output folder
binder_analysis - flags for binder analysis (define if you want binder analysis and the position of binder - binder or binder-second)
prediction_tools - prediction tools to use (colabfold, rosettafold2 or/and ...)

• second round of filtering, based on validation results (af2_plddt, af2_pae, af2_rmsd, ...)
• clustering similar sequences together and prepare final binder sequences to order