PAMmla is a set of machine learning models to predict SpCas9 PAM preference from amino acid sequence as described in the manuscript "Custom CRISPR-Cas9 PAM variants via scalable engineering and machine learning" (Silverstein et al. 2025).
A web tool to facilitate making predictions and running in silico directed evolution is also available.
This code was developed using Python version 3.3.18 and Conda 23.7.4. Package requirements are listed in PAMmla_env.yaml
Installation time: ~ 5 min
-
Clone the github repository
git clone https://github.com/RachelSilverstein/PAMmla_ISDE.git -
Create a conda environment from the provided yaml file
conda env create --name PAMmla_env --file=PAMmla_env.yml -
Activate the conda environment
conda activate PAMmla_env -
Add the PAMmla_env environment to ipykernel (only necessary if using the jupyter notebook to train PAMmla model)
python -m ipykernel install --user –-name=PAMmla_env
This functionality is also provided in the PAMmla web tool.
Three pre-trained neural network models are available:
220924_NN_rand_seed0_ROS, 220924_NN_rand_seed2_ROS, and 220924_NN_rand_seed3_ROS.
These models are three instances of training the AA_to_PAM_NN.ipnb model using different train/test splits and should behave similarly. They correspond to the 3 train/test splits in the manuscript Silverstein et al. 2024.
Predictions from each of these models can be generated using the make_predictions.py script.
First, fill in the INPUTS section at the top of the make_predictions.py script. See parameters section below for details.
The file is currently populated with inputs to generate example predictions for SpG (LWKQQR) and wild type SpCas9 (DSGERT).
Then run the script using python make_predictions.py. This will create a folder in your running directory called 'predictions' which contains the output of the script in the form of a csv file.
pred_run_name : A name to identify the run. Using a previously used name will overwrite previous results in the 'predictions' directory.
saved_model_dir: Directory containing the model to be used to generate predictions. Ex. "./220924_select_rand_seed0_ROS"
choose_muts: If True, create predictions for a list of given mutants. Mutants will either be read from chosen_muts_file or input in chosen_muts parameter.
chosen_muts: List of pCas9 amino acid variants to generate predictions for. Mutants are denoted by the amino acids present at positions 1135, 1136, 1218, 1219, 1337 as a 6-character strings. Ex. To denote WT SpCas9, use ["DSGERT"]
chosen_muts_from_file: True or False. If True, variants are provided in a csv file instead of using the chosen_muts parameter.
chosen_muts_file: Path to a csv file with a single column of variants names as 6 character strings as described above. Only used if chosen_muts_from_file is True.
random_muts: Alternative to choose_muts where random mutants are generated
positions_to_randomize: If random_muts parameter is True, positions listed in this list will be randomized. Must only contain a combination of the following positions: ['D1135', 'S1136', 'G1218', 'E1219', 'R1335', 'T1337']
n_samples: Number of variants to generate predictions for if random_muts is True.
make_plots: True or False. Whether to plot predicted PAM heatmaps in addition to outputting predictions csv.
In silico directed evolution powered by the PAMmla model can be performed using the evolve_vars.py script.
This functionality is also provided in the PAMmla web tool.
First, fill in the INPUTS section at the top of the evolve_vars.py script and fill in the arguments to the evolve() function call at the end of the script.
See parameters section below for details.
The file is currently populated with inputs to evolve a variant to maximize selectivity for NGTN PAMs starting from the wild type SpCas9 amino acid sequence, DSGERT.
Then run the script using python evolve_vars.py. This will create a folder in your running directory called 'predictions' which contains the output of the script in the form of a csv file.
This will generate a folder called evolved in the output directory containing evolution results.
The expected output of this example run can be found in evolved > NGT_selectivity_example > final_variants.csv
pred_run_name : A name to identify the run. Using a previously used name will overwrite previous results.
output_dir : Path to location to output results (where evolved dir will be generated)
saved_model_dirs: List of directories containing the models to be used to generate predictions. If more than one
directory is listed, the predictions from all listed models will be averaged together to perform evolution.
Ex. To average predictions from all 3 pre-trained NN models:
saved_model_dirs = ["./220924_NN_rand_seed0_ROS", "./220924_NN_rand_seed2_ROS", "./220924_NN_rand_seed3_ROS"]
save_traj : True or False. Whether to save the entire evolution trajectory (outputs after each round of selections)
or only save variants from the final round of selections.
selection_function: function used to perform selections. select_custom perfoms all functionalities.
mutations_per_variant: starting number of amino acid changes per variant to generate. This number will decay as evolution progresses.
variants_per_round: number of mutant enzymes to generate per starting variant per round of evolution. Large numbers will increase probability of
obtaining optimal variants but also increase runtime. Suggested to keep below 1000 on normal machines.
n_best_variants_per_round: number of 'best' variants to keep during each round of selections. Large numbers will increase probability of
obtaining optimal variants but also increase runtime. Total number of variants on which to generate predictions
is variants_per_round * n_best_variants_per_round so choosing large values of both of these parameters will cause slow runtimes.
decay_after_n_rounds_plateau: Number of evolution rounds to perform once a fitness plateau is reached before decaying the number of mutations per variant.
pam_to_max: Four nucleotide PAM to maximize activity for (must begin with N) ex. 'NGTG'. If not maximizing activity, set to None.
pam_to_min: Four nucleotide PAM to minimize activity for (must begin with N) ex. 'NGTG'.
If not minimizing activity, set to None.
If both pam_to_max and pam_to_min are provided, the ratio of pam_to_max / pam_to_min will be maximized.
high_pams: PAMs for which you are providing additional constraints on the rate constant required to survive selections.
Activity on these PAMs must remain above the rate constant thresholds specified in high_cutoffs
low_pams: PAMs for which you are providing additional constraints on the rate constant required to survive selections.
Activity on these PAMs must remain below the rate constant thresholds spescfied in low_cutoffs
high_cutoffs: log10 rate constant cutoff values for PAMs specified in high_pams in the corresponding order.
Should be between -1 (highest activity) and -5 (lowest activity).
low_cutoffs: log10 rate constant cutoff values for PAMs specified in low_pams in the corresponding order
Should be between -1 (highest activity) and -5 (lowest activity).
pams_to_max_selectivity: Provide this parameter to sort for maximizing selectivity for PAMs of interest. List of 4nt PAMs beginning in N for which to maximize selectivity (sum of rate constants on PAMs
of interest divided by sum of all rate constants).
This can be interpreted as the fraction of total activity which occurs on the PAMs of interest.
If this parameter is provided, then any parameters provided for pam_to_max or pam_to_min will be overridden.
To leave this parameter blank, input None.
starting_variants: List of variants to initiate evolution form or input 'random' to use a randomly generated list of variants_per_round number of variants.
Ex. to initiate evolution from a pool of two specified sequences, WT SpCas9 (DSGERT) and SpG (LWKQQR):
starting_variants= [["D", "S", "G", "E", "R", "T"], ["L", "W", "K", "Q", "Q", "R"]]
This example evolves variants to maximize selectivity for NGTN PAMs starting from the wild type DSGERT amino acids. 1000 variants are generated per round, each containing 4 amino acid changes relative to the starting sequence. After each round of mutagenesis, the top 10 variants are retained for the next round of mutagenesis and selections.
vars, traj = evolve(selection_function=select_custom,
mutations_per_variant=4,
variants_per_round=1000,
n_best_variants_per_round=10,
decay_after_n_rounds_plateau=4,
saved_model_dirs=saved_model_dirs,
all_pams=pams,
aa_positions=all_positions,
pam_to_max=None,
pam_to_min=None,
high_pams=[],
high_cutoffs=[],
low_pams=[],
low_cutoffs=[],
pams_to_max_selectivity=["NGTG", "NGTA", "NGTT", "NGTC"],
starting_variants= [["D", "S", "G", "E", "R", "T"]]
ramp_up_rounds=0
)
This example evolves variants to maximize activity on NGTG PAMs while requiring that activity on NGGG and NGGA PAMs remain below 10^-4. Here we start from 1000 randomly generated variants instead of specifying a starting sequence.
vars, traj = evolve(selection_function=select_custom,
mutations_per_variant=4,
variants_per_round=1000,
n_best_variants_per_round=10,
decay_after_n_rounds_plateau=4,
saved_model_dirs=saved_model_dirs,
all_pams=pams,
aa_positions=all_positions,
pam_to_max='NGTG',
pam_to_min=None,
high_pams=[],
high_cutoffs=[],
low_pams=['NGGG', 'NGGA'],
low_cutoffs=[-4, -4],
pams_to_max_selectivity=[],
starting_variants= 'random'
ramp_up_rounds=0
)
Source code for pre-processing the training data, training the model, and evaluating the model on the test set is provided in the form of a Jupyter notebook, AA_to_PAM_NN.ipynb.
Data in the manuscript were generated using the neural network based model unless stated otherwise.
The NN models described in the manuscript, "Design of custom CRISPR-Cas9 PAM variant enzymes via machine learning", can be reproduced
by running the notebook as-is with random seed set to either 0, 2, or 3 (representing the three train test splits used for the three provided saved models).
Minor differences compared to saved models may be observed due to differing random states for model weight initiation. To reproduce results exactly, generate predictions from the three saved models:
220924_NN_rand_seed0_ROS, 220924_NN_rand_seed2_ROS, and 220924_NN_rand_seed3_ROS.
The optimized linear model described in the manuscript can be reproduced by running the Jupyter notebook AA_to_PAM_linear.ipynb as-is.
Further exploration of feature importance is also provided.