Artificial intelligence-guided design of lipid nanoparticles for pulmonary gene therapy

Description

This repository contains the data processing code and analysis discussed in the paper "Artificial intelligence-guided design of lipid nanoparticles for pulmonary gene therapy".

---

Table of contents

• Repository structure
• System requirements
• Installation guide
• Demo and Instructions

---

Repository structure

The repository is organized into three main folders - /data, /scripts, and /results.

System requirements

Hardware requirements

To run the analysis code in this repository, it is recommended to have a computer with enough RAM (> 8 GB) to support in-memory operations. Run times will be significantly more efficient with a CUDA GPU, but everything can run with just a CPU.

Software requirements

This code has been implemented with Python version 3.8.19. See below for creating an appropriate conda environment.

OS requirements

This code has been tested on the following systems, although it should work generally for other OS types (e.g., Windows).

• macOS: Ventura 13.2.1

• Linux: Ubuntu 22.04

Installation guide

Install repository

git clone https://github.com/jswitten/LNP_ML.git

Install Python dependencies

Create a conda environnment (lnp_ml) as follows:

conda create -n lnp_ml python=3.8
conda activate lnp_ml
pip install chemprop==1.7.0

Demo and Instructions

For all commands shown below, run them from the /scripts folder.

Making a new dataset (optional)

In order to make a new dataset (and incorporate, for example, your own internal datasets), add a folder in the data/data_files_to_merge directory. The folder should have three files:

• main_data.csv: This should contain two columns, smiles and quantified_delivery, containing the SMILES and measured delivery z-score, respectively, for each row (i.e., LNP formulation).

• formulations.csv: This should contain the columns Formulation (name for the formulation), Cationic_Lipid_Mol_Ratio, Phospholipid_Mol_Ratio, Cholesterol_Mol_Ratio, PEG_Lipid_Mol_Ratio, Helper_lipid_ID, Cationic_Lipid_to_mRNA_weight_ratio, and Form_comment (which can be left blank if there are no relevant notes about it). The molar ratio columns should sum to 100 for each row. You can also provide mass ratios (Phospholipid_Mass_Ratio etc), and, as long as the helper lipid molecular weights are recognized (if not, add an entry in the Component_molecular_weights.csv file), the code will compute molar ratios.

• individual_metadata.csv: This should provide all the other optional data about a measurement: target cell type, organ, etc. See some of the other metadata files for more information.

The rows in these three .csv files should line up, i.e., row #17 for should correspond to the same measurement for all three files. The one exception is that if you have the exact same formulation parameters for each row, you can just specify the formulation parameters once, and the code will assume the formulation is the same for every row.

Also, you will need to add a row in the experiment_metadata.csv file corresponding to your new dataset. The Experiment_ID must match the folder name for your new data folder. The other columns can be specified in experiment_metadata.csv if the value is the same for all measurements in your folder. If not (for example, if you have a dataset with both liver delivery and HeLa cell delivery), the value for that column must be specified in the individual_metadata.csv file.

To incorporate all the data into your new dataset, run the command:

python main_script.py merge_datasets

As a result, all_data.csv in the /data folder will now include your updated dataset. Note that all_data.csv already included in this repository features all the LNP data in this repository. For the subset of data analyzed in the paper, see all_data_for_paper.csv.

Generating train & test splits (optional)

In order to make a train & test split (besides the ones are are already present in the repository, for example, if you are incorporating new data), make a .csv file in data/crossval_split_specs entitled all_amine_split.csv. It should have a format like the current file in the repository:

Data_types_for_component	Values	Data_type_for_split	Train_or_split
Experiment_ID	Whitehead_siRNA	Amine	split
Experiment_ID	LM_3CR	Amine	split
Library_ID,Delivery_target	RM_carbonate, generic_cell	Amine	split
Experiment_ID,Library_ID	A549_form_screen,IR_Reductive_amination	Amine	split
Experiment_ID,Library_ID	A549_form_screen,other	smiles	train
Experiment_ID	Luke_Raj_133_3_form	Random_id	train
Library_ID	RM_Michael_addition_branched	Amine	split
Experiment_ID	IR_BL_AG_4CR	Amine	split
Experiment_ID	IR_BL_AG_3CR_sting	Amine	split
Experiment_ID	Akinc_Michael_addition	Amine	split
Experiment_ID	Li_Thiolyne	Amine	split
Experiment_ID	Liu_Phospholipids	Amine	split
Experiment_ID	Miller_zwitter_lipids	Amine	split
Experiment_ID	Zhou_dendrimer	Amine	split
Experiment_ID	Lee_unsat_dendrimer	Amine	split

Example can be found in data/crossval_split_specs.

The columns for all_amine_split.csv are as follows:

• Data_types_for_component: the data column (in all_data.csv) used to specify a particular subset of the train data.

• Values: values that specify that particular subset (i.e., column identified in Data_types_for_component) of the train data.

• Data_type_for_split: the data column (in all_data.csv) used as the variable along which the split is performed. For example, if you specify Amine, all rows with a particular value for Amine will be grouped together during the train/test/validation splitting. If you specify smiles, all rows with a particular SMILES will be grouped together (there may be multiple rows with the same SMILES string, for example formulation screening with one or more ionizable lipids).

• Train_or_split: specify either train or split. If split is specified, the dataset will be split per the usual train/test/validation approach. If train is specified, all selected data will be put into the train set. This is useful if a data subset is relatively small so measuring and reporting performance on that particular subset is not possible or useful; there is no harm in at least training on these small datasets as it may help the model generalize across different libraries.

To perform a split, run the following command, ensuring to insert the appropriate values for items in {}:

python main_script.py split {split filename} ultra-held-out-fraction {either "-1" or “morgan”} {either nothing or “in_silico_screen_split”}

For example,

python main_script.py split all_amine_split_for_paper.csv -1

This will generate a five-fold cross-validation split according to the split specified in the given split filename. The ultra-held-out-fraction is the fraction of data that you want held out of all of the cross-validation sets. This is useful for reporting error with predictions based on an ensemble average of all the cross-validation splits. If you do not want an ultra-held-out dataset, set it to -1. If you specify morgan then binary Morgan fingerprints (radius 2, 2048 bits) will be included. If you select in_silico_screen_split, then the train and validation sets will be the same, i.e. each model will be trained on 80% of the data with the remaining 20% used as a validation set and also to report (slightly biased) performance of the resulting model. This option should be selected for the actual model used to do in silico screening, since it trains on 80% instead of 60% of the data.

The output will be five folders cv_0, cv_1, cv_2, cv_3, cv_4 and an ultra_held_out folder if that option was specified, within the data/crossval_splits folder.

Training a model

To run a model, run the following command, ensuring to insert the appropriate values for items in {}:

python main_script.py train {name of split folder on which to train}

For example,

python main_script.py train all_amine_split_for_paper

The number of epochs for training can also be specified with --epochs {number of epochs}

If the folder name ends in Morgan: the model will include binary Morgan fingerprints of radius 2, 2048 bits.

Testing model performance

Run the following command, ensuring to insert the appropriate values for items in {}:

python main_script.py analyze {folder with trained model}

For example,

python main_script.py analyze all_amine_split_for_paper

This will create a folder in results/crossval_splits with the name of the folder containing the splits and trained model that contains analyzed results on the five test sets from cross-validation. The folder will have six sub-folders:

• crossval_performance folder will report kendall, spearman, and pearson correlations between predicted and experimental data on the test set, for each test set, and for each individual data subset. It will leave a spot blank if there were < 10 datapoints to reference. It will also report the p value for a significant (Pearson) correlation for each predicted-experimental comparision, along with root mean squared error (RMSE) and n (number of points for that data subset).

• cv_i for i in 0, 1, 2, 3, 4 will contain:

  • predicted_vs_actual.csv, which contains all measurement metadata on the test set along with the columns quantified_delivery and cv_i_pred_quantified_delivery which are the experimental and predicted delivery, respectively.

  • A separate folder for each data subset, which in turn will contain pred_vs_actual_data.csv, the predicted and experimental delivery that particular data subset, and pred_vs_actual.png (a plot of predicted versus experimental delivery).

  • If an ultra-held-out test set was created, results/crossval_splits/{split name} will contain another folder, ultra_held_out. This will contain an individual_dataset_results folder that has individual predicted-experimental plots and data for each data subset, along with ultra_held_out_results.csv which will contain all the statistical characterization for each data subset for the ultra-held-out test set.

Running an in silico screen

Run the following command, ensuring to insert the appropriate values for items in {}:

python main_script.py predict {folder with trained model} {folder with LNPs to screen}

For example,

python main_script.py predict all_amine_split test_screen

The folder with list of SMILES to screen should be in data/libraries/{folder} and should have three files, test_screen.csv, test_screen_extra_x.csv, and test_screen_metadata.csv which hold the SMILES you are screening, the extra_x values (i.e. formulation parameters) and metadata for individual rows of the screen. Note that to run a screen, the formulation parameters must be specified. At this time, we recommend as a baseline selecting the KK formulation (35:16:46.5:2.5 ionizable lipid:DOPE:Cholesterol:PEG-lipid molar ratios) in HeLa cells as this tends to produce the most stable predictions though this may change in the future. If you are doing an expansion of an existing dataset, you should match the conditions from that dataset.

The output will be in results/screen_results/{folder with trained model}/{folder with LNPs to screen}. cv_x_preds.csv will contain predictions for each individual model, while pred_file.csv will contain the averaged predictions along with all the relevant metadata.

Putting it all together

So, an example of a full run-though with small, toy datasets from generating a split to training and analyzing a model to running an in silico screen would be:

python main_script.py split small_test_split.csv 0.2 in_silico_screen_split

python main_script.py train small_test_split_with_ultra_held_out_for_in_silico_screen –epochs 3

python main_script.py analyze small_test_split_with_ultra_held_out_for_in_silico_screen

python main_script.py predict small_test_split_with_ultra_held_out_for_in_silico_screen test_screen

Running all code for training a model based on all data in this repository can take ~3 hours with CUDA GPU or ~24+ hours with CPU only, depending on your computer specifications. The smaller toy dataset discussed immediately above should take significantly less time at ~10 minutes to run with CUDA GPU or ~3 hours with CPU only.