KinDEL: DNA-Encoded Library Dataset For Kinase Inhibitors

KinDEL is a large DNA-encoded library dataset containing two kinase targets (DDR1 and MAPK14) for benchmarking machine learning models.

Usage

Installation

First create environment with dependencies:

curl -fsSL https://pixi.sh/install.sh | bash
pixi install

Benchmarking

Run the following command to train a model:

pixi shell  # activate the environment
PYTHONPATH=. redun -c kindel/.redun run kindel/run.py train \
    --model <model_name, e.g. xgboost_local> \
    --output-dir out \
    --targets ddr1 mapk14 \
    --splits random disynthon \
    --split-indexes 1 2 3 4 5

where <model_name> has to start with one of the following prefixes:

• xgboost
• rf
• knn
• dnn
• gin
• compose

Collecting results

To collect the model performance results after training, you can use the results.py script, providing the path to the model output files:

python results.py --model-path [path]

Datasets

All datasets are located in AWS S3 at the URL: s3://kin-del-2024/data. You can preview the data using 42basepairs here: https://42basepairs.com/browse/s3/kin-del-2024

The recommended training dataset is stored in the {target}_1M.parquet files, which contain top 1M molecules from the DEL screen that were used to train ML models used in our benchmark.

Data splits are generated in the splits/{target}_{random/disynthon}.parquet files, and the training/validation/testing datasets can be loaded using the following code:

from kindel.utils.data import get_training_data

df_train, df_valid, df_test = get_training_data(target, split_index=split_index)

The results in the benchmark are calculated for the held-out testing sets stored in the heldout/{target}_{on/off}dna.csv files, which contain Kd measurements for the on- and off-DNA compounds. Using the in_library argument you can specify if only the in-library or the extended heldout set is returned. This data can be loaded using the following code:

from kindel.utils.data import get_testing_data

data = get_testing_data(target, in_library=True)
print(data['on'])
print(data['off'])

The full dataset can be downloaded using the following code:

from kindel.utils.data import download_kindel

df = download_kindel(target)

Data structure

All dataset files contain the following columns:

• smiles - the SMILES representation of the molecule
• molecule_hash - a molecular hash constructed from the synthons that uniquely identifies the molecule
• smiles_a - the SMILES of the synthon A
• smiles_b - the SMILES of the synthon B
• smiles_c - the SMILES of the synthon C

Some compounds in the heldout set do not contain synthon SMILES strings and the molecule hash. It means that these compounds were picked from outside the DEL (external compounds in the extended set).

Besides the molecular structure information, the heldout datasets contain the kd column with the experimental Kd measurements. The DEL compounds in the training dataset files additionally contain the following columns:

• seq_target_1, seq_target_2, seq_target_3 - sequence counts of the molecules bound to the target in triplicate
• seq_matrix_1, seq_matrix_2, seq_matrix_3 - sequence counts of the molecules bound to the control in triplicate
• seq_load - the pre population of the molecule