scbirlab/steshin-2023-lohi

Lo-Hi Benchmark

Data from Simon Steshin, Lo-Hi: Practical ML Drug Discovery Benchmark, available from the GitHub repositiory. We used schemist (which in turn uses RDKit) to add molecuar weight, Murcko scaffold, Crippen cLogP, and topological surface area.

Dataset Details

From the original README:

Hit Identification

The goal of the Hit Identification task is to find novel molecules that have desirable property, but are dissimilar from the molecules with known activity. There are four datasets simulating this scenario: DRD2-Hi, HIV-Hi, KDR-Hi and Sol-Hi. They are binary classification tasks such that the most similar molecules between train and test have ECFP4 Tanimoto similarity < 0.4.

• data/hi/drd2 -- for DRD2-Hi
• data/hi/hiv -- for HIV-Hi
• data/hi/kdr -- for KDR-Hi
• data/hi/sol -- for Sol-Hi

There are three splits of the datasets. Use only the first split for the hyperparameter tuning. Train your model with the same hyperparameters for all the three splits and calculate mean metric.

Metric: PR AUC.

Lead Optimization

The goal of the Lead Optimization task is to predict how minor modifications of a molecule affect its activity. There are three datasets simulating this scenario: DRD2-Lo, KCNH2-Lo and more challenging KDR-Lo. They are ranking tasks that have clusters in the test set, so that the molecules in each clusters are quite similar with Tanimoto similarity > 0.4 to the central molecules, and each cluster has one similar molecule in the train set, representing known hit.

• data/lo/drd2 -- for DRD2-Lo
• data/lo/kcnh2 -- for KCNH2-Lo
• data/lo/kdr -- for KDR-Lo

There are three splits of the datasets. Use only the first split for the hyperparameter tuning. Train your model with the same hyperparameters for all the three splits and calculate mean metric.

Metric: spearman correlation is calculated for each cluster in the test set and the mean is taken.

Dataset Sources [optional]

• Repository: https://github.com/SteshinSS/lohi_neurips2023
• Paper: https://arxiv.org/abs/2310.06399

Uses

Bechmarking chemical property prediction models.

Dataset Structure

The data are divided into the Hit Identification (hi, binary classification) and Lead Optimization (lo, regression) tasks. Within each are several datasets from a number of assays. Within each of these are three splits of train and test.

Each split is in a separate pair of train and test files. So the files for split 1 are in train_1.csv.gz, test_1.csv.gz and the files for split 2 are in train_2.csv.gz, test_2.csv.gz.

.
├── hi
│   ├── drd2
│   │   ├── test_1.csv.gz
│   │   ├── test_2.csv.gz
│   │   ├── test_3.csv.gz
│   │   ├── train_1.csv.gz
│   │   ├── train_2.csv.gz
│   │   └── train_3.csv.gz
│   ├── hiv
│   │   ├── test_1.csv.gz
│   │   ├── test_2.csv.gz
│   │   ├── test_3.csv.gz
│   │   ├── train_1.csv.gz
│   │   ├── train_2.csv.gz
│   │   └── train_3.csv.gz
│   ├── kdr
│   │   ├── test_1.csv.gz
│   │   ├── test_2.csv.gz
│   │   ├── test_3.csv.gz
│   │   ├── train_1.csv.gz
│   │   ├── train_2.csv.gz
│   │   └── train_3.csv.gz
│   └── sol
│       ├── test_1.csv.gz
│       ├── test_2.csv.gz
│       ├── test_3.csv.gz
│       ├── train_1.csv.gz
│       ├── train_2.csv.gz
│       └── train_3.csv.gz
└── lo
    ├── drd2
    │   ├── test_1.csv.gz
    │   ├── test_2.csv.gz
    │   ├── test_3.csv.gz
    │   ├── train_1.csv.gz
    │   ├── train_2.csv.gz
    │   └── train_3.csv.gz
    ├── kcnh2
    │   ├── test_1.csv.gz
    │   ├── test_2.csv.gz
    │   ├── test_3.csv.gz
    │   ├── train_1.csv.gz
    │   ├── train_2.csv.gz
    │   └── train_3.csv.gz
    └── kdr
        ├── test_1.csv.gz
        ├── test_2.csv.gz
        ├── test_3.csv.gz
        ├── train_1.csv.gz
        ├── train_2.csv.gz
        └── train_3.csv.gz

The column headings of the data are:

• smiles: SMILES string
• value: The assay result. This is True/False for hi and numeric for lo.
• id: Numeric structure identifier
• inchikey: Unique structure identifier
• scaffold: Murcko scaffold
• mwt: Molecular weight
• clogp: Crippen LogP
• tpsa: Calculated topological polar surface area.

The hi datasets also have a cluster column, indicating the structural cluster of the compound.

Dataset Creation

Curation Rationale

To make the Lo-Hi Benchmark readily available with light preprocessing.

Data Collection and Processing

Additional properties were calculated using schemist, a tool for processing chemical datasets.

Who are the source data producers?

Simon Steshin (https://github.com/SteshinSS).

Personal and Sensitive Information

None

Citation

BibTeX:

@misc{steshin2023lohipracticalmldrug,
    title={Lo-Hi: Practical ML Drug Discovery Benchmark}, 
    author={Simon Steshin},
    year={2023},
    eprint={2310.06399},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2310.06399}, 
}

Dataset Card Contact

@eachanjohnson