Skin Doctor CP

Skin Doctor CP is a machine learning model for the classification of small organic compounds into skin sensitizers and non-sensitizers. More specifically, the core of Skin Doctor CP is a random forest binary classifier that is enveloped in an aggregated Mondrian conformal prediction framework. This allows users to define an error significance level (i.e. error rate) for classification. Predictions (i.e. sensitizer or non-sensitizer) are thus only reported for compounds for which the expected reliability reaches or exceeds the error rate defined by the user. The model was trained on a curated data set of 1285 compounds measured in the local lymph node assay (LLNA).

Skin Doctor CP has been developed in collaboration with Beiersdorf AG.

Usage

Skin Doctor CP can be called from the command line. Examples:

# input in SMILES format
skindoctorcp "CCOC(=O)N1CCN(CC1)C2=C(C(=O)C2=O)N3CCN(CC3)C4=CC=C(C=C4)OC"

# significance_level1 to significance_level4 are numbers between 0 and 1
# (default: 0.05, 0.1, 0.2, 0.3)
skindoctorcp --significance_level1 0.05 --significance_level2 0.1 \
    --significance_level3 0.2 -- significance_level4 0.3 \
    "CCN(C)C(=O)OC1=CC=CC(=C1)C(C)N(C)C"

# input can be a file
skindoctorcp molecules.sdf > result.csv

# output format can be specified
skindoctorcp --output sdf molecules.smiles > result.sdf

# more information via --help
skindoctorcp --help

The model can be used in Python. Calling the predict function of the SkinDoctorCPModel class results in a pandas DataFrame containing the prediction results for each input molecule.

from skindoctor import SkinDoctorCPModel

model = SkinDoctorCPModel()

# "predict" method accepts a list of SMILES representations
df_predictions = model.predict(['CCN(C)C(=O)OC1=CC=CC(=C1)C(C)N(C)C'])

# ... or a list of file paths
df_predictions = model.predict(['part1.sdf', 'part2.sdf'])

The result DataFrame contains the columns:

• mol_id: unique number identifying the input molecule
• input: the raw representation provided as input (e.g. OCCCCC)
• input_type: the representation type of the input (e.g. smiles)
• source: the input source (e.g. my_molecules.sdf)
• name: the name of the input molecule (if provided in the input)
• input_mol: the RDKit molecule parsed from the input representation
• preprocessed_mol: the RDKit molecule after preprocessing
• errors: a list of errors that occured during reading or preprocessing the input
• prediction_significance_level1, prediction_significance_level2, prediction_significance_level3, prediction_significance_level4: Prediction of Skin Doctor CP with a reliability fulfilling the selected error significance
• p_value_sensitizing: p-value for the sensitizing class
• p_value_non_sensitizing: p-value for the non-sensitizing class

Contribute

conda env create -f environment.yml
conda activate skin-doctor
pip install -e .[dev,test]
ptw

Contributors

• Anke Wilm
• Steffen Hirte
• Axinya Tokareva