Applicability domain definitions for cheminformatics modelling.

pip install mlchemad

• With molecular fingerprints, prefer the use of the KNNApplicabilityDomain with k=1, scaling=None, hard_threshold=0.3, and dist='jaccard'.
• Otherwise, the use of the TopKatApplicabilityDomain is recommended.

from mlchemad import TopKatApplicabilityDomain, KNNApplicabilityDomain, data

# Create the applicability domain using TopKat's definition
app_domain = TopKatApplicabilityDomain()
# Fit it to the training set
app_domain.fit(data.mekenyan1993.training)

# Determine outliers from multiple samples (rows) ...
print(app_domain.contains(data.mekenyan1993.test))

# ... or a unique sample
sample = data.mekenyan1993.test.iloc[5] # Obtain the 5th row as a pandas.Series object 
print(app_domain.contains(sample))

# Now with Morgan fingerprints
app_domain = KNNApplicabilityDomain(k=1, scaling=None, hard_threshold=0.3, dist='jaccard')
app_domain.fit(data.broccatelli2011.training.drop(columns='Activity'))
print(app_domain.contains(data.broccatelli2011.test.drop(columns='Activity')))

Depending on the definition of the applicability domain, some samples of the training set might be outliers themselves.

The applicability domain defined by MLChemAD as the following:

• Bounding Box
• PCA Bounding Box
• Convex Hull
  (does not scale well)
• TOPKAT's Optimum Prediction Space
  (recommended with molecular descriptors)
• Leverage
• Hotelling T²
• Distance to Centroids
• k-Nearest Neighbors
  (recommended with molecular fingerprints with the use of dist='rogerstanimoto', scaling=None and hard_threshold=0.75 for ECFP fingerprints)
• Isolation Forests
• Non-parametric Kernel Densities