Euclidean Embedding of Molecules and Backbone Refinement Class Features based on Co-Occurrence and Entropy. (Schulz et al.)

The 2-D embedding reflects feature-feature and instance-feature co-occurrence as well as entropy of the features with respect to the target classes.

Usage: (De)activating features are (red) green, (In)active instances (salmon) blue. Point your mouse to a feature (instance). The matching instances (features) will be marked. The brighter a feature, the more significantly it is correlated with the endpoint. Use the mouse wheel to zoom in and out and the left mouse button to drag.

Observations: Separation of target classes along top left to bottom right, many highly descriptive features. Good distribution of features and instances, and well-characterized groups of instances in the outer parts. The data seems suitable for classification tasks.

Links to the datasets used in this study (the thresholds used and switches for fminer are given below):

• CPDB carcinogenicity and mutagenicity datasets
  To reproduce experiments (Fminer arguments) : 95% significance (default), minimum frequency 6 (-f6).
• Yeast anti-cancer datasets
  To reproduce experiments (Fminer arguments) : 95% significance (default), minimum frequency given in the paper.

Data: CPDB salmonella mutagenicity
Thresholds: 95% significance, minimum frequency 6 (Switches: -f 6)

Data: CPDB Multicell Call
Thresholds: 95% significance, minimum frequency 6 (Switches: -f 6)

Data: Mouse Carcinogenicity
Thresholds: 95% significance, minimum frequency 6 (Switches: -f 6)

Data: Rat Carcinogenicity
Thresholds: 95% significance, minimum frequency 6 (Switches: -f 6)

BBRC Feature Validation

CPDB and Large-Scale Datasheet, comparison to Open Trees, Maximal Trees (large scale), Static vs Upper Bound Pruning.

Excel Sheet (.xls) for MS Office 2000/2003/2007 (132K)

• Runtime
• Feature Count
• Leave-One-Out Crossvalidation Accuracy
• Large-Scale Analysis

Repeatability: the thresholds used and switches for fminer are given in the file.

Contact