1.10.1. Defining a Set

DATA >> Add Modeling Test Set was created to make it easy to customize external test sets for use in ADMET Modeler, but they are useful for creating sets in general.

Choosing the Using Sampling... option launches the Create test set using sampling dialog, which allows you to specify a name for the new set (1) and a minimum set size (2). You then need to specify the attribute (3) to be used to sort and partition rows from the spreadsheet. One compound is then chosen from each block of rows in the partition, either drawn at random or because it has the median attribute value for that block (4). You can specify the random number seed to use (5) when the Random option is selected.

The stratified sampling procedure used by this feature is the same as the one used in ADMET Modeler: the observations are sorted by the value of the specified attribute (3), then partitioned into the minimum number of blocks necessary to satisfy the N×min_size/100 blocks where N is the number of observations and min_size is the minimum set size (2). The number actually selected has to be an integer, so the result is rounded up to give the actual number of rows required.

The blocks in the partition are made as nearly equal in size as possible and the differences are spread as evenly as possible, but making all block sizes equal is generally not possible. Taking a 10% sample of 131, for example, entails partitioning the observations into 14 blocks, seven containing 10 observations and seven containing 11 observations. The block size alternates to even out the distribution as much as possible, so the pattern produced is:

11-10-11-10-11-10-11-10-11-10-11-10-11-10

Choosing one example from each block yields a set of 14 rows.

Choosing the Using Selection... option launches the Create test set from selection dialog, which prompts you for a name to give to the new set attribute.

Be aware that if you select an attribute name that already exists, the set attribute will be completely overwritten.