We submitted the following files to Phase I of the PTC for
2000-2001:

	class_blind_0.10_fm_fragment_table
	class_blind_0.01_fm_fragment_table
	class_sens_fm_fragment_table

	class_blind_0.10_mm_fragment_table
	class_blind_0.01_mm_fragment_table
	class_sens_mm_fragment_table

	class_blind_0.10_fr_fragment_table
	class_blind_0.01_fr_fragment_table
	class_sens_fr_fragment_table

	class_blind_0.10_mr_fragment_table
	class_blind_0.01_mr_fragment_table
	class_sens_mr_fragment_table

These files contain tables of frequent linear molecular fragments.
The content of these files should be more or less self-explanatory.
"class blind"/"class sensitive" refers to the way the features
were constructed. "fm" means female mice, "mm" means male mice,
"fr" means female rats and "mr" means male rate.

For the class-blind runs, we set the minimum frequency of the
fragments to 10% and the maximum frequency to 90%. In a second
attempt, we constructed less general features/fragments and
set the minimum frequency threshold to 1% and the maximum
frequency threshold to 99%. The files containing "0.10" in
their names are the files for 10%/90%, and the files containing
"0.01" are those for 1%/99%.

For the class-sensitive runs, we queried for features/fragments
that are significantly over-represented in the active compounds
and under-represented in the inactives. The minimum frequency
on the actives was set to 6, 10, 15 and 20 in turn, and the 
maximum on the inactives was set dependent on the class distribution
such that the occurrence of a fragment in the active class is
significant at a significance level of 0.999 according to the
chi-square statistic. Only very few linear fragments turned
out to be strongly statistically associated with the class
membership. Decreasing the significance level might have
given more of these fragments.

We took only examples with known and valid Smiles strings
for feature construction. Secondly, we generally ignored
the equivocals in feature construction. For class-sensitive
feature construction, compounds with "some evidence" (se) were
counted as positive (carcinogenic).

The details are described in the following papers. Linear molecular
fragments have been introduced in Inductive Logic Programming and
Data Mining in:

Kramer S., Frank E.: Bottom-Up Propositionalization, in: Notes on the
Work-In-Progress Track of the Tenth International Conference on Inductive
Logic Programming (ILP-2000), 2000.

The integration in the inductive database framework (querying
for features of interest) has been described in:

De Raedt L., Kramer S.: The Level-Wise Version Space Algorithm and its
Application to Molecular Fragment Finding, to appear in:
Proceedings of the Seventeenth International Joint Conference on Artificial
Intelligence (IJCAI-01), 2001.

The use of fragments in conjunction with Support Vector Machines has
been proposed in:

Kramer S., Frank E., Helma C.: Fragment Generation and Support Vector
Machines for Inducing SARs, to appear in: SAR and QSAR in Environmental
Research, 2001.

The use of the MolFea, the molecular feature miner (resulting from
the integration of linear molecular fragments and the inductive
database framework of RDM) for feature construction and feature
mining has been documented in the following two papers:
 
Kramer S., De Raedt L.: Feature Construction with Version Spaces for
Biochemical Applications, to appear in: Proceedings of the Eighteenth
International Conference on Machine Learning (ICML-2001), 2001. 

Kramer S., De Raedt L., Helma C.: Molecular Feature Mining in HIV Data,
submitted, 2001. 

If you are interested in more details about this submission or some of
these papers, do not hesitate to send me an email.

Stefan Kramer
Institute for Computer Science
Machine Learning Lab
Albert-Ludwigs-University Freiburg, Germany
http://www.informatik.uni-freiburg.de/~skramer
skramer@informatik.uni-freiburg.de