1. Title: MUSK "Clean2" database 2. Sources: (a) Creators: AI Group at Arris Pharmaceutical Corporation contact: David Chapman or Ajay Jain Arris Pharmaceutical Corporation 385 Oyster Point Blvd. South San Francisco, CA 94080 415-737-8600 zvona@arris.com, jain@arris.com (b) Donor: Tom Dietterich Department of Computer Science Oregon State University Corvallis, OR 97331 503-737-5559 tgd@cs.orst.edu (c) Date received: September 12, 1994 3. Past Usage: (a) Dietterich, T. G., Jain, A., Lathrop, R., Lozano-Perez, T. (1994). A comparison of dynamic reposing and tangent distance for drug activity prediction. Advances in Neural Information Processing Systems, 6. San Mateo, CA: Morgan Kaufmann. 216--223. The clean2 dataset included here is derived from the starting poses employed in this paper. The paper reports the following results: Algorithm: 20-fold XVAL: 1-nearest neighbor (euclidean distance) 75% neural network (standard poses) 75% 1-nearest neighbor (tangent distance) 79% neural network (dynamic reposing) 91% The tangent distance and dynamic reposing technique require computation of the molecular surface, which cannot be done using the feature vectors included in this data set. (b) Jain, A. N., Dietterich, T. G., Lathrop, R. H., Chapman, D., Critchlow, R. E., Bauer, B. E., Webster, T. A., Lozano-Perez, T. Compass: A shape-based machine learning tool for drug design. Accepted for publication in Computer-Aided Molecular Design. This paper describes the dynamic reposing technique in more detail and reports the same result for dynamic reposing as above. The paper also gives a complete description of each of the 102 molecules in the data set. (c) Dietterich, T. G., Lathrop, R. H., Lozano-Perez, T. (submitted) Solving the multiple-instance problem with axis-parallel rectangles. Submitted to Artificial Intelligence. This paper describes a family of axis-parallel rectangle algorithms and compares various approaches to the multiple instance problem. It includes the following table: Algorithm TP FN FP TN errs %correct [CI] iterated-discrim APR 30 9 2 61 11 89.2 [83.2--95.2] GFS elim-kde APR 32 7 13 50 20 80.4 [72.7--88.1] GFS elim-count APR 31 8 17 46 25 75.5 [67.1--83.8] all-positive APR 34 5 23 40 28 72.6 [63.9--81.2] backpropagation 16 23 10 53 33 67.7 [58.6--76.7] GFS all-positive APR 37 2 32 31 34 66.7 [57.5--75.8] most frequent class 0 39 0 63 39 61.8 [52.3--71.2] C4.5 (pruned) 32 7 35 28 42 58.8 [49.3--68.4] key: TP = true positives FN = false negatives FP = false positives TN = true negatives errs = errors = FN+FP %correct = 10-fold cross-validation %correct. CI = 95% confidence interval on proportion of correct predictions. For explanations of the various algorithms, see the paper. C4.5 and backprop were applied ignoring the multiple instance problem (see below) during training, but obeying it during testing. This paper also gives more details on the construction of the data set. 4. Relevant Information: This dataset describes a set of 102 molecules of which 39 are judged by human experts to be musks and the remaining 63 molecules are judged to be non-musks. The goal is to learn to predict whether new molecules will be musks or non-musks. However, the 166 features that describe these molecules depend upon the exact shape, or conformation, of the molecule. Because bonds can rotate, a single molecule can adopt many different shapes. To generate this data set, all the low-energy conformations of the molecules were generated to produce 6,598 conformations. Then, a feature vector was extracted that describes each conformation. This many-to-one relationship between feature vectors and molecules is called the "multiple instance problem". When learning a classifier for this data, the classifier should classify a molecule as "musk" if ANY of its conformations is classified as a musk. A molecule should be classified as "non-musk" if NONE of its conformations is classified as a musk. 5. Number of Instances 6,598 6. Number of Attributes 168 plus the class. 7. For Each Attribute: Attribute: Description: molecule_name: Symbolic name of each molecule. Musks have names such as MUSK-188. Non-musks have names such as NON-MUSK-jp13. conformation_name: Symbolic name of each conformation. These have the format MOL_ISO+CONF, where MOL is the molecule number, ISO is the stereoisomer number (usually 1), and CONF is the conformation number. f1 through f162: These are "distance features" along rays (see paper cited above). The distances are measured in hundredths of Angstroms. The distances may be negative or positive, since they are actually measured relative to an origin placed along each ray. The origin was defined by a "consensus musk" surface that is no longer used. Hence, any experiments with the data should treat these feature values as lying on an arbitrary continuous scale. In particular, the algorithm should not make any use of the zero point or the sign of each feature value. f163: This is the distance of the oxygen atom in the molecule to a designated point in 3-space. This is also called OXY-DIS. f164: OXY-X: X-displacement from the designated point. f165: OXY-Y: Y-displacement from the designated point. f166: OXY-Z: Z-displacement from the designated point. class: 0 => non-musk, 1 => musk Please note that the molecule_name and conformation_name attributes should not be used to predict the class. 8. Missing Attribute Values: none. 9. Class Distribution: Musks: 39 Non-musks: 63