SOMMER: self-organising maps for education and research.

SOMMER is a publicly available, Java-based toolbox for training and visualizing two- and three-dimensional unsupervised self-organizing maps (SOMs). Various map topologies are implemented for planar rectangular, toroidal, cubic-surface and spherical projections. The software allows for visualization of the training process, which has been shown to be particularly valuable for teaching purposes.

Analysis of activity space by fragment fingerprints, 2D descriptors, and multitarget dependent transformation of 2D descriptors.

The effect of multitarget dependent descriptor transformation on classification performance is explored in this work. To this end decision trees as well as neural net QSAR in combination with PLS were applied to predict the activity class of 5HT3 ligands, angiotensin converting enzyme inhibitors, 3-hydroxyl-3-methyl glutaryl coenzyme A reductase inhibitors, platelet activating factor antagonists, and thromboxane A2 antagonists. Physicochemical descriptors calculated by MOE and fragment-based descriptors (MOLPRINT 2D) were employed to generate descriptor vectors. In a subsequent step the physicochemical descriptor vectors were transformed to a lower dimensional space using multitarget dependent descriptor transformation. Cross-validation of the original physicochemical descriptors in combination with decision trees and neural net QSAR as well as cross-validation of PLS multitarget transformed descriptors with neural net QSAR were performed. For comparison this was repeated using fragment-based descriptors in combination with decision trees.

Multi-space classification for predicting GPCR-ligands.

A classification of molecules depends on the descriptor set which is used to represent the compounds, and each descriptor could be regarded as one perception of a molecule. In this study we show that a combination of several classifiers that are grounded on separate descriptor sets can be superior to a single classifier that was built using all available descriptors. The task of predicting ligands of G-protein coupled receptors (GPCR) served as an example application. The perceptron, multilayer neural networks, and radial basis function (RBF) networks were employed for prediction. We developed classifiers with and without descriptor selection. Prediction accuracy was assessed by the area under the receiver operating characteristic (ROC) curve. In the case with descriptor selection both the selection and the rank order of the descriptors depended on the type and topology of the neural networks. We demonstrate that the overall prediction accuracy of the system can be improved by joining neural network classifiers of different type and topology using a "jury network" that is trained to evaluate the predictions from the individual classifiers. Seventy-one percent correct prediction of GPCR ligands was obtained.

Identification of noncanonical melanoma-associated T cell epitopes for cancer immunotherapy.

The identification of tumor-associated T cell epitopes has contributed significantly to the understanding of the interrelationship of tumor and immune system and is instrumental in the development of therapeutic vaccines for the treatment of cancer. Most of the known epitopes have been identified with prediction algorithms that compute the potential capacity of a peptide to bind to HLA class I molecules. However, naturally expressed T cell epitopes need not necessarily be strong HLA binders. To overcome this limitation of the available prediction algorithms we established a strategy for the identification of T cell epitopes that include suboptimal HLA binders. To this end, an artificial neural network was developed that predicts HLA-binding peptides in protein sequences by taking the entire sequence context into consideration rather than computing the sum of the contribution of the individual amino acids. Using this algorithm, we predicted seven HLA A*0201-restricted potential T cell epitopes from known melanoma-associated Ags that do not conform to the canonical anchor motif for this HLA molecule. All seven epitopes were validated as T cell epitopes and three as naturally processed by melanoma tumor cells. T cells for four of the new epitopes were found at elevated frequencies in the peripheral blood of melanoma patients. Modification of the peptides to the canonical sequence motifs led to improved HLA binding and to improved capacity to stimulate T cells.

Impact of different software implementations on the performance of the Maxmin method for diverse subset selection.

Besides the choice of an automated software method for selecting 'maximally diverse' compounds from a large pool of molecules, it is the implementation of the algorithm that critically determines the usefulness of the approach. The speed of execution of two implementations of the Maxmin algorithm is compared for the selection of maximally diverse subsets of large compound collections. Different versions of the software are compared using various C compiler options and Java virtual machines. The analysis shows that the Maxmin algorithm can be implemented in both languages yielding sufficient speed of execution. For large compound libraries the Java version outperformes the C version. While the Java version selects the same compounds independent of the virtual machine used, the C version produces slightly different subsets depending on the compiler and on the optimization settings.

Impact of descriptor vector scaling on the classification of drugs and nondrugs with artificial neural networks.

The influence of preprocessing of molecular descriptor vectors for solving classification tasks was analyzed for drug/nondrug classification by artificial neural networks. Molecular properties were used to form descriptor vectors. Two types of neural networks were used, supervised multilayer neural nets trained with the back-propagation algorithm, and unsupervised self-organizing maps (Kohonen maps). Data were preprocessed by logistic scaling and histogram equalization. For both types of neural networks, the preprocessing step significantly improved classification compared to nonstandardized data. Classification accuracy was measured as prediction mean square error and Matthews correlation coefficient in the case of supervised learning, and quantization error in the case of unsupervised learning. The results demonstrate that appropriate data preprocessing is an essential step in solving classification tasks.