Due both to the rapid growth of genomic data and the inherent complexity and variability of its biological significance, automated pattern recognition is critical for efficiently identifying structural and functional relationships in the genomic databases that can be used to develop effective disease therapies, drugs, and other biotechnology products. However, current analytic techniques for pattern analysis are limited in their ability to identify significant structural and functional relationships, since interpretation of DNA/RNA patterns is required simultaneously on several levels: linear sequence of elements, structural interactions between distant elements, structural motifs, and function. Each of these levels interacts with other levels in complex ways, with especially strong linkages in the sequence/structure domains. Initial efforts with neural networks have been encouraging, but fundamental extensions of network-based techniques are needed. In particular: (1) the interactions between distant elements of a sequence require that networks have a form of """"""""memory""""""""; (2) the very large number of different pattern classes requires a form of adaptive encoding of the underlying sequential information; (3) the multiple levels of interaction of the structural and functional information require the use of more complex architectures involving multiple, interacting networks; and (4) the broad range of pattern sizes and degree of allowed variability within different pattern types require that information presented to the pattern-recognition algorithm be more robust than simple sequence data. The proposed research addresses each of these concerns by the use of modified back-propagation algorithms that allow adaptive sequence encoding and maintenance of sequence memory, by the use of coupled networks, and by the use of both linear and quadratic transforms of the raw sequence data.
