We report an improved unsupervised method for cancer classification by the use of gene-expression profile via sparse non-negative matrix factorization. We demonstrate the improvement by direct comparison with classic non-negative matrix factorization on the three well-studied datasets. In addition, we illustrate how to identify a small subset of co-expressed genes that may be directly involved in cancer.

We use methods from Data Mining and Knowledge Discovery to design an algorithm for detecting motifs in protein sequences. The algorithm assumes that a motif is constituted by the presence of a "good" combination of residues in appropriate locations of the motif. The algorithm attempts to compile such good combinations into a "pattern dictionary" by processing an aligned training set of protein sequences. The dictionary is subsequently used to detect motifs in new protein sequences. Statistical significance of the detection results are ensured by statistically determining the various parameters of the algorithm. Based on this approach, we have implemented a program called GYM. The Helix-Turn-Helix motif was used as a model system on which to test our program. The program was also extended to detect Homeodomain motifs. The detection results for the two motifs compare favorably with existing programs. In addition, the GYM program provides a lot of useful information about a given protein sequence.

In the past few years, pattern discovery has been emerging as a generic tool of choice for tackling problems from the computational biology domain. In this presentation, and after defining the problem in its generality, we review some of the algorithms that have appeared in the literature and describe several applications of pattern discovery to problems from computational biology.

Using TEIRESIAS, a pattern discovery method that identifies all motifs present in any given set of protein sequences without requiring alignment or explicit enumeration of the solution space, we have explored the GenPept sequence database and built a dictionary of all sequence patterns with two or more instances. The entries of this dictionary, henceforth named seqlets, cover 98.12% of all amino acid positions in the input database and in essence provide a comprehensive finite set of descriptors for protein sequence space. As such, seqlets can be effectively used to describe almost every naturally occurring protein. In fact, seqlets can be thought of as building blocks of protein molecules that are a necessary (but not sufficient) condition for function or family equivalence memberships. Thus, seqlets can either define conserved family signatures or cut across molecular families and previously undetected sequence signals deriving from functional convergence. Moreover, we show that seqlets also can capture structurally conserved motifs. The availability of a dictionary of seqlets that has been derived in such an unsupervised, hierarchical manner is generating new opportunities for addressing problems that range from reliable classification and the correlation of sequence fragments with functional categories to faster and sensitive engines for homology searches, evolutionary studies, and protein structure prediction. Proteins 1999;37:264-277.