Artificial Intelligence in Bioinformatics

Species Identification Using Statistical Principal Component Analysis on Genomic Sequence

Traditionally, species identification in Bioinformatics is performed using sequence alignment, which is a very compute intensive process. In this work we exploit the statistical similarities and uniqueness of the genomic sequences of different species to enable automatic identification of a species from its genome sequence with significantly less computation. A set of 64 three-tuple keywords is first generated using the four types of bases: A, T, C and G. These keywords are searched on N randomly sampled genome sequences, each of a given length (10,000 elements) and the frequency count for each of the 4^3 =64 keywords is obtained. Principal component analysis is then employed on the frequency counts for N sampled instances. The principal component analysis yields a unique feature descriptor for identifying the species from its genome sequence. The variance of the descriptors for a given genome sequence being negligible, the proposed scheme finds extensive applications in automatic species identification. Using this technique, given a genome sequence, instead of performing sequence alignment, a feature descriptor can be generated to identify the species.

Key Publications : LNCS (Springer) 2005, JCIS 2005

Topological Clustering of DNA-Descriptors by Self Organizing Feature Map (SOFM)

In this work we strive to map the DNA-descriptors on to a 2D array of neurons by the well-known Self-organizing Feature Map algorithm. Our main interest is to note whether DNA-descriptors of the same species occupy neighborhood neuronal positions and species having close resemblances in their DNA structure form neighborhood clusters. To verify the above, we considered 36 vectors of each of the following 3 species: Mouse,Yeast and E.coli. Naturally, we have 36 × 3=108 vectors to be mapped onto the 2 D array of size (k × k). To perform the experiment, we considered (6× 6) dimensional space for the 2 D array of neurons. Later the maps were created for different dimensions ranging from 4 to 11. After mapping all the 108 input data vectors onto the 2 D array of neurons, it is noted that data from the same species is mapped onto neurons occupying neighboring positions. Hence, it can be inferred that different vectors computed from different samples of the genomic data of a species are close in many respects and hence are mapped onto neighboring spaces on the map,thus forming separate clusters for different species. It can also be claimed that species which are close in characteristics will have similar DNA-descriptors and hence the clusters corresponding to similar species will lie in neighboring positions. Hence, if clustering techniques are applied to DNA-descriptors of a large number of species we will see that species which are similar in many respects e.g. Human and Gorilla will be forming sub-clusters within a super-cluster belonging to their families. Hence, such clustering techniques can help in species identification from genomic sequence.

Key Publications : IICAI 2005, Engineering Letters 2007