mRNA Signalstructure Detection by RIBL
(GMD)

Application domain:	Computational Biology
Source:	Center for Computing Technology (TZI); University of Bremen
Dataset size:	57 signalstructures (618 ground facts)
Data format:	Prolog
Systems Used:	RIBL
Pointers:	Contact Tamás Horváth (tamas.horvath@gmd.de)

The data (GMD)

mRNA is a sequence of nucleotic acids (guanine, adenine, uracil, cytosine). The secondary structure of an mRNA contains special subsequences called signalstructures that are responsible for special biological functions e.g., RNA-protein interactions, poly-A processing, cellular transport etc. For example the signalstructure Iron Responsive Element (IRE) is responsible for

• binding of IRE binding protein (aconitase),
• iron depending control of translation and
• cellular switch between iron uptake and iron storage.

Signalstructures have structural motifes that are composed of different building blocks as e.g., stacking region, hairpin loops, interiror loops etc. There are similar signalstructures that have the same biological function. For example 15 different IRE signalstructures have been described by now. Their similarity roots in a common structure called consensus pattern. The aims are to find

1. new signalstructures of the known classes,
2. new signalstructure classes.

RIBL (Emde and Wettschereck, 96) is a relational instance based learning system developed by GMD. RIBL applies the $k$-nearest neighbor method in a relational representation. As a first step, RIBL collects literals that describe an object. This is a recursive process in which the depth is bounded by a user defined parameter. In the second step, RIBL computes the similarity between objects. Similarity between objects is computed recursively, similarity between attributes is defined same as for the propositional case. Finally, the $k$ most similar instances vote on the class of a query. The vote is weighted by similarity.

We have extended RIBL by a new type for representing sequences. We have also introduced a similarity measure between sequences. The base of this similarity measure is the weighted edit distance. The next four target predicates have been used in this application: ire/3, tar/3, secis₁/3 and secis₂/3.

The fact ire(s_ire₁) means, that the class of the signalstructure s_ire₁ is IRE, it has 6 structure elements and the length of the sequence is 34. The background predicates hairpin/3, stem/4, bulge_left/3 and bulge_right/3 represent the structure elements of the signalstructures. For example the fact hairpin(s_ire₁, 1, [c,a,g,t,g,t]) denotes that the first structure element of the signalstructure s_ire₁ is a hairpin loop with the sequence cagtgt.

Experiments with RIBL (GMD)

The signalstructure database that is currently available for us contains only the above four signalstructure classes and 57 signalstructures with known classes. RIBL has 54 signalstructures correctly classified. The 1 and the 2 nearest neighbors have been the best to classify the queries.

back to index