Epistasis analysis is biologist’s principal tool for reconstruction of genetic networks from experimental data. Normally applied to qualitative phenotypes, we have proved that epistasis can be inferred from global transcriptional phenotypes by automated abductive reasoning in Artificial Intelligence. Our work demonstrates that microarray data can provide a uniform, quantitative tool for large-scale genetic network analysis. The paper was featured in reviews by Tim Hughes (in Nature Genetics) and Orli Bahcall (in Nature Reviews Genetics).
We are developing an open-source data mining framework called Orange (www.ailab.si/orange), which is probably the most comprehensive application of its kind with an interface to popular programming language Python. Together with our international collaborators, Orange was extended to provide modules for data analysis in biology and functional genomics. The system enables users who are not programmers to manage microarray and genomic data flow and to customize their analysis by combining common data analysis tools to fit their needs.
In collaboration with Baylor College of Medicine, Houston, we have developed methods which can infer gene networks from experimental data on mutants (GenePath system, www.genepath.org), using AI techniques and abductive reasoning. The method also proposes experiments which can refine the discovered network. GenePath was reviewed in Science 302(5647), 2003, and enlisted in its NetWatch, Best of the Web directory. Subsequent improvements of the method have been published in Artificial Intelligence in Medicine (2003) in Nucleic Acid Research (2005).
Almost all game-playing programs are based on the minimax algorithm. In practice, decisions based on deeper minimax search are better than those based on shallower search. Surprisingly, the first attempts at analysing this phenomenon showed the opposite: shallower search outperform deeper one. This phenomenon was termed the minimax pathology. Several explanations for the pathology have been proposed, but none of them was general. We developed a minimax model in which position values are real-valued. This model turned out not to be pathological for a wide range of parameter settings.
Argument-based machine learning (ABML) is an extension of machine learning with some concepts from the field of argumentation. ABML allows domain experts to articulate their background knowledge in a convenient way. It was successfully applied to several completely different domains: in a legal application we modeled the social security benefits (published in Artificial intelligence and law journal, 2005), in a medical application we improved a model for bacterial infections, and in game playing it was used for automated knowledge acquisition.