Human-Computer Interaction in a Computational Evolution System for the Genetic Analysis of Cancer

Abstract

The paradigm of identifying genetic risk factors forcommon human diseases by analysing one DNA sequencevariation at a time is quickly being replaced byresearch strategies that embrace the multivariatecomplexity of the genotype to phenotype mappingrelationship that is likely due, in part, to nonlinearinteractions among many genetic and environmentalfactors. Embracing the complexity of common diseasessuch as cancer requires powerful computational methodsthat are able to model nonlinear interactions inhigh-dimensional genetic data. Previously, we haveaddressed this challenge with the development of acomputational evolution system (CES) that incorporatesgreater biological realism than traditional artificialevolution methods, such as genetic programming. Ourresults have demonstrated that CES is capable ofefficiently navigating these large and rugged fitnesslandscapes toward the discovery of biologicallymeaningful genetic models of disease predisposition.Further, we have shown that the efficacy of CES isimproved dramatically when the system is provided withstatistical expert knowledge, derived from a family ofmachine learning techniques known as Relief, orbiological expert knowledge, derived from sources suchas protein-protein interaction databases. The goal ofthe present study was to apply CES to the geneticanalysis of prostate cancer aggressiveness in a largesample of European Americans. We introduce here the useof 3D visualization methods to identify interestingpatterns in CES results. Information extracted from thevisualization through human-computer interaction arethen provide as expert knowledge to new CES runs in acascading framework. We present a CES-derivedmultivariate classifier and provide a statistical andbiological interpretation in the context of prostatecancer prediction. The incorporation of human-computerinteraction into CES provides a first step towards aninteractive discovery system where the experts can beembedded in the computational discovery process. Ourworking hypothesis is that this type of human-computerinteraction will provide more useful results forcomplex problem solving than the traditional black boxmachine learning approach.