UNLABELLED

Gene prioritization refers to a family of computational techniques for inferring disease genes through a set of training genes and carefully chosen similarity criteria. Test genes are scored based on their average similarity to the training set, and the rankings of genes under various similarity criteria are aggregated via statistical methods. The contributions of our work are threefold: (i) first, based on the realization that there is no unique way to define an optimal aggregate for rankings, we investigate the predictive quality of a number of new aggregation methods and known fusion techniques from machine learning and social choice theory. Within this context, we quantify the influence of the number of training genes and similarity criteria on the diagnostic quality of the aggregate and perform in-depth cross-validation studies; (ii) second, we propose a new approach to genomic data aggregation, termed HyDRA (Hybrid Distance-score Rank Aggregation), which combines the advantages of score-based and combinatorial aggregation techniques. We also propose incorporating a new top-versus-bottom (TvB) weighting feature into the hybrid schemes. The TvB feature ensures that aggregates are more reliable at the top of the list, rather than at the bottom, since only top candidates are tested experimentally; (iii) third, we propose an iterative procedure for gene discovery that operates via successful augmentation of the set of training genes by genes discovered in previous rounds, checked for consistency.

MOTIVATION

Fundamental results from social choice theory, political and computer sciences, and statistics have shown that there exists no consistent, fair and unique way to aggregate rankings. Instead, one has to decide on an aggregation approach using predefined set of desirable properties for the aggregate. The aggregation methods fall into two categories, score- and distance-based approaches, each of which has its own drawbacks and advantages. This work is motivated by the observation that merging these two techniques in a computationally efficient manner, and by incorporating additional constraints, one can ensure that the predictive quality of the resulting aggregation algorithm is very high.

RESULTS

We tested HyDRA on a number of gene sets, including autism, breast cancer, colorectal cancer, endometriosis, ischaemic stroke, leukemia, lymphoma and osteoarthritis. Furthermore, we performed iterative gene discovery for glioblastoma, meningioma and breast cancer, using a sequentially augmented list of training genes related to the Turcot syndrome, Li-Fraumeni condition and other diseases. The methods outperform state-of-the-art software tools such as ToppGene and Endeavour. Despite this finding, we recommend as best practice to take the union of top-ranked items produced by different methods for the final aggregated list.

AVAILABILITY AND IMPLEMENTATION

The HyDRA software may be downloaded from: http://web.engr.illinois.edu/∼mkim158/HyDRA.zip.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.