|Résumé||When we identify a gene and describe its function, we are making a shot in the dark, an informed guess about the relations between the gene and various biological processes, that is always incomplete and too often simply wrong. Applying human judgment to the question is time-consuming and error-prone. And it is biased, since we are naturally more inclined to spend time looking for the strange cases, or the odd mistake, than we are to carefully and systematically describe the "ordinary" cases. Manual genome curation, to paraphrase Churchill, is the worst form of genome annotation, except for all the others. What can we do?
The focus on our work to improve the speed and the quality of genome annotation is to understand that it is a dialogue, and that the goal in the asymptote is to reach a consensus. I will describe how we build web-based software tools to help scientists cooperate in finding, refining, and questioning this consensus. I will also show that we can see this incremental refinement as a smooth transition from unsupervised to supervised classification, that works hand in hand with comparative genomics. |