There are a number of teams who regularly participate in CASP. Such teams also participate in the disorder region prediction category that Noguchi participated in. "The teams you could call 'regulars' are two overseas teams. They continue to publish the results of their research that give them the reputation for being the two major groups in this field". There were also a number of teams from Japan. Teams such as those from CBRC and the University of Tokyo participated at CASP7. Noguchi explains, "The two major issues for us were just how closely the Japanese teams could compete with the two pioneering overseas teams, and how close to the top of the table the Japanese teams could rank."
Noguchi's research team was "a crack unit." The main staffs of his team were two excellent researchers. First one was Dr. Shimizu, who was a technical staff. She is a graduate of the Waseda University Faculty of Science and Engineering. The other was Mr. Hirose, who was a collaborator from Pharma Design, Inc. However, to compete with the formidable caliber o£æ teams participating in CASP and achieve a certain degree of success, the type of approach followed up to CASP6 would be insufficient. Within the research team, day after day they continued to consider a "ground design" required for the prediction program framework.
In 2005, the year after CASP6, they began to see that it was possible to improve the prediction accuracy, and decided on a guideline for development of a disorder region prediction method.
Noguchi's team conceived the following strategy. Firstly, they decided to think of disordered regions in terms of "long" and "short". Based on this, they would design a suitable machine learning method mounted on the program for each prediction. Finally, they predicted the disorder regions by the programs. "Although they are called 'disorder regions', the basis for our strategy was that properties differed depending on their length. If their properties differed, then the prediction programs optimized for each type should be developed. We thought that a prediction method with high accuracy could be developed by focusing on the part which had to be learned, and by making up a suitable machine learning method optimized for each type."
The superior point for Noguchi's team was exactly this strategy to assign a suitable machine learning method for each; in other words, to focus on the target one by one. Conventional prediction method for disorder regions uses the technique to delimit the sequence to "Window", which was a set of several residues, and then to predict the center residue in each "Window" in general. On the other hand, Noguchi's team uses several ideas for the prediction for disorder regions. "For instance, in the case of the prediction for long disorder region, they predict whether the whole Window's residues are disorder or not at the first step. This prediction step by sliding the Window continues at the end of the sequence. Next, they predict whether a residue is disorder or not, by using the results based on the prediction of each Window at the fist step. "
Once the entire strategy has been formulated, all that is left is the work. Noguchi reflects that, "Thanks to the fact that we had outstanding team members, we were able to develop the programs in a short time." From the time the process was started, it took half a year to create three versions of the program: "Poodle-L" for predicting long disorder regions; "Poodle-S" for predicting short disorder regions; and "Poodle-W", for predicting whether a whole sequence is disorder or not. Noguchi's research team carried out further research in order to improve the accuracy, and got preparations ready for CASP7.