1997 Fiscal Year Final Research Report Summary
Improvement of Artificial Neural Networks and Its Applications to QSARs.
| Project/Area Number |
08672476
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Physical pharmacy
|
|---|
| Research Institution | Osaka University |
|---|
Principal Investigator |
TAKAGI Tatsuya Osaka University, Genome Information Research Center, Instructor, 遺伝情報実験施設, 講師 (80144517)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
FUJIWARA Hideaki Osaka University, Medical School, Professor, 医学部, 教授 (90107102)
|
|---|
| Project Period (FY) |
1996 – 1997
|
| Keywords | QSAR / Neural Networks / Robust Estimation / Computer Statistics / Shift Test Method |
|---|
| Research Abstract |
In this year, we carried out the quantitative structure activity relationship analyzes of the 2,4-diamino-6,6-dimethyl-5-phynylhydrotriazine derivatives, which show the inhibition activities for dihydrofolatereductase using the previous results about the robust artificial neural network method.
Firstly, we adjusted the number of neurons in the hidden layr by extended shift test method shifting the teacher signals.
The results show that the best number of the neurons in the hidden layr was 25-30 because the good relations between the parent point and the background points were found by the extended shift test method. Therefore, we used the number of neurons in hidden layrs through out this project. The extended shift test method for the input descriptors (pi_2, pi_3, pi_4, MR_2, MR_3, MR_4, SIGMAsigma_<3,4>) showed that there was no deleted one.
We used the two kinds of robust artificial neural network techniques.
1) Firstly, the back propagation learning were carried out using the 90% linearity, and then, the 38 data which showed the large errors were deteted from the learning data. Finally, normal artificial neural networks (0% linearity) were adopted.
2) The linearity was stepwisely decreased from 80% to 0% and the weight, which showed the large errors, was decreased in each step.
In the case of 1), the residual sum of squares resulted in E=23.27. The predictivity of the artificial neural networks was remarkably improved compared with E=30.01 when not using the robust techniques. And in the case of 2), the final root mean residual sum of squares, Ep (=0.49) showed the remarkable improvement compared with Ep=0.69 when not using the robust artificial neural networks.
Otherwise, we carried out the classifying of bioactive chemical substances using livingstone-type 5-layred artificial neural networks and the good results were obtained. These results were reported in the Synposium on Chemical Information and Computer Sciences at Kumamoto, 1997.
|
Research Products
(4 results)
