Neural-network design applied to protein-secondary-structure predictions
- Life Sciences Division, Lawrence Berkeley Laboratories, Berkeley, California 94720 (United States)
The success of neural networks is often limited by a sparse database of training examples, deficient neural-network architectures, and nonglobal optimization of the network variables. The convolution of these three problems has curtailed the application of network models to protein-structure predictions, where homology modeling or information theory approaches are considered better controlled alternatives. This paper introduces our broad objective of disentangling the three degrading features of neural networks cited above, beginning with improved designs of network architectures used in the prediction of protein secondary structure. This work demonstrates that network architecture design considerations greatly improve generalization and more efficiently extract complex sequence-structure relationships from the existing database, as compared to arbitrary architectures with the same size input window.
- Research Organization:
- Lawrence Berkeley National Lab. (LBNL), Berkeley, CA (United States)
- DOE Contract Number:
- AC03-76SF00098
- OSTI ID:
- 27852
- Journal Information:
- Physical Review. E, Statistical Physics, Plasmas, Fluids, and Related Interdisciplinary Topics, Vol. 51, Issue 4; Other Information: PBD: Apr 1995
- Country of Publication:
- United States
- Language:
- English
Similar Records
Combination of deep neural network with attention mechanism enhances the explainability of protein contact prediction
Pose Classification Using Three-Dimensional Atomic Structure-Based Neural Networks Applied to Ion Channel–Ligand Docking