Algorithm Selection for Protein Structure Prediction on 2D AB Off-lattice Model
The present study performs Algorithm Selection for Protein Structure Prediction. The idea is to automatically determine the best algorithm for each given instance through performance prediction. The Protein Structure Prediction problem is concerned with exploring the 3D shape of a protein from any given amino acid sequence. This task is critical as the proteins’ 3D structures can reflect their behaviour. There are already different algorithmic solutions for this problem. However, there is no algorithm that can effectively address varying problem instances, under fair experimental conditions. Focusing on this gap, this work offers a simple yet practical approach benefiting from the existing algorithms, working on the 2D AB off-lattice model representation. A simple set of problem instance features is introduced for performing the selection operations. Besides these cheap and straightforward instance attributes, a suite of existing fitness landscape analysis measures are adopted. The experiments for evaluating the proposed approach is operated with 6 algorithms across 65 problem instances. The resulting analysis showed that Algorithm Selection can successfully specify algorithms in a per-instance basis for PSP, even with simple sequence features. In relation to that, the characteristics affecting the algorithms’ performance are reported. This analysis can show why certain algorithms struggle to solve specific instances while they are doing great on some others.
