Selection-based Per-Instance Heuristic Generation for Protein Structure Prediction of 2D HP Model

Abstract

The present study aims at generating heuristics for Protein Structure Prediction represented in the 2D HP model. Protein Structure Prediction is about determining the 3-dimensional form of a protein from a given amino acid sequence. The resulting structure directly relates to the functionalities of the protein. There are a wide range of algorithms to address Protein Structure Prediction as an optimization problem. Being said that there is no an ultimate algorithm that can effectively solve PSP under varying experimental settings. Hyper-heuristics can offer a solution as high-level, problem-independent search and optimization strategies. Selection Hyper-heuristics operate on given heuristic sets that directly work on the solution space. One group of Selection Hyper-heuristics focus on automatically specify the best heuristics on-the-fly. Yet, the candidate heuristics tend to be decided, preferably a domain expert. Generation Hyper-heuristics approach differently as aiming to generate such heuristics automatically. This work introduces a automated heuristic generation strategy supporting Selection Hyper-heuristics. The generation task is formulated as a selection problem, disclosing the best expected heuristic specifically f or a given problem instance. The heuristic generation process is established as a parameter configuration problem. T he corresponding system is devised by initially generating a training data alongside with a set of basic features characterizing the Protein Structure Prediction problem instances. The data is generated discretizing the parameter configuration space o f a single heuristic. The resulting data is used to predict the best configuration of a specific heuristic used in a heuristic set under Selection Hyper-heuristics. The prediction is performed separately for each instance rather than using one setting for all the instances. The empirical analysis showed that the proposed idea offers both better and robust performance on 22 PSP instances compared to the one-for-all heuristic sets. Additional analysis linked to the selection method, ALORS, revealed insights on what makes the PSP instances hard / easy while providing dis/-similarity analysis between the candidate configurations.