Protein Decoy Generation via Adaptive Stochastic Optimization for Protein Structure Determination

Many regions of the protein universe remain inaccessible by wet-laboratory or homology modeling methods. Elucidating these regions necessitates structure determination in silico. Protein structure determination in the absence of a structural template remains a challenging task with two core problems, known as decoy generation and decoy selection. In this paper, we address the problem of decoy generation, which inherently involves exploring the unknown, vast, and high-dimensional structure space of a given amino-acid sequence in the presence of a finite computational budget for relevant structures. Leveraging a stochastic optimization framework, we first demonstrate how selection pressure can be employed to control the trade-off between exploration and exploitation. Moreover, we then propose a novel algorithm that tunes its behavior towards exploration or exploitation as needed via an adaptive selection mechanism. We present a thorough evaluation on 30 protein targets in a comparative setting, where we compare the proposed adaptive algorithm to state-of-the-art algorithms that include the top ten groups in the two recent CASP competitions. The results show that the proposed algorithm is not only competitive against several of these groups, but it additionally outperforms several of them on many targets, suggesting that adaptive stochastic optimization is a promising framework for decoy generation.