An Energy Landscape Treatment of Decoy Selection in Template-Free Protein Structure Prediction

Abstract

Abstract:The energy landscape, which organizes microstates by energies, has shed light on many cellular processes governed by dynamic biological macromolecules leveraging their structural dynamics to regulate interactions with molecular partners. In particular, the protein energy landscape has been central to understanding the relationship between protein structure, dynamics, and function. The landscape view, however, remains underutilized in an important problem in protein modeling, decoy selection in template-free protein structure prediction. Given the amino-acid sequence of a protein, template-free methods compute thousands of structures, known as decoys, as part of an optimization process that seeks minima of an energy function. Selecting biologically-active/native structures from the computed decoys remains challenging. Research has shown that energy is an unreliable indicator of nativeness. In this paper, we advocate that, while comparison of energies is not informative for structures that already populate minima of an energy function, the landscape view exposes the overall organization of generated decoys. As we demonstrate, such organization highlights macrostates that contain native decoys. We present two different computational approaches to extracting such organization and demonstrate through the presented findings that a landscape-driven treatment is promising in furthering research on decoy selection.

Deep Ranking in Template-free Protein Structure Prediction

“…The latter models and accounts for the interaction energy among the atoms in a given tertiary structure. It is now well-known that scoring/energy functions are inherently inaccurate [2,11,35]. They often drive the exploration of a protein structure space to local minima that contain structures very different from a known native structure.…”

Reducing Ensembles of Protein Tertiary Structures Generated De Novo via Clustering

“…Structure generation algorithms address an optimization problem; they seek tertiary structures that minimize the interaction energy among the atoms of a given target protein. It is now well-known that the energy functions designed in computational laboratories are inherently inaccurate [7][8][9]. In particular, one cannot infer that a lower-energy structure is more similar to the sought native structure than a higher-energy structure.…”

Unsupervised and Supervised Learning over the Energy Landscape for Protein Decoy Selection

“…In this article, we first summarize some of our recent successes in this direction via unsupervised learning. We relate how by leveraging the concept of basins in a methodology that identifies and ranks basins in the energy landscape comprised of thousands of decoys exposes basins rich in near-native decoys [14,15]. We show that utilizing energies yields a distinct, quantifiable improvement over a complementary method that builds over clustering of decoys while ignoring energies [16].…”