**1. Introduction**

There is a growing body of evidence that some proteins act in multiple structural states [1]. It has been demonstrated that the ability of these proteins to switch between distinct structural states may be crucial for their function and regulation [1]. Additionally, a number of key biological functions have been proven to be performed by disordered or partially unstructured proteins [2]. Some proteins fold and obtain their structure only upon binding to their partners, while others form so called "fuzzy complexes" in which both proteins retain a certain degree of disorder [3]. These discoveries modified the core biochemistry principle of "structure determines function". As for now, a consensus has been reached that protein function may be a result of an interplay between protein structure and its dynamics [4,5].

Internal protein motions may be studied both experimentally and with computational methods [6,7]. For example, nuclear magnetic resonance (NMR) spectroscopy is one of the richest sources of information on protein structure and dynamics, especially when accompanied with assisting methods that enhance resolution or provide an additional insight into the dynamics of structures [8]. This approach, however, results in an averaged image of the structural ensemble.

A variety of computational techniques have been developed to assist these challenging experimental studies [7,9]. In the last decades, molecular modeling was dominated by structure-based models or Go-like models (approaches that are biased toward known folded conformations [10,11]). These indeed lead to significant speedup of simulations but may result for example in an unrealistic picture of protein folding, which in reality may also depend on non-native interactions [12–14].

Recent works show that methods combining experimental data and computational approaches may produce the most promising pictures of protein equilibrium dynamics [15,16]. However, the development of these methods poses a number of challenges—both in terms of the validity of the approach and its computationally efficient implementation [17].

Molecular dynamics (MD) has been so far the most widespread computational method for the investigation of protein motions [18]. However, standard all-atom MD implementations are limited to sub-microsecond timescales and may suffer from limited sampling despite recent significant advances in code optimization and hardware [19]. To overcome this problem various MD extensions have been proposed. These extensions include for example replica-exchange MD, meta-dynamics, Markov state models and simulated annealing algorithms [6,20–23].

A number of non-MD sampling methods have also been developed to provide a comprehensive image of protein dynamics using limited computational resources. Of these, Monte Carlo (MC) is perhaps the most commonly used and generally applicable sampling method [11]. Monte Carlo randomly generates conformations and uses an energy-based acceptance criterion that promotes pseudo-trajectory convergence to an energetic minimum. On the expense of losing a direct image of the timescales or kinetics of the ensemble, MC manages to overcome some of the major limitations of MD [24].

Aside from the sampling method, a further extension of effective timescales is possible by using a simplified representation of protein structures to reduce the number of a system's degrees of freedom. The accuracy of the available coarse-grained (CG) models may vary from detailed, almost atomistic representations (Primo [25], Rosetta [26]), medium resolution models (in which a single amino acid is represented by three to five beads: UNRES [27], CABS [28], AWSEM [29], MARTINI [30], PaLaCe [31]), and Scorpion [32]) to significantly simplified models like SURPASS [33,34]. Applications and implementations of these and other CG models are described in detail in a recent review [11].

In addition to the representation and sampling method, the choice of the force field to perform the simulation determines the success of modeling. Traditionally, force fields are divided into two main groups: physics-based, which involve (usually pairwise) interaction terms [35], and those employing a statistical approach; however, most of the successful approaches are usually a mixture of the two. A statistical force field is constructed using the probability of a chosen observable (or a set of observables) in a given ensemble of structures [36]. Early attempts focused on straightforward pairwise contacts [37]; however, with further development, more complex observables were analyzed. This resulted in a generation of knowledge-based force fields, or scores, for various representations, coarse-grained and all-atom: CABS [28], Rosetta [38], DOPE [39], GOAP [40], QUARK [41], Bcl::Score [42] or BACH [36]. Newly developed approaches go a step further and improve the results by combining these methods with experimental data [43,44]. An example of such approach is RosettaEPR [45], which includes distance data from site-directed spin labeling electron paramagnetic resonance experiments. It is generally agreed that statistical force fields frequently allow more accurate scoring than physics-based potentials [11]. The combination of knowledge-based force fields or scores with effective sampling schemes seems to be a promising approach to a number of problems [11], such as protein structure prediction [43,44,46,47], investigation of protein interactions [48] or studies of protein dynamics [17,49–51].

This review briefly describes one of these approaches: an MC-based and knowledge-based interaction scheme for modeling protein–peptide interactions and unfolded states of globular proteins using the CABS coarse-grained protein model. Firstly, the main features of the CABS method will be described, with a focus on their applicability for modeling disordered or unfolded proteins or their fragments. Subsequently, representative case studies will be discussed to provide detailed insights into the modeling results obtained for systems characterized by a varying level of disorder.
