Techniques like X-ray crystallography [1] provide an exquisite picture of the average structure of stable conformations. But challenges remain for many systems including those unsuitable for crystallization, characterizing ensembles, states with low populations, and the transition pathways between these states. Fortunately, many experimental techniques such as Nuclear Magnetic Resonance (NMR), Förster Resonance Energy Transfer (FRET), Double Electron-Electron Resonance (DEER), Paramagnetic Relaxation Enhancements (PREs), chemical crosslinking, and others offer indirect structural information that can be obtained even in challenging situations [2,3].

When enough information is available (data-rich regime), the system is highly constrained, and the data typically define a narrow uncertainty ensemble [4], which is in some cases deposited in the protein data bank (PDB) [5]. Choices such as the computational method and how the data are modeled can affect this ensemble. Increasingly, however, researchers operate in the data-poor regime, where the system is only weakly constrained so that the data alone are insufficient to fully define the structural ensemble. By integrating data from multiple experiments, it is possible to narrow down the possible structures [6]. Even in such scenarios, distinguishing between multiple models requires a hybrid approach involving computational software to explore conformational space and identify possible solutions compatible with the data.

Recently, the wwPDB-dev [7] was developed to accept models originating from integrative/hybrid approaches. Whereas the PDB contains over 200,000 structures, the wwPDB-dev holds just 112 entries as of January 2023. The stark difference arises from difficulty in identifying pipelines to model these challenging systems and the need for better assessment tools to validate the quality of the models [8]. This review focuses on integrative/hybrid modeling in which molecular dynamic (MD) techniques are used to generate models. Excellent reviews focus on other aspects of integrative/hybrid modeling [1,6,9, 10, 11, 12].

MD is seeing a resurgence [13] thanks to improved force fields [14] and enhanced sampling techniques [15] leading to excellent agreement between experiments and simulation techniques [3]. Efforts to synergize with new sources of experimental data provide optimism for their role in integrative approaches [1,6,12]. However, as each experimental technique provides different outputs and uncertainties, there is no unique recipe for hybrid modeling [8,16]. For instance, density maps can be used to restrain residue/atomic positions [9,17], many techniques provide distance or orientation information between parts of the molecule [18,19], and others provide average or ensemble information [20]. MD samples ensembles of conformations that, when processed using statistical mechanics principles, provide insights about the relevant states of the system. Experimental data can be used to bias the ensemble, affecting the probability of sampling different regions in the energy landscape. The resulting ensembles can be processed to learn about states compatible with the data and physics model. Alternatively, data can be used with unbiased MD ensembles to select a subset of structures compatible with the experimental information. The challenges remain similar to other integrative pipelines: determining the number of states represented by the data, how to model uncertainties, or how to validate the ensembles.