Atomistic modeling of protein-DNA interaction specificity: progress and applications

Abstract

An accurate, predictive understanding of protein-DNA binding specificity is crucial for the successful design and engineering of novel protein-DNA binding complexes. In this review, we summarize recent studies that use atomistic representations of interfaces to predict protein-DNA binding specificity computationally. Although methods with limited structural flexibility have proven successful at recapitulating consensus binding sequences from wild-type complex structures, conformational flexibility is likely important for design and template-based modeling, where non-native conformations need to be sampled and accurately scored. A successful application of such computational modeling techniques in the construction of the TAL-DNA complex structure is discussed. With continued improvements in energy functions, solvation models, and conformational sampling, we are optimistic that reliable and large-scale protein-DNA binding prediction and engineering is a goal within reach.

Introduction

Sequence-specific interactions between proteins and nucleic acids play a central role in a wide variety of cellular processes, including transcriptional and translational regulation, DNA replication, and DNA damage repair. The ability to accurately predict and rationally engineer the specificity of these interactions would have a transformative impact on biology and medicine. Here we review recent progress in molecular modeling of protein-DNA interactions, focusing on algorithms for prediction of DNA-binding specificity. In addition to its intrinsic importance for unraveling regulatory interactions, binding specificity prediction provides critical feedback for structure-based design of protein-DNA interactions: indeed, we suggest that a major barrier to successful redesign of DNA binding specificity is the difficulty in modeling the changes in interface structure and/or specificity induced by designed mutations (Fig. 1). Given the polar nature of protein-DNA interfaces, which often feature specific water-bridged interactions, binding specificity prediction is also a stringent test of electrostatics models and representations of aqueous solvation.

Figure 1. The importance of flexibility in protein-DNA modeling.

(A) Shifts in DNA bases (yellow arrows) are evident in template-based modeling of one zinc finger (the target, shown in red and pink) using another (the template, shown in blue). Interfaces are superimposed based on their protein components. Protein sidechains in the target whose DNA contacts would be disrupted by these DNA shifts are shown in stick representation (example taken from [7]). (B) Superposition of design model (pink) and experimental structure (blue) for the I-MsoI ‘-7C’ design of Ashworth et al. [47] reveals a shift of 2.9 Å in the DNA backbone and a 180° flip of a designed Tryptophan residue. These examples suggest that incorporation of backbone conformational flexibility may be necessary for reliable prediction and design of protein-DNA interactions.

Structure-based prediction of protein-DNA binding specificity

Prediction of the DNA binding specificity of a protein of interest (Fig. 2) typically begins with a structure of the protein bound to a DNA target site. Models of the protein bound to alternate sites (often all single base-pair mutants of the target site) are built from this structure, and relative binding affinities are computed for these sites by comparing the energies of the starting and mutant complexes under a suitable energy function. For benchmarking purposes, these predicted binding affinities may be compared with corresponding experimental values, or a position weight matrix (PWM) model can be constructed from the calculated energies and compared with one estimated from known target sites or high-throughput experimentation [1-3, 4*].

Figure 2. Flowchart of structure-based calculation of the protein-DNA binding specificity.

Step 1. A protein-DNA complex structure, taken from either experiments (X-ray or NMR) or molecular modeling (homology modeling or docking), serves as the input of the method. Here the Zif268 zinc finger protein is shown on the left in cartoon representation as an example (PDB ID: 1AAY). Step 2. The DNA base-pairs in the target site are mutated, perhaps exhaustively to all possible sequences (4^L sequences for a site of length L), or more commonly to all single-base mutants (3L sequences). A superposition of four possible base pairs at a given position is shown in stick representation on the left. Step 3. The protein-DNA binding energy change caused by the DNA sequence mutation(s) is evaluated, using either a knowledge-based potential or a molecular mechanics force field. For methods that use molecular mechanics force fields, the binding energy difference is typically evaluated using the thermodynamic cycle shown on the left. Here the protein-DNA binding energy difference caused by the DNA sequence mutation (ΔΔG) is obtained by subtracting the unbound DNA energy (ΔG_U) from the bound protein-DNA energy (ΔG_B). Step 4. When single base pair mutations are performed, based on the approximation of DNA position independence, the binding energies of the different complexes can be converted into a position weight matrix and represented as a sequence logo [48] to show the binding specificity graphically (such as the logo shown on the left). When the DNA sequence space is exhaustively searched, a collection of high-affinity DNA sequences (such as those with a predicted binding affinity within some cutoff value of the optimal sequence) can also be represented as a sequence logo, although it should be noted that these sequences contain information regarding inter-position correlation that can be represented by higher-order graphical representations [49].

Classification scheme

Here we review a representative subset of recently described approaches to structure-based specificity prediction, organized according to the energy functions and conformational sampling approaches employed. To aid in describing and differentiating these methods, and highlight key issues in structure-based specificity prediction, we provide in Table 1 the answers to the following five questions for each of the discussed methods.

Table 1.

Structure-based specificity prediction studies, with answers to the five questions introduced in the text (see “Classification scheme”).

Citation	Method or software	Energy function		Relaxation of mutants		Unbound DNA	B-P independence		Tested w/templates?
			intra-DNAa	Proteinb	DNAc		scoringd	samplinge
Angarica [20]	DNAPROT	KB	harmonic	N	N	N	N	Y	Y
Alamanova [22]	Statistical potential	KB	-	N	N	N	Y	Y	Y
Donald [21]	Statistical potential	KB	-	N	N	N	Y	Y	N
AlQuraishi [23]	Compressed sensing	KB	-	N	N	N	Y	Y	N
Havranek [7]	Rosetta	MM	-	fixbb-rot	N	N	N	N	Y
Morozov-stat [24]	Rosetta	MM	harmonic	N	N	N	N	Y	Y
Morozov-dyn[24]	Rosetta	MM	harmonic	fixbb-rotmin	min	Y	N	Y	Y
Alibes [25*]	FoldX	MM	MM/KB	fixbb	fixbb	N	N	Y	N
Siggers [29*]	Tinker	MM	MM	fixbb-rotmin	fixP-rotmin	Y	N	N	Y
Rahi [31]	Amber	MM	MM	cst-min	cst-min	Y	N	Y	N
Liu [32]	CHARMM	MM	MM	MD	MD	Y	N	Y	N
Beierlein [33]	Amber	MM	MM	MD	MD	Y	N	Y	N
Seeliger [34*]	Gromacs	MM	MM	MD	MD	Y	N	Y	N
Moroni [36]	CHARMM/NAMD	MM	MM	min or MD	min or MD	N	N	N	N
Temiz [37*]	Experimental contact energies	KB	-	N	N	N	N	N	Y
Zakrzewska [13*,14*]	ADAPT	MM	MM	fixbb-min	min	Y	N	N	N
Yanover [39*]	Rosetta	MM	MM	MC	MC	Y	N	N	Y

^aThe intra-DNA energy function; ‘KB’ denotes knowledge-based potential; ‘MM’ denotes molecular mechanics force field; ‘harmonic’ indicates a model for DNA strain [11] that is quadratic in the base-step (and possibly base-pair) parameters.

^bConformational relaxation applied to protein after DNA mutation; ‘N’: none; ‘fixbb’: backbone is fixed; ‘rot’: rotameric sidechain sampling; ‘min’: gradient-based energy minimization; ‘cst’- energetic restraints on the atomic positions; ‘MD’: molecular dynamics relaxation; ‘MC’: Monte Carlo optimization.

^cConformational relaxation applied to DNA after mutation; ‘N’, ‘min’, ‘MD’, ‘MC’, ‘cst’: as above; ‘fixbb’: sugar-phosphate backbone is fixed; ‘fixP’: phosphorus atoms are fixed; ‘rot’: small-perturbation DNA rotamers.

^dAre DNA binding energies position independent? ‘Y’: yes or ‘N’: no.

^eDoes the DNA sequence sampling protocol assume independence between positions (e.g., all single-based mutants of the crystallized target site)? ‘Y’: yes or ‘N’: no.

1. What type of energy function is employed? Here we focus exclusively on studies that evaluate binding energies using atomically detailed descriptions of complexes. Recent studies can be roughly divided according to whether they use knowledge-based (KB) or molecular mechanics (MM) energy functions. In knowledge-based approaches, the shape of the distance-dependent atomic interaction energy curve is estimated from an experimental dataset such as a collection of protein-DNA complex structures. MM energy functions, on the other hand, are composed of independent, physicochemical energy terms that capture van der Waals and Coulombic interactions, solvation effects, and other atomic interaction energies. While the weights on these terms may be fit to experimental data, the functional forms of the individual terms are more constrained than for KB approaches. For the purposes of this review, we include in the MM category physics-based force fields such as FoldX [5] and Rosetta [6,7] that may incorporate knowledge-based energy components (such as rotamer probability estimates) in addition to standard MM terms.

2. When building structural models of mutant DNA target sites, what degree of conformational relaxation is allowed after base mutation? DNA mutations are typically performed so as to preserve the plane of the base, the geometry of the glycosidic linkage, and the conformation of the sugar phosphate backbone. If no further conformational relaxation is allowed, this procedure may introduce atomic overlaps and/or unfavorable electrostatic interactions – either internal to the DNA or in the protein-DNA interface – that have the potential to bias the specificity profile toward the DNA site present in the starting structure. When starting from a co-crystal structure of the native protein bound to the high-affinity consensus target site, this effect may enhance performance; in template-based specificity predictions or when assessing the quality of a large-scale specificity redesign, on the other hand, it seems likely that some degree of conformational relaxation is necessary (Fig. 1). For example, sequence-dependent fluctuations in minor groove width, which can contribute to binding specificity for certain families [8], have been successfully modeled using Monte Carlo simulations [9]. For the classification given in Table 1, we consider whether protein and/or DNA backbone flexibility is allowed, and whether local energy minimization or more intensive conformational optimization is used. The latter distinction is motivated by the observation that energy minimization (which involves a search for a nearby local energy minimum) may not sample the discrete rearrangements of interface protein sidechains that are accessible to rotamer-based optimization protocols.

3. Is the energy of the unbound DNA site considered? Binding affinities involve energy differences between the bound and unbound states. Evaluating only the energies of the bound complexes risks missing indirect readout mechanisms, which depend on the sequence-dependent deformability of DNA and are determined in part by energetic differences between the bound and unbound state. For example, the greater flexibility of pyrimidine-purine base steps (TpA, CpA, and CpG) is thought to reflect weaker stacking interactions for these steps in unbent B-form DNA [10]. Capturing this effect requires either scoring of unbound models of the DNA sequence, or the explicit inclusion of a knowledge-based scoring term that measures these geometric preferences (such as harmonic base-step potentials derived from crystal structures [11] or molecular dynamics (MD) simulations [12]). Explicit treatment of the unbound state is also critical when there are differences in the internal energy of base pairs or base steps of different sequences, as occurs when intra-DNA interactions are evaluated with an MM potential.

4. Is independence assumed between DNA base pairs? Positional independence may be a feature of the energy calculations themselves, or may follow from the extent of DNA sequence sampling that is performed to construct a binding specificity model. In the absence of intra-DNA energetics and conformational relaxation of mutant complexes, the binding affinity contribution of a DNA base at one position in the target site is independent of the base identities at other positions. Even when calculated energies are not strictly pairwise independent, however, many studies choose to sample only single base mutants from a consensus target site. The ADAPT method [13*,14*], by contrast, evaluates binding energies for all possible variants of the target site, and thus is able to capture energetic correlations between DNA sequence positions. Although positional independence has been shown to be a reasonably good approximation to binding energetics overall [15], there are clear cases where this assumption breaks down [16], which may be associated with sequence-dependent DNA bending or protein sidechain rearrangements at the interface.

5. Is the approach tested with homologous templates? We expect that performance evaluations based only on predictions made from structures of native, high-affinity complexes will tend to reward protocols that bias toward the crystal structure sequence, whereas such protocols would be expected to perform less well when predicting binding specificity from the structure of a template with divergent binding preferences or when assessing the quality of a large-scale specificity redesign. Given the limited experimental coverage of protein-DNA complexes, the ability to make accurate predictions starting from structures of homologous protein-DNA complexes [17], or unbound protein structures via protein-DNA docking [18,19], is highly desirable; such applications are, however, likely to require significant backbone and sidechain conformational sampling.

Knowledge-based potentials

Several groups have explored the use of knowledge-based potential energy functions in the context of a static model in which little or no conformational adjustment is performed after base mutation. The DNAPROT method [20] uses a potential energy function that combines Olson’s harmonic model of DNA flexibility [11] with three atomic preference matrices (for hydrogen bonds, water-mediated hydrogen bonds, and hydrophobic interactions) derived from a large database of protein-DNA interfaces. This method performed well at PWM prediction from native structures, and even generated reasonable PWM predictions for two targets whose structures were modeled by homology (with rotamer-based optimization of sidechains at the protein-DNA interface in the context of the template DNA sequence). Donald et al. [21] compared the performance of several statistical potential functions and found that the quasichemical approach outperformed other potentials, including the Amber99 MM force field. Alamanova et al. [22] derived a distance-dependent all-atom statistical potential from protein-DNA crystal structures and successfully applied it to predict PWMs for transcription factors (TFs) in the p53 and NF-κB families. Compressed sensing methods were used by AlQuraishi and McAdams [23] to learn a “de novo” interatomic energy potential from protein-DNA interface structures and binding energies. In this approach, protein-DNA binding interactions are viewed as mesoscopic sensors of an underlying microscopic signal, the atomic interaction potential, which is fit by a regularized logistic regression procedure that enforces sparsity of the components in order to avoid over-training. This approach gave improved results over the quasichemical and DNAPROT methods (as well as the MM approach of Morozov et al. [24], see below) in recovery of crystal structure DNA sequences for the helix-turn-helix superfamily; transferability to other families and applicability of the potential to prediction of specificity from homology models or models where relaxation upon mutation is performed remain to be demonstrated.

Molecular mechanics force fields and conformational relaxation

Havranek et al. [7] developed a simple physicochemical model for prediction and design of protein-DNA interactions that featured fixed-backbone, rotameric sampling of interface sidechains together with a frozen DNA approximation. Accurate specificity predictions from native structures were demonstrated, and binding preferences for a set of zinc finger mutants were predicted by homology modeling with good experimental agreement for most targets. Morozov et al. [24] compared specificity prediction approaches with (“dynamic model”) and without (“static model”) conformational relaxation of mutant complexes, and found that the static model gave better performance overall, noting however that this model “…is more likely to fail when DNA conformational change is required to predict novel sites by homology.” The potential function combined MM terms with a harmonic model for DNA strain inspired by Olson et al. [11]; during conformational relaxation of mutant complexes the protein backbone and protein-DNA interface were fixed while protein sidechains and DNA dihedral angles were adjusted by Monte Carlo rotamer optimization and gradient-based minimization. Serrano and co-workers have used the FoldX force field [5] to predict PWMs from native crystal structures using a protocol that features limited conformational rearrangement of mutant complexes, with promising benchmark results [25*]. The FoldX model has been used to predict the effect of protein and DNA mutations on DNA-binding affinities of the Pax6 paired domain [25*], as well as in a variety of structure-based specificity redesign applications [26,27].

Using a molecular mechanics potential based on the Amber [28] force field, Siggers and Honig [29*] found that fixed-backbone, template-based specificity predictions for C₂H₂ zinc fingers are highly sensitive to the degree of structural divergence between target and template interfaces. Their protocol featured sidechain relaxation of mutant complexes (using a protein conformer library, small-perturbation DNA rotamers, and gradient-based energy minimization) as well as explicit evaluation of unbound DNA energies; it was able to generate highly accurate predictions when given the structure of the target protein (Zif268) or a structurally similar template. Conformational relaxation was found to be important for correctly predicting binding affinity for a series of target sites (as opposed to recapitulating a consensus sequence). Using constrained energy minimization of mutant complexes with the Amber force field and a generalized Born solvation model [30] (and explicit scoring of unbound DNA sites), Rahi et al. [31] successfully recapitulated experimental ΔΔG values for the PurR TF and single amino acid variants, and examined the relative contributions of direct and indirect readout mechanisms to DNA recognition by this TF.

Molecular dynamics simulations and free energy calculations

MD simulations have been widely used to study the energetics of macromolecular binding events, and several groups have used MD simulations and free energy calculations to obtain relative binding affinities of target proteins for alternate DNA sites. In so-called alchemical free energy simulations, the free energy difference between two states is calculated by sampling along an artificial “alchemical” transformation coordinate (λ) that links the two states (e.g., the wild-type and mutant base pairs), using simulations at multiple, fixed values of λ or in which λ is varied continuously between the states. Liu and Bader [32] used the thermodynamic integration (TI) approach and a dual topology representation (in which both states are simultaneously present, with interactions scaled by λ) of the mutated base pair to calculate relative binding affinities for all single base mutations of a TF target site. Calculations for the homeodomain Mat-α2 and bZIP GCN4 showed good agreement with experimental binding data, and analysis of the trajectories suggested an important role for water molecules in sequence-specific recognition. Beierlein et al. [33] investigated the specificity of C-protein:DNA recognition at a highly conserved G:C base pair using TI with a dual topology representation and a soft-core treatment of non-bonding interactions (to avoid singularities at the endpoints of the alchemical transformation). In agreement with experiment, all three base mutations were predicted to destabilize binding. Seeliger et al. [34*] applied both an equilibrium approach (using simulations at fixed λ-values) and a non-equilibrium approach (based on work distributions from many short simulations during which λ is varied between the two states) to compute relative binding affinities of the TF Zif268 for single mutants of its preferred target site. The spacing of intermediate states along the transformation coordinate was chosen so as to maximize phase space overlap between neighboring states. Agreement with experiment was encouraging, with deviations from experimental ΔΔG’s averaging less than 1 kcal/mol. For the non-equilibrium method, insufficient sampling of the endpoint states was found to be a major source of error. Taken together, the results of these alchemical free energy calculations are highly encouraging, although further benchmarking is necessary to assess performance in more challenging applications such as template-based specificity prediction and prediction of binding specificity changes induced by protein mutations. In addition, due to the high computational cost of MD simulations [32,33,35], advances in computational efficiency will likely be necessary for large-scale applications of these methods.

Moroni et al. [36] evaluated the use of MD conformational sampling in endpoint free energy calculations using a variety of implicit solvent models (distance dependent dielectric, solvent accessible surface area-based, and Generalized Born). Energetic contributions from van der Waals and Coulombic interactions and solvation effects were independently weighted to optimize the fit between predicted and experimental ΔΔG values for 52 single base-pair mutations to the lambda repressor target site. The best agreement between prediction and experiment was found when minimizing and scoring mutant complexes using a potential energy function incorporating standard MM terms as well as the Generalized Born plus solvent accessible surface area (MM/GBSA) solvation model, together with a term counting protein-DNA hydrogen bonds. Using MD conformational sampling instead of energy minimization resulted in poorer correlations with experiment, which was attributed by the authors to the presence of large conformational fluctuations in the MD trajectories.

Temiz and Camacho [37*] have described a novel approach for predicting the binding affinities of proteins in the C₂H₂ zinc finger family, one that incorporates MD sampling and homology modeling into an empirical scoring framework based on analysis of protein-DNA hydrogen bonds and atomic desolvation. In this model, hydrogen bond strengths are modulated by a weighting factor that reflects the solvent accessibility of the interacting atoms (computed by explicitly solvating the interface models). Energies computed with this empirical framework show very good agreement with experimental binding data for several zinc finger proteins.

Enhanced sampling of sequences and structures

The ADAPT approach of Lavery and co-workers [38] has recently been extended [13*,14*] to allow large-scale sampling of DNA sequence space with simultaneous conformational relaxation of the DNA target and neighboring protein sidechains. In this approach, the binding site is broken into overlapping five base segments, and binding energies are calculated for all 4⁵ = 1024 possible DNA sequences in each segment. To evaluate the binding energy of a single sequence segment, bound and unbound energy minimizations are performed in which nucleotides within that segment and nearby protein sidechains are flexible while the protein backbone as well as nucleotides in other segments are held fixed. The binding energy for any full-length DNA target site can then be computed by summing segment binding energies, with a weighting scheme that accounts for residue-pair interactions present in multiple overlapping segment minimizations. This approach was applied to the DNA-binding proteins TBP and SRY and, using a grid-computing-enabled approach, to the nucleosome (147 DNA base pairs wrapped around a histone octamer). The use of MD snapshots rather than experimental structures as starting points for the protocol was explored with promising results, and the predicted nucleosome binding energy matrix was used to scan the sequence neighborhood of transcription start sites in the yeast genome, recapitulating features of the experimentally observed nucleosome distribution such as a well-known nucleosome-depleted region upstream of transcription start sites.

To enhance conformational sampling in template-based specificity prediction, a recent study [39*] using the Rosetta software package [6] extended fragment assembly techniques for monomeric protein structure prediction to protein-DNA complexes by adding DNA-duplex and protein-DNA interface fragment replacement moves. In this approach, models of protein-DNA complexes are built in a two-stage procedure consisting of an initial coarse-grained, fragment-assembly simulation followed by an all-atom refinement stage with a high-resolution MM force field. DNA sequence space is explored by the use of Monte Carlo sequence mutation moves in which a randomly selected base pair is mutated and nearby bases and sidechains are re-optimized. Starting from randomized target sites, energy-biased acceptance of these sequence moves focuses sampling in high-affinity regions of sequence space, so that binding models in reasonable agreement with experiment can be extracted from the final, optimized DNA sequences. This approach has been tested on the C₂H₂ zinc finger family with promising results; transferability to other protein families remains to be assessed.

TAL effector structures and simulations

Transcriptional activator-like (TAL) effectors are bacterial DNA-binding proteins that contain multiple, tandemly-repeated copies of a highly conserved 33-35 amino acid consensus sequence. TAL effectors have been shown to recognize their DNA target sites by a remarkable, modular recognition process in which individual repeat units map to successive DNA base pairs, with the identity of the amino acids at positions 12 and 13 of the repeat unit (termed the repeat-variable diresidue, or RVD) showing a high degree of correlation with the identity of the recognized base pair [40,41]. In a development with relevance to protein-DNA modeling as well as rational engineering of protein-DNA binding specificity, the first structures of TAL effectors were recently solved [42*,43*], revealing the molecular basis for their modular DNA recognition properties (Fig. 3). Notably, the crystal structure of the TAL effector PthXo1 was solved by molecular replacement using structure prediction models of the protein-DNA complex built by the fragment-assembly techniques described above [39*]. In addition, de novo models built prior to the availability of experimental data have also been published [44*], allowing a direct comparison with experiment (Fig. 3B). These studies suggest that protein-DNA interface modeling techniques may play a growing role in experimental structure determination.

Figure 3. TAL effector-DNA interactions.

(A) Structure of the 23.5-repeat TAL effector PthXo1 bound to its target site [42*] reveals the mechanistic basis of modular DNA recognition by TAL repeats. Successive repeats form a left-handed helical bundle that wraps around the DNA duplex paralleling the major groove. Inset panels show a subset of the specificity-determining contacts formed between repeat residue 13 and its associated base-pair in the target site. (B) A de novo TAL-effector model [44*] is superimposed onto three repeats of the dHax3 [43*] TAL effector (1.2 Å Cα-RMSD over 102 residues). Whereas structural modeling played a critical role in structure determination for PthXo1, the dHax3 structure was solved by standard techniques and can serve as an independent reference for assessment of the structure predictions.

Conclusions

Atomistic molecular modeling represents one promising avenue for the rational prediction and reengineering of protein interactions. Success depends on potential energy functions that are able to accurately balance the contributions of electrostatic, hydrogen-bonding, and non-polar interactions in the presence of solvent and ions, as well as conformational sampling algorithms that can recapitulate the structural rearrangements that accompany sequence changes in binding proteins or their partners. Given the increasing availability of standardized, high-throughput experimental datasets on protein-DNA interactions [1-3], the many structural templates for these interactions in the PDB [45,46], and the largely polar (and hence challenging) nature of protein-DNA interfaces, prediction of protein-DNA binding specificity is emerging as an excellent in silico test of interface modeling approaches. In this review, we have described a representative collection of recent approaches to structure-based prediction of protein-DNA interactions, all of which depend on atomically detailed structural models of the binding interface. We conclude that the majority of these approaches give good results when predicting consensus sequences from native, high-affinity structures. In the context of this somewhat artificial test, conservative approaches with little or no conformational flexibility appear to give the best results [21,23,24,25*]. In template-based specificity, and more generally when modeling interactions not directly visualized by the input structure, there is evidence that conformational flexibility improves prediction accuracy [20,29*,39*], although the lack of comprehensive benchmarks that demand extensive sampling hinders comparison of different approaches. Moving beyond the native-recapitulation tests that dominate these studies should advance method development in this area. Looking ahead, we identify as promising research directions the accurate modeling of interfacial waters, the incorporation of cofactor binding influences by explicit modeling of multi-protein complexes, and the further development and application of backbone sampling techniques in prediction and design.

Highlights.

• We review structure-based approaches for protein-DNA binding specificity prediction.

• Interface flexibility is necessary for accurate template-based predictions.

• Assessments on native complexes are likely biased toward conservative approaches.

• Applications to structure determination and protein design are discussed.