PDA-Pred: Predicting the binding affinity of protein-DNA complexes using machine learning techniques and structural features

Abstract

Protein–DNA interactions play an important role in various biological processes such as gene expression, replication, and transcription. Understanding the important features that dictate the binding affinity of protein-DNA complexes and predicting their affinities is important for elucidating their recognition mechanisms. In this work, we have collected the experimental binding free energy (ΔG) for a set of 391 Protein-DNA complexes and derived several structure-based features such as interaction energy, contact potentials, volume and surface area of binding site residues, base step parameters of the DNA and contacts between different types of atoms. Our analysis on relationship between binding affinity and structural features revealed that the important factors mainly depend on the number of DNA strands as well as functional and structural classes of proteins. Specifically, binding site properties such as number of atom contacts between the DNA and protein, volume of protein binding sites and interaction-based features such as interaction energies and contact potentials are important to understand the binding affinity. Further, we developed multiple regression equations for predicting the binding affinity of protein-DNA complexes belonging to different structural and functional classes. Our method showed an average correlation and mean absolute error of 0.78 and 0.98 kcal/mol, respectively, between the experimental and predicted binding affinities on a jack-knife test. We have developed a webserver, PDA-PreD (Protein-DNA Binding affinity predictor), for predicting the affinity of protein-DNA complexes and it is freely available at https://web.iitm.ac.in/bioinfo2/pdapred/

1. Introduction

Protein-DNA interactions are essential components for all biological systems. These interactions control/regulate replication, repair, methylation, transcription and maintain the genome stability [1]. Numerous studies have been carried out to understand the structure–function relationship of protein-DNA complexes such as identification of DNA binding proteins, predicting the binding site residues, binding affinity of protein-DNA complexes and recognition mechanism of protein-nucleic acid complexes [2–7].

The function of a protein-DNA complex is dictated by its binding affinity [8], which is quantitatively measured using dissociation constant (K_d) and binding free energy (ΔG). Understanding the affinity of protein-DNA interactions is important to study the effects on gene expressions, delineate the recognition mechanism based on interactions between protein and DNA and so on [9,10]. It has a broad spectrum of applications such as designing complexes with the desired affinities, developing prediction methods for the target sites, and quantitative simulation of gene regulation networks.

Experimentally, binding affinities of protein-nucleic acid complexes are studied with electrophoretic mobility shift assay, filter binding assay, fluorescence spectroscopy, isothermal titration calorimetry, and surface plasmon resonance. These experimental data are compiled in ProNAB database, which contains more than 20,000 entries on binding affinities of protein-nucleic acid complexes and their mutants [11].

Computationally, Protein-DNA binding affinities are studied through molecular dynamics simulations, statistical methods and machine learning techniques [12]. Barissi et al. [13] developed a physics-based machine learning method, DNAffinity, for predicting the binding affinity of transcription factors using the features derived from molecular dynamics stimulations. Rastogi et al. [10] proposed a method to quantify the sequence-affinity relationship by analyzing data from DNA sequencing experiments to comprehensively probe protein-DNA interactions. Yang et al. [14] developed a structure-based machine learning method, PredDBA, for predicting the binding affinity of protein-DNA complexes using a limited set of non-redundant complexes. Most of these methods are focused on a specific protein-DNA complex or a small set of data. In addition, the performance of available methods is not uniform in different structural and functional classes of protein-DNA complexes. On the other hand, the availability of protein-DNA binding affinity data in ProNAB database aid to develop reliable methods for binding affinity prediction. Understanding the binding affinity is also crucial for structure modeling of protein-DNA complexes [15].

In this work, we constructed a non-redundant dataset for binding affinities of protein-DNA complexes with known three-dimensional structures. We classified the complexes based on their structural and functional classes, and derived features at three levels such as protein, DNA and protein-DNA interactions. These features are related with binding affinity using machine learning techniques. Our method showed a correlation and mean absolute error of 0.78 and 0.98 kcal/mol, in jack-knife test, respectively. The salient features of important structural parameters and performance of the method based on structural and functional classifications are discussed.

2. Materials and methods

2.1. Dataset

We obtained the binding affinity of protein-DNA complexes from ProNAB database using the following criteria: (i) experimentally known binding affinity (ΔG), (ii) known 3D structure and (iii) non-redundant complex structures. ProNAB is a well-curated database with the highest number of protein-DNA binding affinity data. In some complexes, multiple binding affinity values were reported for the same protein-DNA complex, determined from different experimental methods or conditions such as pH, temperature, or buffer. In such cases, the affinity determined at physiological conditions (pH: 7; T: 298 K) was selected. The average ΔG was considered for the complexes where multiple data were reported in the literature. Using the above criteria, we obtained the binding affinity of 391 complexes. These complexes were grouped into different classes based on their structure (all-α: 113; all-β: 45; αβ: 181 and others: 36), function (Regulatory: 187 and others: 152) and percentage of binding site residues. The structural and functional classification were obtained from NDB [16].

2.2. Development of features

We derived a set of 117 features from protein-DNA complex structures. Features were grouped as (i) protein-based, (ii) DNA-based and (iii) interactions between protein and DNA.

2.3. Protein-based features

1. Accessible surface area of polar and non-polar atoms in the interface of the protein using NACCESS [17].

2. Residue depth of binding site residues using the python package, Bio.PDB.

3. Volume and surface area at the binding site of the protein using 3 V server [18].

4. Percentage of helix, sheet and coil at the binding interface of the protein.

5. Total Accessible surface area of residues in the interface of the apo-protein and in the complex using NACCESS

6. Number of polar, non-polar and charged residues in the binding interface

7. Number of C, N, O, S atoms in the interface.

2.4. DNA-based features

1. Number of P, O, N, C atoms in the interface.

2. Accessible surface area of nucleotides in the interface of apo-DNA and in the complex using NACCESS

3. Base step parameters obtained using the webserver w3DNA [19], which includes shift, slide, rise, tilt, roll and twist of the DNA.

2.5. Features based on protein-DNA interactions

1. Interaction energy between protein and DNA, which includes van der Waals and electrostatic energies.

2. Interaction energy due to main chain-main chain (mc-mc), side chain-side chain (sc-sc), main chain-side chain (mc-sc) and side chain - main chain (sc-mc) of protein and DNA.

3. Number and percentage of atom contacts between protein and DNA (OC, NO, ON, NP, CP, CC, NC, OP, OO, CN, NN, SN, SC, SO and SP).

4. Residue based contact potentials of protein-DNA complexes, obtained with a cut off distance of 3.5 Å and 6 Å.

5. Atom based contact potentials with a cut off 3.5 Å.

6. Accessible surface area of interface residues in a protein, DNA and overall.

7. Accessible surface area of polar and non-polar residues in the interface of protein-DNA complexes, and ASA difference of interface residues between uncomplex and complex forms.

8. Energetic parameters of protein-DNA complexes obtained using FOLDX [20].

2.5.1. Interaction energy

The interaction energy of protein-DNA complexes was calculated using the formula [21]

$$\mathit{\text{Interaction}}\mathit{\text{energy}}=\sum \left[\left(\frac{A_{ij}}{r_{ij}^{12}}-\frac{B_{ij}}{r_{ij}^6}\right)+\frac{q_i{q}_j}{\epsilon r_{ij}}\right]$$ (1)

$$A_{ij}={\epsilon }_{ij}^{*}{\left(R_{ij}^{*}\right)}^{12},B_{ij}=2{\epsilon }_{ij}^{*}{\left(R_{ij}^{*}\right)}^6,R_{ij}^{*}=\left(R_i^{*}+R_i^{*}\right),{\epsilon }_{ij}^{*}={\left({\epsilon }_i^{*}{\epsilon }_j^{*}\right)}^{1/2}$$

The total interaction energy is calculated as the sum of van der Waals and electrostatic interaction energies. We also calculated the interaction energy contributed by the main chain and side chain atoms such as interaction energy due to main chain-main chain, main chain-side chain, side chain-main chain and side chain-side chain atoms of protein and DNA respectively.

2.5.2. Development of contact potentials

We constructed a non-redundant dataset of 989 protein-DNA complexes with a sequence identity of ≤ 25% for developing atom–atom and amino acid residue-nucleotide contact potentials.

2.5.3. Atom propensity

We computed the preference of the interacting atoms at the interface of a protein-DNA complex as described in Shanmugam et al. [22]. It is given by

$$\text{Propensity}(\text{i},\text{j})=\frac{N_{ij}/N_{PD}}{\left(n_i/{n_p}^{*}{n}_j/n_d\right)}$$ (2)

where, N_ij is the number of contacts of type ‘i’ in protein and ‘j’ in DNA. n_i and n_j are total number of atoms of type i and j in protein and DNA, respectively. N_PD, n_p and n_d are the number of interacting atoms in the protein-DNA interface, number of atoms in proteins and number of atoms in DNA, respectively.

2.5.4. Residue propensity and contact potentials

We have computed the propensity of amino acid-nucleotide contacts between protein and DNA using the equation [23]:

$$\mathit{\text{Propensity}}(i,j)=\frac{N_{ij}/\sum N_{ij}}{T_j/\sum _j{T}_j}$$ (3)

where N_ij is the total number of contacts between base i and amino-acid residue j, and T_j is the total number of residue j in the whole data set. We have computed the propensities for the distance cut-off of 3.5 and 6 Å, which are widely used in the literature.

The propensity has been converted into potentials using the equation [23]:

$$\text{Potential}(i,j)=-\text{RT}\text{ln}[\text{propensity}(i,j)]$$ (4)

where R is the gas constant and T is the temperature.

2.6. Development of multiple regression models

We have developed multiple regression equations to relate the binding affinity with different features developed from protein-DNA complexes. It is defined as,

$$y_i={\beta }_0+{\beta }_1{X}_{i1}+{\beta }_2{X}_{i2}+\cdots +{\beta }_p{X}_{ip}+\epsilon$$ (5)

where, i is the number of observations, y_i is the dependent variable (binding affinity), X_i are the structure-based parameters, β₀, β₁ … β_p are regression coefficients and ε is the error term of the model.

2.7. Procedure for feature selection

We have carried out an exhaustive systematic search for all possible combinations of four features and selected the best combination based on correlation. Further, additional features were added until there was no increase in correlation (r) and the number of features reaches the square root of the number of data points. The selected features were used for the development of the final model. We have utilized the python machine learning package scikit-learn for constructing linear regression models [24]. A similar procedure was used to develop models for all the classes.

2.8. Assessment of prediction performance

The performance of the method was assessed using Pearson (r) and Spearman correlation coefficients, which explain the relationship between the experimental and predicted binding affinities, and mean absolute error (MAE) to examine the absolute difference between predicted and experimental affinity values using the SciPy package of python [25]. The model was further validated using a jack-knife test, where regression equations were developed using (n − 1) data points and used the same to predict the ΔG of the left-out complex, recursively.

3. Results and discussion

3.1. Analysis of binding affinity of protein-DNA complexes

We have analyzed the binding affinity of 391 protein-DNA complexes and the distribution of ΔG values is shown in Fig. 1. We observed that ΔG ranges from −20 kcal/mol to −4 kcal/mol and 70% of the complexes have the ΔG of −11 to −7 kcal/mol. The complexes 1P47 (Zif268-DNA complex) and 4DTM (RB69 DNA Polymerase-DNA complex) have the highest and lowest affinities of −20.0 kcal/mol and −3.9 kcal/mol, respectively.

Fig. 1.

Distribution of experimental binding affinity (ΔG) in a set of 391 protein-DNA complexes.

Further analysis revealed that 67% of residues in Zif268 are interacting with DNA in 1P47 whereas only 20% are at the interface of 4DTM. In addition, 1P47 has twice the number of charged residues at the interface as well as more interface area compared to 4DTM. Fig. 2 shows the protein atoms, which are in contact with the DNA in both high (1P47) and low affinity (4DTM) complexes. Interestingly, 1p47 has more polar atoms and spanned with larger interface area compared to 4DTM, which has more non-polar residues and less interface area.

Fig. 2.

Examples for high and low affinity complexes, showing the interactions between protein and DNA: (a) high affinity complex (1P47) with larger interface area and more charged contacts (blue) and (b) low affinity complex (4DTM) with more non-polar contacts (green) and comparatively less dispersed area of interactions.

We explored the binding affinities of protein-DNA complexes based on structural and functional classes and the results are presented in Supplementary Fig. S1. The analysis revealed that the range of ΔG for all-α, all-β and mixed class is −13.1 to −5.9 kcal/mol, −14.4 to −5.8 kcal/mol and −20.0 to −3.9 kcal/mol, respectively. Based on functional classification, the range is −20.0 to −4.7 kcal/mol for regulatory and −17.4 to −3.9 kcal/mol for other complexes.

3.2. Atomic contact potentials

We have computed the propensity of atomic contacts at the interface of protein-DNA complexes (Eqn. (2)).These propensity values were converted to contact potentials (Eqn. (4)) and the results are presented in Supplementary Table S1. We observed that N-O atom pairs are the most preferred ones with the lowest energy of −0.97 kcal/mol, which are involved in electrostatic interactions. Further, contacts between C and O atoms between protein and DNA, occur more frequently at the interface of the protein-DNA complexes and are energetically favorable with a contact potential of −0.28 kcal/mol. Interestingly, similar results are also reported for protein-carbohydrate complexes, with preferred CO contacts between proteins and carbohydrates [22].

3.3. Amino acid-nucleotide contact potentials

We have computed the preference of amino acid-nucleotide contacts between protein and DNA and the results obtained with a distance cutoff of 3.5 Å are presented in Supplementary Table S2. We observed that all the four bases (A, T, G, C) prefer to interact with the positive charged amino acids Arg (20–25%) and Lys (14–16%), along with higher preference for Arg compared to Lys, which is similar to the observations reported in [23]. Further, Cytosine and Adenine tend to prefer most of the polar and charged amino acids such as Arg, Lys, Tyr, Ser, Thr and Asn at the interface. In addition, Thymine occurred most frequently at the protein-DNA interface, which agrees well with the previous result reported by Meysman et al. [26]. In addition, Cys, Asp, Met are less favored at the interfaces of protein-DNA complexes.

We calculated amino acid-base propensities from the number of contacts, and further obtained the contact potentials, and the results are presented in the Table 1. Overall, Thymine-Trp interaction is energetically favorable at the interface with −2.25 kcal/mol, showing the importance of π-π interactions [27]. Further, Trp-Thymine has the strongest interaction among other aromatic amino acids Phe and Tyr. Frequently occurring amino acids Arg and Lys have the interaction energy of −0.93 and −0.91 kcal/mol, respectively with Thymine.

Table 1.

Amino acid-nucleotide contact potential for protein-DNA complexes.

DNA bases Amino acids	G	C	A	T
Ala	−0.65	−0.46	−0.73	−1.00
Arg	−0.81	−0.63	−0.85	−0.93
Asn	−0.98	−0.83	−1.12	−1.25
Asp	−0.90	−0.92	−0.93	−0.87
Cys	−1.89	−1.59	−1.66	−1.66
Gln	−0.89	−0.93	−1.15	−1.23
Glu	−0.62	−0.67	−0.76	−0.84
Gly	−0.91	−0.74	−0.79	−1.01
His	−1.44	−1.19	−1.38	−1.65
Ile	−0.70	−0.72	−0.89	−0.90
Leu	−0.39	−0.31	−0.42	−0.83
Lys	−0.73	−0.61	−0.80	−0.91
Met	−1.31	−1.11	−1.56	−1.62
Phe	−0.97	−0.85	−1.12	−1.34
Pro	−0.99	−0.77	−1.14	−1.26
Ser	−0.81	−0.66	−0.85	−0.99
Thr	−0.92	−0.69	−0.96	−1.07
Trp	−1.47	−1.49	−1.93	−2.25
Tyr	−1.03	−0.89	−1.27	−1.40
Val	−0.74	−0.52	−0.84	−0.94

3.4. Relationship between structure-based features and the binding affinity

We have related the parameters derived from structures of proteins, DNA and protein-DNA complexes (Materials and methods) with binding affinities to explore the relationship between them. Interestingly, although the direct correlation between each property and binding affinity is poor, we observed an increase in specific properties with different ranges of binding affinities.

The features related to binding interface of protein-DNA complexes such as contact potentials and total number of polar residues directly relate with protein-DNA binding affinity. Fig. 3 shows the data obtained with specific parameters such as (i) contact potentials at 3.5 Å (Fig. 3a) and (ii) number of polar residues in the interface of protein-DNA complexes (Fig. 3b). We observed that the average contact potentials (3.5 Å) of the complexes decreases from −31.95 to −8.22 kcal/mol with the decrease in the affinity range of the complex from −20 to −16 kcal/mol and −8 to −4 kcal/mol (Fig. 3a). For the complexes with the binding affinity of −20 to −16 kcal/mol and −8 to −4 kcal/mol, the average number of polar residues are 82 and 43 (Fig. 3b), respectively. Inspection of results revealed that the average number of polar residues and contact potential decreases approximately two and four folds with the decrease in affinity in the range of −20 to −16 kcal/mol to −8 to −4 kcal/mol.

Fig. 3.

Variation of structural features of protein-DNA complexes at various ranges of binding free energies as (a) contact potentials at 3.5 Å and (b) number of polar residues in the interface of protein-DNA complexes.

3.5. Prediction of binding affinity

3.5.1. Single property correlation

We related the binding affinity of 391 protein-DNA complexes with each feature derived from protein-DNA complexes and obtained a maximum correlation (r) of 0.21. Earlier studies showed that the classification of protein–protein/protein-nucleic acid complexes based on structure and function improved the correlation for understanding the relationship between sequence/structural features with binding affinity [14,28–30]. Hence, we classified the dataset based on functional and structural classes of proteins as well as number of binding site residues with the condition that sufficient number of data are present in each class. Table 2 shows the correlation between structural features and binding affinity of protein-DNA complexes belonging to of different classes.

Table 2.

Best single property correlation for different classes of protein-DNA complexes.

DNA strand	Structural class	Functional class	% Binding site residues	Feature	N	Maximum correlation (r)
Single stranded DNA (ssDNA)				Sidechain Hydrogen bond energy	16	0.75
Double stranded DNA (dsDNA)	all-α	Regulatory	High	Tilt of each strand of DNA	48	0.44
	all-α	Regulatory	Low	Percentage of NN contacts between the protein and DNA	36	0.45
	all-α	Not-regulatory		Percentage of beta sheet in the binding site	29	0.51
	all-β	Regulatory		Contact potential calculated with binding site cut off 3.5 Å	24	0.58
	all-β	Not-regulatory		Number of Interface Residues Clashing	21	0.54
	αβ	Regulatory	High	Electrostatic interaction	44	0.61
	αβ	Regulatory	Low	van der Waals clashes among the protein residues at the interface of the complex	35	0.43
	αβ	Not-regulatory	High	Contact potential calculated with binding site cutoff 6 Å	49	0.41
	αβ	Not-regulatory	Low	Percentage of atom contacts between N and C atoms of protein and DNA	53	0.35
	Others			van der Waals clashes among the DNA residues at the interface of the complex	36	0.41

N: number of complexes; αβ includes α + β and α/β classes of proteins.

The classification of protein-DNA complexes based on structure, function and percentage of binding site residues improved the correlation in the range of 0.35–0.75. In single-stranded DNA, we obtained a maximum correlation of 0.75 between energy due to side chain hydrogen bonds and binding affinity (Fig. 4). Interestingly, previous study showed that hydrogen bonds at the interface are the major contributors to protein–ssDNA binding affinity [31]. In addition, electrostatic interactions, which are favored by the highly negatively charged surface of DNA with positive charged amino acid residues also showed a correlation of 0.61 in the αβ-regulatory with high percentage of binding site residues. In all-β-regulatory complexes contact potentials showed the highest correlation of 0.58.

Fig. 4.

Relationship between side chain hydrogen bond energy and experimental binding affinity of protein-single stranded DNA complexes.

3.5.2. Multiple regression technique for predicting the binding affinity

We have utilized multiple regression technique to combine the features for predicting the binding affinity and the results are presented in Table 3. The selected features and regression equations obtained for each class are presented in Supplementary Table S3 and Supplementary Table S4, respectively. We observed that the correlation lies in the range of 0.77 to 0.98 with MAE in the range of 0.16 to 1.27 kcal/mol in all the sub classes in the training set. Further, on evaluating the model with jack-knife test, we obtained a correlation in the range of 0.64 to 0.96 with MAE of 0.24 to 1.53 kcal/mol. Overall, our method could relate the binding affinity of 391 complexes with a correlation of 0.86 and MAE of 0.76 kcal/mol on training, and a correlation of 0.78 and MAE of 0.98 kcal/mol, in jack-knife test. The relationship between experimental and predicted binding affinities are shown in Fig. 5, and for all the individual classes of protein-DNA complexes are provided in Supplementary Fig. S2. Further, we have examined the performance using 10-fold cross validation and the results are presented in Supplementary Table S5. Our method showed an average correlation and MAE of 0.75 and 1.02 kcal/mol, respectively.

Table 3.

Prediction of experimental ΔG using multiple linear regression technique.

Classes	N	Features	Self-consistency				Jack-knife test
			r	MAE	p-value	spearman	r	MAE	p-value	spearman
ssDNA	16	4	0.98	0.16	9.01E-11	0.94	0.96	0.24	6.99E-09	0.93
dsDNA
all-α-regulatory-high	48	7	0.81	1.02	4.85E-12	0.75	0.72	1.22	6.81E-09	0.65
all-α-regulatory-low	36	6	0.84	0.74	1.27E-10	0.79	0.75	0.94	1.77E-07	0.68
all-α-not regulatory	29	5	0.81	0.84	8.70E-08	0.77	0.68	1.11	0.000047	0.68
all-β-regulatory	24	5	0.93	0.64	3.66E-11	0.93	0.81	1.04	0.000002	0.86
all-β-not regulatory	21	5	0.96	0.43	1.27E-11	0.89	0.92	0.61	5.41E-09	0.84
αβ-regulatory-high	44	7	0.89	0.79	4.69E-16	0.76	0.80	1.01	7.14E-11	0.67
αβ-regulatory-low	35	6	0.86	0.72	2.62E-11	0.83	0.79	0.89	2.36E-08	0.76
αβ-not-regulatory-high	49	7	0.77	1.27	1.54E-10	0.70	0.64	1.53	8.27E-07	0.63
αβ-not regulatory-low	53	7	0.81	1.1	1.85E-13	0.75	0.74	1.32	2.44E-10	0.69
Others	36	6	0.82	0.67	7.66E-10	0.77	0.75	0.83	1.84E-07	0.67

Fig. 5.

A scatter plot showing the relationship between experimental and predicted binding affinities on jack-knife test.

3.6. Analysis of features important for binding affinity of protein-DNA complexes

The features selected in each class of protein-DNA complexes to predict the binding affinity are presented in Supplementary Table S3. Further, we grouped these features based on interaction energy, structural features, percentage of secondary structures, accessible surface area and base step parameters, and the results are shown in Table 4 under different categories such as protein-based, DNA-based and interaction-based. We observed that interaction-based features are important in most of the models (except all-α-regulatory-high and αβ-regulatory-low). Single-stranded DNA class mainly depends on interaction-based features whereas DNA-based features are dominant in αβ-regulatory-low class.

Table 4.

Comparison of features used in each classification model.

Single Stranded-DNA	all-α-regulatory-high	all-α-regulatory-low	all-α-not-regulatory
Interaction-based features Ionization Energy Electrostatic energy Number of interface atoms van der Waals energy	Protein-based features Volume ASA % secondary structure Helix dipole DNA-based features Base step parameters	Protein-based features Volume DNA-based features Base step parameters Interaction-based features Torsional clash Atomic contacts	Interaction-based features Torsional clash Protein-based features % secondary structure DNA-based features Base step parameters
all-β-regulatory	all-β-not regulatory	αβ -regulatory-high	αβ -regulatory-low
Interaction-based features Atomic contacts ASA Contact potential	Interaction-based features ASA Protein-based features % secondary structure	Interaction-based features ASA Electrostatic energy Atomic contacts Protein-based features Residue depth	DNA-based features Base step parameters
αβ-not regulatory-high	αβ-not regulatory-low	others
Interaction-based features Contact Potential ASA DNA-based features Base step parameters Protein-based features % secondary structure	Interaction-based features Atomic contacts van der Waals energy Solvation energy Interaction energy Ionization Energy Hydrogen bond	Interaction-based features Interaction Energy Solvation energy Atomic contacts DNA-based features ASA

^*The features present in 2 or more classes are shown in italics.

We observed that base step parameters and accessible surface area are important to predict the binding affinity in most of the classes of protein-DNA complexes. In addition, atomic contacts and contact potentials are identified in mixed class proteins (αβ). Interestingly, non-regulatory proteins belonging to mixed class is dominated with interaction energies. This analysis reveals the importance of specific features common to several classes of protein-DNA complexes. Detailed analysis of each class is provided in Supplementary information.

3.7. Comparison of our method with PreDBA

Yang et al. [14] developed a machine-based learning model, PreDBA, to predict the protein-DNA binding affinity, which was trained on 100 non-redundant complexes. To compare our predictions with PreDBA, we tested the model developed in Yang et al. [14] with our dataset. We excluded the redundant protein-DNA complexes that are used in their training dataset and obtained a test dataset of 225 complexes. PreDBA provides the binding affinity for each chain of the protein against each strand of DNA. Hence, protein-DNA complexes with more than one protein or DNA chain, we computed the affinity for all the possible combinations of DNA and protein pairs and considered the closest prediction to the experimental binding affinity to estimate the correlation/MAE. Fig. 6 shows the relationship between experimental and predicted binding affinities obtained with PreDBA (Fig. 6a) and in the present work (Fig. 6b).

Fig. 6.

Prediction of protein-DNA binding affinity in a set of 225 non-redundant complexes (a) PreDBA and b) present work.

We observed that our method predicted with an average correlation of 0.76 and MAE of 0.98 kcal/mol, while average correlation of PreDBA was 0.32 with MAE of 1.86 kcal/mol. The improved performance in our method is mainly due to the classification of complexes and the availability of experimental affinity for a large number of protein-DNA complexes in ProNAB database [11]. Supplementary Table S6 shows the results for each class of protein-DNA complexes based on structure and function.

3.8. Prediction on the web

We have developed a web server, PDA-PreD, for predicting the binding affinity of protein–DNA complexes. It takes the protein–DNA complex structure in PDB format or the PDB ID as input. The output includes the predicted value of binding affinity ΔG and the dissociation constant (K_d) computed using the equation, ΔG = -RT ln(K_d). The web server is freely accessible at https://web.iitm.ac.in/bioinfo2/pdapred/.

3.9. Case study with DNA Polymerase I -DNA complex

We have explored the binding affinity of DNA Polymerase I-DNA complex (2HHU) using structural features identified in this work. It belongs to αβ non-regulatory class with binding site residues less than the average. It has the experimental binding affinity of −4.03 kcal/mol and our model predicted the affinity as −4.33 kcal/mol, which shows a good agreement. Fig. 7 shows the interactions between the sugar-phosphate backbone of the DNA and the protein. We analyzed the individual features used in the model to predict the binding affinity of the complex. Our analysis showed that it has high percentage of CO atom contacts with preferred atomic contact potentials, backbone hydrogen bond energy of −10.6 kcal/mol and favorable van der Waals and total interaction energies of −24.5 kcal/mol and −162.3 kcal/mol, respectively. The combination of these parameters could correctly predict the binding affinity of the complex with a deviation of 0.3 kcal/mol.

Fig. 7.

a) Structure of DNA Polymerase I-DNA complex (2HHU), b) Interaction between the sugar-phosphate backbone of the DNA (in orange) and amino acid residues (in yellow) in the protein. The backbone phosphate is highlighted in magenta, and the hydrogen bond interactions are shown in red dotted lines.

4. Conclusions

We have systematically analyzed the relationship between structural features of protein, DNA as well as interaction energy, and binding affinity of protein-DNA complexes. Our analysis showed that interaction energy, contact potential, number of interface atoms, ASA of protein and DNA, surface area and volume of the protein, and DNA base step parameters are important to understand the binding affinity. Further, classification of complexes based on structure, function and binding site residues are important to account the binding affinity. We have developed multiple regression-based methods for predicting the binding affinities of the protein-DNA complexes and our method showed an average correlation of 0.78 and MAE of 0.98 in a jack-knife test, which are better than other existing methods in the literature. We have developed a web server for predicting the binding affinity and it will be useful for large scale analysis and developing strategies for therapeutic targets.

Data availability

Data will be made available on request.