Learning Sequence Determinants of Protein:protein Interaction Specificity with Sparse Graphical Models

Abstract

In studying the strength and specificity of interaction between members of two protein families, key questions center on which pairs of possible partners actually interact, how well they interact, and why they interact while others do not. The advent of large-scale experimental studies of interactions between members of a target family and a diverse set of possible interaction partners offers the opportunity to address these questions. We develop here a method, DgSpi (Data-driven Graphical models of Specificity in Protein:protein Interactions), for learning and using graphical models that explicitly represent the amino acid basis for interaction specificity (why) and extend earlier classification-oriented approaches (which) to predict the ΔG of binding (how well). We demonstrate the effectiveness of our approach in analyzing and predicting interactions between a set of 82 PDZ recognition modules, against a panel of 217 possible peptide partners, based on data from MacBeath and colleagues. Our predicted ΔG values are highly predictive of the experimentally measured ones, reaching correlation coefficients of 0.69 in 10-fold cross-validation and 0.63 in leave-one-PDZ-out cross-validation. Furthermore, the model serves as a compact representation of amino acid constraints underlying the interactions, enabling protein-level ΔG predictions to be naturally understood in terms of residue-level constraints. Finally, as a generative model, DgSpi readily enables the design of new interacting partners, and we demonstrate that designed ligands are novel and diverse.

1 Introduction

The molecular machinery of the cell is driven largely by protein:protein interactions. Traditional high-throughput technologies [5] provide evidence for the existence of interactions that existing computational systems biology techniques utilize to build global networks of interacting proteins. However, finer-grained methods are necessary in order to better understand, predict, and control these interactions. Fortunately, appropriate experimental methodologies are rapidly developing, e.g., using protein microarrays to isolate numerous pairs of possible partners, and fluorescence polarization to assess their interaction strength (Fig. 1, left). Several large-scale studies have been conducted using such techniques for particular families of interacting proteins, including PDZ domains and their peptide ligands [4, 40], and human basic-region leucine zippers (bZIPs) and their coiled-coil partners [6, 8]. In lieu of large-scale studies, the aggregation of a large number of smaller-scale experiments can also yield extensive amounts of detailed binding data, e.g., for major histocompability complex (MHC) and ligands [29, 27, 41, 2, 43], and serine proteases and inhibitors [21, 18].

Figure 1.

DgSpi: Data-driven Graphical models of Specificity in Protein:protein Interactions. A graphic model of PDZ:peptide interactions encapsulates the amino acid constraints conferring the strength and specificity of the interactions in an input dataset. (left) The dataset has ΔG values (shades of green) or “non-interacting” indications (‘X’s) for some PDZ (blue) peptide (red) pairs, (middle) We learn a graphical model with bipartite nodes for some residues in the PDZ (blue) and peptide (red), with edges (green) encapsulating and providing a probabilistic interpretation for amino acid constraints, (right) We use the model to predict novel interactions as well as to design novel peptide partners for PDZs.

As one particular example, consider the specific recognition between PDZ domains and their peptide ligands. PDZs are small peptide recognition modules that bind specific C-terminal peptides of other proteins (Fig. 1, right), in order to mediate protein:protein interactions (e.g., in signaling networks). Early studies of PDZ:peptide recognition developed consensus motifs to capture the common amino acids comprising the ligands of different PDZ “classes” (e.g., class I = S/T-X-Φ vs. class II = Φ-X-Φ, where Φ is a hydrophobic residue). More recent studies yielded more refined statistical binary interaction predictors (interact or not?), based on analysis of amino acid pairs (across the PDZ:peptide interface) in curated datasets of experimentally identified PDZ:ligand partners [3, 38]. MacBeath and colleagues then made the leap to large-scale quantitative data, determining the ΔG of binding for 829 PDZ:peptide pairs from 96 PDZs (from mouse, fly, and worm) against a panel of 259 possible peptide partners [36]. They used this data to develop a binary interaction predictor, based on the constituent PDZ:peptide amino acid pairs like the predictors mentioned above, but taking advantage of the quantitative and negative data [4]. More recently, Bader and co-workers used the MacBeath data to train a type of support vector regression model for predicting ΔG of binding for PDZ:peptide pairs [33].

Motivated by the exciting growth in quantitative studies of protein:protein interactions, we seek to develop a data-driven, sequence-based model that directly and compactly reveals and represents the amino acid interactions underlying experimentally measured ΔG values of binding (henceforth just ΔG for short) and enables efficient, accurate, robust, and transparent prediction of ΔGs for new pairs of possible partners (Fig. 1). To do so, we employ a graph-structured model (which we refer to simply as a graphical model) that explicitly models amino acid interactions and provides a probabilistic interpretation for them. Sequence-based graphical models of protein families have been used to capture amino acid interactions in order to predict protein structure [26, 10, 12], function [37, 1] and design new proteins [39, 15]. We build here on our earlier work on sequence-based models of interacting protein families for binary prediction of interaction [38], significantly extending that approach to incorporate quantitative data and to predict ΔG.

We call our approach DgSpi, for Data-driven Graphical models of Specificity in Protein:protein Interactions. Using the PDZ data from MacBeath and co-workers, we demonstrate that DgSpi is highly predictive of ΔG, obtaining predicted-experimental correlation coefficients of up to 0.69 in a ten-fold cross-validation and 0.63 in leave-one-PDZ-out cross-validation. This performance is essentially equivalent to that obtained by Bader and colleagues, but importantly, our approach provides a readily interpretable model of the amino acid contributions underlying specific interactions. Furthermore, since our graphical models are generative, they can be used in designing new interacting partners (again, interpretable in terms of the amino acid contributions), and we show that there is a diversity of novel peptides that are predicted to bind well against any given PDZ and thus provide worthwhile hypotheses for experimental testing.

2 Methods

DgSpi takes as input (Fig. 1, left) two sets of protein sequences; for simplicity but without implications about function, we refer to one set as the “receptor” and the other as the “ligand”; e.g., the PDZ protein recognition modules as receptors and corresponding peptides as ligands. In addition to the sequences, there are experimental binding measurements for some of the pairs (one from each set); the measurement is either a ΔG value or an indication of “non-interacting” within the sensitivity of the experiment. Our goal is to be able to predict the ΔG of interaction for a previously untested receptor:ligand pair and to design new ligands partners for a specified receptor (Fig. 1, right). To do this, we seek a method that admits explanation of predictions in terms of the underlying amino acid-level interactions conferring specificity of interaction. Thus we employ a graph-structured, or graphical, model (Fig. 1, middle) with nodes for the receptor and ligand residues, and bipartite edges capturing the amino acid constraints between receptor and ligand residues.

We first summarize a graphical model to predict ΔG from a pair of sequences, and then the algorithms to construct a model from training data of sequence pairs with observed ΔG.

2.1 A graphical model of binding free energy

We assume the receptors have been multiply aligned to p informative (non-gappy) columns, and the ligands likewise to q residues. Let X = {X₁, X₂, …, X_p} be a set of p random variables representing the receptor amino acid composition, with X_i a discrete random variable for the amino acid type at position i. Each X_i takes values in A = {ala, arg,…, val, -}, corresponding to the 20 amino acid types and an additional ‘-’ for a gap in the multiple sequence alignment. Similarly, let Y = {Y₁, Y₂,…, Y_q} be a set of q random variables representing the ligand composition.

Given a receptor sequence x = {x₁, x₂, ….,x_p} (i.e., a set of amino acid values for the random variables in X), along with a ligand sequence y = {y₁, y₂, …., y_q}, we want to predict the strength of a possible interaction between the two proteins, ΔG_pred(x, y). Our goal here is to develop a robust predictive model that is interpretable in terms of the amino acid interactions driving specific protein:protein recognition. Therefore we model ΔG_pred with a bipartite graphical model, with nodes for X and Y representing the amino acids and edges ε ⊂ X × Y representing their dependencies. Nodes x_i, y_j have associated $\left|\mathcal{A}\right|\times 1$ vectors V_i[a], V_j[b] to capture position specific environment effects to binding. Edge (i, j), between nodes x_i and y_j, has an associated $\left|\mathcal{A}\right|\times \left|\mathcal{A}\right|$ matrix W_i,j[a, b] of weights for a, b ∈ , holding the position-specific contributions to binding for each possible pair of amino acids, intended to capture electrostatics, van der Waals, hydrogen bonding, and other such interactions, which depend on the composition of the amino acids involved. We point out that these physical justifications for the parameters of our model are descriptive rather than prescriptive: they guide the intutition for learning and interpreting the model, but we make no assumptions on sources of these interactions beyond what we can learn from data.

In summary, then, given two protein sequences x and y, we predict their binding free energy as:

$$\mathrm{\Delta }{G}_{\mathit{\text{pred}}}(\mathbf{x},\mathbf{y})=\sum _{i=1}^p{V}_i[x_i]+\sum _{j=1}^q{V}_j[y_j]+\sum _{(i,j)\in \epsilon }{W}_{i,j}[x_i,y_j]$$ (1)

2.2 Training objective

Our data-driven approach to modeling protein:protein interaction specificity uses experimental data to learn the parameters V and W that define the model in Eq. 1. We assume that the experimental data is partitioned into interactions ℐ₊, with ΔG values, and non-interactions ℐ₋, where the binding was weaker than ΔG_max, a maximum experimentally detectable ΔG value. Thus each member of ℐ₊ is of the form (x, y, ΔG), giving a pair of sequences and the measured ΔG value, while each member of ℐ₋ is simply an (x, y) pair.

We take as our primary objective minimizing the squared error between the observed and predicted ΔG values for members of ℐ₊. For the non-interactions in ℐ₋, we incorporate a penalty for an incorrect prediction, i.e., for ΔG_pred better than ΔG_max. In particular, we use a one-sided squared penalty for noninteractions predicted as interactions. Compared with the hinge-loss commonly used in SVMs, this tends to penalize small differences to a lesser extent, which is a desirable property in cases such as ours where the focus is on the regression error and not the misclassification cost. The one-sided square error has no points of discontinuity, making optimization easier as well.

Thus our objective function for a specific set of parameters V, W is:

$$L(V,W)=\sum _{(\mathbf{x},\mathbf{y},\mathrm{\Delta }G)\in {\mathcal{I}}_{+}}{(\mathrm{\Delta }{G}_{\mathit{\text{pred}}}(\mathbf{x},\mathbf{y};V,W)-\mathrm{\Delta }G)}^2+\sum _{\begin{array}{c}(\mathbf{x},\mathbf{y})\in {\mathcal{I}}_{-}s.t\\ \mathrm{\Delta }{G}_{\mathit{\text{pred}}}(\mathbf{x},\mathbf{y};V,W)<\mathrm{\Delta }{G}_{\mathit{max}}\end{array}}{\gamma }_{-}\cdot {(\mathrm{\Delta }{G}_{\mathit{\text{pred}}}(\mathbf{x},\mathbf{y};V,W)-\mathrm{\Delta }{G}_{\mathit{max}})}^2$$ (2)

where we emphasize the dependence of ΔG_pred in Eq. 1 on V and W by including them as parameters. The parameter γ₋ sets the relative weighting between the contributions from interactions and non-interactions.

2.3 Block-sparse Regularization

A suitable model can be learned from the data by minimizing the objective function in Eq. 2. However, directly optimizing this function is likely to lead in overfitting as there are usually far more parameters in the model than there are data points available with which to fit them. To circumvent this problem, we instead optimize a regularized objective function. Regularization is usually described as a penalty to the objective function; an alternate but equivalent view of the regularization is that it is a Bayesian prior on the models that biases the learning method towards models consistent with the prior. Protein:protein interactions can be reasonably expected to display structural sparsity —due to spatial restrictions, only a few of all possible bipartite interactions between the partners are likely to be important in biochemical interactions. Motivated by this prior belief, we employ block-L1 regularization, a form of regularization that penalizes the number of non-zero edges (or “blocks”), so that each edge (i, j) is penalized unless all parameters within the edge, W_i,j, are zero. This promotes a sparser structure promoting interpretability; furthermore, by reducing the number of non-zero parameters in the model, it helps avoid overfitting. For our model, the block-L1 regularization term is:

$$R_{1,2}(V,W)=\frac{1}{\left|\mathcal{A}\right|}(\sum _{i=1}^p\Vert V_i\Vert +\sum _{j=1}^q\Vert V_j\Vert )+\sum _{(i,j)\in \epsilon }\Vert W_{i,j}\Vert$$ (3)

where ‖ · ‖ refers to the vector two-norm of the corresponding set of parameters and the fraction $\frac{1}{\left|\mathcal{A}\right|}$ (number of amino acids plus gap) is a correction factor to account for the different degrees of freedom in V and W. Our learning objective is then:

$$\underset{V,W}{argmin}(L(V,W)+{\lambda }_{1,2}\cdot R_{1,2}(V,W))$$ (4)

where λ_1,2 sets the relative weight between the learning objective and the regularization term.

2.4 Learning algorithms

Schmidt et al. [32] developed a Limited-Memory Projected Quasi-Newton (PQN) approach suitable for squared error objectives. We customize their method for our graphical models of protein-protein interactions. A constrained optimization to incorporate the block-L1 regularization is performed by a projected gradient method that iterates between unconstrained gradient descent updates to the parameter values, and constrained projections of the parameter values onto the constrained space.

While this approach can be used to learn the structure and parameters of the model (i.e., which vertices and edges, along with their weights), in practice, the resulting procedure can result in biased weights for the non-zero parameters, despite identifying the correct structure [1, 23]. To avoid this, after learning the structure of the model, we re-learn the non-zero parameters with L2 regularization. That is, in a second stage, we restrict the optimization to the vertices and edges contributing in the first stage, but reoptimize their weights using a modified version of Eq. 4, replacing R_1,2 with:

$$R_2(V,W)=\sum _{i=1}^p{\Vert V_i\Vert }^2+\sum _{j=1}^q{\Vert V_j\Vert }^2+\sum _{(i,j)\in \epsilon }{\Vert W_{i,j}\Vert }^2,$$ (5)

weighted by a corresponding λ₂. Since R₂(V, W) penalizes the square of the vector 2-norms, each element of each parameter vector is penalized independent of any group membership; the regularization is thus independent of the degrees of freedom in the corresponding groups. This two-stage approach finds sparse models with small edge weights, regularizing a pseudo-likelihood objective similarly to the approach of [1]. We find in practice that this approach yields models that are both interpretable and predictive of ΔG.

3 Results

Our goal is to make quantitative predictions of the ΔG of PDZ:peptide interactions, interpretable in terms of underlying amino acid constraints. This is in contrast to the approach of MacBeath and co-workers [4], who studied the ability of a computational method to classify interaction vs. non-interaction.¹ It is also in contrast to the Support Vector Regression approach of Bader and co-workers [33], in that while our method achieves comparable predictive accuracy, it has the added benefit of being able to automatically identify the amino acid-level interactions with the greatest impact, and directly characterize their contributions. These interactions not only allow us to characterize the sequence determinants of binding affinity and specificity, but also allow us to design new interacting partners based on the derived “rules” of good interactions.

We apply DgSpi to the extensive PDZ dataset collected by MacBeath and colleagues [36, 4]. To enable appropriate comparison of results, we use the processed version of the dataset provided by Bader and colleagues [33]. The dataset includes 82 mouse PDZs and 217 peptides, with a reported 560 interactions and 1167 non-interactions. A structure- and sequence-based multiple sequence alignment of 225 columns was obtained for the PDZs; the peptides were represented by 5 C-terminal residues each. We then removed highly conserved and highly gap-ful columns, reducing the alignment to 114 PDZ positions and 5 peptide positions.

3.1 ΔG prediction

10-fold cross-validation

To test the ability of our model to predict the affinity of PDZ-peptide interaction, we first performed a ten-fold cross-validation (i.e. we learned the model with 90% of the data and tested it on the left-out 10%, doing this with each 10% left out). This represents the scenario in which data are available for some interactions, and we want to make predictions for others.

Our learning approach has three main parameters: γ₋, a parameter trading off the relative importance of positive and negative interactions in the objective function; λ_1,2, the strength of the block-L1 regularization used to determine the non-zero parameters of the model and λ₂, the regularization weight used to estimate the values of the non-zero parameters. γ₋ and λ₂ were set to 0.05 and 1 respectively based on our initial experiments using one train-test split. The small value of γ₋ reflects the relative abundance of non-interactions in our dataset and our emphasis on modeling interactions comprehensively since they are biologically more interesting. For each training split, we varied λ_1,2 generating multiple models spanning the spectrum from models with no interactions to models where nearly all possible interactions were allowed.

Fig. 2 summarizes trends over values of λ_1,2. The left characterizes the increase in number of interactions with decreasing λ_1,2 (top), and the corresponding increase in the average Pearson correlation coefficient (bottom), from 0.49 when there are no edges in the model and all contributions are due to the V_i, V_j, to 0.66 when ∼ 60 interactions are included, to a maximum of 0.69 when ∼ 300 interactions are included. The relatively large increase in model accuracy when the number of edges increases from 0 to 60 suggest that these edges make important contributions to binding affinity and specificity. In contrast, the relatively small increase in accuracy of the model as the number of edges increases beyond 60 to being a completely connected model suggests that the edges introduced later have relatively low importance. The right panel shows the average strength of each edge, calculated as the norm of W_i,j, as a function of λ_1,2. Each line represents a separate unique interacting pair of residues; interactions that have high weight at λ = 25 are highlighted in color, while the remaining interactions are shown in black. When the model is sparse (high λ_1,2), there are few, strong, interactions; as the density of the model increases (low λ_1,2), most interactions have non-zero strength but are very weak.

Figure 2.

Trends with varying regularization weight (parameter λ_1,2), with higher values yielding sparser models. (left-top) Number of edges in models learned with varying λ_1,2. The red star indicates a model with edges fixed according to contacts in a crystal structure. (left-bottom) Test correlation coefficient as a function of the average number of edges in trained model. (right) Regularization path, showing each edge's strength as a function of λ_1,2.

Fig. 3 shows the prediction results for one 10-fold repetition at λ_{1, 2} = 20. The overall correlation coefficient across the dataset was 0.67 while the root mean square error between experiment and prediction was 0.62. Most errors were equally distributed around zero, and actually within typical experimental error. However, there were a few clear outliers where the model under-predicted binding energies.

Figure 3.

Example prediction results, combined across 10 splits in one repetition at λ_1,2 = 20. (left) Scatterplots of experimental vs. predicted ΔG. Pearson Correlation Coefficient across entire test-split was 0.67. (right) Histogram of prediction errors.

Contact-based model structure

When an experimentally determined 3D structure of the protein-protein interface is available, an alternate approach to determining the structure (edges) of the graphical model could be to restrict the non-zero interactions to the pairs of residues close to each other in the 3D structure. The parameters of this model with fixed structure can then be readily learned with L2 regularization, as before. Macbeath and colleagues identified 38 contacts between 16 PDZ residues and 5 peptide [4]. We repeated our 10 fold cross-validation experiments, using these 21 positions and 38 contacting residues as the set of vertices and edges in the model (instead of identifying them using the block-L1 penalty), and estimated their parameters with L2 regularization. The average correlation coefficient of the contact-based models is 0.60, which, while good, is lower than the 0.66 correlation obtained by models with about 60 interactions. Could the difference in accuracy be due to the difference in the number of interactions? The bottom-left panel in Fig. 2 highlights the accuracy of this model (shown as a red star), compared to the correlation coefficients obtained by varying λ_1,2. We see that the models with learned structure can achieve accuracy similar to the contact-structure model but using fewer interactions; alternativeely, a model with learned structure and a comparable number of interactions to that of the contact structure achieves higher correlation. Thus our data-driven approach to learning model structure can identify important interactions beyond those that might be inferred by inspection of the 3D structure.

Leave-one PDZ out

To test the scenario where the model is applied to make predictions for a new PDZ, we performed “leave-one-PDZ-out” cross-validation following the approach of Bader and colleagues [33]. We held out data for each of the 23 PDZ domains with at least 10 interactions, training the model on the remaining data and testing on the held-out domain. Since the effect of λ_1,2 on the sparsity of the model depends on the number of sequences in the training set, instead of choosing the same value of as selected by ten-fold cross validation, we performed a grid search on λ_1,2 and used the value that gives a model of similar sparsity as the cross-validated models. This process allows us to parameterize the model by the number of edges as opposed to the less natural λ_1,2. Using this procedure we obtained an average correlation coefficient of 0.61 across the 23 PDZs that had at least 10 interactions. Again, allowing for denser models by changing the regularization weight slightly improved the average correlation coefficient to 0.63 which is comparable to the 0.65 obtained by Bader and colleagues by Support Vector Regression [33]; when restricting to contact edges, we obtain 0.54 (about the same as the 0.56 of Bader and colleagues).

3.2 Model analysis

A key feature of DgSpi is that a model can be easily “opened up” to characterize the amino acid determinants of binding. To illustrate, we characterized the models trained at λ_{1, 2} = 25 across the 10 folds, computing the average strength of the vector 2-norms for the protein positions (i.e., V_i), peptide positions (i.e., V_j), and potentially interacting pairs (i.e., W_{i, j}). Fig. 4-top shows these values: the strengths of the vertex terms appear along the axes (x-axis for PDZ positions and y-axis for peptide positions), while the strengths of the PDZ:peptide edge terms appear in the heat map. As might be expected for interaction affinities, the position-based terms are relatively weaker, with most being less than 0.2. In contrast, more than 40 interaction terms have norms larger than this value with a large fraction of them between position 4 of the peptide and the protein. Fig. 4-bottom-left overlays these strong interactions on the the structure of the murine al-syntrophin PDZ (colored blue to light pink according to position) complexed with the peptide KESLV (colored in red). Fig. 4-bottom-right plots the edge strength (y-axis) against the distance of the corresponding residue pairs (x-axis) Interestingly, while most of the strong edges tend to be between positions less than 15 Å apart in the crystal, there are a few edges that are at a longer-range that appear consistently.

Figure 4.

Model analysis. (top) Average strength of the vector 2-norms for the PDZ positions (i.e., V_i), peptide positions (i.e., V_j), and potentially interacting pairs (i.e., W_i,j) in the model trained at λ_1,2 = 25. (bottom-left) Strong interactions highlighted in top panel, displayed on the NMR structure of the alpha syntropin PDZ (pdb id: 2PDZ). Color scheme same as above. (bottom-right) Average edge strength across 10 training splits plotted against distance in the 3D structure.

Despite the fact that no 3D structure information was used in learning the model, our method identifies several contacting residues as important for determining interaction specificity. This suggests that our data-driven approach might be capturing physically important interactions. To test this hypothesis further, we determined the average weight assigned to each possible amino acid pair for the top three interacting residue pairs across the models for the 10 training folds at λ_{1 2} = 25. Fig. 5 shows these weights with strong negative energies (i.e., favoring binding) in shades of blue and strong positive energies in shades of red. Darker shades correspond to stronger effects in both cases. The strongest interacting residue pair (PDZ position 54 : peptide position 2) strongly favors interactions between oppositely charged Arginine/Lysine in the PDZ and Glutamate in the peptide, while strongly penalizing Aspartic Acid/Glutamate : Glutamate between pairs of negatively charged residues, suggesting a strong electrostatic effect between these positions. Similar effects are seen in the other two interactions with Glutamate:Lysine favored between 48:1 and Aspartic Acid:Threonine penalized between 12:4. Our learning method can thus not only provide structural information as well as insights into the biochemical determinants of binding affinity.

Figure 5.

Average weights for amino acid pairs for the top three interacting residue pairs.

In summary, our results suggest that a large fraction of the binding affinity is due to interactions between a relatively small set of positions, not all spatially adjacent to the binding pocket. A larger set of weak interactions might have an additional small effect on binding; these might effect particular sub-families of PDZs or might reflect allosteric affects related to alternate conformational states of the protein previously described in this family [20].

3.3 From Sequence Determinants to Sequence Optimization and Design

The accuracy and simplicity of our model allows us to rapidly evaluate the binding affinity of any PDZ-peptide pair. We demonstrate the utility of this approach by “designing” optimal binders for a given PDZ sequence. Using a model learned from the entire training set with λ_1,2=25, we searched all five residue peptides and determined the top 10 peptides by their predicted ΔG for each PDZ sequence. Fig. 6-left shows the density of predicted binding energies of these PDZ:peptide pairs in blue; the density of natural PDZ-peptide pairs binding energies in our training dataset is shown in red. The predicted binding affinities of designed sequences are considerably lower than those of the natural sequences.

Figure 6.

(Left) Density of the predicted PDZ-peptide ΔG for the designed peptides (blue) and experimental ΔG for natural PDZ-peptide pairs (red). (Right) Sequence logos for the top 10 peptide designs for SHANK1 (top), CHAPSYN (middle), and PSD95 (bottom).

While the designed sequences include the natural substrates (at close to their predicted affinities, as discussed above), they also include a diverse array of alternatives. Fig. 6-right shows sequence logos of the top 10 designed peptides for three different PDZs. Even among these sets of top predicted binders, we see interesting diversity among the peptides, suggesting novel designs potentially worthy of experimental evaluation.

4 Discussion and Conclusion

We have developed a graphical model which is highly predictive of the ΔG of binding in protein:protein interactions, while providing an interpretable and designable basis for its predictions. The notion of modularity is fundamental to the idea of a graphical model. Hence these models form a powerful and natural tool to solve problems involving complex probability distributions over many random variables, like the ones here. Due to the natural equivalence between the graph structure of a model and the structure of spatial interactions in proteins, graphical models have seen considerable use in modeling various aspects of proteins: in recognizing structural motifs [19, 24, 25], in protein structure alignments [42], and in modeling dynamics [30]. A growing body of work using graphical models to capture correlated mutations in protein families has also seen substantial success in predicting residue-residue contacts in the protein structure [10, 12, 22, 26, 28], highlighting the power of these models.

While basing the modeling of ΔG on sequence and data is fundamentally different from structure-based predictors, which employ physics-based models and analysis of side-chain (and potentially backbone) conformations to assess interactions (e.g., [9, 16, 35]), we note that structure-based undirected graphical models have been used to predict ΔG [13, 14]. The integration of the structure-based approach and the sequence+data-based approach provides an interesting future direction. Our preliminary work on such integration for individual proteins [11] provides evidence that the two viewpoints can be complementary and enable better prediction than either alone.

The method we developed here could be applied to any pair of interacting protein families with a similar extent of quantitative binding data. Due to their size and easy availability, PDZ domains form a “model system” for studying protein-protein interactions, [4, 17, 34, 40]. They are involved in formation of protein complexes which are involved in cellular signal transduction and neural circuitry [34] and so make an interesting test case from the point of view of protein-engineering [7] and drug design [31].

We demonstrated that our models can be used to design novel peptides that interact strongly with a given PDZ domain. This approach could be extended using sampling or other inferential techniques to design a desired interaction, rather than only the peptide, and to scale up to larger sets of involved residues.