PDZ domain binding selectivity is optimized across the mouse proteome

Abstract

PDZ domains have long been thought to cluster into discrete functional classes defined by their peptide-binding preferences. We used protein microarrays and quantitative fluorescence polarization to characterize the binding selectivity of 157 mouse PDZ domains with respect to 217 genome-encoded peptides. We then trained a multidomain selectivity model to predict PDZ domain-peptide interactions across the mouse proteome with an accuracy that exceeds many large-scale, experimental investigations of protein-protein interactions. Contrary to the current paradigm, PDZ domains do not fall into discrete classes; instead, they are evenly distributed throughout selectivity space, which suggests that they have been optimized across the proteome to minimize cross-reactivity. We predict that focusing on families of interaction domains, which facilitates the integration of experimentation and modeling, will play an increasingly important role in future investigations of protein function.

Fig. 1

(A) Strategy for constructing a multidomain selectivity model for mouse PDZ domains. Protein microarrays were used to test all possible interactions between 157 mouse PDZ domains and 217 genome-encoded peptides. Array positives were retested and quantified by FP, thereby correcting array false positives. The resulting data were used to train a predictive model of PDZ domain selectivity. The model highlighted putative array false negatives, which were tested by FP, and the corrected data were used to retrain the model. After three cycles of prediction, testing, and retraining, the refined model was used to predict PDZ domain–protein interactions across the mouse proteome. (B) Representative images of protein microarrays, probed with fluorescently labeled peptides. PDZ domains were spotted in quadruplicate in individual wells of 96-well microtiter plates. (Four wells were required to accommodate all of the domains.) The red images (Cy5) show the location of the PDZ domain spots. The green images show arrays probed with a promiscuous peptide derived from Kv1.4 (left) and a selective peptide derived from ephrin B1/2 (right). (C) FP titration curves obtained for the array positives identified in (B).

Fig. 2

(A) Graphical view of the training-set data. K_d’s of FP-confirmed positives are represented by colors, ranging from high affinity (red) to low affinity (light blue). Array negatives are shown in black, and FP-confirmed negatives are shown in dark blue. Numerical values are provided in table S3. (B) Performance of the MDSM on the training set, with m set to 5. True positives are shown in red, false positives in green, true negatives in blue, and false negatives in yellow. (C) Graphical representation of the MDSM parameters, θ_i,p,q. Positive contributions to discriminative binding are graded from black to yellow, and negative contributions are graded from black to light blue. Numerical values are provided in table S4. Single-letter abbreviations for the amino acid residues are as follows: A, Ala; C, Cys; D, Asp; E, Glu; F, Phe; G, Gly; H, His; I, Ile; K, Lys; L, Leu; M, Met; N, Asn; P, Pro; Q, Gln; R, Arg; S, Ser; T, Thr; V, Val; W, Trp; and Y, Tyr. (D) Tight clusters embedded in the bipartite interaction network between the 74 PDZ domains and the 217 training-set peptides. (E) ROC curves for three versions of the MDSM, obtained with the test set of 48 peptides. The best performance was obtained after smoothing over both PDZ domains and amino acids. The performance of each version of the MDSM with m set to 5 is indicated with an arrow.

Fig. 3

(A to C) Correlations between z scales and model parameters at position −4 for three PDZ domains. (A) z₁ positively correlates with θ_−4,q for Dlgh3 (1/1). (B) z₂ negatively correlates with θ_−4,q for Magi-1 (4/6). (C) z₃ negatively correlates with θ_−4,q for MUPP1 (10/13). (D) Correlation matrix between the model parameters for all 74 PDZ domains at positions −4, −3, −2, and −1 and the first three z scales of the amino acids. (E) Percentage of variance in the correlation matrix that is explained by the 12 principal axes identified through singular-value decomposition. (F) Graphical representation of the first three principal axes, used to define PDZ domain selectivity space. (G) Distribution of the 74 PDZ domains in selectivity space. Selected PDZ domains are shown, representing class I domains [PSD-95 (1/3) and Shank3 (1/1)], class II domains [Grip1 (6/7) and PDZ-RGS3 (1/1)], and class III domains [nNOS (1/1)]. Erbin (1/1), which has been described as a dual-specificity domain, lies between the class I and class II domains. (H) Correlation between pairwise sequence divergence of PDZ domains and their pairwise distances in selectivity space. Sequence divergence was obtained from pairwise alignments performed with Vector NTI version 8 (InforMax, Invitrogen Life Science Software, Frederick, Maryland), using the blosum62mt2 matrix. Pairwise distances in selectivity space are Euclidean distances obtained from the three-dimensional plot in (G).