The interaction between two proteins is a geometric and electrostatic match between two polypeptide surfaces that results in a stable set of bonds between amino acid side chains or backbone atoms. The interacting amino acids are often part of conserved sequence features such as domains or short linear motifs that constitute the interaction site between the two proteins. Despite the increased coverage and sensitivity of experimental techniques for detecting protein interactions [1–6] (reviewed in [7]), elucidating the precise interacting residues remains experimentally difficult. In most cases, all that is known about an interaction is the identity of the two interacting proteins, with little information about the underlying binding site. However, detailed knowledge of interaction specificity is important for understanding reaction mechanism, interaction prediction, and drug development.

Interacting domains are autonomous structural elements that exhibit distinct binding specificity to a multitude of target polypeptides. Such domains act as independent elements that can be “plugged” into a new protein and thereby introduce new functionality to the emerging protein [8]. From an evolutionary perspective, such rearrangements and the multiplication of existing conserved domains is a likely mechanism by which organisms generate new proteins, pathways, and novel functionalities [9,10]. Several protein interaction prediction methods exploit the conservation of protein-binding interfaces by identifying domain pairs that consistently co-occur in interacting proteins or coevolve, which are then used to predict new interactions [11–16]. Structure-based prediction methods use known protein complexes to model interactions between proteins that are homologous to the complex components [17,18]. Other prediction methods use integrative approaches that incorporate interaction experiments with additional functional information such as correlated expression level, common functional annotation [19,20], and cross-species comparisons [21]. Alternative approaches attempt to identify correlated sequence motifs that represent generic interacting sequence elements that may or may not be components of conserved domains [22–25]. In a few limited cases, detailed experimental data are used to generate high-resolution definition of domain binding profiles; however, such information is available only for a small number of domains [26,27].

Our primary objective is to predict interaction between proteins strictly from sequence information. Our approach is based on identifying the binding specificity of interacting domains that can then be used to predict new interactions. Here, we use existing physical interaction data to derive sequence profiles of the binding sequences that are presumed to determine the binding specificity of interacting domains. Our method, called domain–motif interactions from structural topology (D-MIST), is based on a two-step approach. First, potential domain-binding motifs are extracted from structural data. Second, these motifs are converted to sequence profiles in the form of position-specific scoring matrices (PSSMs). These PSSMs are derived using a subset of experimentally determined binary interactions that contain the domain of interest (Figure 1). Gibbs sampling, seeded with the motif extracted from structural data, is used to generate a PSSM from similar sequences that occur in a subset of established interacting proteins. We used the domain-binding profiles to predict protein interactions in yeast. The predictions were compared to a hidden set of known interactions reported in the literature, and several predicted interactions were confirmed directly by in vivo coprecipitation experiments.

The library of 3-D structures of protein complexes contains a detailed description of the binding interfaces between interacting proteins that include atom contacts and residue side-chain interactions [28]. Using more than 10,000 structural complexes, we identified the domains in the binding sites and extracted their associated sequence motifs on the opposing chain. Interacting residues were defined as two residues on opposite polypeptide chains separated by a maximum of 5 Å (Figure 1A). On average, each domain had two spatially separated interacting sequence motifs per interaction. Most domains were present in multiple 3-D structures in a variety of conformations, resulting in varied interacting sequence motifs with different levels of similarities.

The binding specificity of a domain is determined by a combination of physiochemical properties and structural constraints at the binding site that can be satisfied by multiple variations of the consensus sequence motif [29]. The interacting sequence motifs extracted from the protein structures represent a first approximation of the binding specificity of the interacting domains, but do not represent the full evolutionary variations of the residue–residue interactions available in one binding topology. A more informative representation of the possible motif variations is a sequence profile in the form of a PSSM that captures the compositional variance by assigning probabilities to each amino acid at each position. These sequence variations of the binding profiles can be learned from proteins that are known to interact through the same domain.

We collected a set of 87,894 nonredundant protein interactions from four databases containing binary protein interactions from multiple species. Interactions derived from structural studies were excluded to preclude self-identification, as well as high-throughput protein complexes identification experiments [30,31] (see Methods). Gibbs sampling [32] was used to learn the PSSM binding profiles for a specific domain by sampling positions in the set of proteins that interact with proteins that contain the domain of interest. The majority of the proteins in the learning set are assumed to interact through the common domain, and the generated PSSM will represent its binding profile (Figure 1B). Gibbs sampling enables the incorporation of prior knowledge about the length and composition of the binding profiles. The motifs identified in the 3-D structural analysis were used as prior knowledge in seeding the profile detection step to bias the sampling towards similar sequence regions. The result is a set of sequence PSSMs that represent the binding profiles of the interacting domains (Text S1).

The learned PSSMs were used to predict interactions for 703 yeast proteins with domains for which we successfully derived binding profiles. A physical interaction was predicted between proteins containing interacting domains and proteins with one or more of the interacting profiles associated with those domains (Figure 1C). A total of 18,459 interactions were predicted between 2,313 proteins (Dataset S1). We compared the predicted interactions to a comprehensive list of physical and genetic yeast interactions extracted from the literature [33] and found that 609 predicted interactions have reported experimental evidence (∼3%; p = 1.0 × 10⁻¹³; Figure S1). We note that 591 predicted interactions were found in both the 87,894 set of interactions used for the PSSM derivation and in the set of yeast literature curated interactions (∼32,000). However, none of the 609 predicted interactions that have supporting evidence in the literature overlap with those common 591 interactions. We did not incorporate additional experimental information such as cellular localization, functional annotation, surface accessibility, or gene expression data that would likely improve our prediction accuracy given that our primary goal was to predict novel interactions exclusively from sequence information.

Experimental verification of a subset of the predicted interactions was performed by a one-step immunoaffinity purification of one of the two interaction partners, followed by mass spectrometric identification of associated proteins (IP-MS) as previously described [31]. The IP-MS method confirmed 37 predicted interactions, including 23 novel interactions (Figure 2). As a second means to experimentally verify our predictions, we immunoprecipitated one protein in the interacting pair, followed by antibody detection of the second protein (IP-western), also as described in [31]. The IP-western method reaffirmed five of the interactions confirmed by IP-MS (yellow edges; Figure 3) and identified an additional four novel interactions (green edges; Figure 3). We note that six interactions confirmed by the IP-MS approach were not detected by IP-western (red dashed edges; Figure 3); this discrepancy may be due either to nonspecific interactions detected by IP-MS or to interference of the second epitope tag with some interactions and/or expression levels in vivo. Of the 18 predictions that were tested by IP-western, nine novel interactions were confirmed, and a total of 30 new interactions were identified by both the IP-MS and IP-western methods.

As noted previously, we excluded additional experimental evidence, such as localization and expression data from our prediction method. Although additional experimental information and functional annotation would likely improve prediction accuracy, it may also limit predictions only to those proteins with prior experimental or functional information. In addition, the use of functional annotation such as Gene Ontology terms (assigned by human experts or predicted computationally) in a prediction method will penalize predicted interactions between proteins with unrelated functions. Therefore, it restricts the ability to predict interactions between apparently unrelated proteins that could illuminate new cellular functions [43].

The D-MIST method for identifying domain-binding modules is currently limited in a number of ways. The first limitation is the availability of detailed binding information, as attained primarily through structural studies and peptide-based approaches such as phage display [44] and random peptide libraries [45]. In addition, several studies have concluded that the repertoire of protein structures in the Protein Data Bank is significantly biased in that trans-membrane and disordered domains are underrepresented due to limitations in structure determination [46,47]. Consequently, D-MIST analysis that depends on structural representation of protein interactions is similarly biased. The existing detailed examples of interactions are therefore sparse and noncomprehensive, with only a small subset of all possible domains that is represented. The second limitation is that the derived motifs do not represent the entire repertoire of all possible domain-binding sequences, even for those domains where structural data exist. The third limitation arises from the statistical framework of the Gibbs sampling method that requires a sufficient number of proteins to sample from in order to converge towards a meaningful PSSM. We restricted the analysis to domains with five or more putative interactors, thereby excluding domains that are infrequently found in our set of protein interactions. Fourth, some domains are not amendable to this type of analysis due to the diverse nature of their binding motifs that lack sequence conservation [29]. Last, many interactions are governed by posttranslational modifications or precise physiological states, which may also hamper the accuracy of D-MIST predictions. Despite the above limitations, we have shown that novel protein interactions can be predicted strictly from primary sequence information. D-MIST not only predicts interactions between proteins but also provides sequence level predictions about the binding sites that can be verified experimentally. Predicting protein interactions without the need for additional information or prior experiments is particularly valuable when studying uncharacterized proteins and for predicting interactions in poorly studied organisms where typically only sequence information and predicted open reading frames are available. The sole dependence on sequence information allows for interaction prediction in other organisms without further modifications to the method or input datasets. With the advent of structural genomics initiatives [48], the power of the D-MIST approach will certainly increase.