A reference map of the human binary protein interactome

Generation and characterization of HuRI

The newly established human ORFeome v9.1 covers 17,408 protein-coding genes (Extended Data Fig. 1a, b, Supplementary Table 1, 2), forming a search space (Space III) that encompasses over 150 million pairwise combinations (Fig. 1a). This search space more than doubles the space screened to generate HI-II-14 and represents the most comprehensive search space systematically screened for human PPIs. Limitations in PPI assay sensitivity can be overcome by using different assays^15,16 or different versions of the same assay^17,18. To maximize sensitivity we used three Y2H assay versions (Fig. 1b, Supplementary Table 3, 4) that showed good sensitivity and low false positive rates when benchmarked against gold-standard positive and random reference sets (PRSv1 and RRSv1, respectively)¹⁵, while detecting complementary PPI sets (Extended Data Fig. 1c, d, Supplementary Table 5). We assessed assay performance using a test space of ~2,000 by ~2,000 human genes² (Extended Data Fig. 1e, Supplementary Table 6). The Y2H versions were complementary, in that three screens for each of three versions doubled the number of detected PPIs and proteins relative to nine screens using a single version (Extended Data Fig. 1f, Supplementary Table 7).

To map the reference interactome, we performed nine screens of Space III, followed by pairwise verification by quadruplicate retesting and sequence confirmation. PPIs verified by two orthogonal binary PPI assays, MAPPIT¹⁹ and GPCA²⁰, were recovered at rates on par with high-confidence binary PPIs from the literature (each having ≥2 pieces of experimental evidence, with at least one from a binary assay type; Lit-BM) over a large range of score thresholds (Fig. 1c, Extended Data Fig. 1g, h, Supplementary Table 8). Each additional screen identified novel PPIs and proteins, with the largest gains obtained by switching assay versions (Fig. 1d, Extended Data Fig. 1i). The dataset, versioned HI-III-20 (Human Interactome obtained from screening Space III, published in 2020), contains 52,569 verified PPIs involving 8,275 proteins (Supplementary Table 9). Although our knowledge of the interactome remains incomplete, we refer to HI-III-20 as a reference map of the human binary protein interactome (HuRI) given its systematic nature, extensive coverage, and scale.

Molecular mechanisms can be more readily inferred from direct than indirect PPIs, yet the fraction of PPIs reported in various human protein interactome maps that are direct remains unknown. Structures from protein complexes with at least three subunits²¹ show that, within those complexes, more PPIs in HuRI correspond to direct biophysical contacts than do PPIs from Lit-BM (90% vs. 81%, P = 0.019, two-sided Fisher’s exact test, n = 121 and 410 for HuRI and Lit-BM, respectively) or from protein complex interactome maps (< 50%, P < 0.001 for all tested maps, two-sided Fisher’s exact test; Fig. 1e, Supplementary Table 10). Combining HuRI with all previously published systematic screening efforts at CCSB yields 64,006 binary PPIs involving 9,094 proteins (HI-union; Supplementary Table 11), which is approximately five-fold more PPIs than the entirety of high-quality binary PPIs curated from the literature (Fig. 1f, Extended Data Fig. 2, Supplementary Table 12–14). The union of Lit-BM and HI-union represents the most complete collection of high-quality direct PPI data available to date (http://interactome-atlas.org).

Complementarity of the three Y2H versions might stem from steric constraints that differ between protein fusions used in the assays. Integrating HuRI with structures of PPIs²¹, we observed reduced sensitivity of Y2H assays where the interaction interface was close (< 20Å) to whichever terminus was fused to the Gal4 activation domain (AD; Extended Data Fig. 3a, b, Supplementary Table 15). PPIs found in multiple screens had larger interaction interfaces (P = 0.03, two-sided permutation test, n = 234) and corresponded more often to direct PPIs within, rather than between, protein complexes (P = 3 × 10⁻¹⁸, two-sided Fisher’s exact test, n = 1,817), however, HuRI PPIs found in a single screen were observed to have precision as high as those found in multiple screens (Extended Data Fig. 3c–g, 4, Supplementary Table 16–18, Supplementary Note 1). These results reinforce previous observations^12,22 that the protein interactome might be dominated by weaker and more transient PPIs that are difficult to detect, as indicated by the fact that the majority of PPIs in HuRI were found in only one screen (Extended Data Fig. 3h, i). This is reflected in an increased estimate of the binary protein interactome size (Extended Data Fig. 3j), so that HuRI is estimated to represent 2–11% of the binary protein interactome²³ (Supplementary Note 2).

Functional relationships in HuRI

Based on the observation that HuRI is enriched in direct PPIs, we also hypothesize that proteins in HuRI with similar interaction interfaces tend to share interaction partners. For example, retinoic acid receptors RXR-γ and -β (Extended Data Fig. 5a, left panel) share previously reported interaction partners involving binding to retinoic acid receptor RAR types²⁴ and oxysterol receptors NR1 group H types²⁵. We derived a profile similarity network (PSN) from HuRI (Supplementary Table 19), and found that the number of pairs of proteins in HuRI with similar interaction profiles is higher than random (P < 0.01, one-sided empirical test; Extended Data Fig. 5b) and proteins of overall higher sequence identity tend to exhibit higher interaction profile similarities (P < 0.01, one-sided empirical test; Extended Data Fig. 5c). We also found that proteins with a tendency to share interaction partners often have interaction interfaces that are similar, as opposed to complementary, and therefore tend not to interact with one another, except where proteins originate from a common ancestor that could self-interact²⁶ (Extended Data Fig. 5d). Indeed, only 5% of the protein pairs found to interact in HuRI share more than 10% of their interaction partners. Both HuRI PPIs and the PSN are enriched for links between proteins of similar function (P < 0.01, one-sided empirical test; Fig. 2a, Supplementary Table 20, Extended Data Fig. 5e, f) and both contain more functional modules²⁷ than our previously published interactome maps^2,14 (Fig. 2b, Extended Data Fig. 5g). Thus, HuRI and the PSN are complementary maps of functional relationships between proteins.

As shown above, global sequence identity between two proteins can be indicative of shared interaction interfaces, however, it likely fails to identify pairs of proteins whose shared interaction interface is small. Indeed, 50% (502) of all protein pairs in HuRI with interaction profile similarities ≥ 0.5 exhibit ≤ 20% sequence identity, so that functional relationships identified by the PSN are not necessarily identifiable by sequence identity. One such protein pair is the endoplasmic reticulum (ER) transmembrane protein TMEM258 and the uncharacterized protein C19ORF18, which have only 10% sequence identity but share 80% of their interactors (Extended Data Fig. 5a, right panel). TMEM258 catalyzes the first step in N-glycosylation of proteins in the ER and might play a role in protein translocation across the ER²⁸. Roles in protein transport and ER function have also been ascribed to two of the four shared interaction partners, ARL6IP1²⁹ and IER3IP1³⁰, suggesting that C19ORF18 and potentially the other two shared interaction partners (MFSD6 and AC012254.2) contribute to ER-related functions of protein maturation and transport.

Uncharted disease-related interactions

Unlike Lit-BM, HuRI was generated by systematically testing protein pairs for interaction. While Lit-BM is highly biased towards the most studied genes², HuRI covers the genome-by-genome space more uniformly and at increased depth compared to Lit-BM and our previous screening efforts (Fig. 3a, Extended Data Fig. 5h). Interestingly, we find that the agreement between Lit-BM and HuRI is highest among the best-studied genes, where Lit-BM is most complete and where ~40% of the PPIs in HuRI have been previously identified (Fig. 3b). HuRI substantially expands the number of biomedically interesting genes for which high-quality direct PPI data is available (Fig. 3c, Extended Data Fig. 5i), and finds new interaction partners for these genes in previously uncharted regions of the protein interactome (Fig. 3d, Extended Data Fig. 5j).

Essential proteins were often found to have significantly more interaction partners³¹. However, correlation between two variables does not necessarily imply causality, especially where there are other confounding variables (Extended Data Fig. 5k). We find that protein popularity (measured by publication count) and endogenous expression level strongly correlate with: i) each other; ii) the number of interaction partners (‘degree’) in protein complex and literature-curated protein interaction networks but not in HuRI; as well as iii) gene properties such as essentiality, age³², fitness effect³³ and loss-of-function intolerance⁶ (Extended Data Fig. 5l, Supplementary Table 21). After correcting for popularity and expression level, we found substantially reduced correlations between interaction degree and other gene properties, including gene essentiality (Fig. 3e, Extended Data Fig. 5m, n). In line with previous observations³⁴, this suggests that correlations between degree and essentiality, age, loss-of-function intolerance, and fitness effects are confounded by underlying expression and study biases, and thus may not reflect causal relationships. These results highlight the value of HuRI as a uniformly-mapped reference for the study of systems properties of genes and networks.

Compartment-specific roles of PPIs

Proteins are localized to specific compartments, carrying out functions that depend both on the subcellular environment and the local PPI network. Despite available proteome-wide datasets on the localization of individual proteins⁵, experimental determination of cellular localization-specific PPI networks remains challenging. We find that proteins localizing to a diverse range of subcellular compartments are evenly represented in HuRI (Extended Data Fig. 6a), suggesting that cellular localization-specific PPI networks can be inferred for many different cellular compartments via integration of HuRI with available protein localization data.

Extracellular vesicles (EVs) have been studied intensively using proteomics approaches³⁵. However, our understanding of the molecular mechanisms that lead to protein recruitment into EVs and subsequent secretion remains limited. The subnetwork of interactions between EV proteins (Extended Data Fig. 6b) shows significantly higher connectivity in HuRI than in degree-controlled randomized networks (P < 0.001, one-sided empirical test; Extended Data Fig. 6c), enabling prediction of EV recruiters using the number of EV interaction partners. Seven of the top 21 most connected proteins in this EV network have established roles in EV biogenesis or cargo recruitment³⁶. SDCBP (syntenin-1) functions in ESCRT-dependent EV generation and its knockout shows reduced EV production³⁷. SDCBP has 48 PPIs with other EV proteins and is frequently detected in EVs, suggesting that it regulates recruitment of interacting proteins into EVs. To test this hypothesis (Fig. 4a), we knocked out SDCBP in the U373vIII cell line (Extended Data Fig. 6d). Of six SDCBP partners detected in the U373vIII EV proteome, three (CALM1, CEP55, HPRT1) displayed significantly reduced protein levels in EVs in the SDCBP knockout line (P < 0.05, one-sided empirical test, depletion < 0.66; Fig. 4b). In contrast, only 15% of the non-interaction partners of SDCBP were reduced (P = 0.042, one-sided empirical test; Extended Data Fig. 6e). Thus, SDCBP may play a role in the recruitment of proteins into EVs, highlighting the potential value of HuRI in studying protein function within specific subcellular contexts.

Proteins often change subcellular localization, e.g. as part of their maturation process or in response to external or internal signaling events. These dynamics are difficult to comprehensively capture at proteome-scale, yet existing efforts have already revealed numerous proteins that localize to multiple subcellular compartments with some pairs of compartments exhibiting significantly more crosstalk than others⁵. Despite a tendency for interactions in HuRI to link proteins localized to the same compartment (P < 0.01, one-sided empirical test; Fig. 2a), a considerable number of interactions were identified between proteins never reported to co-localize. To explore whether lack of co-localization of interacting proteins in HuRI might originate from incomplete localization data, we assessed whether these non-co-localized interacting proteins tend to reside in compartments with significant crosstalk and observed a significant positive correlation (P < 0.001, one-sided empirical test; Extended Data Fig. 6f–h). This suggests that HuRI could prove useful in predicting protein localization dynamics.

Principles of tissue-specific function

Despite recent advances in systematic genome-wide identification of tissue-preferentially expressed (TiP) genes⁴ (Extended Data Fig. 7, Supplementary Table 22), we lack a concrete understanding of how the surprisingly small set of TiP genes operate together and coordinate their activity with the core ‘housekeeping’ machinery to mediate tissue-specific functions. Insights can be obtained from investigation of the tissue-specific network context of TiP proteins, inferred from integrating protein interactome data with tissue transcriptomes (Supplementary Table 23). However, we should not expect uniform coverage of TiP proteins with PPIs using experimental methods that demand expression of both partners within a single immortalized cell line or result from screening an incomplete ORFeome. Indeed, contrary to protein complex^11–13 and literature-curated interactome maps³⁸ as well as our previously published binary PPI datasets^2,14, we find that TiP proteins are well-represented in HuRI (Fig. 5a, Extended Data Fig. 8a).

Restricting HuRI to PPIs between proteins expressed in the same tissue, we observe that TiP proteins engage in as many PPIs and are as central as more uniformly expressed proteins (Extended Data Fig. 8b), contrary to previous observations using literature-curated PPI networks^39,40. This result, paired with the fact that PPIs mediated by a TiP protein are effectively also tissue-specific, leads to the finding that the fraction of tissue-specific PPIs in the protein interactome as characterized by HuRI is higher than that of tissue-specific genes in the expressed genome, indicating that substantial information on tissue-specific functions can only be obtained from the interactome. The opposite is observed for Lit-BM, likely owing to its bias against TiP genes (Extended Data Fig. 8c).

To investigate the local network neighborhoods of TiP proteins within their respective tissue contexts, we used HuRI to derive protein interactome maps for 35 tissues^4,41, each of which contains about 25,000 PPIs (Supplementary Table 24, Extended Data Fig. 8d). Within each tissue PPI network, we focused on the interactions involving at least one TiP protein (Fig. 5b). The TiP PPI networks show extensive interactions between TiP and non-TiP proteins, but with few TiP-TiP PPIs. Despite significant enrichments for HuRI to link proteins that work in the same biological process, TiP-TiP PPIs, as highlighted for brain in Fig. 5c, are not enriched, nor is the average shortest path among TiP proteins shorter than in degree-controlled randomized networks (P > 0.05, empirical test). Using either metric, TiP proteins were found to be significantly close to each other in only six of 35 tissues. In four of these six tissues, enrichment for network proximity was driven by clusters of specifically expressed keratins or late-cornified envelope proteins (Extended Data Fig. 8e). These results support a model in which tissue-specific functions emerge through interactions between TiP proteins and more uniformly expressed members of the basic cellular machinery, presumably modulating and adapting common cellular processes for cellular context-specific needs⁴².

One biological process with both cell-type and developmental stage-specific homeostatic roles is apoptosis. We used HuRI to identify proteins with interaction partners that were enriched for known apoptosis regulators (Supplementary Table 25). Five proteins among the top ten predictions had previously demonstrated roles in apoptosis (Supplementary Note 3). Among three TiP genes predicted to be implicated in apoptosis (Extended Data Fig. 8f, g), we further examined OTUD6A. Abundance of OTUD6A negatively correlated with time-of-death after addition of TRAIL (TNF-related apoptosis-inducing ligand; P = 0.012, two-sided, empirical test, n = 40 cells), but not after expression of OTUD6A alone (Extended Data Fig. 8h, Supplementary Table 26), suggesting that OTUD6A participates in the apoptosis pathway but is not an inducer of cell death. This and other evidence (Extended Data Fig. 8f, i, Supplementary Note 3) suggests that OTUD6A exerts an apoptosis sensitization function via transcriptional activation in a haematopoietic cellular context.

Mechanisms of tissue-specific diseases

Many Mendelian diseases display tissue-specific phenotypes, rarely explained by tissue-specific expression of genes with disease-associated mutations⁴³ (Fig. 5d, Extended Data Fig. 9a). Such mutations broadly or specifically affect PPIs involving the mutated protein⁴⁴. Perturbations of PPIs between uniformly expressed disease-associated proteins and TiP proteins in the affected tissues have been suggested to underlie the tissue-specific phenotypes of those diseases⁴³. In HuRI-derived tissue PPI networks, we find 130 such PPIs involving 63 distinct non-TiP disease-causal proteins and 94 TiP proteins. Although we see no enrichment for PPIs between causal proteins and TiP proteins (Extended Data Fig. 9b, Supplementary Note 4), this does not rule out the possibility that perturbations of some of these interactions mediate tissue-specific phenotypes of Mendelian diseases.

To explore this hypothesis, we experimentally tested whether pathogenic variants associated with Mendelian diseases were able to perturb these PPIs. Of ten causal proteins tested, seven showed perturbation of PPIs to preferentially expressed interaction partners in the corresponding “disease tissues” (Fig 5e, Extended Data Fig. 9c, d, Supplementary Table 27, 28). One example is PNKP (polynucleotide kinase 3’-phosphatase) whose mutations have been associated with microcephaly, seizures, and developmental delay. The pathogenic PNKP mutation Glu326Lys affects neither the DNA kinase nor DNA phosphatase activity of PNKP, rendering the mechanism of pathogenicity unclear⁴⁵. We observed that Glu326Lys perturbed PPIs with two partners preferentially expressed in the brain, SYNGR1 and TRIM37, whereas a benign control mutation Pro20Ser⁴⁶ did not affect any PNKP PPIs (Fig. 5f, Extended Data Fig. 9c, e). TRIM37 facilitates DNA repair⁴⁷, suggesting a potential mechanism by which perturbation of this interaction could affect the brain-specific DNA repair function of PNKP. In other examples, HuRI identified CTNNA3 and SUCLA2 to have respective TiP interaction partners TRIM54 and ARL6IP1 (Extended Data Fig. 10), which cause similar diseases with overlapping symptoms^48,49, supporting the relevance of these interactions in the physiopathology. Overall, this study yields hypotheses of molecular mechanisms for otherwise unexplained tissue-specific phenotypes of seven Mendelian diseases (Extended Data Fig. 9d) and demonstrates the utility of HuRI as a reference to study biological mechanisms within specific disease contexts.

Discussion

Here, we present HuRI, a systematically generated human protein interactome map with more than 50,000 PPIs of high biophysical quality. While HuRI displays highly significant overlap with known functional relationships, the cellular function of most individual PPIs remains to be elucidated. We show that follow-up studies on the function of proteins and PPIs can be guided by integration of HuRI with contextual genome, transcriptome and proteome data to infer the cellular context in which subnetworks of PPIs operate together to mediate a function. With advances in single cell transcriptomics⁸ as well as systematic determination of subcellular protein localization, inference of functional PPI subnetworks that are specific for a given cellular state will further increase in precision. However, a priori removal of PPIs from HuRI because they are not currently known to function together in a physiological cellular context could discard data that can help, e.g., to understand the functional consequences of dysregulated gene and protein expression causing disease.

Despite our extensive screening, many PPIs remained undetected as Y2H, like all characterized assays^15,16,18, has limited sensitivity, detecting no interactions for half of the tested proteins. This negative result can guide the design of future interactome mapping efforts to target these proteins. Due to the limitations of Y2H, we expect HuRI to be depleted for PPIs that depend on post-translational processing of human proteins that the yeast cell is unable to catalyze or that require additional partners to stabilize the interaction. Screening only one isoform per gene also misses interaction partners specific to alternative spliceforms⁵⁰. Accurate estimation of the total size of the interactome remains challenging. PPIs display a continuum of binding strength or stability that, along with other factors, could underlie a continuum of detectability, as this study suggests. Furthermore, the results obtained by us and others¹² indicate that very stable and functionally conserved PPIs constitute a minority of the interactome.

While incomplete, HuRI’s proteome and interactome coverage and uniformity enable its use as a reference for the study of most aspects of human cellular function. Efforts to further complete this reference will require development of new technologies as well as integration with complementary reference maps of protein complex assemblies¹³. Although multiple challenges remain to be solved for a complete and context-specific map of protein functions, interactions, and higher-level organization, HuRI provides an unbiased genome-scale scaffold with which to coordinate this information as it emerges.

Extended Data

Supplementary Material