A protein interaction landscape of breast cancer

Abstract

Cancers have been associated with a diverse array of genomic alterations. To help mechanistically understand such alterations in breast invasive carcinoma, we have applied affinity-purification mass spectrometry to delineate comprehensive biophysical interaction networks for 40 frequently altered breast cancer (BC) proteins, with and without relevant mutations, across three human breast cell lines. These networks identify cancer-specific protein-protein interactions (PPIs), interconnect and enrich for common and rare cancer mutations, and are substantially rewired by the introduction of key BC mutations. Our analysis identified PIK3CA-interacting proteins which repress AKT signaling and has uncovered USP28 and UBE2N as functionally relevant interactors of BRCA1. We also show that the PP1 phosphatase regulatory subunit, Spinophilin, interacts with and regulates dephosphorylation of BRCA1 to promote DNA double-strand break repair. Thus, PPI landscapes provide a powerful framework for mechanistically interpreting disease genomic data and can identify valuable new therapeutic targets.

One Sentence Summary

Large-scale protein interaction maps of breast cancer genes provide a framework to recognize previously unidentified oncogenic drivers.

Breast cancer (BC) is the most common malignancy in women and the second leading cause of cancer‐related death in the United States, where an estimated 281,550 women and 2,650 men will be newly diagnosed in 2021 (1). The disease has been divided into different subtypes, based largely on the presence or absence of three key proteins: estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2/ERBB2). Despite this and much additional heterogeneity at the molecular level, the majority of BC patients are treated using untailored chemotherapy or hormone therapies. Therefore, there is an urgent need to develop targeted therapies matched to the specific molecular alterations in a tumor, with the goal of achieving better efficacy and avoiding unnecessary treatment.

Advances in DNA sequencing technology have enabled the widespread analysis of breast tumor genomes, creating a catalog of genetic mutations that may initiate or drive tumor progression (2, 3). In addition to common mutations in well-known cancer genes, such as TP53 and PIK3CA, breast cancers harbor many additional mutations, each of which are seen rarely across the patient population. A key question is how these less common alterations, dispersed across a multitude of genes, elicit pathologic consequences and patient outcomes. An important answer may lie in understanding how individual gene mutations converge on multi-gene functional modules, including the signaling pathways orchestrating cell proliferation and apoptosis and DNA repair complexes (4–15). To broadly enable a pathway-based understanding of cancer, a prerequisite is to generate general and comprehensive maps of cancer molecular networks in relevant malignant and premalignant cell contexts.

Protein-protein interaction mapping of breast cancer drivers

Using the TCGA analysis of BC tumors (2, 3), we collected a panel of genes that are associated with molecular alterations, in terms of cell growth, proliferation, and genome stability in BC and used this list to guide the selection of 40 proteins for generation of protein-protein interaction (PPI) networks. Our selected targets included proteins with well-known roles in BC (e.g., TP53, PIK3CA, CDH1, and BRCA1) as well as less-well characterized proteins with recurrent mutations (e.g., CHEK2, MLH1, SMARCB1, and XPC) (16–19). Importantly, 93% of BC tumors in TCGA harbor an alteration in one or more of these 40 genes (Fig. 1A). Three breast cell lines derived from human mammary epithelium – MCF7 (ER+, luminal A subtype), MDA-MB-231 (ER−, PR−, HER2− triple-negative TN subtype) and MCF10A (non-tumorigenic mammary epithelial cells) – were selected for the PPI analysis because they have been shown to replicate therapeutically relevant responses found in BC tumors (20), their RNA profiles are highly correlated with those of BC tumors (21) and ER+ and TN subtypes together account for approximately 90% of BC patients (22). We reasoned that comparing protein networks among ER+, TN and non-tumorigenic models would allow study of how PPI networks are altered between normal and tumorigenic backgrounds as well as influenced by different mammary epithelial lineages.

Figure 1. Overview of protein-protein interaction mapping in breast epithelial cells.

(A) The gene alteration frequencies from the breast invasive carcinoma (TCGA Firehose Legacy) dataset for the 40 genes selected as AP-MS baits in this study. Each vertical grey column represents a patient, such that various genetic alterations of 40 genes in a given patient are indicated as seen at the bottom. In total, 93% (1028 out of 1108) of BC patients have non-synonymous mutation, chromosomal copy-number alteration (CNA), or mRNA/protein expression alteration in one or more of these 40 genes. Genes analyzed for both wild-type and mutant proteins are highlighted in yellow. Existing gene alterations in MCF7 and MDA-MB-231 are shown on the right. (B) The experimental workflow in which each bait was expressed in biological triplicate in 3 cell lines and subjected to AP-MS analysis. (C) Majority (78%) of the high-confidence PPIs identified in this study are not represented in a panel of public PPI databases (CORUM, BioPlex 2.0, IMEx and BioGRID low throughput & multivalidated). (D) The frequency of non-synonymous mutations, chromosomal CNVs, or mRNA expression alterations of 10,000 random size-matched permutations taken from the list of genes detected in the global protein abundance analysis. The white circle indicates the median of the random sampling, and the grey bar represents ±1 standard deviation. The frequency of alterations found in the prey retrieved in our PPI dataset is indicated in the black circle. (E) Venn diagram illustrating the overlap of PPIs (PPI score ≥ 0.9) across the 3 cell lines. PPI score is an average of the PPI confidence scores calculated from compPASS and SAINTexpress (see methods for details). The frequency of non-synonymous mutations of the prey genes in each sector of the Venn diagram was compared to those of 10,000 random size-matched permutations as in (D). The p-values for mutation enrichment in each prey set were shown in a color scale, where a stronger red represents more significant mutation enrichment.

To generate PPI maps, “bait” proteins were cloned into triple FLAG-tagged lentiviral vectors (table S2), individually transduced into each cell line and expressed in biological triplicate via a doxycycline inducible promoter (Fig. 1B). Cells were harvested after approximately 40 hr doxycycline-induction, and anti-FLAG tag-based affinity purification was performed followed by mass spectrometry to detect interacting “prey” proteins in an unbiased manner. We employed two PPI scoring algorithms to quantify high-confidence interacting proteins: compPASS (23, 24) and SAINTexpress (25) (Materials and Methods). Although overexpression of tagged baits may generate possible artifacts, it permits the capture of PPIs in a highly sensitive and reproducible manner with a relatively uniform background signal, which can be removed by the appropriate algorithms. We and many others have used overexpression AP-MS to characterize thousands of physiologically relevant PPIs (26–30). Using this approach, we identified a total of 589 high-confidence PPIs involving 493 prey proteins (Fig. 1C, fig. S1A–B, tables S3 and S4). Collectively, 78% of the BC PPIs we identified were not previously reported in protein-protein interaction databases (CORUM, BioPlex 2.0, IMEx and the low throughput and multivalidated BIOGRID) (Fig. 1C). The high percentage of new interactions may reflect cell type-enriched PPIs as nearly all systematic protein-protein interaction analyses to date have been performed in HEK293T or HeLa cell lines (24, 26, 27, 31, 32).

PPIs often suggest functional relationships among proteins that work together to regulate a specific cellular process. Previously, we and others had found a significant enrichment of frequently mutated proteins in large PPI repositories (8–12, 33–35). Similarly, we investigated whether our BC PPI network showed enrichment for three major types of alterations – non-synonymous mutations, chromosomal CNVs, and mRNA expression alterations – documented in the BC TCGA cohort. Accordingly, we calculated the average frequency of each alteration for prey proteins detected in our PPIs, compared to background expectation (fig. S1C and table S5). We observed that BC-associated mutations were significantly enriched in BC PPIs, but that CNVs and mRNA expression alterations were not (Fig. 1D), a trend that was also observed in head and neck squamous cell carcinoma (Swaney et al., accompanying paper). Furthermore, we found enrichment of tumor mutations in preys detected specifically in either of the two cancer cell lines (MCF7, MDA-MB-231) among BC patient tumors but not in the preys of non-cancerous MCF10A cells (Fig. 1E and table S4). This result supports the notion that the interaction partners of frequently altered cancer proteins are also under positive pressure for mutations. Along this line, in a prior study, we compared the preys of frequently mutated baits to those of non-frequently mutated baits from the BioPlex dataset and found the same trend to emerge for almost all cancer types (33).

Out of 589 PPIs identified, 81% were not shared with other cell lines, reflecting high cell type-specificity of PPIs in different genetic contexts (Fig. 1E). We speculated that differential protein abundance across cell lines might provide one explanation for cell type-enriched PPIs. However, while some changes in interaction could be explained by changes in protein abundance, we also found many cases with the opposite behavior, in which a gain in interaction was observed with a concomitant decrease in protein abundance (table S5 and fig. S1D).

Cell type-enriched interactions

To compare PPIs across cell lines, we defined a cancer-enriched differential interaction score (DIS) as the probability of the PPI being present in a cancer cell line (either MCF7 or MDA-MB-231) but absent in the normal cell line (MCF10A) (Materials and Methods). Given that each PPI has a specific DIS, a continuous color scheme was used to represent cancer versus non-cancer differential interactions. We used the results of this differential scoring analysis to visualize the entire BC PPI network showing PPIs that are (1) enriched in a cancer cell line, (2) enriched in non-cancerous MCF10A cells, or (3) conserved in the two cancer cell lines but absent in the non-cancerous context (Fig. 2 and table S6).

Figure 2. Differential interaction analysis of the BC-enriched interactome.

(A) Interactome of the union of all high-confidence PPIs detected across all cell lines. Edges are colored based on their differential interaction, with pink edges representing PPIs that are enriched to BC cell lines (unique to either MDA-MB-231 or MCF7) as compared to MCF10A cells (shown in teal edges). Dotted line represents the physical protein-protein association (validated in other studies) with high Integrated Association Stringency score. (B) PPIs connecting HRAS, STK11, and CDH1. HRAS and STK11 have several interactors including FANCI and MMS19 in BC cells involved in cellular response to DNA damage stimulus. STK11 and CDH1 interact with PKP4 and PLEKHA7 in a cell type-specific manner, implying differential regulation of cell adhesion and cell-cell junction in non-BC and BC cells. (C) Representative images of MCF7 cells from the PLA between HRAS and FANCI antibodies. Maximum intensity projection images are shown to represent total PLA interactions. PLA with only one of the two primary antibodies was performed as negative control. PLA spots (white), HCS CellMask Green stain (Green) and DAPI (blue). Scale bar = 10 μm. (D) Total PLA spots per cell using HRAS and FANCI antibodies were quantified in MCF7, MDA-MB-231, and MCF10A cells. n= total number of cells analyzed in each condition. **** p-value ≤ 1.0×10⁻⁴. (E) Representative images of MDA-MB-231 cells from the PLA between STK11 and MMS19 antibodies. (F) Total PLA spots per cell using STK11 and MMS19 antibodies were quantified in MCF7, MDA-MB-231, and MCF10A cells. **** p-value ≤ 1.0×10⁻⁴, ** p-value ≤ 1.0×10⁻². (G) Percent nuclear PLA spots in each PLA condition. (H) High-confidence PPIs that are commonly detected only in two cancer cell lines (MDA-MB-231 and MCF7) but not in non-cancerous MCF10A cells. Node and edge styles and colors as seen in A. (I) STK11 forms a heterotrimeric complex with CAB39 and STRADA to activate its kinase activity and phosphorylate downstream kinases including AMPK and SIKs for regulating energy homeostasis and cell cycle. (J) STK11 kinase activity was monitored by measuring total and phosphorylation levels of its known downstream substrates (AMPK and SIK2) as well as itself. The following phospho-epitopes were detected by antibodies: pSTK11 (pS428), pAMPKα (pT172), pSIK2 (pT175). (K) Knockdown of STK11 and its interacting protein (STRADA) by two individual small-interfering RNAs reduces phosphorylation of STK11 (S428), AMPK (T172), and SIK1 (T182).

Among interactions enriched to a cancer cell line, we found the HRAS proto-oncogene and the tumor suppressor kinase STK11 (also known as LKB1) to interact with a set of DNA damage response (DDR) proteins (PDS5A, FANCI, MMS19, GPS1) in MCF7 and MDA-MB-231 cells, respectively (Fig. 2B). The HRAS-FANCI and STK11-MMS19 interactions were further assessed using a proximity ligation assay (PLA). Abundant PLA spots were observed when both HRAS and FANCI antibodies were incubated with fixed MCF7 cells, but either antibody alone generated only a background level of signal (Fig. 2C–D and fig. S2A), confirming the interaction seen by AP-MS. Also consistent with the mass spectrometry results, a significantly higher number of HRAS-FANCI PLA spots were detected in MCF7 compared to MDA-MB-231 and MCF10A cells (Fig. 2D). Again consistent with the cell type specificity uncovered via AP-MS, STK11-MMS19 PLA spots were observed with consistently higher numbers in MDA-MB-231 than in MCF7 and MCF10A cells (Fig. 2E–F and fig. S2A). Intriguingly, we also observed the HRAS-FANCI and STK11-MMS19 interactions in two additional BC cell lines (T47D and SKBR3) by PLA (fig. S2B–C). Analysis using 3D segmentation revealed that the PLA spots are present in the cytoplasm as well as the nucleus (Fig. 2G and fig. S2B–C), indicating nuclear roles of HRAS and STK11, potentially involving DNA damage/repair functions based on their PPIs. Consistent with our findings, the Human Protein atlas (www.proteinatlas.org) reveals that HRAS and STK11 are present both in the nucleus and cytosol based on immunofluorescence microscopy (36). Given the previous observations that silencing of HRAS and STK11 lead to defective DNA repair and genome instability (37), these interactions may provide insights into direct effectors by which HRAS and STK11 modulate DDR. STK11 also interacted with cell adhesion factors in MCF10A cells (PLEKHA7 and PKP4, Fig. 2B), consistent with its role in cell autonomous polarization (38) and actin filament assembly at the cellular leading edge (39).

Interestingly, CDH1 but not STK11 was found to interact with cell adhesion factors in MDA-MB-231 cells. CDH1 plays critical roles as a master regulator of cell-cell adhesion via adherens junctions, cell polarity and cell migration (40), and abrogation of CDH1 expression is a hallmark of the epithelial-to-mesenchymal transition (41). The observed interaction patterns suggest that STK11 may contribute to cell polarity and focal adhesion through a physical interaction with PLEKHA7 and PKP4, but that it requires the cellular ability to form adherens junctions. This may explain the lack of interaction of STK11 with PLEKHA7 and PKP4 in MDA-MB-231, which do not express CDH1 due to promoter hypermethylation (42).

We also found that STK11 interacts with STRADA and CAB39 (also known as MO25) preferentially in the two cancer cell lines (Fig. 2H). STRADA and CAB39 form a heterotrimeric complex with STK11 (43) to properly position the activation loop of STK11 in an active conformation (44), enabling STK11 to phosphorylate and activate downstream kinases, including AMP-activated protein kinases (AMPKs) and salt-inducible kinases (SIKs) involved in energy homeostasis and cell cycle regulation (Fig. 2I) (45–47). The increased associations among STK11, CAB39, and STRADA (regardless of protein abundance in these cells) suggested that STK11 activity is generally augmented in cancer. Consistent with this hypothesis, we found that both total and activated STK11 (phosphorylated at Ser428) are more abundant in MCF7 and MDA-MB-231 than in MCF10A cells (Fig. 2J). Furthermore, phosphorylation of STK11 downstream targets, including SIK2 and AMPK, was higher in either of two BC cell lines (Fig. 2J). Knockdown of STRADA by two different individual small-interfering RNAs (siRNAs) significantly reduced phosphorylation of STK11 (S428), AMPK (T172), and SIK1 (T182) in MDA-MB-231 cells (Fig. 2K), providing further evidence the association of STK11 with STRADA contributes to downstream signaling of activated STK11. Increased STK11 activity may support cellular fitness by balancing energy production with anabolic metabolism, as previously seen in hepatocellular carcinoma (48).

Previously unidentified regulators of PIK3CA signaling

PIK3CA and AKT activating mutations and copy-number amplifications are frequently found in many cancer types including BC (49–51), indicating that the PI3K/AKT pathway is a key signaling module for cancer cell proliferation, and thus an attractive target for therapeutic intervention (52). However, given the substantial role of the PI3K pathway in tumorigenesis, mechanisms of regulation or tuning by interacting proteins remains largely unknown as prior research predominantly centers around how mutations and alterations in the PIK3CA and AKT genes themselves regulate pathway activity.

Activation of PIK3CA via receptor tyrosine kinase (RTK) or oncogenic mutations leads to membrane recruitment and activation of AKT (Fig. 3A) (53, 54). In BC, mutations at E545 and H1047 residues are most frequently found (Fig. 3B). Using AP-MS, we identified 20 prey proteins which interact with PIK3CA, 18 of which were observed in MCF7 cells (table S4). Of the 18 proteins, only 4 (IRS1 and PIK3R1/2/3) were previously known interactors (55–58). Many of these new PPIs are significantly decreased, and in some cases completely abolished, by different PIK3CA mutations (Fig. 3C). To determine whether these PIK3CA interactors modulate the PI3K/AKT pathway, we tested whether depletion of each target by siRNAs affects downstream AKT activation in MCF7 cells by measuring cellular phospho-AKT (pS473) levels in an in-cell western assay (59). Four independent siRNAs per target gene were pooled and transfected for knockdown (Materials and Methods). Non-targeting control siRNAs (NTCs) as well as siRNAs targeting PIK3CA (a positive regulator) and PTEN (a negative regulator) were included as controls (60). As expected, knockdown of PIK3CA in MCF7 cells significantly diminished pAKT signal, while knockdown of PTEN augmented it (Fig. 3D, fig. S3A, and table S7). Knockdown of the PIK3CA interactors BPIFA1 and SCGB2A1 (also named PLUNC and Mammaglobin-B, respectively) increased pAKT activity to a degree higher than or the same as the PTEN knockdowns, indicating that these two proteins are negative regulators of the PI3K/AKT pathway (Fig. 3D, fig. S3A, and table S7). To verify these in-cell western results, we performed pooled and individual siRNA-mediated knockdown of BPIFA1 and SCGB2A1 and confirmed the increase of AKT pS473 by standard western blot analysis in MCF7 cells (Fig. 3E and fig. S3B). Although BPIFA1 and SCGB2A1 were below the level of detection by the PIK3CA AP-MS in MDA-MB-231 cells (table S3), pooled and individual siRNA-mediated knockdown of BPIFA1 and SCGB2A1 in MDA-MB-231 cells also led to an increase in AKT pS473 (Fig. 3F and fig. S3C). These results suggest that BPIFA1 and SCGB2A1 may act as regulators of the PI3K/AKT pathway in multiple cellular contexts. Consistent with this, the interaction of BPIFA1 with 3xFLAG-tagged PIK3CA (WT) was confirmed by PLA in not only MCF7 but MDA-MB-231 cells (fig. S3D), albeit at lower levels. Furthermore, abundant PLA spots between BPIFA1 and endogenous PIK3CA were also observed in T47D and SKBR3 cells (fig. S3E). To verify a negative role of BPIFA1 and SCGB2A1 in PIK3CA signaling, we measured their effect on the PIK3CA kinase activity in a reconstituted in vitro assay using recombinant proteins (PIK3CA/PIK3R1, BPIFA1 and SCGB2A1) and lipid substrate (phosphoinositol-4,5-bisphosphate) (Materials and Methods). Consistent with the AP-MS results (Fig. 3C), increasing amounts of recombinant BPIFA1 and SCGB2A1 preferentially inhibit WT PIK3CA kinase activity towards lipid substrate when compared to the two mutant forms of PIK3CA (E545K and H147R) (Fig. 3G–H). These data both confirm the validity of the PPI and provide further verification of the WT vs. mutant specificity revealed by the AP-MS.

Figure 3. Comparative interactome analysis of WT and mutant PIK3CA proteins.

(A) Overview of the receptor tyrosine kinase (RTK)-PI3K signaling cascade leading to the phosphorylation (T308 and S473) and activation of the AKT pathway. (B) A lollipop plot representing the sites of PIK3CA mutations and the number of BC patients bearing a given PIK3CA mutation from TCGA (Firehose Legacy) study. (C) Relative quantification of the abundance of high-confidence preys observed from pull-down of PIK3CA (WT and mutants) in MCF7 cells. Preys detected only in WT are represented in deep blue while preys detected only in mutants are in deep red. All three PIK3CA mutants were expressed and detected at a similar level. ND, not detected. (D) The level of AKT S473 phosphorylation (as proxy of activation) was measured by in-cell western analysis upon siRNA-mediated knockdown of PIK3CA interacting preys and control genes (PTEN, PIK3CA, and PIK3R1) in MCF7 cells. The intensity of AKT pS473 was normalized to total AKT as well as cell viability. *** p-value ≤ 1.0×10⁻³, ** p-value ≤ 1.0×10⁻². (E-F) Increase of AKT S473 phosphorylation upon knockdown of BPIFA1 and SCGB2A1 was confirmed by western blot in both MCF7 and MDA-MB-231 cells, respectively. (G-H) PIK3CA (WT, E545K, and H1047R) kinase activity towards phosphatidylinositol-4, 5-bisphosphate (PIP₂) was measured in vitro in the presence of increasing amounts of recombinant BPIFA1 and SCGB2A1, respectively. Results are from at least 3–6 independent experiments. **** p-value ≤ 1.0×10⁻⁴.

Effect of pathogenic mutations on the BRCA1 interactome

BRCA1 is a major hereditary cancer susceptibility gene (61, 62) that plays critical roles in DNA repair by homologous recombination (HR) (63) in addition to other processes, such as regulation of transcription, RNA splicing and cell cycle (64, 65). BRCA1 carries out its functions in concert with other proteins (63), leading to many studies of BRCA1-containing complexes and their roles in DNA repair (66, 67). To date, many of these findings have been based on either immunoprecipitation with antibodies against the WT BRCA1 protein or interrogation of pairwise protein interactions with the yeast two-hybrid system. Moreover, these analyses were done mainly using WT BRCA1 protein and did not systematically capture how different mutations in BRCA1 might affect its protein interactions.

To comprehensively catalog the BRCA1 interactome and how pathogenic BRCA1 mutations alter these interaction profiles, we performed AP-MS on WT and pathogenic variants reported in cancer patients, including C61G and R71G in the N-terminal RING domain (68, 69) and S1655F, 5382insC and M1775R in the C-terminal tandem BRCT domain (70, 71) (Fig. 4A). Given that alternative splicing in cancer often generates BRCA1 isoforms lacking exon 11, which confers residual HR activity and therapeutic resistance (72), an isoform (isoform 6; UniProt identifier P38398–6) lacking exon 11 with distinct C-terminus was also included in the analysis. The I26A separation of function mutation in the RING domain, which abrogates E3 ubiquitin ligase activity but retains BARD1 binding, was also analyzed (73). We induced the expression of these BRCA1 proteins in all three breast cell lines; however, only MDA-MB-231 cells (harboring the TP53 R280K mutation) supported ectopic 3xFLAG-BRCA1 expression. These observations were consistent with previous studies which have shown that ectopic overexpression of BRCA1 (both WT and mutants) is not stably maintained without a compensatory TP53 mutation (74). AP-MS experiments in MDA-MB-231 cells identified 128 high-confidence interactions from 8 BRCA1 constructs (WT and 7 mutants, PPI score ≥ 0.65; fig. S4A and table S8); of these interacting proteins, 70 showed ≥ 8-fold change (Fig. 4B and fig. S4B).

Figure 4. Quantitative analysis of the effect of mutations on the BRCA1 interactome.

(A) Functional domains in the BRCA1 gene and the location of mutations analyzed by AP-MS. (B) Relative quantification of the abundance of prey proteins (PPI score ≥ 0.65, ≥ 8-fold change) identified by BRCA1 AP-MS in MDA-MB-231 cells. All prey abundance values were normalized by 3xFLAG-tagged BRCA1 levels in their respective AP-MS experiments. Preys detected only in WT are represented in deep blue while preys detected only in mutants are in deep red. A group of proteins involved in HR repair (boxed in green) are clustered together, wherein RING domain and BRCT domain BRCA1 mutants show distinct PPI abundance profiles. Spinophilin has not previously been known to have a function relevant to HR repair. UBE2N is boxed in sky blue. ND, not detected. (C) PPIs of selected proteins with BRCA1 (WT or C61G mutant) were confirmed by co-immunoprecipitation with anti-FLAG antibody followed by western blot analysis. (D) Box plot shows that the patient group (enrolled in the I-SPY 2 clinical trial) with pCR to veliparib (PARP inhibitor) and carboplatin (VC) had pre-treatment tumors with significantly lower UBE2N mRNA expression (LR p-value = 0.034) than those of non-responding patients. In contrast, BC patient tumors in the control arm did not show any significant difference in UBE2N expression between pCR and no pCR groups. (E) Mosaic plot shows that BC patients who did pCR to VC in addition to standard chemotherapy had 2.9 times more likely had lower mRNA expression of UBE2N in their pre-treatment tumors (Odds Ratio = 2.9). In the control arm, there is no significant difference in pCR between low and high UBE2N expression groups. Numbers in each block represent the patient sample size. Column width indicates the relative proportion of the UBE2N low and high expression group on the patient population. (F) A schematic of the HR reporter assay. The DR-GFP reporter contains two defective copies of the GFP gene, one disrupted by an I-SceI site and the other lacking a promoter. I-SceI cutting of the first copy generates a DSB, and repair by HR with the second copy as a template leads to restoration of a functional GFP gene. siRNA-mediated knockdown of HR-related genes leads to reduction of GFP+ cells (a proxy of HR activity) compared to NTC experiments. (G) HR activities upon depletion of USP28 relative to NTC (set to 100%). Depletion of BRCA1 and RIF1 was included and analyzed as controls. Data shown are the means from 3–6 independent experiments for each siRNA. Error bars represent standard deviations (SDs). **** p-value ≤ 1.0×10⁻⁵, ** p-value ≤ 0.01, * p-value ≤ 0.05.

These data revealed a number of previously unidentified BRCA1-interacting proteins, along with known interactors, many of which were differentially affected by mutations in different domains of BRCA1. For example, HR proteins previously known to interact with BRCA1 (including BRIP1 and RBBP8) (71, 75) had a similar pattern of interaction loss (boxed in green in Fig. 4B) associated with BRCT domain mutants (S1655F, 5382insC, M1775R), whereas RING domain mutants (I26A, C61G, R71G) maintained these interactions. In a separate pattern, we found that the C61G RING domain mutant abolishes interaction with BARD1 (Fig. 4B), as previously reported (66). Several interactions could be confirmed by co-IP/western blot analysis (Fig. 4C). These results suggest that RING domain mutants are hypomorphic and may retain some BRCA1 functions, which could explain at least in part why the BRCA1 C61G variant is only moderately sensitive to cisplatin and poly (ADP-ribose) polymerase (PARP) inhibitors and becomes readily resistant to these drugs (76, 77).

A ubiquitin E2-conjugating enzyme, UBE2N (also known as UBC13), was found to interact with WT BRCA1, but to a lesser degree with mutant forms of BRCA1 (PPI score < 0.6) (boxed in sky blue in Fig. 4B). For example, consistent with reports from yeast two-hybrid studies (78), we found a six-fold reduction in UBE2N associated with the I26A mutant compared to WT, suggesting that UBE2N interacts with BRCA1 through the RING domain. Notably, the M1775R mutation in the BRCT domain also dramatically reduced the interaction with UBE2N (Fig. 4B), suggesting that M1775 residue in the BRCTs domain may also contribute to the interaction with UBE2N, although the underlying mechanism is currently unclear. Depletion of UBE2N shows HR defects including altered RAD51 filament formation and E3 Ub ligase function of BRCA1 (79), indicating a critical role of UBE2N in HR repair. Consistent with the cell line models, we found that baseline UBE2N mRNA expression was significantly lower in patients who achieved pathologic complete response (pCR) to the PARP-inhibitor (veliparib)/carboplatin (80) in comparison to non-responsive patients (p = 0.034, Fig. 4D) in I-SPY 2, a neoadjuvant, adaptive clinical platform trial for high risk early stage breast cancer (81). In contrast, BC tumors in the control arm did not show any significant difference in UBE2N expression between pCR and no pCR groups. Although the average expression of UBE2N is numerically lower in TN than HR+HER2-tumors, the difference is not statistically significant (p=0.1, fig. S4C). Moreover, the association of low UBE2N expression to pCR is seen more strongly in the HR+HER2- subset (p=0.0012, fig. S4D) than in the population as a whole (p=0.034, Fig. 4D), suggesting that the association is not merely a proxy for the TN subtype. BRCA1/2 germline mutation status is available for 112 out of 115 patients in the VC and concurrent control arms, and 15 out of 112 (13%) are BRCA1/2-. UBE2N trends toward lower expression levels in BRCA1/2 mutation carriers (p=0.07, fig. S4E). However, in the HR+HER2- subset (n=55 with BRCA1/2 mutation status data), where lower UBE2N levels associate with response to VC, there are only 4 mutation carriers and no difference in UBE2N levels by BRCA1/2 status (p=1). Furthermore, only 2 HR+HER2- BRCA1/2 mutation carriers were treated by VC (only one of these achieved pCR), implying that association between UBE2N and pCR in this subset may not be driven by ‘spurious’ association with germline BRCA1/2 mutation (fig. S4F). These results indicate that expression of UBE2N may serve as a biomarker of response to PARP inhibitors and other DNA repair targeted therapies (Odds Ratio = 2.9, Fig. 4E). Importantly, there are several observations that not all of BRCA1-interacting HR proteins show a response to PARPi upon depletion (by CRISPR or siRNA) when tested in multiple cell line models, as seen in a recent CRISPR screen (82). Furthermore, the clinical response of many BRCA1-interacting proteins to PARP inhibition has never been explicitly tested in a clinical setting. Therefore, identification of UBE2N as a potential biomarker for PARPi response could be clinically valuable to help stratify patients with UBE2N alteration for targeted therapy.

Intriguingly, a deubiquitinase USP28 was found to interact more strongly with the BRCA1 mutants C61G and 5382insC than WT by AP-MS (boxed in brown in Fig. 4B), and this differential interaction was confirmed by co-IP/western blot analysis (Fig. 4C). USP28 was previously known to stabilize multiple DDR proteins (e.g. CHEK2, CLSPN, NBS1, and 53BP1) in response to DNA damage (83, 84), and its recruitment to DNA double-strand breaks (DSBs) was shown to be dependent on the tandem BRCT domain of TP53BP1 (85). To further delineate the role of USP28 in DNA DSB repair, we analyzed the effect of USP28 knockdown on HR activity. In this assay, DNA DSBs were induced by I-SceI endonuclease, which cleaves non-functional GFP cassettes engineered in the genome of U2OS reporter cell lines (DR-GFP) (86). DSB repair that depends on the HR mechanism restores a functional GFP gene, yielding a readout tied to fluorescent signal intensity (Fig. 4F). Upon USP28 knockdown by siRNA, HR activity was significantly reduced when compared to NTC (Fig. 4G and fig. S5A), suggesting USP28 is regulating proper HR repair. In the same assay, knockdown of BRCA1 greatly decreased HR as expected (87), while depletion of RIF1, a protein functioning in an alternative non-homologous end joining DNA repair pathway did not. To identify proteins potentially regulated by USP28 via deubiquitination, we performed a global ubiquitome analysis upon knockout of USP28 in MDA-MB-231 cells by CRISPR/Cas9 ribonucleoprotein (RNP) strategy (88). Polyclonal USP28 knockout cells with ~85% knockout efficiency (crRNA#3, fig. S5B) as well as control (NTC) cells were lysed and ubiquitinated peptides were subsequently enriched and analyzed by mass spectrometry. Functional enrichment analysis of 275 ub-enriched proteins observed in USP28 knockout cells compared to control cells (log₂FC ≥ 1, p-value ≤ 0.05; table S9) revealed that 19 proteins (7%) belong to the cellular response to stress category (GO: 0033554) (fig. S5C), and subsequent analysis further categorized proteins involved in cellular response to DNA damage stimulus (8 proteins) (fig. S5D–E). Notably, these include ubiquitylation-dependent DDR signaling proteins (UBA1, RAD18, DDB2) (89–91) and DNA replication helicases (MCM3, MCM6) (92), providing further evidence for a role of USP28 in responding to DNA damage and replication stress in cells.

Spinophilin is a previously unidentified BRCA1-interacting protein

Another protein interacting with BRCA1 in a mutation-dependent manner is Spinophilin (encoded by PPP1R9B), a neuronal scaffolding protein that regulates synaptic transmission through its ability to target protein phosphatase 1 (PP1) to dendritic spines where it inactivates glutamate receptors (93). Binding of Spinophilin to BRCA1 was abolished by BRCT domain mutations similar to the pattern observed earlier for HR proteins (Fig. 4B). Reciprocal AP-MS was performed using 3xFLAG-tagged Spinophilin in MDA-MB-231 cells, which confirmed the interaction of Spinophilin with BRCA1 as well as with PP1 catalytic subunits (PPP1CA, PPP1CB, PPP1CC) (Fig. 5A). MSstats analysis of differential interactions between BRCA1 WT and BRCT domain mutants demonstrated that an intact BRCT domain is required for the BRCA1-Spinophilin interaction (fig. S6A). In fact, Spinophilin was previously observed but unexplored in a systematic analysis of proteins interacting with the BRCT domain of BRCA1 (94). The BRCA1 BRCT domain binds to a pS-x-x-F motif on target proteins (95), and such recognition has been reported for BRCA1 interactions with FAM175A, BRIP1, and RBBP8 (71, 75, 96). Importantly, Spinophilin has three potential pS-x-x-F recognition motifs at S212, S248, and S635. To map the BRCT domain-binding motif on Spinophilin, we mutagenized Serine residues into Alanine at each pS-x-x-F site and transfected each Spinophilin (HA-tagged) construct individually along with 3xFLAG-BRCA1 DNA construct into HEK293T cells. The association of each Spinophilin mutant with BRCA1 was monitored after immunoprecipitation of Spinophilin using anti-HA beads followed by western blotting with anti-FLAG antibody. This experiment showed that the S635A mutant interacted significantly less with BRCA1, demonstrating that Spinophilin requires the pS⁶³⁵PTF motif for binding the BRCT domain of BRCA1 (Fig. 5B and fig. S6B). The F451A mutation, which resides in the interface with PP1 RVXF binding groove, also significantly diminished the interaction of Spinophilin with PP1 (PPP1CA) (Fig. 5B and fig. S6B), as observed previously (97).

Figure 5. Spinophilin interacts with BRCA1 and regulates DNA damage response via dephosphorylation.

(A) AP-MS of 3xFLAG-tagged Spinophilin (SPN, encoded by PPP1R9B) identifies BRCA1 (highlighted in a red edge) and other DDR-related proteins as well as PP1 catalytic subunits (PPP1CA, PPP1CB, and PPP1CC) in MDA-MB-231 cells. (B) HA-tagged SPN (WT, S212A, S248A, S635A, or F451A) was transfected with 3xFLAG-BRCA1 into HEK293T cells. After pulldown with anti-HA magnetic beads, co-associated 3xFLAG-BRCA1 was monitored. S635A mutation significantly diminished BRCA1 pulldown, while F451A mutation abolished the association with PP1 catalytic subunit (PPP1CA). Empty vectors were used as negative control. (C) HR activities upon depletion of SPN relative to NTC (set to 100%) were measured as in Fig. 4G. Data shown are the means from 3–9 independent experiments for each siRNA. Error bars represent standard deviations (SDs). **** p-value ≤ 1.0×10⁻⁵, ** p-value ≤ 1.0×10⁻². (D) A schematic of the SA-GFP reporter assay. The SA-GFP reporter contains a 5’-fragment of GFP (5’-GFP) and a 3’-fragment of GFP (Sce3’-GFP) that contains an I-SceI site. Repair of the DSB in Sce3’-GFP using 266 nt homology by single-strand annealing (SSA) restores a functional GFP gene. (E) SSA activities upon depletion of SPN relative to NTC (set to 100%). Depletion of BRCA1 and BRCA2 was included and analyzed as controls. Data shown are the means ± SDs from six independent experiments for each siRNA. **** p-value ≤ 1.0×10⁻⁴, *** p-value ≤ 1.0×10⁻³. (F-G) CMV promoter-driven SPN (WT, S635A, or F451A) expression DNA construct was transfected into U2OS DR and SA-GFP reporter cells, and the effect of SPN overexpression on HR and SSA activities was monitored, respectively. **** p-value ≤ 1.0×10⁻⁴, ** p-value ≤ 1.0×10⁻². (H) Selective peptides derived from various proteins including BRCA1 and H2AX as well as non-DNA repair proteins (INCENP and BCAR1) were individually mixed with lysates from either SPN KO or parental cells and subsequently monitored for phosphorylation by measuring ATP consumption in each reaction. Net peptide phosphorylation values are net changes in ATP concentrations between SPN knockout cells and parental control cells (subtraction of parental runs from SPN KO runs). Mean value of two independent runs was shown in the y-axis. Units are arbitrary. Error bars represent standard deviations (SDs). * p-value ≤ 0.05, ** p-value ≤ 1.0×10⁻². (I) SPN KO and parental cells were treated with 2.5 μM etoposide (Eto) for 16 hr and changes in the phosphorylation level of BRCA1 S1423 (pS1423) and H2AX S140 (γ-H2AX) were monitored at 0, 1, 2, and 4 hr post Eto treatment with fresh medium. (J) Model for the role of SPN in regulating DDR. See text for details.

The AP-MS experiment found that Spinophilin interacts with additional proteins involved in DNA replication and repair, including MCM10 and WDR48 (Fig. 5A). These results suggest that Spinophilin may participate in and/or regulate DNA repair by interacting with various DNA repair and replication proteins, including BRCA1. To explore this hypothesis, we analyzed the effect of Spinophilin knockdown on DNA repair by HR and single-strand annealing (SSA). Upon Spinophilin knockdown, HR activity was significantly reduced compared to NTC siRNA (Fig. 5C and fig. S6C). Similarly, knockdown of Spinophilin significantly reduced SSA activity, comparable to BRCA1 depletion, while BRCA2 depletion dramatically increased SSA as seen previously (87) (Fig. 5D–E and fig. S6D), implying that Spinophilin promotes both HR and SSA-mediated DSB repair.

According to the breast invasive carcinoma patient cohort study by TCGA (2), there is no clear evidence that the Spinophilin (PPP1R9B) gene is recurrently silenced by deep DNA copy loss or mutation. Using the interface to the TCGA methylation data (http://maplab.imppc.org/wanderer/), we do not observe evidence for methylation of the Spinophilin (PPP1R9B) promoter, either. However, among the sequenced TCGA BC patient tumors that have alterations in the Spinophilin (PPP1R9B) gene (73 out of 1084), the majority (69 out of 73) of the alterations are amplification, suggesting that higher expression of Spinophilin may be a pathogenic driver of breast cancer. Thus, we also analyzed the effect of Spinophilin overexpression (mimicking amplification) on HR and SSA DSB repair. CMV promoter-driven overexpression of WT Spinophilin in the U2OS reporter cell lines led to significant reduction in both HR and SSA activities, compared to empty vector control (Fig. 5F–G). Consistent with these observations, a recent study identified Spinophilin as a BC oncogene (98). However, the effect of overexpression of the S635A mutant (incapable of interacting with BRCA1) on HR was significantly mitigated, and more importantly, overexpression of the F451A mutant (incapable of interacting with PP1) did not affect HR or SSA at all (Fig. 5F–G and fig. S6E–F), implying that the inhibitory effect of WT Spinophilin overexpression is likely exerted by PP1-mediated dephosphorylation and/or BRCA1 interaction (at least for HR).

To further unravel the biological function of Spinophilin, we knocked out Spinophilin in MDA-MB-231 cells using a CRISPR/Cas9 RNP method (88). Using genomic PCR/sequencing and western blot analyses, at least two independent Spinophilin knockout clones were identified (fig. S7). We hypothesized that Spinophilin targets PP1 to specific DNA repair proteins for dephosphorylation. To uncover potential dephosphorylation targets under this model, we employed a high-throughput peptide phosphorylation assay platform (99). This system uses a collection of peptide sequences derived from biological targets of multiple kinases, which serves as phosphorylatable probes in a large-scale ATP-consumption assay (100). In this assay, we measured changes in phosphorylation (i.e., ATP-consumption) of peptide substrates derived from various proteins, including BRCA1 and the DSB-associated histone H2AX as well as proteins unrelated to DNA repair (e.g., INCENP, BCAR1), in Spinophilin knockout (KO) and parental cells. We found that BRCA1 residues at T509, S1387, and S1423, as well as H2AX at S140 (γ-H2AX), were significantly increased in phosphorylation in Spinophilin KO cell lysates compared to lysates from parental cells, and, in fact, were among the top 20 most increased sites (Fig. 5H and fig. S8A). BRCA1 pT509 enhances nuclear localization and transcriptional activity of BRCA1 (101), and pS1387 and pS1423 sites in the BRCA1 SQ-cluster region are critical for HR repair and cell-cycle checkpoint functions (102–104). γ-H2AX is a hallmark of DNA DSB (105) and initiates a signaling cascade to recruit various DSB repair proteins to properly repair the DNA damage (106). These results were in contrast to phosphorylation of the INCENP and BCAR1 peptides which were not significantly changed by Spinophilin disruption. To validate these results, Spinophilin KO and parental cells were treated with 2.5 μM etoposide (Eto) for 16 hr to induce BRCA1 pS1423 and γ-H2AX, and persistence of these phosphosites was monitored by western blot at 0, 1, 2, 4 hr post Eto treatment. Indeed, these phosphosites remained at a significantly higher level and were maintained for longer in Spinophilin KO cells compared to parental cells (Fig. 5I). Consistently, we also observed higher levels of BRCA1 pS1423 and γ-H2AX after Eto treatment in Spinophilin knockdown U2OS cells (fig. S8B).

To unbiasedly identify targets that are potentially dephosphorylated by Spinophilin, we performed a phospho-proteomic analysis using the two Spinophilin KO clones and parental MDA-MB-231 cells. In this experiment, we found that 473 phospho-peptides (from 407 proteins) were more than 4-fold enriched (log₂FC ≥ 2, p-value ≤ 0.05) in both Spinophilin KO clones compared to parental cells, while 328 phospho-peptides (from 280 proteins) were more than 4-fold depleted (log2FC ≤ −2, p-value ≤ 0.05) (fig. S9A–B and table S10). Notably, proteins known to be directly connected to BRCA1 via physical and/or functional interactions were significantly enriched in the group of up-regulated phospho-proteins seen in Spinophilin KO cells (fig. S9C). Functional enrichment analysis of these 407 up-regulated phospho-proteins revealed that 32 proteins (8%) belong to the DNA repair category (GO: 0006281) (fig. S9D), and subsequent analysis further categorized 15 proteins involved in double-strand break repair (fig. S9E). Importantly, the 32 DNA repair proteins include not only BRCA1-interacting proteins (e.g. BRIP1) but also contain key regulators of DNA repair pathways (e.g. MDC1, TP53BP1), and helicase/nuclease (e.g. MRE11, WRN), suggesting broad roles of Spinophilin in modulating DNA repair and genome integrity (fig. S9F–G). Taken together with the previous data, these results indicate that Spinophilin regulates BRCA1 and DDR signaling by dephosphorylation (Fig. 5J).

DISCUSSION

The cell is comprised of a series of protein complexes, or “machines” (107), that function together in an elaborate network of pathways. Mutations, such as seen in cancer, perturb the machines and therefore the network of pathways (108). Understanding the machines and networks in healthy and diseased states is crucial for a deeper understanding of disease biology and ultimately the discovery of new therapeutic strategies and the application of precision medicine (109). Using this premise as an underlying motivation, we generated comprehensive interaction maps for 40 frequently altered BC proteins. This large-scale study of biophysical interactions in BC and across three cell lines of human breast tissue origin provides a PPI resource to study BC biology and to contextualize uncharacterized mutations within signaling pathways and protein complexes. Approximately 78% of PPIs we identified have not been previously reported (Fig. 1C), and 81% are not shared across cell lines (Fig. 1E). The cell line PPI specificity we observe in this study and our accompanying manuscript (Swaney et al.), as well as in a recent large-scale AP-MS study (27), speaks to the complexity of PPI networks and their heterogeneity among different cell types, calling for more studies to deeply characterize these networks in a wider range of cellular contexts. Given that each BC subtype arises by distinct tumorigenic mechanisms due to different genomic alterations, PPIs from cell lines representing other BC subtypes will most likely capture additional sets of interactors reflecting their distinct tumorigenic mechanisms. Our results also suggest that protein abundance in a cell line is not the sole mechanism for PPI specificity (fig. S1D). Presumably other features, such as differential posttranslational modifications (PTM), cellular compartmentalization, and/or mutational status of proteins may contribute to cell type-specificity. Notably, prey proteins enriched to either of the two BC cell lines are more frequently mutated in breast tumors than preys from non-tumorigenic cells (Fig. 1E), implying that proteins interacting with cancer drivers may also contribute to the onset of cancer.

Using the systematic proteomic approach described in the manuscript, we have identified previously unidentified interactors of PIK3CA that negatively regulate the PI3K/AKT pathway (Fig. 3D–H). Interestingly, RNA expressions of several of these interactors (e.g. BPIFA1, BPIFB1, and MUC5B) were found to be correlated in human airway epithelial cells (110), implying that they may be controlled under the same transcriptional regulatory program and/or functionally related. Our results reveal that they play a role in the PI3K/AKT signaling cascade in BC context. BPIFA1 is a lipid-binding protein with antimicrobial and immunomodulatory functions (111). It is significantly downregulated in nasopharyngeal carcinoma (112) and its single-nucleotide polymorphisms are associated with increased susceptibility to this tumor type (113). BPIFA1 is known to increase the expression of PTEN via downregulating the miR-141 oncogene (114), thus knockdown of BPIFA1 could indirectly activate PI3K/AKT. The BPIFA1-PIK3CA interaction we identified and the inhibition of WT PIK3CA kinase activity by BPIFA1 in vitro suggests that BPIFA1 may also directly modulate PI3K/AKT via PPI, which warrants a structural study of the complex. Another PIK3CA interactor, SCGB2A1, is a small secreted protein highly differentially expressed in multiple types of cancer including breast and endometrium (115, 116). Previous studies have shown that SCGB2A1 is expressed at lower levels in luminal BC compared to histologically normal breast epithelium (117), and that decreased SCGB2A1 expression in endometrial tumors is significantly correlated with higher grade, advanced stage, and worse overall survival (118). Neither the BPIFA1 or SCGB2A1 gene harbors recurrent focal deep deletions, missense mutations, or truncating mutations in their coding regions among 996 breast tumors in TCGA analysis (2). However, SCGB2A1 harbored loss of heterozygosity (LOH) in 10% of breast tumors (7% LOH in all tumors) according to a recent study (119), which suggests that its allelic imbalance may contribute to the development of tumors.

We also uncovered a number of previously unidentified interactors of the well-studied BRCA1, which is tightly connected to DNA repair processes, including Spinophilin. Interestingly, both knockdown and overexpression of Spinophilin led to significant impairment in DSB repair by both HR and SSA pathways (Fig. 5C, E–G), establishing that this protein has a defined role in DNA repair. An intriguing question is how the alterations of Spinophilin expression interfere with HR and SSA repair activity. Given the role of Spinophilin in dephosphorylation, one plausible explanation is that prolonged phosphorylation or prevailed dephosphorylation of BRCA1 and other DDR proteins is inhibitory to multiple steps during DNA repair, including DSB-end resection which is a prerequisite for HR and SSA. In agreement with this hypothesis, continuous DNA damage signaling and phosphorylation of several DDR proteins (including H2AX, NBN, RPA2, and CHEK2) induced by short double-stranded DNA molecules (mimicking DNA DSB) was shown to disorganize the cellular DNA repair system and inhibit DSB repair (120). Alternatively, but not exclusively, Spinophilin may play a role in initiating the DSB repair process by removing constitutive phosphorylations that inhibit the function of DDR proteins. Supporting this scenario, a phospho-proteomic study revealed that over one-third of the captured phospho-peptides were dephosphorylated within minutes of DNA damage (121). Additionally, Spinophilin may be involved in counteracting DSB-induced phosphorylation events, thus promoting recycling of DDR proteins as DNA damage is being repaired, possibly through its interaction with BRCA1.

Finally, while the approach we described in this study was applied to BC, it is equally powerful against other cancers, including head and neck squamous cell carcinoma (HNSCC) (Swaney et al., accompanying manuscript). Efforts such as this will ultimately lead to hierarchical maps of protein complexes and systems in both healthy and diseased cells (Zheng et al., accompanying manuscript), which, based on the mutational landscape, can be predictive for use of known treatments and can also be used to uncover previously unidentified therapeutic strategies across a multitude of disease areas.

Data and materials availability:

All raw data files and search results are available from the Pride partner ProteomeXchange repository (122). For AP-MS and protein abundance data from MCF10A, MCF7, and MDA-MB-231 cells, these can be accessed under the PXD019639 identifier. For PTM (ub and ph) analysis data, these are available under the PXD025931 identifier. PPI networks can be accessed on the NDEx platform (https://bit.ly/3y10JFY). DOIs and links for each PPI network are in table S1. Expression vectors used in this study are readily available from the authors for biomedical researchers and educators in the nonprofit sector under MTA. Further information and requests for resources and reagents should be directed to and will be fulfilled by NJK (nevan.krogan@ucsf.edu).