Rapid, Optimized Interactomic Screening

Abstract

We must reliably map the interactomes of cellular macromolecular complexes in order to fully explore and understand biological systems. However, there are no methods to accurately predict how to capture a given macromolecular complex with its physiological binding partners. Here, we present a screen that comprehensively explores the parameters affecting the stability of interactions in affinity-captured complexes, enabling the discovery of physiological binding partners and the elucidation of their functional interactions in unparalleled detail. We have implemented this screen on several macromolecular complexes from a variety of organisms, revealing novel profiles even for well-studied proteins. Our approach is robust, economical and automatable, providing an inroad to the rigorous, systematic dissection of cellular interactomes.

High-throughput DNA sequencing facilitates whole genome characterization within weeks^1,2. Likewise, advances in mass spectrometry (MS)^3,4 are enabling cellular proteomes to be defined. However, we have yet to exhaustively map any interactome – the cell’s comprehensive biomolecular interaction network^5,6; we may have identified less than 20% of the protein interactions in humans, not counting dynamic, tissue- or disease-specific interactions^7–9.

A main approach for interactomic exploration is affinity capture^10,11. For this, cells are broken and their contents extracted into a solution that ideally preserves each target macromolecular complex. Complexes are then specifically enriched from the cell extract using affinity reagents – usually antibodies – that recognize the target, either directly or through an epitope tag, permitting subsequent characterization of the complex. However, one of the foremost challenges facing affinity capture studies is the precise optimization of the extraction conditions, because no single condition is optimal for the preservation of the many different types of interactions found in macromolecular complexes^12–14. As a result, affinity capture experiments either require time-consuming optimization on a case-by-case basis, or a compromise must be made by using un-optimized conditions; the latter is a common strategy but often results in sparse coverage of protein-protein interactions and error-prone data^15–17. A variety of advanced bioinformatics tools¹⁸ and databases of common contaminant proteomes^19,20 have attempted to mitigate this problem^21–24, but cannot fully substitute for optimized sample preparation¹⁵. Because any given extraction solution influences the complement of copurifying proteins, multiple extractant formulations are required if one intends to broadly sample the interactome, as underscored by a recent high-throughput study of membrane protein interactions in yeast²⁵.

The problem of maintaining post-extraction protein complex stability is comparable to that which once hindered protein crystallographic efforts. Crystallography requires the empirical determination of conditions promoting interactions that permit efficient crystal growth. Similarly, affinity capture requires the empirical determination of conditions that support the retention of in vivo interactions and minimize the in vitro artifacts. For crystallography, the answer came with the development of massively parallel crystallization optimization screens^26,27 that allow hundreds of conditions to be simultaneously explored²⁸. Inspired by this, we have developed improved methods for the rapid processing of cellular material in conjunction with parallelized, multi-parameter searches of extraction conditions. Our approach is compatible with both standard lab scale investigations and high-throughput robotics, and facilitates the systematic exploration of the interactome of any given protein in a cell.

Results

Designing a large-scale interactomics screen

Our strategy (Fig. 1), starts with the distribution of cryomilled cell material^29,30 to a multi-well plate. To enable the uniform delivery of frozen cell powder to each well in the plate, we designed dispensing manifolds (Fig. 2a,d and Supplementary Fig. 1). After dispensing, the powder in the wells is thawed by addition of an array of distinct extractants. The resulting extracts are clarified of insoluble material using a clog-resistant filtration device (Fig. 2b,d) that provides a filtrate matching the quality of centrifugally clarified cell extract (Fig. 2c). The remainder of the procedure implements commercially available supplies and equipment (Online Methods and Supplementary Protocol 1).

Figure 1.

Schematic representation of the parallelized affinity capture procedure. (i) cells expressing a tagged protein of interest are mechanically disrupted at cryogenic temperature to produce a micron-scale powder and precise aliquots of the frozen powder are deposited into the wells of a multi-well plate using a dedicated manifold. A diverse set of extraction solvents are rapidly added in parallel and complete re-suspension with concomitant extraction of the powder is ensured using immediate brief mechanical agitation and/or low-power sonication; (ii) rapid removal of insoluble material is achieved by either centrifugation or using a novel deep-bed filtration device (this study); (iii) affinity capture is then performed on the clarified extract using magnetic beads coupled with an affinity reagent for the tag used, and a multipronged magnet separator in register with the multiwell plate, allowing subsequent washing steps; (iv) elution of the complexes is followed by SDS-PAGE, Coomassie blue staining, and (as desired) MS analysis. The resulting copurification profiles are catalogued and cross-compared to infer interactomes and determine preparative conditions appropriate for further biochemical and analytical means.

Figure 2.

Dispensing manifold and filtration device. (a) Schematic representations of the manifold used to dispense a calibrated amount of frozen cell powder into a 96-well plate. A set of adapters and volume displacing prongs are used to deliver the required amount of cell material. (b) Schematic representation of the filtration device used to clarify crude yeast cell extracts. Each well contains a composite filter comprised of multiple filtration elements, including: a coarse pre-filter that retards the flow of highly aggregated material; a diatomaceous earth depth filter that permits the passage of soluble material, and traps the insoluble debris that can clog submicron filters; and a 0.2 micron membrane filter, which provides a uniform final clarification. (c) Coomassie blue stained SDS-PAGE analysis of Nup53p-SpA affinity capture (100mg cell material resuspended with 600uL extraction solvent: 40 mM TRIS-Cl, pH 8, 250 mM trisodium citrate, 150 mM NaCl, 1% v/v Triton X-100) comparing extract clarification by centrifugation at 14k rpm for 10 min (“Centrifuged”) and filtration at 3.5k rpm for 5 min (“Filtered”), MW – molecular weight standard. Duplicate experiments produced identical results (not shown). Proteins labeled in accordance with Figure 4a. (d) Pictures of the actual devices from left to right: adjustable volume dispensing manifold (as in a), shown bottom up; dispensing manifold with 96-well deep-well plate atop, cell material transfer is achieved upon inversion of this assembly; a 96-well filtration device (as in b) atop a 96-well, deep well collection plate.

The bandwidth of our screen allowed us to thoroughly explore the most common reagents used in affinity capture experiments: salts, buffers, and detergents (Fig. 3a, Supplementary Table 1)^12,31. Salts are frequently classed as kosmotropes (e.g. sodium citrate or ammonium acetate) or chaotropes (e.g. sodium perchlorate), in accordance with their tendency to respectively stabilize or disrupt protein structures in aqueous solution (the so-called Hofmeister series³²). The mechanism of these effects is still largely uncharacterized^33,34 and cannot be predicted a priori. The chemical character of the buffering agent can also contribute to the efficacy of affinity capture in unpredictable ways beyond simple pH control¹². Detergents, used to extract membrane-anchored complexes and inhibit aggregation of all complexes, also exhibit unpredictable behaviors and require empirical optimization³¹.

Figure 3.

Extraction condition design and copurification pattern analysis. (a) Flow-chart representation of mixtures of components in select extraction solvent formulations. The main components are a pH buffer, 1 or 2 salts, an additive and a detergent. Some examples of useful formulations discovered through screening are indicated (refer to Fig. 5 to view the associated copurification patterns, indicated by Roman numerals). (b) Comparison of SDS-PAGE (“Gel”) and LC-MS/MS (“MS”) clustering analysis of Nup1p-SpA 96-well purification. For a visual comparison the MS data is represented as a pseudo-gel, where each band corresponds to a protein above a certain intensity threshold (see Methods). Known Nup1p interacting proteins are indicated with blue bands, the rest are labeled black. Co-clustering conditions with identical or highly similar components producing distinct copurification profiles are highlighted in blue (low ammonium acetate or low potassium acetate), orange (high sodium citrate/high ammonium acetate with Triton X-100) and green (sodium citrate/potassium acetate with CHAPS) boxes. See Supplementary Figure 3 for lane labels.

Evaluation of affinity capture profiles

There are two commonly used approaches for analyzing affinity captured samples: SDS-PAGE with dye-based visualization^35,36 and subsequent MS of select conditions; or, direct MS of the samples^10,11. We compared these two approaches for the 96-well purification of Nup1p, a component of the yeast nuclear pore complex (NPC), exploring a diverse set of extraction conditions. Two replicates were carried out for the comparison: one set resolved by SDS-PAGE and stained with Coomassie blue (Supplementary Fig. 2), the other set processed for LC-MS/MS (Supplementary Table 2, see Online Methods for details).

Gel images were segmented into lanes, aligned, and the intensity of bands ranked. The lanes were clustered based on the intensity rank of bands exhibiting similar apparent molecular mass (Fig. 3b, Supplementary Fig. 3). MS data were filtered of exogenous and endogenous contaminants (Supplementary Table 3), and proteins exhibiting intensities below 10% that of the most intense species detected were removed. The results for each set were clustered based on the presence of common proteins and represented as a dendrogram (Fig. 3b, Supplementary Fig. 3). The cophenetic correlation coefficient between the two dendrograms was found to be 0.53 with a p-value of < 1 × 10⁻⁷ (Supplementary Fig. 4), indicating that both analytical methods describe the data similarly to a high degree of significance – revealing that the various extractions yield complexes with a number of distinct protein compositions and a high degree of redundancy; and also confirming that individual well failures do not compromise the conclusions of the whole experiment. Three proteins were revealed in the majority of conditions: Kap95p, Kap60p and Nup1p-SpA (Fig. 3b, green box), and a larger number of proteins were observed in e.g. low acetate (Fig. 3b, blue box) or high citrate/acetate (Fig. 3b, orange box).

Our conclusions are twofold. First, SDS-PAGE provides a representative readout of the composition of affinity captured samples, faithfully revealing the effects of changing affinity capture conditions, in a rapid, robust and inexpensive fashion. Second, with sufficient resources, direct sample-to-MS analytical approaches may be utilized. Both conclusions bolster the prospects for automated, unsupervised high-throughput approaches for optimizing affinity capture and dissecting interactomes (more below).

Exploring of the molecular organization of a 50 MDa complex

The yeast NPC is ~50 MDa in size and consists of multiple copies of ~30 different proteins. It presents an excellent test bed for our screen because it comprises a diverse physicochemical landscape and has a modular architecture consisting of subcomplexes of different sizes. Moreover, an extensive catalog of already existing affinity capture results³⁷, enabled us to assess our findings and the quality of results produced by the screen. Based on the above-described initial results, we further modified our conditions matrix to test other reagents and applied it collectively to SpA-tagged NPC proteins Nup1p, Nup53p and Pom152p.

For all three proteins the screen revealed novel conditions that exhibited improvements in yield, background, and hierarchical coverage, compared to the best results previously obtained³⁷ (Fig. 4a). Interestingly, all three proteins responded similarly to the extraction conditions presented (Fig. 4a): condition (i) gave small complexes with few interactors for all three, and no common components; condition (ii) gave more complicated profiles that partially overlapped in composition with each other; and condition (iii) gave the most complicated and highly similar profiles, representing almost the entirety of the NPC (Supplementary Data). These profiles were in agreement with the previously determined arrangement of proteins in the NPC, constituting interaction shells of increasing size and degree of overlap³⁷ (Fig. 4b). These overlaps demonstrated our method’s ability to “walk” from protein to protein via common complex components through the entire NPC, showing how reagent screening can be a tool for comprehensive interactomic mapping of different macromolecular complexes.

Figure 4.

NPC purification – from single proteins to macromolecular assemblies. (a) A representative SDS-PAGE image and MS analysis of affinity capture of 3 nucleoporins – Nup1p, Nup53p and Pom152p. Composition of affinity isolation solvents: i – 50 mM trisodium citrate, 300 mM NaCl, 10 mM Tween 20, 2 mM EDTA, 40 mM TRIS, pH 8; ii – 250 mM trisodium citrate, 10 mM Brij58, 0.3 mM Sarkosyl, 40 mM TRIS, pH 8; iii - 1.5 M ammonium acetate, 15 mM Triton X-100. The protein bands identified by MS are marked on the gel. The table below contains the list of identified proteins. Affinity tagged nups are labeled red, the remaining NPC constituents are black and non-NPC proteins are gray. Each protein identified by MS is marked by a dot under the corresponding lane. The brackets indicate comigrating proteins identified in a single band. (b) Section through the density map of the NPC³⁷ with one spoke enlarged and minimum Chimera representations of NPC subcomplexes in (a) for 1 spoke.

Simultaneous mapping of distinct interaction networks

We tested whether the screen was equally applicable for many different types of macromolecular complexes. We examined four proteins with two different tags acquired from commercial collections. These proteins (Arp2p-GFP, Csl4p-TAP, Snu71p-TAP and Rtn1p-GFP) exhibit distinct subcellular localization patterns and functions. Each protein was subjected to a 32-condition screen (Supplementary Figs. 5–8), allowing us to assay multiple proteins within the same 96-well plate. High quality copurification profiles were obtained (Fig. 5a, Supplementary Data), including the observation of novel and distinctive copurification patterns for proteins already extensively subjected to affinity capture MS strategies.

Figure 5.

Affinity capture strategy implementation on different protein complexes, affinity tags and model organisms. Affinity capture profiles: (a) S. cerevisiae; (b) E. coli; (c) H. sapiens. Representative SDS-PAGE profiles are shown. The cell schematics indicate the localization of the tagged proteins; the different tagged proteins screened are indicated in black; each lane corresponds to a different purification and is assigned an arbitrary Roman numeral (see Supplementary Table 1 for extraction conditions, except when specified). Some of the newly identified putative interactors were subjected to affinity capture (labeled brown) and the resulting profiles are indicated by arced arrows originating from the profiles in which they were identified. Copurifying protein bands identified by MS are marked next to each profile (see Supplementary Data). Protein names marked in blue are previously characterized physical interactors, those in red are select novel physical interactors or proteins of interest discussed in the main text, and those in grey are contaminants or proteins of indeterminate specificity based on their high frequency of copurification²⁰, or as determined by I-DIRT analysis (Fig. 6a). The bands labeled with an “*” indicate heavy and light chains of the antibody used for affinity capture. Ent2p and End3p extraction solvent – 40mM TRIS-Cl, pH 8, 150 mM NaCl, 250 mM trisodium citrate, 10 mM deoxy-BigCHAP, Tcb1p, Tcb2p, Tcb3p, and Dpm1p extraction solvent - 40 mM TRIS-Cl, pH 8, 150 mM NaCl, 50 mM trisodium citrate, 5 mM CHAPS.

The Arp2/3 complex is a conserved actin nucleator that participates in multiple actin-dependent processes, including endocytosis^38,39. Screening revealed a putative novel macromolecular assembly comprised of clathrin (Chc1p) with its adaptor protein (Ent2p) and actin (Act1p) with its recruiting and activating proteins⁴⁰ (Pan1p, End3p and Arp2/3 complex; Fig. 5a, Arp2p-GFP, i). Of these, however, only Act1p and Pan1p are known to physically interact with Arp2p^41,42. As End3p is linked to Arp2p genetically³⁸ and no direct links between Arp2p and Ent2p have been demonstrated, we chose to perform a secondary affinity capture for GFP-tagged versions of these proteins. The results (Fig. 5a, Ent2p-GFP, Ent3p-GFP and Arp2p-GFP, see also Supplementary Note 1), supported the role of Pan1p as a core scaffold, coupling actin and clathrin with the rest of the endocytic machinery⁴², in an interaction network involved in the early stages of actin-dependent clathrin-mediated endocytosis^42–46.

Csl4p is a component of the eukaryotic exosome, a modular multi-protein ribonuclease with compartment-specific components^47,48. Screening revealed conditions that selectively destabilized the compartment-specific components Rrp6p, Lrp1p, and Ski7p (Fig. 5a, Csl4p-TAP, iii and iv, compared with the canonical exosome, profile ii), while retaining the component Dis3p. Due to the relative stabilities of these different components in established purification conditions^48–50, strains with genetic deletions have been necessary to obtain comparable complexes (e.g.⁵⁰). In a separate profile, components of the cytoplasmic exosome cofactor Ski complex Ski2p/Ski3p were observed (Fig. 5a, Csl4p-TAP, i), which are considered recalcitrant to copurification^51,52.

Snu71p is a component of the nuclear localized U1 snRNP complex, a constituent of the spliceosome^53,54. Of interest, we purified U1 snRNP (Fig. 5a, Snu71p-TAP, ii) with nuclear mRNA associated proteins Sto1p and Pab1p and the major coat protein of ScVLA virus, which is known to covalently bind the mRNA cap⁵⁵ (Fig. 5a, Snu71p-TAP, i). We also noted a profile demonstrating a direct interaction between Snu71p and Prp40p (Fig. 5a, Snu71p-TAP, iii). Despite more than a decade of research on the composition of this complex, this dimer was only unambiguously shown recently by the introduction of a deletion mutation to a third constituent of this RNP that contains 17 distinct components⁵⁶. Here, it was obtained within a single screen.

Rtn1p, an integral endoplasmic reticulum (ER) membrane component, simultaneously embodies many of the challenges to affinity capture approaches: it is spread between multiple localizations and functionalities, is expected to form particularly dynamic or transient complexes^57–59, and as a membrane protein is among a class of proteins often refractory to interactomics⁶⁰. Although it is important in numerous cellular processes^58,59,61 and has been subjected to an affinity capture screen intended for membrane proteins²⁵, there are comparatively few validated physical interaction data available for Rtn1p. Lacking prior knowledge of the expected interactions, we selected a condition from our screen using our experience-based SDS-PAGE profile criteria (see Discussion). The copurifying proteins (Fig. 5a) included known and uncharacterized putative Rtn1p-interacting partners. Among these was the ER membrane-associated protein, Dpm1p, known to have a negative genetic interaction with Rtn1p⁶². We used secondary affinity capture to validate this interaction with Dpm1p, as well as with the ER membrane tricalbins (Tcb1-3p), which also copurified but have no previously demonstrated physical links to Rtn1p; all four GFP-tagged proteins copurified Rtn1p, the Tcb proteins each copurifed one another⁶³ and also yielded Dpm1p.

We repeated the Rtn1p-GFP affinity capture experiment, implementing isotopic differentiation of interactions as random or targeted (I-DIRT) analysis to distinguish interactions formed in vivo from those likely to be in vitro artifacts⁶⁴. I-DIRT analysis (see Supplementary Table 4 for unprocessed and analyzed data) indeed confirmed that most of the strong bands in our optimized affinity capture represent protein interactions with Rtn1p that were present in vivo (Fig. 6a), with a few prominent bands corresponding to common contaminants (see also Supplementary Note 1). Among putative Rtn1p interactors were 5 out of the 6 proteins known to tether ER to plasma membrane, namely Tcb1-3p, Ist2p, and Scs2p. Notably, Rtn1p, Tcb1p, and Tcb3p were identified in Sac1p (lipid phosphatase) and Scs2p (ER-plasma membrane tether) purifications⁶⁵. Given its proposed function to stabilize curved membranes^58,66, we suggest that Rtn1p may help stabilize membrane curvature at ER-plasma membrane contact sites where lipid transfer/modification occurs.

Figure 6.

An in-depth analysis of Rtn1p affinity capture. (a) Frequency distribution of I-DIRT ratios – light protein intensity/total protein intensity, normalized to 100% – from Rtn1p-GFP affinity capture experiment (extraction condition as in Fig. 5a, Rtn1p-GFP). Putative in vivo interactors are represented with shaded bars (≥85%), likely post-extraction associations – with open bars (<85%). Representative proteins are shown above each bar, proteins labeled in Figure 5a (Rtn1p-GPF) and known interactors of Rtn1p are bolded. The proteins with <85% and ≥85% I-DIRT ratio were separately analyzed for the subcellular localization and molecular function. The corresponding pie charts are plotted above the I-DIRT ratio distribution (see Methods for details). (b) SDS-PAGE and MS comparison of standard and optimized affinity capture. 4g of Rtn1p-TAP powder was processed essentially as previously described²⁵ using Triton X-100 as a detergent in a 2-step affinity capture experiment and 0.4g of Rtn1p-TAP was processed in a 1-step affinity capture experiment using conditions revealed in the present study (Fig. 5a, Rtn1p-GFP). Half of the elution was analyzed by SDS-PAGE followed by Coomassie staining and the other half was analyzed by LC-MS/MS. The distribution of subcellular localizations and molecular functions was analyzed as in (a) and is plotted bellow the corresponding gel lanes.

Could a standard affinity capture approach have reproduced these results? To address this question, we executed a side-by-side comparison of the popular tandem affinity purification procedure, recently tuned for membrane proteins²⁵, with an optimized procedure emerging from our screen (see Supplementary Table 4 for MS analysis). The results illustrate that the classic approach cannot compete with our screening strategy, either in terms of quality or efficiency (Fig. 6b).

Adaptability of the screen for diverse model organisms

Different model organisms often exhibit idiosyncrasies associated with affinity capture experiments. Our screen allows alternatives to be explored at each step. For example, an issue found with E. coli was the high viscosity of cell extracts due to high concentrations of released genomic DNA. We therefore modified our procedure to include a short low energy sonication using a multi-tip probe, sufficient to re-suspend the frozen cell powder during extraction and reduce viscosity to levels compatible with affinity capture. Similarly, low quantities of starting material may present an issue when working with tissue culture cells. Thus, we also modified the screen for a 24-well format in order to economize on cell usage. In our procedure, the mass of yeast cell pellet required per purification was reduced from the typical order of grams^25,67–69 to the range of tens to hundreds of milligrams; similarly, we consume only 50 mg cryomilled human cells per profile, an ~8-fold reduction over contemporary high-throughput studies^21,24. These modifications therefore enable economical interactomic screens in diverse model organisms.

From E. coli we purified the RNA polymerase (RNAP)⁷⁰ complex corresponding to the σ70 containing holoenzyme in complex with RapA⁷¹ and RpoC-SpA isolated away from RNAP (Supplementary Data, Fig. 5b, RpoC-SpA i and ii) demonstrating that the implemented modifications provide affinity capture results comparable in quality to those from yeast (Supplementary Fig. 9). Utilizing human cells, we revisited the RNA exosome, conducting purifications via a 3xFLAG-tagged hRRP6 (EXOSC10) – adding another common tag variety to those tested thus far (Supplementary Fig. 10). Among our observations (see also Supplementary Note 1), we noted the stable retention of SKIV2L2 (hMTR4) in numerous interaction profiles (see Supplementary Data). SKIV2L2 is also a member of the nuclear-specific human exosome cofactor NEXT complex^72,73, along with ZCCHC8 and RBM7. We readily observed another member of this complex, ZCCHC8⁷², in human exosome profiles (e.g. Fig. 5c RRP6-3xFLAG, i and ii), raising the question as to whether SKIV2L2 is single or multiple copy in the combined exosome/NEXT containing fractions. To extend our exploration we applied this screen to the NEXT complex itself, purified via LAP-tagged RBM7 (Supplementary Fig. 11). Among our findings, we observed NEXT in association with NCBP1 (CBP80), SRRT (ARS2), and ZC3H18 (NHN1) (Supplementary Data, Fig. 5c, RBM7-LAP, i; see also Supplementary Note 1) and made the novel observation of a direct interaction between RBM7 and ZCCHC8 (Fig. 5c, RBM7-LAP, compare ii to iii), demonstrating that these interactors form a stable dimer. We validated the above interactors extensively in a parallel study⁷⁴ via both MS-based label free quantitation and a suite of functional assays – revealing a new pathway of RNA surveillance utilizing the mRNA cap-binding complex, the NEXT complex and the exosome complex.

Robotic automation and gel curation for higher throughput

Our pipeline is designed for easy bench-scale execution. However, translation to automation has several advantages, including increased throughput and reproducibility. Using a liquid handling robot, we developed a version of the screen that includes automated production of extractant matrices and sample handling from the addition of affinity medium to clarified extracts through to the final wash. Given the intriguing results observed through the course of this study using trisodium citrate during manual screening (e.g. Fig. 4a), we implemented automation to systematically explore the effect of this reagent on Nup53p-SpA affinity capture profiles over 48 increments from 50 to 300 mM. We observed three distinct profiles (Fig. 7, detailed in Supplementary Note 1). These results demonstrated that the automated procedure was precise and revealed systematic changes in the copurification pattern specific to trisodium citrate, which involved the loss of Kap121p and the increased retention of a large number of NPC components as the concentration increased.

Figure 7.

Robotic implementation. SDS-PAGE analysis of one Nup53p-SpA purification screen performed using a Hamilton STAR liquid handling workstation, testing 50 – 300 mM trisodium citrate (40mM TRIS-Cl, pH 8, 1% v/v Triton X-100 are common to all lanes). The purifications form bracketed lanes were manually repeated and analyzed by MALDI-TOF-MS (Supplementary Data). Three distinct profiles are observed: copurification with Nup170p and Kap121p (i); dimer with Nup170p (ii); and a larger subcomplex of the NPC: Nup192p, Nup188p, Nup170p, Nup116p, Nup120p, Nsp1p and Nic96p (iii).

Because large amounts of data were generated during screening, we developed a web portal to assist in affinity capture data management. Our software (described in Supplementary Note 2 and publicly accessible at www.copurification.org) accepts images of gels along with experimental metadata. Gel images are automatically sectored lane-by-lane and annotated with the conditions applied to each, respectively. The lanes are also clustered according to protein banding pattern similarity, to ease the discovery of lane-to-lane differences and trends (Fig. 3b). This database provides a platform for the work of different experimentalists to be compared side-by-side, with instantaneous access to the respective experimental conditions for ease of reproduction.

Discussion

The solvent environment of the extractant plays a crucial role in dictating the stability of both real and artifactual protein-protein interactions during affinity capture^12–14. The difficulty in finding extractants that maximally explore the real interactome while minimizing artifacts^15–17 has limited high-throughput screens^{21–24,68,69,75}. Our approach addressed this limitation, providing a fast, efficient and cost-effective means to scan many conditions for their ability to preserve physiological interactions and minimize noise. This is particularly important for studies that hope to go beyond protein identifications and further obtain high quality protein preparations for biochemical and structural studies¹⁶.

Doubtless because of the huge diversity of interaction types, we have not yet found one set of conditions that works well for all the protein interactomes we have studied, underscoring the need for our screen. However, encouragingly, our results suggest that the optimal set of extraction conditions determined for a subset of constituents in a given complex, will suffice for the interactomic exploration of all the components in that complex (e.g., Fig. 4a). Moreover, we observed that during the secondary affinity capture of putative interactors identified in a primary screen (i.e. biochemical validation), copurification profiles containing both overlapping and distinct proteins were frequently obtained (see e.g. Fig. 4a, Fig. 5a, Arp2p-GFP and Rtn1p-GFP, and Fig. 5c RRP6-3xFLAG and RBM7-LAP) – highlighting the potential of this screen to uncover local (sub)complexes as well as the broader interactome. These combined attributes are particularly important given the current efforts to create a “human proteome encyclopedia⁷⁶” – which will undoubtedly require rigorous attention by any investigator to the preparation of the highest quality samples for affinity capture MS analyses.

For general purposes, we favor SDS-PAGE with protein staining for sample quality assessment, followed by band excision and MS to determine protein identities. SDS-PAGE is a proven, fast, parallel and quantifiable assay giving the approximate number, size, and amount of each band on the gel³⁵. Our findings, time and again, reinforced the notion that high quality affinity capture experiments are typified in SDS-PAGE profiles by a discrete pattern of sharp, abundant, and roughly stoichiometric bands as well as a paucity of background staining from other fainter bands (see e.g.^37,54, Figs. 4a and 5). The existence of increasingly sensitive general protein stains provides gel-based visualization options even for very low abundance samples⁷⁷. These criteria in turn allow for the judicious application of MS analyses to only the most potentially informative samples.

When tens to hundreds of SDS-PAGE protein copurification profiles are viewed in parallel, patterns of common and changing proteins and their solvent dependencies typically become readily apparent, and several promising conditions reveal themselves. A typical 96-well screen, as a consequence of being thorough, may yield many gel lanes with comparable banding profiles (Fig. 3b), and many that do not meet the criteria for further analysis. To modulate between throughput and screen breadth the total number of conditions can be adjusted (presented here at multiplicities of 96, 32, and 24).

While a promising protein copurification profile accompanied by high quality MS-based protein identifications provides the basis for a strong hypothesis regarding the existence of a physically associated protein complex that exists in vivo, such data should encourage the design of orthogonal experiments intended to rigorously test this hypothesis, including the importance of affinity capture optimization and the utility of I-DIRT in revealing high confidence targets for in vivo experimentation⁷⁸. Hence, in one sense, the presented screen can be considered a rapid and efficient hypothesis generation machine for physical interactions.

Our procedure is also compatible with direct sample-to-MS analyses (Fig. 3b) and can be implemented using robotic automation (Fig. 7), greatly enabling throughput. Future data mining opportunities will include the development of unsupervised, machine-based classification schemes to further improve our ability to identify promising samples, greatly augmenting high-throughput interactomic studies.

Online Methods

Affinity capture

All cell lines/strains utilized in this study are listed in Supplementary Table 5. Yeast, E. coli and Human cell lines were cultured using standard procedures and cryomilled and affinity captured essentially as previously described^29,30, except adapted for 96-well plates as described in text and elaborated step-wise in the Supplementary Protocol 1. Human cell lines have not been subjected to mycoplasma testing during the course of the study. Rabbit IgG used for purifying TAP and SpA tagged proteins was purchased from Innovative Research. Anti-GFP polyclonal antibodies were prepared and conjugated as previously described³⁰, except the concentration of ammonium sulfate used during the conjugation was 1.5 M. In all cases, proteins were eluted from the affinity medium by the addition of 1x NuPAGE LDS sample loading solution (Life Technologies); elution of GFP-tagged proteins included incubation at 70°C for 10 min. Extraction solvent working solutions were mixed from concentrated stock solutions in 2.5 ml deep-well plates (VWR) manually, using a Formulator (Formulatrix), or using a Hamilton STAR liquid handling workstation (program files provided in Supplementary Protocol 2). Supplementary Figure 12 contains detailed engineering diagram of the powder dispensing manifold. Resuspension of powders in extraction solvents included sonication with an ice water chilled microplate horn (yeast) or 8-tip micro-probe (bacteria & human) (Qsonica). Yeast lysates were also vortexed with steel beads to aid rapid homogenization. Supplementary Figure 13 displays the bead dispensing manifold utilized in yeast affinity capture experiments. Custom manufactured filters (Fig. 2, Orochem Technologies) were used to clarify yeast cell extracts for screens, otherwise extracts were clarified by centrifugation at 14k RPM and 4°C, for 10 min in a bench top microfuge. To ensure reproducibility all purifications presented (and processed for MS) were repeated individually in microfuge tubes using an otherwise comparable procedure except that extracts were clarified via centrifugation. Polyacrylamide gels were stained with either a homemade colloidal Coomassie brilliant blue G-250 solution⁷⁹ or with Imperial Protein Stain (Thermo Fisher Scientific). Gel images were recorded in TIFF using a Fujifilm LAS-3000 or an Epson Photo v700. In addition to the cited publications, detailed protocols for many preparatory procedures utilized in this study can be obtained at http://www.ncdir.org/public-resources/protocols/.

Standard mass spectrometric identification of proteins

The major bands observed in SDS-polyacrylamide gels were excised and analyzed either by MALDI-TOF-MS essentially as previously described³⁰, or nanoLC-ESI-MS/MS on an LTQ Orbitrap XL, Orbitrap Velos, Q Exactive Plus or Orbitrap Fusion mass spectrometer (Thermo Fisher Scientific). For analysis of excised protein bands using the LTQ Orbitrap XL or Orbitrap Velos the dry peptides were resuspended in 0.5% v/v acetic acid and pressure loaded on a self-packed C18 column and subjected to a 10 min gradient: 8 min 0 – 43%, 2 min 43 – 100% solvent B (solvent A = 0.1M acetic acid, solvent B = 0.1M acetic acid, 70% acetonitrile, 100 nl/min). The eluted peptides were analyzed with the following settings: top 10 peaks were selected for fragmentation, without dynamic exclusion. For the analysis on Q Exactive Plus and Orbitrap Fusion the peptides were resuspended in 0.1% formic acid, and separated using a 10 minutes gradient (8 min 0 – 30%, 2 min 30 – 100%, 1 min 100% solvent B) on an EASY-Spray column (Thermo Fisher Scientific) using an EASY-nLC 1000 (Thermo Fisher Scientific; solvent A = 0.1 % v/v formic acid, solvent B = 0.1 % v/v formic acid in acetonitrile, flow rate 300 nl/min). The 3 most abundant ions were selected in each full scan and sequentially fragmented by HCD with Q Exactive Plus. With Orbitrap Fusion a fixed duty cycle of 3 seconds was used. The RAW files were converted to MZXML format with the MM File Conversion tool (http://www.massmatrix.net/mm-cgi/downloads.py) or MGF format by ProteoWizard⁸⁰ and searched against the yeast protein database with X! Tandem⁸¹.

For the analysis of whole affinity captured fractions in Figures 3b and 6b, the protein samples were run ~4–6mm into an SDS-polyacrylamide gel (gel plug), and gels were Coomassie blue stained. Stained gel regions were excised, cut into 1 mm cubes, de-stained, and digested for 6 hours with 120 μl of 3.1 ng/μl trypsin (Promega) in 25 mM ammonium bicarbonate. An equal volume of 2.5 mg/ml POROS R2 20 μm beads (Life Technologies) in 5% v/v formic acid, 0.2% v/v trifluoroacetic acid was added, and the mixture incubated on a shaker at 4°C for 24 hr. Digests were desalted on C18 tips, eluted, dried by vacuum centrifugation, resuspended in 0.1% formic acid, and separated using a 10 minute gradient on an EASY-Spray column (as above). The 12 most abundant ions were selected in each full scan and sequentially fragmented by HCD (Q Exactive Plus); dynamic exclusion was enabled. RAW files were converted and searched as above. In order to determine the molecular functions of constituent proteins (Fig. 6b) we searched the gene descriptions for key words (as for I-DIRT analysis, see below).

Gel and MS data clustering and correlation analysis

The details of the gel image analysis are provided in the Supplementary Note 2; the source code is publicly available at https://github.com/FenyoLab/copurification. Once the lanes were sectored and bands identified, quantified and assigned an apparent molecular weight they were categorized as dark, light or, not observed (Fig. 3b, Supplementary Fig. 3).

Supplementary Table 2 contains the unfiltered search results of 96 LC-MS/MS runs. For each sample we extracted the intensities of all the hits, filtered out exogenous and endogenous contaminants (Supplementary Table 3) and considered the hits that were at least 10% as intense as the most intense hit (after initial contaminant filtering). We used a modified version of a source code available at the GPM repository ftp://ftp.thegpm.org to output the resulting protein sets as a pseudogel (Fig. 3b, Supplementary Fig. 3).

For both data sets, the Ward method was used for hierarchical clustering with distance between data points calculated as Euclidean distance⁸². To perform the correlation, the cophenetic distance was calculated for the gel and MS dendrograms⁸³. The cophenetic correlation was then calculated, which is defined as the Pearson correlation between the cophenetic distance matrices of the dendrograms⁸³. A p-value was obtained by a permutation test: the labels of the MS dendrogram were shuffled 10 million times, and a correlation calculated between the MS and gel dendrograms for each random shuffle (see Supplementary Fig. 4 for frequency distribution). A p-value of < 1 × 10⁻⁷ was calculated as a proportion of the random distribution equal to or greater than the actual correlation (0.53).

Graphical representation of NPC subcomplexes

We used the density maps for individual nucleoporins available at http://salilab.org/npc/ and the UCSF Chimera package⁸⁴ to graphically represent the NPC subcomplexes.

I-DIRT data analysis

I-DIRT was carried out essentially as previously described⁶⁴: the Rtn1-GFP strain was grown in synthetic complete minimal medium lacking lysine and supplemented with 50 mg/L of isotopically light lysine and a wild type DF5α strain was grown in the same medium but supplemented with 50 mg/L of isotopically heavy lysine (L-Lysine:2HCl 13C6, Cambridge Isotopes), both were frozen, mixed and cryomilled. Rtn1-GFP was affinity captured from the mixed powder extracted in 40 mM Tris-Cl pH 8, 50 mM trisodium citrate, 150 mM NaCl and 5 mM CHAPS. The eluted sample was reduced, alkylated and precipitated with methanol/chloroform⁸⁵. The precipitate was resuspended in 50 mM ammonium bicarbonate, 0.1% w/v RapiGest (Waters) via bath sonication with heating (20 min at 70°C, followed by 2min at 95°C). The proteins were digested with trypsin (Promega) overnight. Rapigest was depleted following the manufacturer’s instructions and the digest was desalted over C18 Omix tips (Agilient Technologies). The eluted fractions were analyzed on LTQ Orbitrap XL as described before except with 1h gradient and dynamic exclusion enabled. The output data was processed with MaxQuant⁸⁶ (http://maxquant.org/) using essentially the default parameters (for the light sample all amino acids were set to light, for the heavy sample 6 Da heavy lysine was selected; the yeast translated ORF sequences – http://www.yeastgenome.org/ – reversed sequences and contaminants database were searched) to identify and measure the intensity of heavy and light peptides. The “Evidence.txt” file containing all the peptide identifications and heavy/light measurements was used in the final analysis. Peptides mapped to contaminants, constituents of a reversed sequence database or those containing no lysine were excluded from the analysis. We further excluded non-unique peptides and peptides with a single MS/MS fragmentation event. For the remaining peptides I-DIRT ratios were calculated by dividing the intensity of the peptide with light lysine by the total intensity (light and heavy). To calculate the I-DIRT ratio of proteins the I-DIRT ratios of its constituent peptides were averaged. Proteins with a single peptide contributing to the I-DIRT ratio measurement were excluded from the analysis as unreliable. For proteins with 4 or more peptides contributing to the I-DIRT ratio measurement, those peptides with outlying I-DIRT ratios were filtered out (no more than 1 peptide removed per protein) using the following criteria: if the calculated I-DIRT ratio for a peptide was < Q1 (first quartile) - 1.5 x IQR (interquartile range) or > Q3 (third quartile) + (1.5 x IQR), that peptide was excluded. All statistical calculations and plotting were done with R (http://www.R-project.org)⁸⁷. To assess the normality of protein I-DIRT ratio distribution a Q-Q plot was constructed, which revealed a notable deviation from the y = x line implying that the data was not normally distributed (Supplementary Fig. 14). To assess the shape of the distribution the data were binned in 5% intervals (Supplementary Fig. 15). Despite a low bimodality coefficient^88,89 (0.2956), the distribution deviates significantly from unimodal by Hartigan’s dip test⁹⁰ (p-value = 0) and has a positive Akaike’s information criterion difference⁹¹ (0.1061) – suggesting bimodality. We used Mixtools package in R⁹² to fit 2 normal distributions to the data (Supplementary Figs. 15). We accepted a cut off of mean ± 2 standard deviations of the second distribution as stable interactors of Rtn1p (≥85%). We considered anything below 85% to constitute interactions indistinguishable from those formed post-extraction. In order to determine the molecular functions of constituent proteins we searched the gene descriptions for key words. The following are the categories and key words searched: Endoplasmic reticulum – “ er”/“endoplasmic reticulum”; Sugar metabolism – “glycolysis”/“gluconeogenesis”/“glucose”/“glycolytic”/“pentose”; Vacuole – “vacuol”; Ribosome/Translation – “ribosom”/“ translat”; Lipid metabolism – “lipid”/“fatty acid”/“choline”/“sterol”/“ceramide”; Mitochondrion – “mitochond”; Other - everything that didn’t match. All searches were case insensitive. Once a gene description matched a keyword it was put into the corresponding category, allowing us to count the number of proteins in each category and construct a pie chart of the distribution of molecular functions/localization for a given protein set (Fig. 6).