Tapioca: A Platform for Predicting De Novo Protein-Protein Interactions in Dynamic Contexts

Abstract

Protein-protein interactions (PPIs) drive cellular processes and responses to environmental cues, reflecting cellular state. Here, we develop Tapioca, an ensemble machine-learning framework for studying global PPIs in dynamic contexts. Tapioca predicts de novo interactions by integrating mass spectrometry interactome data from thermal/ion denaturation or co-fractionation workflows with protein properties and tissue-specific functional networks. Focusing on the thermal proximity coaggregation (TPCA) method, we improved the experimental workflow. Finely-tuned thermal denaturation afforded increased throughput, while cell lysis optimization enhanced protein detection from different subcellular compartments. This Tapioca workflow was next leveraged to investigate viral infection dynamics. Temporal PPIs were characterized during the reactivation from latency of the oncogenic Kaposi’s sarcoma-associated herpesvirus (KSHV). Together with functional assays, NUCKS was identified as a proviral hub protein, and a broader role was uncovered by integrating PPI networks from alpha- and beta-herpesvirus infections. Altogether, Tapioca provides a web-accessible platform for predicting PPIs in dynamic contexts.

Protein-protein interactions (PPIs) are at the core of cell health and disease. These interactions are highly dynamic, varying with cell state and providing the means to respond to a range of external stimuli. The global PPI network resulting from these interactions can be used to describe processes occurring in a cell at any given moment^1,2. Given the essential nature of PPIs in cellular biology, their study has been at the core of numerous biological and method development studies. These studies have led to the formation of several large PPI repositories, such as CORUM³, BioGRID⁴, STRING⁵, and IntAct⁶. A critical next step is to understand the interdependence of PPI regulation at a systems level, which represents signatures for different biological contexts (Fig. 1A). For example, a viral infection induces the temporal remodeling of PPIs at a global scale, resulting in the formation and dissolution of both known and de novo PPIs not yet contained within a repository⁷. Hence, to understand the complicated global coordination of PPIs in diverse dynamic contexts, experimental and computational methods for systems-level dynamic PPI characterization are needed.

Fig 1. Tapioca, a machine learning method for predicting global protein-protein interaction networks in dynamic contexts.

A, PPI repositories often represent static PPI networks, aggregated across many experimental conditions, while true cell states are dynamic subsets of these static PPI networks. B, Schematic representation of Thermal Proximity Coaggregation (TPCA) and Ion-based Proteome-Integrated Solubility Alteration (I-PISA) methods and resulting melting curves. C, Schematic representation of the Co-Fractionation (CF) method and produced protein distribution curves. D, The data modalities used by Tapioca to predict PPIs; ‘n’ represents the number of features used by Tapioca derived from the given data type (features listed in methods and supplementary tables S1G-J). E, Diagram of Tapioca’s eight sub-models, the features they use, and integration of sub-model predictions into the final Tapioca predicted protein-protein interaction network (see equations in methods section). F, Tapioca and sub-model performance by area under precision-recall curve (AURPC) and one minus the false positivity rate (1-FPR), evaluated on 48 datasets from 6 studies^{22,26,37–40}; the solid dot represents the mean value and the shaded region represents the 95% confidence interval. G, Examples of analyses enabled by Tapioca’s ability to accurately predict PPIs in dynamic contexts.

Various experimental methods have been designed to study PPIs from either a local or global perspective ⁸. Microscopy-based and yeast two-hybrid assays have provided the means to study interactions between sets of chosen proteins ^9,10, while protein microarrays have been aimed at constructing broader views of PPIs¹¹. A powerful set of methods for studying PPIs has been used in conjunction with mass spectrometry (MS), including affinity purification, proximity labeling, and crosslinking, allowing for the detection of PPIs without the need for prior knowledge ^12,13.

A well-established methodology is co-fractionation (CF) MS. In CF-MS, cells are gently lysed, preserving PPIs, and this lysate is then fractionated by an attribute, such as size in size exclusion chromatography (SEC) (Fig. 1C). Each of these fractions is analyzed by MS, resulting in an abundance versus fraction distribution curve for thousands of proteins. These distributions can be used to predict PPIs through several computational methods ^14–16. CF-MS allows the detection of proteins in multiple complexes, and the resolution of the method can be scaled by increasing the number of fractions¹⁷.

Another method for global PPI profiling is thermal proximity coaggregation (TPCA) MS. TPCA relies on the principle that interacting proteins tend to stabilize each other, thereby responding similarly when subjected to thermal denaturation ^18–20. In TPCA, intact cells are subjected to thermal denaturation along a temperature gradient (Fig. 1B). Each fraction along this gradient is lysed and the soluble portions are labeled using tandem mass tagging (TMT), allowing for their multiplexed MS-based analysis. This results in log-logistic shaped abundance versus temperature curves for thousands of proteins. These curves are used to predict PPIs, traditionally by measuring the Euclidean distance between all pairs of curves and using a distance cutoff to determine interactions ²¹. Although not yet specifically used for studying PPIs, a valuable recent adaptation of this method is Ion-based Proteome-Integrated Solubility Alteration (I-PISA), which monitors ion-based precipitation of proteins resulting in similar log-logistic shaped abundance versus concentration curves (Fig. 1B) ²². TPCA and I-PISA have higher throughput than CF-MS, being suitable for experiments involving many conditions and replicates. Additionally, TPCA and I-PISA can probe in vivo and ex vivo PPIs, while CF-MS only detects in vitro PPIs.

A challenge in assaying dynamic protein-protein interactions is the yet unmet need for general, accurate, and robust computational and analysis frameworks for these complex datasets, including for the accurate identification of de novo PPIs. A difficulty in accurately predicting PPIs in the context of TPCA and I-PISA is that the curves of many proteins are often in close proximity to other proteins that they do not interact with. Avoiding a high false positivity rate requires a strict distance cutoff, resulting in the missed detection of many real interactions. Since the distance cutoff depends on the distribution of distances between all pairs of protein curves, this threshold varies between datasets. Hence, the direct comparison or integration of PPI networks from different experiments is challenging. Together, these limitations make it difficult to predict de novo interactions, thus, TPCA work has primarily focused on the dynamics of known interactions (e.g., present in CORUM³).

Here, we developed Tapioca, an integrative machine learning based computational pipeline for de novo prediction of dynamic PPIs. Tapioca integrates MS curve data from TPCA, CF-MS, or I-PISA, with protein properties and tissue-specific functional networks. We demonstrate superior prediction performance compared to Euclidean distance-based methods. Using insights gained from generating Tapioca, we improved the experimental workflow of TPCA for PPI prediction. Using in silico and experimental methods, we optimized the temperature range for thermal denaturation, reducing the number of temperature points and increasing throughput, as well as cell lysis conditions, increasing the number and variety of proteins (localization) detected and PPI prediction quality. Finally, we applied Tapioca and our optimized TPCA workflow to study the dynamic context of a viral infection. Focusing on an important oncogenic virus, the gamma-herpesvirus Kaposi’s sarcoma-associated herpesvirus (KSHV) ^23–25, we created a temporally resolved global PPI network during KSHV reactivation from latency.

To date, few studies have implemented TPCA to investigate viral infections. However, previous work on herpesviruses^26,27 and SARS-COV-2²⁸ have demonstrated the value of TPCA for unbiased insights into the biology and pathogenesis of viral infections. The limited implementation of this method likely derives from difficulties connected to untangling the complex data and the computational skills and platforms needed. This challenge is exacerbated by the fact that virus infections drive the formation of numerous de novo PPIs, and therefore the monitoring of known protein complexes is not sufficient for understanding processes regulating infection. In this study, Tapioca-driven analysis led to the identification of de novo host-host, host-virus, and virus-virus PPIs during KSHV reactivation from latency. Further analysis and experimental validation using an orthogonal IP-MS approach led to the discovery of NUCKS, encoded by the gene nuclear ubiquitous casein and cyclin-dependent kinase substrate 1 (NUCKS1), as a cellular hub protein targeted by viral proteins. Tapioca-mediated integration of our KSHV, HSV-1, and HCMV datasets, followed by experimental virology assays, pointed to NUCKS as a broad-spectrum herpesvirus proviral factor.

RESULTS

Tapioca for PPI prediction in dynamic contexts

To address the need for computational tools that allow determining the dynamics of protein-protein interactions (PPIs) at a global scale, we developed Tapioca, a logistic regression-based ensemble machine learning framework. Tapioca provides the means to integrate curve-based dynamic PPI data from thermal proximity coaggregation (TPCA), Ion-based Proteome-Integrated Solubility Alteration (I-PISA), or co-fractionation (CF) mass spectrometry (MS) data with static interaction data to accurately predict PPIs in dynamic contexts (Fig. 1D). The prior interaction knowledge leveraged by Tapioca consists of sequence predicted protein physical properties, domain information from the PFAM database²⁹, and tissue-specific functional networks. These functional networks are constructed using a Bayesian probabilistic framework based on biological information from thousands of integrated -omics datasets (e.g., gene co-expression, transcription factor binding, and protein-protein interactions)^30–32. The Bayesian machine learning model automatically upweights datasets that are informative for the tissue of interest, thereby achieving tissue-specificity in the integration^33–36. By identifying closely connected proteins that share functional associations, albeit not necessarily via direct PPIs, these functional networks provide maps of predicted protein behavior in diverse cellular pathways.

Tapioca consists of eight sub-models, each utilizing unique combinations of static interaction and dynamics data derived from MS analyses (Fig. 1E). Each of these sub-models make predictions about the PPIs present within a given dataset (e.g., a TPCA dataset). As detailed in the next section, the integration of the scores derived from these sub-models was optimized to obtain a balance between accuracy of PPI prediction and maintenance of system dynamics, through calculating a Dynamics Correction Score (Fig. 1E). Tapioca was trained on six TPCA datasets from Tan et al. 2018 ²¹, and evaluated using a collection of 48 independent datasets, consisting of 30 TPCA^22,26,37,38, 16 CF^39,40, and 2 I-PISA²² datasets representing 11 tissue/cell types, using 5-fold cross validation of gold standard protein interactions (Fig. ED1A-B, tables S1A-F). The final version of Tapioca, which was used for all further analyses in this study, was trained and tested on a 70% train, 30% test split (Fig. 1F). We evaluated the Tapioca workflow using different machine learning algorithms—Logistic Regression, Naïve Bayes, and Random Forest. We found logistic regression offered an appropriate balance between PPI prediction accuracy and capture of system dynamics (Fig. ED2C-F). Tapioca greatly outperforms Euclidean distance (Fig. 1F), the traditional methodology of predicting PPIs from TPCA or I-PISA data, and performs well across a broad range of experimental methodologies and biological contexts.

As shown in this study, Tapioca predictions can be used for diverse downstream analyses, including the integration of dynamic PPI networks, aiding the understanding of protein roles across biological contexts, and the discovery of interactome similarity to predict functional homology for poorly characterized viral proteins (Fig. 1G). Through such applications, Tapioca enriches the ability to obtain biological discoveries from global PPI network experimental data. To allow for broad utilization of Tapioca, we provide this pipeline on GitHub (https://github.com/FunctionLab/tapioca) and through a user-friendly website (https://tapioca.princeton.edu/) for submitting data to be analyzed by Tapioca.

Balancing robust PPI prediction with system dynamics capture

Tapioca was trained using as a gold standard a set of well-defined CORUM complexes that have been observed in various experimental contexts⁴¹. The use of this gold standard assumes that these complexes, and the PPIs they represent, should be present in all datasets being trained on and evaluated against. We make use of this static gold standard because there is, to the best of our knowledge, no experimentally validated dataset of global PPI dynamics large enough and with varied enough biological contexts to be used as an alternative dynamic gold standard. A drawback of a static gold standard is that there is a risk of diminishing the ability of sub-models trained on it to capture system dynamics. This is due to the sub-models’ use of static interaction data. Sub-models may put greater weight on features associated with prior knowledge (static data), effectively diminishing the contribution of dynamics data (e.g., TPCA, I-PISA, or CF data) to sub-model predictions.

To understand the effect of increasing the amount of static interaction data used by a sub-model on its ability to capture system dynamics, we compared the predicted interactomes of fibroblasts that were either uninfected or infected with herpes simplex virus type-1 (HSV-1). HSV-1 infection induces global changes in host cell proteomes and PPI networks, and the protein interactomes at 15 hours post infection (HPI) (i.e., a late stage of infection) should reflect such changes (Fig. 2A). We observed that, while PPI prediction quality generally improved as the amount of prior knowledge used by a sub-model increased, its ability to capture system dynamics decreased (Fig. 2B). This indicates that, as anticipated, the sub-models generally put too much weight on static interaction data when making predictions, underrepresenting system dynamics. Tapioca’s integration of these sub-models with the dynamics data balances accurate PPI predictions with the capture of biological perturbations in the system.

Fig 2. Tapioca balances accurate PPI prediction with the preservation and capture of system dynamics.

A, Schematic representation of the assessment of system dynamics capture using HSV-1 infection (SD = standard deviation). B, Assessment of Tapioca and its sub-models by PPI prediction quality and capture of system dynamics in HSV-1 infection. C, Schematic of different methodologies for integrating the PPI predictions of Tapioca sub-models (𝑌 = model score, 𝐴 = AUC or AUPRC, 𝑟 = Pearson’s correlation) D, Assessment of PPI prediction quality and capture of system dynamics in HSV-1 infection on different methodologies for combining sub-model PPI predictions.

We further compared Tapioca’s approach for integrating sub-models with alternatives (Fig. 2C). Starting with simple integration methods, such as using the average of sub-model predictions (mean, median) or of their performance on our gold standard (AUC, AUPRC), we indeed observed some restoration of system dynamics (Fig. 2D). However, these approaches are static solutions to a dynamic problem, combining predictions the exact same way regardless of the data (mean, median) or they are combined by how well they meet the expectations of a static gold standard (AUC, AUPRC). Thus, we formulated a dynamics-driven solution, taking the Pearson’s correlation of the sub-models scores with Euclidean distance-derived scores. This correlation was used for the weighted integration sub-model scores. This method, which Tapioca employs for sub-model integration, provided the best performance of balancing system dynamics with reliable PPI prediction (Fig. 2D).

Tapioca generalizes to heterogeneous dynamics data

Tapioca was designed to be a generalizable tool for predicting PPIs from curve-based dynamics data, with feature generation being agnostic to curve resolution or shape. Trained initially on TPCA datasets, Tapioca accurately predicts PPIs from I-PISA data and, more significantly, from CF data (table S1A) despite the significant differences in curve shape between TPCA and CF data (Fig. 1B, 1C). To further test Tapioca’s generalizability, we evaluated its performance on heterogeneous dynamics data (HDD), which were generated by combining the curves of published CF and TPCA datasets (CF-TPCA) or I-PISA and TPCA datasets (I-PISA-TPCA) from either similar (CF-TPCA; Heusel et al. 2020⁴⁰ & Becher et al. 2018³⁸) or identical (I-PISA-TPCA; Beusch et al. 2022²²) biological contexts (Fig. 3A). We observed that HDD datasets generally performed on par with or better than their parent datasets (Fig. 3B). Tapioca generally assigned more conservative scores to interactions obtained from HDD than parent datasets, and hence fewer PPI predictions (Fig. 3C). Thus, at the cost of potentially losing some interactions captured specifically by only one of the experimental methods, integration of heterogeneous datasets can help to highlight subsets of high confidence PPI predictions.

Fig 3. Tapioca reliably predicts PPIs from heterogeneous dynamics data and identifies unique subsets of PPIs.

A, Schematic representation of the generation of heterogeneous dynamics data from TPCA, CF, and I-PISA datasets, and subsequent analysis by Tapioca. B, Evaluation of heterogeneous dynamics and source datasets by 1-FPR at different Tapioca score thresholds, the error bands (shaded region) represent the 95% confidence interval. C, The number of PPIs that were assigned scores equal to or greater than a given Tapioca score for heterogeneous dynamics and source datasets, the error bands (shaded region) represents the 95% confidence interval. D, Comparison of scores for a given PPI that were derived from either heterogeneous dynamics datasets or the corresponding source datasets (i.e., TPCA, CF, or I-PISA). E, UpSet plots comparing Tapioca predicted PPIs on heterogeneous dynamics and source datasets (PPIs called at independent score cutoffs with < 0.1 FPR for each dataset). F, Fraction of proteins with at least one unique PPI predicted by heterogeneous dynamics data and not captured by source data, versus the fraction of PPIs predicted by heterogeneous dynamics data and not captured by source data. G, The relative enrichment of known PPIs of the set of PPIs identified uniquely by CF-TPCA (not predicted in the corresponding CF and TPCA source datasets) compared to the random sub-sampling of all CF-TPCA predicted PPIs (repeated 10,000 times; see methods). H, The relative enrichment of known PPIs of the set of PPIs identified uniquely by I-PISA-TPCA (not predicted in the corresponding I-PISA and TPCA source datasets) compared to the random sub-sampling of all I-PISA-TPCA predicted PPIs (repeated 10,000 times; see methods). I, GO Term enrichment for proteins with at least on unique PPI predicted by CF-TPCA and not captured by CF and TPCA sources. M# (e.g., M1) represents a functional module identified by HumanBase⁶³ and the rest of the GO Terms associated with a module can be found in supplemental table S4A. J, GO Term enrichment for proteins with at least on unique PPI predicted by I-PISA-TPCA and not captured by I-PISA and TPCA sources (modules listed in supplemental table S4B).

To gain a better understanding of the PPI networks captured by using either HDD or parent datasets, we assessed the common and unique interactions identified with similar confidence in these datasets. As expected, the scores derived from HDD analysis showed modest correlation with those derived from its parent sources (Fig. 3D), suggesting that unique PPI predictions can also be obtained from HDD datasets. When considering the total PPI predictions, HDD datasets offered an intermediate number of interactions when compared to their parent datasets (Fig. 3E). In agreement with our observation above (Fig. 3D), we found that HDD datasets provided a considerable percentage of unique PPIs (40–50% of total PPIs) that were not predicted in either of the parent dataset, with ~65–85% of proteins having at least one HDD unique PPI. To understand whether these derived from true interactions or a failure of Tapioca to handle HDD, we looked at the enrichment of experimentally validated PPIs from BioGRID⁴, MINT⁶, and REACTOME⁴². HDD unique PPIs were enriched in known PPIs when compared to randomly selected subsets of all HDD PPIs (Fig. 3G-H). GO term enrichment on proteins with HDD unique PPIs displayed a broad representation across biological processes (Fig. 3I-J). Altogether, these results demonstrate that Tapioca can leverage HDD datasets to both predict PPIs more confidently and identify PPIs that would be missed using only homogeneous dynamics data.

Tapioca-driven optimization of TPCA thermal denaturation

During the development of Tapioca, we identified differential relative importance of temperature points for PPI prediction (Fig. 4A). Temperature points closer to the prototypical melting point displayed outsized learned weights compared to other points. Hence, we conducted in silico experiments to rank the relative importance of each temperature point. We created a collection of synthetic five-temperature TPCA datasets through the combinatorial sampling of ten-temperature datasets (Fig. 4B). The quality of PPI predictions from these synthetic datasets was evaluated (using AUC and AUPRC values), and these scores informed on the relative importance of each temperature point (Fig. 4C). This analysis indicated that a 37–55°C range would be sufficient for conducting thermal denaturation.

Fig 4. Optimizing the thermal denaturation and lysis conditions in TPCA for studying protein-protein interactions.

A, Schematic representation of the relative importance for accurate protein-protein interaction prediction of each temperature point on a TPCA melting curve learned by early proto-Tapioca models. B, Schematic representation of the in silico generation of synthetic five temperature TPCA datasets. C, The relative importance of each temperature for accurate protein-protein interaction prediction (using only Euclidean distance) from TPCA data determined from evaluation of in silico datasets, n=3 independent experiments; from each independent experiment 126 in silico datasets were generated and used. Error bars represent the 95% confidence interval. D, Evaluation by the area under a true positive versus false positive rate curve (AUC) of the original and optimized (Opt) thermal denaturation ranges. The use of ten (Opt10) and five (Opt5) temperature points within the optimized temperature range was evaluated. For original and Opt10, n=3 biologically independent samples. For Opt5, n=4 biologically independent samples. For boxplots, boxes show median, 25^th and 75^th percentile values, with the line within the box representing the median value, whiskers represent +/− 1.5 interquartile range, and points are outliers. Euclidean distance was used to predict protein-protein interactions. Exact values of data points can be found in table S5A. E, Schematic representation of lysis condition optimization experiments. F, Evaluation by number of identified proteins and AUC of the six tested lysis conditions; the solid dot represents the mean value, and the shaded region represents the 95% confidence interval. Euclidean distance was used to predict protein-protein interactions. Exact values of data points can be found in table S5B. G, Differential detection of proteins within different subcellular localizations; the ratio is the number or proteins identified using the TTD lysis conditions over the number of proteins identified using the mechanical (Mech) lysis condition.

To experimentally evaluate this computational prediction, we performed TPCA experiments using different temperature points and ranges (Fig. 4D). Cells were subjected to ten temperature points in either the original 37–64°C (Original) or the revised 37–55°C (Opt10) range. Improved PPI prediction quality was observed for the Opt10 range. We next evaluated our optimized five temperature range (37–55°C, Opt5), and found comparable performance to Opt10. Hence, the Opt5 range provides both improved PPI prediction quality and throughput of analysis.

Optimizing TPCA lysis conditions

The lysis conditions used in TPCA affect the number and types of proteins detected and the quality of the measured melting curves. For example, membrane bound proteins frequently exhibit irregular curve shapes in TPCA data, partly due to their limited solubility with routinely used lysis methods. Given that the melting curves used to predict PPIs are generated at the thermal denaturation step, TPCA may be amenable to the use of more stringent lysis conditions. To maximize the number of solubilized proteins in our biological system, while retaining PPI information, we performed TPCA using six different lysis conditions and evaluated the number of unique proteins identified and the quality of PPI predictions (Fig. 4E). Among the less stringent lysis methods, we found that syringe-based mechanical lysis yielded better PPI predictions, while NP40 identified more proteins. Lysis using CHAPS, a zwitterionic detergent known to protect the native state of proteins⁴³, as well as the non-ionic detergent DDM, used to solubilize membrane bound proteins⁴⁴, outperformed NP40. Two buffers optimized for immunopurification^45,46, referred to here as TT and TTD, gave the best performance in terms of the number of unique proteins identified and the quality of PPI prediction (Fig. 4E). These lysis conditions both use the non-ionic detergents Triton X-100 and Tween-20, and TTD additionally contain the ionic detergent sodium deoxycholate. We observed that the TTD buffer allowed the detection of proteins from multiple subcellular compartments, including membrane bound proteins from the ER, Golgi, and mitochondria. Altogether, these improved lysis methods afford a more complete understanding of global cellular PPI networks.

Tapioca study of KSHV reactivation from latency

Having optimized the TPCA protocol for improved PPI predictions, we applied our Tapioca framework to study the reactivation from latency of Kaposi’s sarcoma-associated herpesvirus (KSHV), an oncogenic gammaherpesvirus (Fig. 5A). TPCA was performed at 0, 12, 24, 48, and 72 hours post reactivation (HPR) in biological triplicate (Fig. ED2A). These times broadly represent latency, early reactivation, viral genome replication, and the assembly and egress of new virions. A total of ~7,000 unique proteins were identified, and Tapioca predicted ~45,000 to ~80,000 PPIs at each measured time point (Fig. ED2B). Throughout infection, numerous CORUM complexes showed temporal assembly and disassembly. For example, the MRE11-RAD50-NBN complex, involved in DNA repair and cell cycle checkpoint signaling, and the CNOT1-CNOT2-CNOT3 complex, involved in the regulation of transcription and gene silencing, became assembled primarily at 48 HPR (Fig. ED2C), a time point associated with KSHV genome replication.

Fig 5. Leveraging Tapioca to characterize global protein-protein interaction network dynamics during KSHV reactivation from latency.

A, Schematic representation of the reactivation from latency of Kaposi's sarcoma-associated herpesvirus (KSHV); reactivation and a temporal cascade of immediate early, delayed early, and late gene expression, genome replication, and virus assembly and egress occurring over 72 hours. B, The temporal dynamics of the relative abundance and number of participating protein-protein interactions of all detected KSHV proteins. C, Host proteins ranked by the number of unique KSHV proteins that they are predicted to interact with throughout KSHV reactivation. D, Tapioca scores of KSHV proteins that interact with NUCKS during at least one time point. E, Schematic of immunoaffinity purification (IP) MS study of the NUCKS interactome at 48 HPR, performed by using a series of experimental conditions—two different antibodies, two different lysis conditions, and two different strength nucleases. F, Relative abundance of NUCKS, RIR1, and RIR2 in IP-PRM using rabbit IgG (control) and NUCKS antibody, n=3 biologically independent samples. Error bars represent the standard error. Significance was determined by two-tailed Student’s t test (*= P < 0.05, **= P < 0.01, ***= P < 0.001, ****= P < 0.0001), NUCKS P=3E-4, RIR1 P=7E-5, RIR2 P=1.1E-2. G, The NUCKS KSHV interactome. Grey edges represent Tapioca/IP identified interactions between NUCKS and KSHV proteins, blue edges represent Tapioca identified interactions between KSHV proteins, and the dotted edge represents a previously known interaction. The green concentric circles represent the method(s) by which the NUCKS-KSHV PPI was identified. The ‘*’ proteins were determined as interactors by IP-PRM. H, GO Term enrichment of NUCKS 48 HPR host protein interactors identified by IP-MS. M# (e.g., M1) represents a functional module identified by HumanBase⁶³ and the rest of the GO Terms associated with a module can be found in supplemental table S8C. I, The temporal dynamics of GO term enrichment of host interactors of NUCKS; the solid line represents the mean value, and the shaded region represents the 95% confidence interval. J, Relative number of intracellular viral DNA (genomes) and number of infectious units after reactivation from latency of KSHV in CRISPR-mediated NUCKS1 knockout iSLK.219 cells compared to control (Scr) iSLK.219 cells, n=3 biologically independent samples. Error bars represent the standard error. Significance was determined by two-tailed Student’s t test (*= P < 0.05, **= P < 0.01, ***= P < 0.001, ****= P < 0.0001), Virus Production P=1.2E-2, Genome Amplification P=8.7E-3.

Throughout reactivation, 53 KSHV proteins were detected at time points corresponding to their temporal expression profile (immediate early, delayed early, and late) (Fig. 5B). Tapioca predicted that KSHV proteins participate in a combined total of ~200–1000 interactions at each time point throughout reactivation (Fig. ED2B), including known KSHV-KSHV interactions (or predicted by homology with other herpesviruses. An example is the RIR1 (ORF61) and RIR2 (ORF60) interaction, which forms the viral ribonucleoside-diphosphate reductase complex⁴⁷. Other examples include the tegument proteins LTP (ORF64) and ITP (ORF63), whose HSV-1 homologs interact to tether the outer and inner tegument proteins with the capsid^48–51, and TRX2 (ORF26) and CVC2 (ORF19), which make up structural components of the KSHV capsid^52,53.

Performing network-based module analysis using Lovein community clustering, we analyzed the temporal regulation of host processes by KSHV proteins during reactivation and replication. For each time point, we projected the set of host proteins predicted to interact with KSHV proteins onto a kidney-specific functional network (to reflect the iSLK.219 cells used) and identified coherent modules, with GO term enrichment analysis of each module identifying specific processes impacted. We then identified clusters of host processes with similar dynamics by K-means clustering of biological-process-specific vectors that represented the number of host proteins in that process interacting with viral proteins at each time point. Highlighted were terms associated with regulation of DNA replication, chromosome segregation, and transcription factor activity, which increased in representation between 24–48 HPR (Fig. ED3). Hence, KSHV reactivation drives the formation of diverse virus-host interactions and rewiring of CORUM complexes, peaking at 48 HPR, a time point associated with genome replication.

Identifying NUCKS as a Proviral Factor in KSHV Replication

An effective mechanism through which viruses can modulate and control cellular processes is by targeting host proteins. Throughout reactivation, a host protein targeted by the highest number of unique KSHV proteins (13) was NUCKS, encoded by the gene nuclear ubiquitous casein and cyclin-dependent kinase substrate 1 (NUCKS1) (Fig. 5C). NUCKS is a chromatin-associated protein with roles in DNA repair, transcription regulation, chromatin remodeling, metabolism, and cytokine expression^54,55. NUCKS plays a proviral role in HIV-1 infection via its interaction with the viral protein Tat⁵⁶, a transcriptional activator of viral gene expression, and is associated with several diseases, including cancer and Parkinson’s disease^54,55.

Tapioca predicted that an increase in NUCKS-viral protein interactions occurred at 48 HPR, which included RIR1, RIR2, and TK that play roles in nucleotide metabolism^47,57,58, and ORF45 that functions in late viral transcription⁵⁹ (Fig. 5D). Given the role of NUCKS in processes linked to genome replication and the relevance of the 48 HPR time point in KSHV genome replication, we aimed to validate these interactions using immunoaffinity purifications (IP) of endogenous NUCKS. As IP-MS analyses are known to be influenced by the affinity and specificity of the antibodies, as well as the lysis buffer used, we compared two antibodies, two lysis conditions (i.e., IP optimized versus TPCA buffers), and two strengths of nucleases (Fig. 5E, Fig S4). When considering KSHV proteins, nine proteins were detected, with RIR1 being the main viral protein identified in all conditions. Nevertheless, despite being predicted by Tapioca to interact with NUCKS (along with its known interactor RIR2), RIR1 failed to reach a statistically significant threshold in the NUCKS IP datasets (compared to control IPs). Hence, we designed a targeted mass spectrometry assay to obtain a more accurate quantification of the presence of RIR1 and RIR2 viral proteins in the IP datasets (Fig. 5F). Altogether, these untargeted and targeted analyses validated three of the four Tapioca predicted NUCKS-KSHV interactions at 48 HPR (RIR1, RIR2, and TK) and pointed to additional associations. These additional viral proteins were functional or binding partners of interactors predicted by Tapioca at 72 HPR, being linked to genome replication, nucleotide metabolism, gene expression, and immune regulation (Fig. 5G).

We next compared Tapioca and IP-MS identified host interactors of NUCKS. We found that the higher the Tapioca score, the more likely a protein was to be identified as an interactor by IP-MS (Fig. ED5C). A total of 61 IP-MS identified interactions overlapped with Tapioca prediction (Fig. ED5A, ED5D, see methods section “Considerations for comparing IP-MS and TPCA data”). Overall, despite differences between IP-MS and TPCA, we observed that most IP-MS unique interactions (83%) had known interactions (by BioGRID⁴, REACTOME⁴², and MINT⁶⁰) with Tapioca interactions (Fig S5D). Indicating a non-random intersection, this observation contrasted the 2% overlap seen between the IP-MS datasets and the broader Tapioca interactomes for all other proteins in the KSHV TPCA dataset. GO Term enrichment on the NUCKS IP data highlighted interactions with host proteins involved in mRNA processing, chromosome organization, and nucleotide metabolism (Fig. 5H).

Two of the main features of the Tapioca identified NUCKS interactome were 1) the large number of interactions (top 99.5^th percentile), suggesting that NUCKS may act as a hub protein, and 2) the temporal regulation, with no interactions being maintained across all time points and about a half detected at four time points. In particular, NUCKS interactions with proteins involved in innate immunity decreased throughout reactivation, while interactions with proteins involved in DNA repair and chromosome organization were retained (Fig. 5I, S6A). Given the targeting of NUCKS by KSHV proteins and its association with host factors involved in processes critical for an infection outcome, we sought to determine the effect of NUCKS on KSHV replication. We performed CRISPR-mediated knockout of NUCKS1 in iSLK.219 cells and measured the effect on the production of KSHV genomes and infectious virions using qPCR at 48 HPR and titering at 96 HPR, respectively. A statistically significant decrease was observed for both genomes and infectious virions upon NUCKS1 knockout (Fig. 5J). These results establish NUCKS as a proviral factor during KSHV infection and point to a role likely occurring at or before KSHV genome replication.

NUCKS is a broad-spectrum proviral factor for herpesviruses

Given the proviral role of NUCKS for KSHV and its interactions with regulators of critical processes for viral replication, including nucleotide metabolism, DNA replication, and innate immunity, we investigated whether NUCKS may have a broader role in herpesvirus infections. We used Tapioca to reanalyze previously published temporally resolved TPCA datasets during infections with HSV-1²⁶, an alphaherpesvirus, and human cytomegalovirus (HCMV)²⁷, a betaherpesvirus. When assessing NUCKS interactions with viral proteins, either no (HSV-1) or few (HCMV) interactions were predicted, with the only two HCMV interactors predicted being the poorly understood proteins UL22A/UL20A and IR10. Hence, to understand the function of NUCKS in these contexts, we designed a different analysis pipeline. We asked whether there are similarities between the observed temporal profiles of interaction networks that can point to functional homologies between viral proteins from these different viruses. Leveraging the temporal dimensions of these TPCA datasets we predicted “interactome similarity” scores between HSV-1, HCMV, and KSHV proteins. Viral proteins (from distinct viruses) that share correlated temporal dynamics in their interactions with the same host proteins are assigned high scores and considered as potentially having some level of functional homology (Fig 6A). This analysis identified specific subsets of HSV-1 and HCMV proteins that had high “interactome similarity” scores with KSHV proteins predicted to interact with NUCKS (Fig. 6B). Specifically, the resulting interactome homology network contained 33 HSV-1 and six HCMV proteins. Similar to the KSHV NUCKS interactors, these HSV-1 and HCMV proteins have known roles in genome replication, nucleotide metabolism, gene expression, and immune evasion.

Fig 6. NUCKS plays a proviral role across herpesvirus infections.

A, Schematic representing the creation of an interactome homology network for a given pair of viral infections. B, The interactome homology network of KSHV proteins that interact with NUCKS. C, The GO term enrichment of NUCKS host interactors throughout HSV-1 and HCMV infections. M# (e.g., M1) represents a functional module identified by HumanBase⁶³, and the rest of the GO Terms associated with a module can be found in supplemental tables S10A-S10B. D, Relative number of intracellular viral DNA (genomes) and number of infectious units in CRISPR-mediated NUCKS1 knockout compared to scramble (Scr) after infection of HFF cells with HSV-1 and MRC5 cells with HCMV, n=3 biologically independent samples. Error bars represent the standard error. Significance was determined by two-tailed Student’s t test (*= P < 0.05, **= P < 0.01, ***= P < 0.001, ****= P < 0.0001), Virus Production: HSV-1 P=4E-5, HCMV P=4E-4, Genome Amplification: HSV-1 P=1.5E-3, HCMV P=2.4E-2. E, Schematic of Tapioca-enabled analyses for systems level biological questions.

Given the ability of our method to identify viral clusters that display high interactome similarities, we ask whether this approach may also aid in predicting the function of poorly characterized viral proteins. For example, the KSHV protein RIR2, which we validated to interact with NUCKS, showed high interactome similarity with the uncharacterized HSV-1 nuclear protein UL4. In agreement with the known role of KHSV RIR2 in nucleotide metabolism, HSV-1 UL4 clustered with KSHV and HSV-1 proteins involved in genome replication and virion assembly and egress. Leveraging HVint2.0⁶¹, a database of herpesvirus-herpesvirus PPIs predicted from homology, experimental data, and consensus clustering, we observed that UL4 has predicted associations with HSV-1 proteins involved in nucleotide metabolism, genome replication, virion assembly and egress, and regulation of gene expression. The large overlap in biological processes linked to UL4 by either Tapioca-derived interaction similarity network or HVint2.0, distinct methodologies, indicates that our approach may be useful for predicting the functions of poorly characterized viral proteins.

The identification of similar viral functional clusters within the NUCKS interactomes for these three infections suggests that NUCKS is positioned to modulate a number of critical host and viral pathways across infections. Furthermore, when considering its interactions with host proteins during HSV-1 and HCMV infections, NUCKS associates with the same processes observed during KSHV reactivation, e.g., nucleotide metabolism and immune regulation (Fig. 6C). These viral and host interactions suggest that NUCKS may be positioned to modulate critical pathways during these infections. Hence, we performed HSV-1 and HCMV infections in NUCKS1-KO fibroblasts. Matching our findings for KSHV infection, statistically significant decreases in viral genome amplification and virus production were observed for both viruses upon NUCKS1-KO, indicating a broader proviral role for NUCKS in herpesvirus infections (Fig. 6D).

DISCUSSION

Here, we developed Tapioca, a machine learning framework for global PPI network prediction in dynamic contexts (Fig. 6E). We demonstrate the broad applicability of Tapioca within multiple MS-based workflows, implementing it for the analysis of global PPI data obtained from TPCA, CF, and I-PISA MS workflows and the integration of these results with auxiliary information. This automatic integration provides robust, high-confidence interaction predictions and reliably captures PPI dynamics. Additionally, we show that CF, TPCA and I-PISA datasets can be directly integrated using Tapioca to enhance the robustness of PPI prediction, as well as to expand the subsets of identified PPI. We demonstrate Tapioca’s applicability for analyzing MS data across diverse cell types, experimental conditions, and species. To facilitate the accessibility of this user-friendly platform, we provide Tapioca on GitHub (https://github.com/FunctionLab/tapioca) and have generated a website (https://tapioca.princeton.edu/) for easy implementation of Tapioca for the analysis of user-generated datasets, as well as for access to the temporal PPI dynamics datasets during HSV-1, HCMV, and KSHV infections.

In addition to providing a powerful computational framework for data analysis, we experimentally optimized the TPCA workflow, increasing throughput and improving the ability to access proteins within different subcellular compartments. Importantly, we show that five temperature point thermal denaturation in an optimized range provides the necessary information to predict high confidence PPIs. Thus, it is possible to reduce runtime and/or the cost of TPCA experiments through greater multiplexing, facilitating the application of TPCA to a broader range of experimental conditions. Of note, our temperature optimizations were performed in human cell culture. Other species or systems may exhibit different thermal denaturation ranges, as recently reported ⁶². It is also relevant to note that this abbreviated thermal denaturation temperature range will result in the loss of some information (e.g., melting temperature) for more thermostable proteins. The methodology described here provides a framework for performing similar optimizations in other systems.

Leveraging this optimized Tapioca platform, we experimentally characterized processes that underlie the reactivation from latency of an important oncogenic virus. We defined the temporal dynamics of global PPI networks during KSHV reactivation, uncovering de novo host-host, virus-host, and virus-virus protein interactions. Tapioca analysis led to the recognition of the cellular protein NUCKS as playing a proviral role in KSHV infection and characterized its interactions with KSHV proteins. This rich dataset can be explored in future studies to better understand the many facets of KSHV infection, such as uncovering other cellular factors critical for KSHV replication, understanding host responses to infection, and determining the functions of poorly characterized KSHV proteins and drivers of virus pathology. The extension of Tapioca analysis to HSV-1 and HCMV infections discovered that NUCKS, a protein previously not implicated in herpesvirus infections, is a broad-spectrum proviral factor in herpesvirus replication.

Overall, the integrated computational and experimental approach presented here advances the field’s ability to predict global PPI network dynamics. Additionally, the experimental temporally resolved TPCA dataset of KSHV reactivation from latency can enable further studies of this important oncogenic virus.

ONLINE METHODS

Tapioca feature generation

Tapioca uses four different classes of features: mass spectrometry-based dynamics data (e.g., TPCA, CF, I-PISA), protein physical properties, PFAM database data, tissue-specific functional networks.

Dynamics data.

Normalized dynamics data (curves) is used to calculate a total of 23 features. The first 13 of these values encompass the absolute and relative distance between the curves, and the integrals and derivatives of these curves, of the two given proteins. The next 7 values are the Pearson’s correlation between the curves, and the integrals and derivatives of these curves. The last three values are the minimum, mean, and maximum relative Euclidean distance z-score between the curves. The relative Euclidean distance z-score is calculated individually for each curve in the pair. For each curve in the pair, the z-score is calculated such that it represents how close curve A is to curve B relative to the distance between curve A and all curves in the dataset, and vice versa for curve B. Thus, the unlike all prior distance calculations, the relative Euclidean distance z-score is non-symmetrical. See supplemental table S1G for more details.

Protein physical properties.

The physical property information is represented by 18 values which encompass values pertaining to length, gravy (grand average of hydropathy), molecular weight, aromaticity, instability index, isoelectric point, and predicted fraction of helix, turn, and sheet secondary structures. For each of these properties, the average value and absolute difference of values for the given pair of proteins are used as features. Se supplemental table S1H for more details.

PFAM data.

The PFAM database information is represented by 12 values which encompass the minimum, mean, max of the co-occurrence for all possible pairs do domains, families, or clans from assigned, by PFAM, to the given pair of proteins. The co-occurrence of a given pair of domains, families, and clans was calculated as the percentage of time that the pair appears for in a PPI versus compared to all PPIs that the given domains, families, and clans participate in. The PPIs used were those with experimental validation from the BioGRID⁴, Reactome⁴², and MINT⁶⁰ databases. The total number of domains, families, and clans shared by the given pair of proteins are also used as features. See supplemental table S1I for details.

Tissue-specific functional networks.

The score assigned to a given pair of proteins from a tissue-specific functional network (networks are from HumanBase⁶³, https://hb.flatironinstitute.org/). Pairs not represented in the network are assigned a score of zero. The score is used as a single-value feature in Tapioca. See supplemental table S1J for more details.

Tapioca model structure

Tapioca consists of eight logistic regression sub-models made using the Python Scikit-Learn package’s SGDClassifier (Logistic Regression), RandomForestClassifier (Random Forest), or GaussianNB (Naïve Bayes) and CalibratedClassifierCV classes. All the logistic regression sub-models use curve-based dynamic MS data (i.e TPCA, I-PISA, or CF data) as a feature. They additionally all take unique combinations of the tissue specific functional network, physical properties, and PFAM feature sets.

The dynamic MS data and functional network inputs are automatically combined by Tapioca with sequence-predicted physical protein properties (ex. molecular weight and isoelectric point) and information from the PFAM²⁹ database (domains, families, and clans). Tapioca itself consists of eight distinct logistic regression models that take unique combinations of these features as inputs. The predictions of these eight models are combined into a final interaction score for each pair of proteins in a given dataset.

Each of the eight sub-models can be represented by the equation:

$$Y_{sub-model}=\frac{e^{B+{\sum }_{i=1}^n{X}_n{W}_n}}{1+e^{B+{\sum }_{i=1}^n{X}_n{W}_n}}$$ Eq. 1

Where the predicted probability of a PPI is $Y_{sub-model}$, the set of inputted features are $X$, the set of learned weights are $W$ (weights are learned independently for each sub model), the number of features, $n$, and a bias term $B$. To integrate the predictions of the sub-models into the final Tapioca score, the Pearson’s correlation between the scores from the sub-models and Euclidean distance (representing a “dynamics score”) is calculated as:

$$r_{sub-model}=\frac{{\sum }_{i=1}^n\left(Y_{sub-model,i}-{\overline{Y}}_{sub-model}\right)\left(Y_{euclidean,i}-{\overline{Y}}_{euclidean}\right)}{\sqrt{{\sum }_{i=1}^n{\left(Y_{sub-model,i}-{\overline{Y}}_{sub-model}\right)}^2{\sum }_{i=1}^n{\left(Y_{sub-model,i}-{\overline{Y}}_{sub-model}\right)}^2}}$$ Eq. 2

Where the correlation is $r_{sub-model}$, the score of a single PPI of the sub-model is $Y_{sub-model,i}$, its dynamics score is $Y_{euclidean,i}$, the mean PPI score of the sub-model is ${\overline{Y}}_{sub-model}$, the mean dynamics score is (${\overline{Y}}_{sub-model}$ and $n$ is the total number of PPI interactions. The final Tapioca score for a single PPI is given by:

$$Y_{Tapioca}=\frac{{\sum }_{i=1}^8{Y}_i{r}_i}{{\sum }_{i=1}^8{r}_i}$$ Eq. 3

Where the Tapioca score is $Y_{Tapioca}$, and the score and Person’s correlation from each of the eight sub-models is $Y_i$ and $r_i$, respectively.

Each of the eight models were trained and tested independently using 5-fold cross validation. Tapioca uses the following python packages: Numpy⁶⁴, Scikit-learn⁶⁵, Scipy⁶⁶, Pandas^67,68, and Biopython⁶⁹.

Generating gold standard

The gold standard described in Stacey et al. 2018⁴¹ was used as a foundation on which to construct a gold standard to train and evaluate Tapioca. The proteins that made up CORUM complexes in the Stacey gold standard were separated into training set (70%) and a test set (30%) proteins (70/30 gold standard), such that there was no protein appeared in both the train and test set. Additionally, a separate 5-fold cross-validation was performed, again assuring that no protein appeared in both the train and test set. All nonredundant pairs of proteins that occurred within the same CORUM complex were used as true examples of PPIs. All pairs of proteins within the train and test sets that were not being used as true PPIs (true positive examples) were used as proto-false PPIs. This set of proto-false PPIs were first trimmed such that any pair of proteins that occurred in any of the same subcellular localizations, using the human protein atlas database ⁷⁰ for protein localizations, were discarded from the proto-false PPI set. Proteins with no localization data were preserved. The set human PPIs from the BioGRID⁴, REACTOME⁴², MINT⁶⁰, and STRING⁵ databases were used to identify proto-false PPIs with some level of evidence that indicated they were true interactions. Those with any evidence of being a true PPI were removed from the proto-false PPI set, and those that remained were used as the true false examples in the gold standard.

Training and evaluating Tapioca

Tapioca was trained using six TPCA datasets representing six cell types (K562, A375, HCT116, HEK293T, HL60, MCF7) from Tan et al. 2018²¹. Each of the eight sub-models of Tapioca were trained independently on the same data. Tapioca was evaluated using a collection 48 datasets from Becher et al. 2018³⁸, Heusel et al. 2020⁴⁰, Justice et al. 2021²⁶, Skinnider et al. 2021³⁹, Beusch et al. 2022²², and George et al. 2023³⁷, where a single dataset is the TPCA, I-PISA, or CF data from a single condition and replicate. Training and evaluation were performed for both the 70/30 gold standard and by 5-fold cross-validation. The area under a precision versus recall curve (AUPRC) and area under a true positive versus false positive curve (AUC), and false positivity rate at a score of 0.5 were used in evaluation of model performance. During evaluation, only true interactions from the gold standard that were supported by two or more publications were considered. Additionally, to standardize the evaluation, only proteins commonly detected across the Tan and Justice datasets were considered.

Tapioca performance on randomized dynamics data

To understand the effects of randomized dynamics data on PPI prediction, Tapioca was evaluated on the same 48 datasets in which all curves (dynamics data) was randomized prior to their input into Tapioca. Evaluation was performed using the 70/30 gold standard. See Fig S6B for results.

Generation of non-logistic regression-based Tapioca models

Additional Tapioca models were constructed in which the only change was that the sub-models were made using the Python Scikit-Learn package’s GaussianNB (Naïve Bayes), RandomForestClassifier (Random Forest), and AdaBoostClassifier (AdaBoost) instead of the SGDClassifier (Logistic Regression). These models were evaluated both by the 70/30 gold standard and by 5-fold cross validation.

Assessing the capture of systems dynamics

To assess the capture of systems dynamics by Tapioca, its sub-models, and different ways to integrate sub-model predictions, these models were used to predict PPI networks at both 0 hours post infection (HPI) and 15 HPI in herpes simplex virus type-1 infection (data from Justice et al. 2021²⁶). For all PPIs achieved a score of at least 0.5 at either 0 or 15 HPI the PPI score at 15 HPI was subtracted from the score at 0 HPI. The standard deviation of the resulting distribution was used as a metrics to assess systems dynamics, with higher standard deviations generally being expected to represent a greater capture of system dynamics. Relevantly a very high standard deviation could indicate instability in a model’s predictions. (i.e., the model has failed to generalize), though absence a dynamics gold standard it is unclear at which point a high standard deviation value transitions from good systems dynamics capture to unreliable predictions.

Sub-model integration methods

Multiple methods of sub-model prediction integration were performed. For unweighted integration, the mean or median of the scores assigned by each sub-model to an individual PPI was calculated and this value was used as the final score for the given PPI. For weighted integration, the AUPRC or AUC of a sub-model was calculated then, for every individual PPI, the AUPRC or AUC values were used as weights when taking the weighted average of sub-model scores for a given PPI. For the Tapioca method of sub-model prediction integration see the “Tapioca model structure” section of the methods.

Difference in PPI predictions between Tapioca and sub-models

To understand the effect of sub-model integration on the predicted interactomes we analyzed the differences in interactomes predicted by sub-models compared to Tapioca. Across all 48 datasets used to evaluate Tapioca, the Tapioca score was subtracted from each individual sub-model score assigned to a given PPI, for all PPIs that achieved a score of at least 0.5 by either Tapioca or the given sub-model. We found that the sub-models assign higher PPI scores than Tapioca by ~0.2 (Fig. ED6C). This finding suggests that the sub-models may over predict PPIs, explaining why Tapioca generally has a higher 1-FPR than its sub-models (Fig. 1F). When assessing the PPIs predicted uniquely by the sub-models (i.e., not by Tapioca; table S3A-I), we observe that these proteins were largely sub-model specific (Fig. ED4B, ED6D, ED7, tables S3J-Q for GO term enrichment). Hence, Tapioca provides conservative PPI predictions that retain the dynamics context.

UpSet Plots

An UpSet plot can be thought of as a bar chart-based representation of a Venn diagram. The horizontal bar chart shows the total number of items assigned to a given category (e.g., the total number of proteins found in a given condition). The vertical bar chart represents the number of items shared between (the intersection) a given set of conditions. The matrix below the vertical bar chart shows which set of conditions is being compared for each given bar of the bar chart. A black dot means the given condition is included in the comparison.

HumanBase GO term enrichments

GO term enrichment was performed using HumanBase⁶³ (https://hb.flatironinstitute.org/), which identifies functional clusters, or module, within the set of genes submitted. HumanBase creates graphical representations of these modules, which have be used here in figures. In these figures, the nodes represent proteins and are colored by the module that the given protein belongs to. The size of the node indicates the relative number of other proteins with which the given protein has a high degree of functional relatedness, with a larger node indicated a greater number of proteins. For each of the graphical representations, a supplementary excel table has been provided detailing the names each protein in each module, as well as the complete list of GO terms enriched in that cluster.

Generating heterogeneous dynamics data

Heterogeneous dynamics data was created by appending TPCA data to non-TPCA data (e.g., appending TPCA curves to CF curves). Curves were appended at the individual protein level and the TPCA values were uniformly adjusted such that the first value of the TPCA curve equaled the last value of the non-TPCA curve. Proteins that were not present in both original datasets were not included in the heterogeneous dataset.

Calculating the relative enrichment of known PPIs

Relative enrichment was calculated by first ranked PPIs by score from highest to lowest, and then the first 50,000 PPIs are kept. Each PPI is then assigned a value using the following equation:

$$scor{e}_n=\frac{\left(\frac{{\sum }_{i=1}^n{T}_i}{n}\right)}{\left(\frac{L_t}{L_n}\right)},T_i=\left\{\begin{array}{ll}1,\hfill & P_i\in K\hfill \\ 0,\hfill & P_i\notin K\hfill \end{array}\right.$$ Eq 4

Where the enrichment score, $scor{e}_n$, assigned to the n^th PPI, is equal to the sum of all PPIs from the first PPI up to n^th PPIs, where each PPI, $P_i$, is assigned a score, $T_i$, which is 1 if the PPI is in the set of known PPIs, $K$, else is 0, divided by $n$, all divided by the length of the set of known PPIs, $L_t$, divided by the length all PPIs, $L_n$. BioGRID⁴, MINT⁶, and REACTOME⁴² were used as the source of known PPIs. The enrichment score of each was then plotted against the respective rank of the PPI. The area under this resulting curve was calculated, giving the relative enrichment score.

In silico thermal denaturation temperature optimization

In silico optimization of thermal denaturation temperatures was performed by first creating a set of synthetic five-temperature datasets using ten-temperature datasets from Justice et al.²⁶ and Hashimoto et al.²⁷. The synthetic datasets were created such that for each original dataset there existed a set of synthetic datasets which contained all unique combinations of five temperature points from the original datasets. PPIs were predicted using Euclidean distance from each of these synthetic datasets and these predictions were evaluated by AUC and AUPRC. Each of these datasets were then ranked by their summed AUC and AUPRC scores, with those having a higher score receiving a lower rank. Each temperature in the original ten-temperature dataset was assigned the average rank of the set of synesthetic datasets that used that given temperature.

TPCA MS sample preparation

Samples were prepared as in Justice et al. 2021²⁶, with modification to the thermal denaturation temperatures and lysis conditions, which are described in the ‘Optimization of thermal proximity coaggregation methodology’ section of the methods. In the KSHV experiments, thermal denaturation was performed using ten temperatures between 37°C and 55°C (37°C, 38.5°C, 42°C, 43.3°C, 45°C, 47°C, 48.5°C, 52°C, 53.4°C, 55°C) and lysis was performed using 1% Triton-X100 and 0.1% Tween-20 with 0.5% Sodium Deoxycholate in 20 mM HEPES, 110 mM KOAc, 2 mM MgCl₂, 1 μM ZnCl₂, and 1 μM CaCl₂, at pH 7.4. Briefly, cells were trypsinized, pelleted, and then resuspended in 1X PBS and then aliquoted at 50 μL per temperature in PCR tubes. Cells were then subjected to thermal denaturation with a thermal cycler (Thermo Fisher Scientific) for 3 minutes then held at 4°C for three minutes before 100 μL of 1.5X lysis buffer (75 mM HEPES pH 7.5, 15 mM MgCl2, 1.5X HALT protease and phosphatase inhibitor, 3 mM tris(2-carboxyethylphosphine) (TCEP)) was added to each tube. Samples were snapped frozen in liquid nitrogen and sored at −80°C until MS analysis. Samples were thawed at room temperature and lysed. Heat denatured proteins were pelleted at 20,000 x g for 20 minutes at 4°C. The supernatants, containing the soluble proteins, were transferred to a second clean 1.5 ml low bind tube before being reduced and alkylated with 25 mM TCEP and 25 mM chloroacetamide for 30 minutes at 55°C. Proteins were precipitated by methanol chloroform then resuspended in 160 μL 100 mM HEPES (pH 8.3) buffer to produce a protein concentration of 0.5 mg/ml in the 37°C sample, and a typically lower concentration at higher temperatures, before being digested with 1 μg of sequencing grade trypsin (Thermo Fisher Scientific PI90059) at 37°C for 16 hours. Following digestion 40 μL 100 mM HEPES (pH 8.3) and 40 μL of 100% anhydrous acetonitrile was added. Peptide samples were labeled with the isobaric labeling TMT 11-plex reagent (Thermo Fisher Scientific, 90406 and A34807) for 1 hour at room temperature while shaking at 1000 rpm. Ten TMT channels (126–131N) were used to label the heat denatured samples, and the 131C was used to label a reference sample (lysed by 5% SDS). In the KSHV samples, the reference sample was an equal volume mixture of samples from 0–72 HPR samples. TMT labeling was quenched with 0.33% hydroxylamine (Sigma-Aldrich, 467804–10ML) for 10 minutes. A test mix (2 μL of each labeled sample in 1% trifluoroacetic acid (TFA, Thermo Fisher Scientific, 28904)) was generated. The test mix was desalted by C18 stagetip desalting then dried down by speedvac before being resuspended in 5 μL of 1% formic acid (FA) and 1% acetonitrile. A total of 2 μL of this test mix was analyzed on a Q-Exactive HF mass spectrometer (Thermo Fisher Scientific). A quantitative mix was generated by combining each of the labeled samples in ratios obtained from derived ratios from the test mix calculated from searching the RAW file in Proteome Discoverer v2.4 which were additionally fitted to a sigmoidal curve to correct for decreased protein solubility at higher temperatures and to minimize technical variation between samples. This quantitative mix was dried down by speedvac, resuspended in 0.1% TFA, and separated by basic reverse phase fractionation into eight fractions using a Pierce High pH Reversed-Phase Peptide Fractionation column (Thermo Fisher Scientific, 84868). Fractionated samples were dried down by speed-vac, resuspended in 6 μL of 1% FA and 1% acetonitrile and analyzed on a Q Exactive HF mass spectrometer.

NUCKS Immunopurification DDA and PRM MS sample preparation

For immunopurification, 5 μg of NUCKS antibodies (Thermo Fisher Scientific, 12023–2-AP or Thermo Fisher Scientific, A301–507A, referred to as antibody 1 and 2 respectively, 1:5 dilution in 20 mM HEPES pH 7.4, 110 mM KOAc, 2mM MgCl₂, 0.1% Tween-20, 1 μM ZnCl₂, and 1 μM CaCl2) were conjugated to magnetic protein A/G beads (Thermo Fisher Scientific, 88802) for one hour at 4°C in a buffer of 20 mM HEPES pH 7.4, 110 mM KOAc, 2mM MgCl₂, 0.1% Tween-20, 1 μM ZnCl₂, and 1 μM CaCl₂. The beads were then washed in the same buffer and then resuspended in the same lysis buffer used for the cells. Simultaneously, the iSLK.219 cell pellets (snap frozen in liquid nitrogen at 48 HPR and stored at −80°C) were lysed in IP buffer (10 mM HEPES pH 7.4, 0.5% Triton-X100, 137 mM NaCl, 2 mM CaCl2₂, 3 mM KCl, 2mM MgSO₄, 1.2 mM NaH₂PO₄, and 1x HALT protease and phosphatase inhibitor) or in TPCA buffer (1% Triton-X100 and 0.1% Tween-20 with 0.5% Sodium Deoxycholate in 20 mM HEPES, 110 mM KOAc, 2 mM MgCl₂, 1 μM ZnCl₂, and 1 μM CaCl₂, at pH 7.4 and 1x HALT protease and phosphatase inhibitor) on ice for 20 minutes. Pierce universal nuclease or micrococcal nuclease, referred to as strong and weak nuclease respectively, was added at a 1:2500 dilution ratio and the samples were incubated for 10 minutes on ice. Samples were pelleted at 10000 x g for 5 minutes at 4°C. The supernatant was then added to the conjugated beads and the mixture was incubated for one hour at 4°C while rotating. The beads were then pelleted using a magnet and washed four times with lysis buffer (not containing protease and phosphatase inhibitor). Protein was eluted once with 20 μL TES buffer for 5 minutes at 70°C, then again with 50 μL 106 mM Tris HCl, 141 mM Tris base, 2% SDS, 0.5 mM EDTA (TES) buffer for 5 minutes at 70°C. The elutions were then pooled and prepped for mass spectrometry analysis. Samples were reduced and alkylated with 25 mM TCEP and 25 mM chloroacetamide for 30 minutes at 55°C, phosphoric acid was added to a concentration of 1.2%. Next, 165 μL of 100 mM TEAB, pH 7.1 in 90% methanol, was added and the sample was mixed by flicking. The sample was added to an S-Trap Micro Spin Columns (Protifi) and then centrifuged at 4000 x g until all volume had passed through the column (about 30 seconds). The column was then washed 5 times in 165 μL of 100 mM TEAB, pH 7.1 in 90% methanol. Next, 100 μL of 50 mM TEAB, pH 8 containing trypsin at a 1:50 ratio was added to the column, then digestion was performed for 16 hours at 37°C. Peptides were eluted with 40 μL of 25 mM TEAB followed by a second elution with 40 μL 0.2% aqueous formic acid. The elutions were combined, dried down by speed vac, then resuspended in 10 μL of 1% FA and 1% acetonitrile and analyzed on a Q Exactive HF mass spectrometer.

Peptide liquid chromatography tandem mass spec

LC-MS/MS data acquisition and methods generation was performed using Xcalibur v4 (Thermo Scientific), and peptides were resolved for nanoscale liquid chromatography–MS (nLC-MS) analysis with a Dionex UltiMate 3000 nRSLC (Thermo Fisher Scientific) equipped with a self-packed column with reproSil-Pur 120 C18-AQ, (1.9 μm particle size, 75 μm diameter, 500 mm length; Dr. Maish) or an EASY-Spray C18 column (2 μm particle size, 75 μm diameter, 500 mm length; Thermo Fisher Scientific, ES903), using a mixture of solvent A (0.1% formic acid, 99.9% LC-MS H₂O) and solvent B (97% LC-MS acetonitrile, 2.9% LC-MS H₂O, 0.1% formic acid ) for all mass spectrometry experiments. A Q Exactive HF instrument (Thermo Scientific) was used for all mass spectrometry experiments.

TPCA TMT liquid chromatography-tandem mass spectrometry (LC-MS/MS) analysis.

For the test mix samples, peptides were resolved in a linear gradient of 6%-18% solvent B over 60 minutes, then 18–29% B over 30 minutes at a flowrate of 250 nl/min. MS and tandem MS (MS/MS) spectra were automatically obtained in a full MS/data-dependent MS2 method. Full MS operated with a resolution of 120,000, a target automatic gain control (AGC) of 3e⁶, a maximum injection time (MIT) of 30 ms, and a scan range of 350–1800 m/z. MS/MS scans was performed for the top 20 most intense precursor ions, with a resolution of 45,000, a target AGC of 1e⁵, a 72 ms MIT, a 1.2 m/z isolation window, a fixed first mass of 100 m/z and a normalized collision energy (NCE) of 34. Dynamic exclusion was used with a 25 second exclusion time. The same methodology was used for the Quant mix samples, except peptides were resolved on a linear gradient of 6–18% solvent B for 80 minutes followed by a 18–29% solvent B gradient for 30 minutes, and an isolation window of 0.8 m/z was used.

IP LC-MS/MS analysis.

The same methodology as was used for the TPCA TMT test mix was used with the following changes. Peptides were resolved on a 3–30% solvent B gradient for 60 minutes. In MS/MS scans a resolution of 15,000, a MIT of 42 ms, a fixed first mass of 150 m/z, and a NCE of 27 were used.

PRM LC-MS/MS analysis.

Peptides were resolved in a linear gradient of 3–30% over 60 minutes. Full MS operated with a resolution of 120,000, AGC of 3e⁶, a MIT of 15 ms, and a scan range of 400–2000 m/z. PRM scans were obtained with a resolution of 30,000, AGC of 1e⁵, MIT of 60 ms, isolation window of 1.2 m/z, a fixed first mass of 125 m/z, and a NCE of 27. PRM method included 19 PRM scans to insure at least 10 points across the peak quantification. PRM data was run as acquired with a scheduled method. (See supplementary table S5)

Peptide identification and quantification

TPCA TMT data.

Peptide identification and quantification were performed as in Justice et al. 2021²⁶. Briefly, all TPCA TMT MS/MS spectra were compared to protein sequences to a human reference proteome (downloaded in January 2021), a KSHV reference proteome (downloaded in March 2021) obtained from the UniPort-SwissProt database, and common contaminants using the SEQUEST algorithm in Proteome Discoverer v2.4 (Thermo Fisher Scientific). For offline recalibration of the mass accuracy a spectral recalibration node was used. The database was limited to only fully tryptic peptides with a maximum missed cleavage of 2 and a static cysteine carbamidomethylation modification and dynamic TMT labeling on the peptide N terminus and lysine (test mix search) and static TMT labeling at the same sides (quantitative analysis). Methionine oxidation, asparagine deamination, and protein N-terminal methionine loss and acetylation were additionally included dynamic modifications. Precursor mass tolerance was set to four parts per million (ppm) and fragment ion mass tolerance was set to 0.02 Da. The false discovery rate (FDR) of the matched spectra was determined using Percolator (FDR of 1%) using a reversed sequence database search. An integration tolerance of 10 ppm was used for the reporter quantifier node and the integration, method was set to the most confident centroid. The co-isolation threshold and average reporter S/N threshold was set to 30 and 8, respectively, in the consensus workflow reporter ions quantifier node.

IP data.

As above, all MS/MS spectra were searched using human and KSHV reference proteomes and common contaminants using the SEQUEST algorithm in Proteome Discoverer v2.4. Offline recalibration of the mass accuracy was performed using a spectral recalibration node, The database was limited to only fully tryptic peptides with a maximum missed cleavage of 2, a static cysteine carbamidomethylation modification dynamic methionine oxidation, asparagine deamination, and protein N-terminal methionine loss and acetylation modifications. Precursor mass tolerance and fragment ion mass tolerance were set as above. The false discovery rate (FDR) was determined as above. An integration tolerance for the reporter quantifier node was set as above, and the method was set to the most confident centroid. The co-isolation threshold and average reporter S/N threshold in the consensus workflow reporter ions quantifier node was set as above.

PRM data.

Unique peptides for KSHV protein RIR1 were previously curated in Kennedy et al. 2022⁷¹, while unique peptides for human NUCKS were identified in the whole proteome DDA runs described previously. Peptides for KSHV protein RIR2 were not monitored in Kennedy et al. and were not identified by DDA, so unique peptides for RIR2 were screened in unscheduled PRM analyses to identify reliable targets to monitor during data acquisition. After acquisition of PRM data targeting these peptides using a scheduled PRM method, raw files containing PRM spectra were imported into Skyline⁷² for manual assessment. Peak quality for all peptides were monitored and compared to a reference spectral library generated by searching raw files using the SEQUEST algorithm in Proteome Discoverer v2.4. Peptides without convincing spectra or spectra with excessive interference were manually discarded. Following quality control, peptide abundance was calculated by summing the total peak area of the top three most abundant fragment ions for each peptide. Peptide quantification was exported as a csv file for further analysis.

IP-MS data analysis

The results of the IP-MS experiments are visualized as volcano plots (Fig. ED4), and a table of the proteins and their unnormalized abundances and spectral counts can be found in supplemental table S8A-F. We used a fold change cutoff of 2 and a p-value cutoff of 0.05 as significance thresholds for calling interactions. The data were normalized by total protein abundance per condition per replicate. For missing values, the minimum normalized value for that protein observed amongst all the remaining replicates for the given condition was used.

Considerations for comparing IP-MS and TPCA data

The IP-MS and TPCA approaches probe for protein interactions using fundamentally distinct workflows, thereby having different impacts on the specific and non-specific associations detected (Fig. ED8). These significant differences can lead to the identification of different true and false (contaminant) interactions. In IP-MS, the stringency of lysis conditions must be carefully balanced to ensure the survival of true interactions and the prevention of non-biologically relevant (contaminant) interactions that can be formed in the cell lysate. Contaminate proteins can also form non-specific associations with the beads and antibodies used in immunoaffinity purification. The specificity of antibodies for particular protein isoforms can also influence the interactome observed downstream ¹³.

In contrast, in TPCA lysis conditions have play a lesser role in the observed interactome, and no beads, antibodies, or similar tools are used in the workflow. PPIs are “detected” during the thermal denaturation step, which occurs prior to cell lysis, preventing the creation of non-biologically relevant interactions. However, in TPCA there is a bias for the detection of PPIs of high stoichiometry (e.g., the main interactor of a target protein will be detected while a rare interactor will be missed). Proteins with curves randomly overlapping the curve of a given target protein serve as contaminates in this methodology, leading to the prediction of false interactions. Furthermore, as a protein’s melting curve is influenced by all types of interactions it forms, including with nucleic acids and lipids, DNA binding proteins such as NUCKS can display atypical melting curves (Fig. ED9A-D). Thus, it can be expected that IP-MS and TPCA workflows will identify substantially different interactomes for a given target protein, likely probing different subpopulations of the target protein’s interactome, with unique contaminates. Relevantly, Tapioca addresses the stated limitations in TPCA PPI prediction through the integration of auxiliary data modalities and use of relative distance model features.

IP-MS and Tapioca identified NUCKS interactors analysis

To determine if IP-MS and Tapioca unique NUCKS interactors might be interacting with one another, we generated a list of all possible unique pairs of proteins from these subsets. We cross-referenced this list with the set of all experimentally validated human PPIs in the BioGRID⁴, REACTOME⁴², and MINT⁶⁰ databases. This analysis determined that 80% of IP-MS unique NUCKS interactors had at least one interaction (median of eleven) with a NUCKS Tapioca interactor. To understand the specificity of this analysis, we performed it again, now replacing the Tapioca NUCKS interactome with the Tapioca determined interactome of a random protein. This was repeated for all proteins identified in the KSHV TPCA dataset. In this comparison, a median of 2% of IP-MS NUCKS interactors (all, not just unique) had known interactions with non-NUCKS Tapioca interactomes.

PRM data analysis

Peptide abundances were normalized to the average abundance across all conditions, bringing all peptide abundances to a similar scale so that multiple peptides for each protein can be compared. Next, the protein abundance was calculated by averaging the normalized peptide abundances. All protein abundance values were scaled to the average abundance in the control IgG condition, giving the fold-change for NUCKS interacting proteins compared to control.

Processing of TPCA data

TPCA data was normalized as in Justice et al. 2021²⁶. Briefly, TPCA data was normalized using median of median normalization for within plex normalization. The median value of all proteins in the reference channel within a single 11-plex was calculated. Next, the median value of these median values was calculated. Lastly, all proteins across all plexes were divided by this median value. For cross plex normalization the average values of the reference channel for each individual protein across plexes was calculated (an individual average value for each individual protein). A scaled reference value is calculated for each protein for each plex by dividing an individual protein’s reference channel value in a given plex by the protein’s and plex’s average value calculated in the previous step. All values of a protein within a plex were divided by the respective protein’s and plex’s scaled reference value. The curve for each protein was fit to a three-parameter log-logistic equation $f\left(y\right)=c+\frac{1-c}{1+e^{b\left(\ln \left(y\right)-\ln \left(a\right)\right)}}$.

Considerations for protein melting curve shape in TPCA data

In TPCA, a protein’s observed melting curve is influenced not only by its protein-protein interactions, but by any interaction that the protein participates in (e.g., protein-DNA, protein-RNA, protein-lipid, or protein-ligand interactions) ^20,38,73. Interactions with non-proteins can have major effects on the shape of a protein’s observed melting curve, depending on both the stoichiometry, strength, and stability of these interactions. Much of this alteration occurs due to insoluble cellular components, such as lipids and DNA, sequestering their protein interactors during the centrifugation-based separation of insoluble and soluble proteins following lysis. At higher temperatures, these protein-non-protein interactions are more likely to be disrupted, resulting in an atypical melting curve wherein the relative abundance of a protein is higher at higher temperatures than at the lowest temperature. The strength of this effect depends on the relative stoichiometry and strength of these interactions as compared to all other interactions that a protein participates in, as well as the directness of the protein-non-protein interaction. For example, a protein that participates in a complex that interacts with DNA, in which the given protein itself does not interact with DNA, may have exhibit a curve that that temporally begins to increase in relative abundance as temperature increases, but that ultimately reaches a lower relatively abundance than that observed at the lowest temperature. This overall concept is graphically depicted in Fig. ED9A.

Throughout KSHV reactivation, NUCKS exhibited an irregular curve shape suggesting that non-protein interactions may play a large role in NUCKS’s observed melting curve. Given that NUCKS is known to be chromatin associated, it is likely that interactions with DNA may be resulting in its irregular curve shape. Indeed, three histone proteins (H4, H2AC20, and H2BC18) that NUCKS interacts with (by Tapioca predictions and by IP-MS experiments), exhibit similar, though less drastic curve shapes (Fig. ED9B). This suggest that indeed, DNA interactions are at least partially responsible for NUCKS’s non-melting status.

We observed that NUCKS interactors often displayed typical melting curves at timepoints in which they were not predicted to interact with NUCKS. At timepoints in which they were predicted to interact with NUCKS, their melting curves became stabilized, sometimes becoming non-melters, such as in the case of KSHV protein RIR1 (Fig. ED9C). We also examined the curve shapes of several CORUM complexes in which Tapioca predicted NUCKS interacts with at least 60% of the complex’s subunits and at least one of the subunits were identified in our NUCKS IP-MS experiments. These protein complexes exhibited similar highly thermally stabilized melting curves, with fluctuations that appeared to be mostly complex specific. The vast majority of these complexes, and of all NUCKS interactors, were amongst the most thermally stabilized proteins compared to the background distribution of all detected proteins, as observed in example complexes in Fig. ED9D.

Considerations for interpreting PPIs from TPCA data

It is important to note that a byproduct of TPCA-derived melting curves being some average of all of a protein’s interactions (including with non-proteins) is that this creates what could be consider an “interaction gradient”. Consider a 3-protein system, in which: i) the majority of protein A is bound to protein B to form complex AB; ii) the majority of protein C is also bound to protein B, forming complex CB; and iii) complexes ABC and AC do not exist. The melting curves of A and C will represent a mixture of free-floating A and C, respectively, as well as complexes AB and CB, respectively. The melting curve of protein B will represent a mixture of free-floating B and complexes AB and CB. As the reviewer correctly pointed out, this will result in the melting curve of B being located somewhere in between that of A and C. This will also result in the curves of proteins A and C appearing closer to one another than expected for two non-interacting proteins. Thus, observing only the melting curves, one might predict that complex ABC or AC exist, when in reality only complexes AB and CB exist. Extrapolated out to thousands of proteins with closely packed melting curves, as observed in TPCA data, this creates a sort of “interaction gradient”, where proteins that interact with one another can be erroneously predicted to interact with their interactors’ interactors (i.e., it is likely to predict the existence of AC or ABC because the complexes AB and CB do exist). Tapioca largely addresses this problem through the use of static interaction knowledge and more in-depth analysis of curve pairs (see tables S1G-S1J for more details), with most proteins having on average less than 12 confidently predicted interactions. This would suggest that for proteins with regularly sized interactomes this concept of interaction gradients may not be applicable for Tapioca.

However, once interactomes become very large, interaction gradients do become an issue. This is exemplified by NUCKS, which at 48 HPR interacts with a total 441 unique proteins, placing in the top 99.5^th percentile in terms of number of interactions. Thus, we consider NUCKS to be a hub protein, a protein that serves as a bridge between many proteins within the global PPI network. We found that these hub proteins often had highly interconnected interactomes, with many hub protein interactors having a large number of interactions between one another. While the interactions of the hub proteins are being reliably predicted, as exemplified by the NUCKS IP-MS experiments, their highly connected interactomes make it difficult to identify discreate protein complexes. To better identify functionally related interacting protein subsets within the NUCKS interactome we visualized the NUCKS 48 HPR interactome (Fig. ED9E), only including edges between proteins that that score in the top 97^th percentile (calculated from the distribution of scores derived from the NUCKS 48 HPR interactome, not all scores; proteins that had no remaining edges were dropped from the visualization for clarity). We performed Leiden community clustering on this network and observed coherent GO term enrichment within some of these communities, as well as identified a few discrete two-member complexes.

Optimization of thermal proximity coaggregation methodology

Optimization of TPCA methodology was performed in HEK293T cells. In the lysis optimization experiments, TPCA thermal denaturation was performed using five temperatures between 37°C and 55°C (37°C, 40.7°C, 44.6°C, 52.8°C, 55.3°C). Lysis was performed using (1). 0.15% NP40, (2). 0.1% CHAPS in 30 mM Tris and 150 mM NaCl, (3). 1% n-dodecyl-β-D-maltoside (DDM) in 50 mM Tris and 200mM NaCl, (4). 0.6% Trition-X100 and 0.1% Tween-20 in 20 mM HEPES, 110 mM KOAc, 2 mM MgCl₂, 1 μM ZnCl₂, and 1 μM CaCl₂, at pH 7.4, (5). 1% Triton-X100 and 0.1% Tween-20 with 0.5% Sodium Deoxycholate in 20 mM HEPES, 110 mM KOAc, 2 mM MgCl₂, 1 μM ZnCl₂, and 1 μM CaCl₂, at pH 7.4, or (6). by passage through a 21-gauge needle ten times and a 26-gauge needle six times. All lysis conditions were followed by three rounds of freeze thaws. In the temperature optimization experiments, thermal denaturation was performed using ten temperatures between 37°C and 64°C (36.9°C, 40.2°C, 43.9°C, 46.6°C, 48.6°C, 52.7°C, 55.3°C, 58.5°C, 61.2°C, 64°C) or using either five (37°C, 40.7°C, 44.6°C, 52.8°C, 55.3°C) or ten (37°C, 38.5°C, 42°C, 43.3°C, 45°C, 47°C, 48.5°C, 52°C, 53.4°C, 55°C) temperatures between 37°C and 55°C. Lysis was performed using 1% Triton-X100 and 0.1% Tween-20 with 0.5% Sodium Deoxycholate in 20 mM HEPES, 110 mM KOAc, 2 mM MgCl₂, 1 μM ZnCl₂, and 1 μM CaCl₂, at pH 7.4.

CORUM complex analysis

CORUM complexes were assigned scores by taking the average of the scores of all possible protein pairs of the subunits of that complex. To be considered detected, a CORUM complex had to have a minimum of 50% of their subunits detected. Those that met this criterion and had a minimum Tapioca score of 0.5 were considered both detected and assembled.

Temporal GO term enrichment and clustering

To understand the temporal changes in biological processes we analyzed GO term enrichment on sets of proteins that interacted with other proteins of interest (ex. viral proteins) at different time points. GO term enrichment was performed using HumanBase⁶³ using kidney tissue specific background. The number of proteins associated with a specific GO term was tracked at each timepoint, creating a curve. K-means clustering was then performed on these curves to group GO terms with similar temporal behaviors.

Calculating relative Z-Scores

Some proteins have many high scoring interactions, while others have very few. To better sort through and prioritize Tapioca predictions we calculated z-scores from the set of interactions scores relative to a single protein. The relative z-scores are calculated as such:

$$Z{S}_{p_1,p_2}=\frac{S_{p_1,p_2}-mean\left(S_{p_1}\right)}{std\left(S_{p_1}\right)}$$ Eq. 5

The z-score between protein one and protein two, relative to protein $1,Z{S}_{p_1,p_2}$, depends on the Tapioca score between protein one and two, $S_{p_1,p_2}$, and the distribution of all Tapioca scores pertaining to protein one, $S_{p_1}$. While Tapioca scores are symmetric between pairs of proteins, the relative z-score are asymmetric.

Calculating interactome similarity scores

The interactome similarly scores (ISSs) between viral proteins were calculated by first representing each viral protein as such:

$$V=\left\{I_j:j\in AllHostProteins\right\},I=\left\{\frac{1}{\left|R\right|}\sum _{i=1}^{\left|R\right|}{S}_{k,i}:k,i\in T,R\right\}$$ Eq. 6

Where the viral protein, $V$, is represented as the set of temporally resolved interactions with host protein, $I$, for all host proteins detected in the dataset. The interactions with a single host protein, $I$, is represented by the set Tapioca scores assigned to the virus-host interaction at each observed timepoint, $T$, averaged across replicates, $R$. To avoid influence from non-interactions (low Tapioca scores) all host proteins with a maximum score of less than 0.5 and relative z-score of less than 5.0 were discarded from the respective viral protein. Using the filtered sets, the ISSs were calculated as such:

$$ISS\left(V_1{V}_2\right)=\frac{1}{n}\sum _{i=1}^{n}corr\left(I_{1,i},I_{2,i}\right)$$ Eq. 7

Where the ISS between viral protein one, $V_1$, and two, $V_2$, is the average correlation between the set of temporally aligned Tapioca scores for a given host protein with viral protein one, $I_1$, and two, $I_2$. The ISS is only calculated between viral proteins from different viruses, and only if those viral proteins share in common a minimum of three interactions with the same host proteins.

Cell lines and primary culture

U2OS human bone osteosarcoma cells (ATCC HTB-96), MRC5 primary human lung fibroblasts (ATCC CCL-171), HFF (ATCC SCRC-1041) and HEK-293T (ATCC) were cultured in complete growth medium (DMEM supplemented with 10% fetal bovine serum [FBS]) at 37°C and 5% CO₂. iSLK.219 cells harboring latent KSHV (a gift from Dr. Britt Glaunsinger, University of California, Berkeley) were grown in complete growth medium supplemented with 500 μg/ml hygromycin (Thermo Fisher Scientific, 10687010) and 1% penicillin-streptomycin at 37°C and 5% CO₂. For titering of virus supernatants, reporter plates consisting of U2OS (HSV-1), MRC5 (HCMV), or U2OS (KSHV) cells were maintained in complete growth medium at 37°C and 5% CO₂.

Virus strains and infections

Wild-type HSV-1 strain 17+ (a gift from Dr. Beate Sodeik, Hannover Medical School, Hannover, Germany) was produced and propagated as previously described Diner et al., 2015⁷⁴. Briefly, P0 stocks were produced by electroporating pBAC-HSV-1 into U2OS cells. P1 stocks were then generated from the P0 stock by infecting U2OS cells at a low multiplicity-of-infection (MOI) and virus was collected ~3 days later when cells displayed 100% cytopathic effect. In a similar manner, wild-type HCMV stain TB40/E (a gift from Dr. Thomas Shenk) were produced from BAC electroporation into MRC5 cells and P1 stocks were generated by infecting MRC5 cells at a low MOI, then virus was collected ~14 days later when cells displayed 100% cytopathic effect. In both cases, cell-associated virus was released by sonication, combined with supernatant virus, then concentrated by ultracentrifugation in a SW28 swinging bucket rotor (Beckman Coulter) (20,000 rpm, ~ 53,664 x g, 2 hours, 4°C over a 10% ficoll cushion for HSV-1; 16,000 rpm, ~ 34,345 x g, 1.5 hours, 16°C over a 20% sorbitol cushion for HCMV) before being resuspended in 1/10^th original volume. Working stock virus titers were determined by plaque assay for HSV-1 or tissue culture infectious dose (TCID₅₀) for HCMV and experimental infections were performed at an MOI of 1 and 3, respectively. For HSV-1 infections, working stock virus was diluted into DMEM media containing 2% FBS and added to U2OS cells. For HCMV infections, working stock virus was diluted into complete growth medium and added to MRC5 cells. In both cases, cells were incubated in virus inoculum at 37°C for 1 hour with regular rocking, then inoculum was replaced with complete growth medium. KSHV infections were performed by reactivating iSLK.219 cells with 1mM sodium butyrate (Sigma-Aldrich, B5887) and 1 μg/ml doxycycline (Sigma-Aldrich, D9891), which resulted in 100% reactivation after 72 hours.

HSV-1 titering by plaque assay

Supernatant virus was collected at 24 hours post-infection (HPI), then diluted into DMEM media with 2% FBS and added to a reporter plate containing confluent monolayers of U2OS cells. After 1 hour, viral media was replaced with a DMEM solution containing 1% methocel and cells were incubated at 37°C for 3–4 days. Cells were subsequently rinsed with PBS and fixed for 15 minutes in methylene blue stain (50% methanol, 0.5% methylene blue). Plaques were counted, and virus titers were calculated as plaque-forming units per mL (PFU/mL) based on the dilution factor. Three biological replicates were collected.

HCMV titering by IE1 staining assay

Supernatant virus was collected at 120 HPI, then diluted in fresh complete growth media and added to a reporter plate containing confluent monolayers of MRC5 cells. After 24 hours, the reporter cells were washed three times with PBS and then fixed in −20°C methanol at −20°C for 15 minutes. Cells were then washed three times in PBS, blocked in 3% bovine serum albumin (BSA) in PBS with 0.2% Tween-20 (PBST), and incubated with α-IE1 primary antibody (a gift from Dr. Thomas Shenk; 1:100 dilution in 0.3% BSA in PBST). Cells were subsequently washed three times with PBST and incubated with secondary antibody Alexa 488 goat anti-mouse IgG (Invitrogen, A-11001; 1:1000 in 0.3% BSA in PBST) and DAPI (Thermo Fisher Scientific, 62248; 1:1000). Finally, cells were washed three times in PBST and imaged on the Operetta imaging system (Perkin Elmer) to count IE1-positive cells. The total number of IE1-positive cells was used to calculate infectious units/mL (IU/mL) based on the dilution factor. Three biological replicates were collected.

KSHV titering by supernatant transfer assay

Supernatant virus was collected at 96 HPR, passed through an 0.45 μm filter, then diluted in complete growth media with 1mM sodium butyrate and 8 μg/ml polybrene (Millipore Sigma, TR-1003-G) and added to a reporter plate containing confluent monolayers of U2OS cells by spinfection (1,000 x g for 30 minutes at room temperature). After 48 hours, cells were resuspended in growth medium and infected (GFP-positive) cells were measured on a Countess II automated cell counter equipped with a GFP light cube (Thermo Fisher). The total number of GFP-positive cells was used to calculate infectious units/mL (IU/mL) based on the dilution factor. Three biological replicates were collected.

qPCR of intracellular viral genomes

To quantify intracellular viral genomes, infected cells were collected at the indicated timepoints by washing with PBS, then scraping and collecting in complete growth medium before being flash frozen in liquid nitrogen. Three biological replicates were collected. Samples were thawed slowly on ice, then sonicated to release intracellular DNA. Cell lysates were added to DNA Resuspension Buffer (400 mM NaCl, 10 mM Tris pH 8.0, 10 mM EDTA) and cellular protein was digested by treating with 20 μg Proteinase K and 0.2% SDS overnight at 37°C. Following digestion, DNA was extracted from the samples using phenol-chloroform extraction followed by isopropanol precipitation at room temperature and one wash with 70% ethanol. Precipitated DNA was resuspended in nuclease-free water for qPCR analysis. Genome abundances were quantified by qPCR (ViiA7 Real-Time PCR System) using SYBR green PCR master mix (Thermo Fisher Scientific, 4368706) with primers targeting the HCMV IE1 gene (Forward primer: 5ʹ-TCGTTGCAATCCTCGGTCA-3ʹ; Reverse primer: 5ʹ-ACAGTCAGCTGAGTCTGGGA-3ʹ), the HSV-1 UL30 gene (Forward primer: 5ʹ-GCGAAAAGACGTTCACCAAG-3ʹ; Reverse primer: 5ʹ-GGAGACGGTATCGTCGTAA-3ʹ), or the KSHV ORF26 gene (Forward primer: 5’-CTCGAATCCAACGGATTTGAC-3’; Reverse primer: 5’-TGCTGCAGAATAGCGTGCC-3’), with the nuclear β2-microglobulin (B2M) gene serving as a loading control (Forward primer: 5ʹ-TGCTGTCTCCATGTTTGATGTATCT-3ʹ; Reverse primer: 5ʹ-TCTCTGCTCCCCACCTCTAAGT-3ʹ). Absolute viral genome counts were determined by comparison to an absolute standard curve of pCR-TOPO plasmid containing the relevant amplicon.

Generation of CRISPR-Cas9 KO cells

NUCKS1 KO cell lines were generated via the TrueCut system by Invitrogen with matched scramble controls. Briefly, cells were seeded the evening before transfection at low cell density. NUCKS1 sequence-specific TrueGuide single-guide RNA (sgRNA) (Invitrogen #A35533) targeting 5’-AAATGTGCGCCAACAACGGC-3’ and scrambled negative-control sgRNA (Invitrogen #A35526) were combined in equal molar ratios with TrueCut Cas9 Protein V2 (Invitrogen #A36499) in Opti-MEM (Thermo Fisher Scientific #31985062) and 2:1 CRISPRMAX transfection reagent (Thermo Fisher Scientific #CMAX0001). Editing efficiency was determined by Western Blot for all cell lines (Figs. S10A, B, C).

Cell viability assessment of CRISPR-Cas9 KO cells

Apoptosis-induced cell death after CRISPR-mediated knockout of NUCKS1 in HFF, MRC5, and iSLK.219 cells was monitored by terminal deoxynucleotidyl transferase dUTP nick-end labeling (TUNEL) assay using the in situ Cell Death Detection Kit (Roche, 12156792910 – TMR red) as per manufacturer recommendations. Cells were imaged on the Operetta imaging system to count TUNEL-positive cells using TMR red signal intensity in the nucleus. See Figs. S10D, E, F for bar plots of results.

Quantification and statistical analysis

Data processing and large-scale analyses were performed using Python 3.10.8, utilizing the Python libraries Numpy⁶⁴, Scikit-learn⁶⁵, Scipy⁶⁶, Pandas^67,68, Biopython⁶⁹, Seaborn⁷⁵ and Matplotlib⁷⁶. Some network visualizations were generated using Cytoscape⁷⁷. GraphPad Prism 9 was used to perform statistical analysis, with significance being determined by two-tailed Student’s t test (*= P < 0.05, **= P < 0.01, ***= P < 0.001, ****= P < 0.0001), and two degrees of freedom were used. For calculating statistics, in all cases measurements were taken from distinct samples which themselves were measured once. For boxplots, boxes show median, 25^th and 75^th percentile values, with the line within the box representing the median value, whiskers represent +/− 1.5 interquartile range, and points are outliers. For violin plots, the white dot represents the median, the thick black bar represents the +/− 1.5 interquartile range, and the thin grey line represents the total range, excluding outliers. Figures were created in GraphPad Prism and Microsoft PowerPoint.

Extended Data

Extended Data Fig 1

Extended Data Fig 2

Extended Data Fig 3

Extended Data Fig 4

Extended Data Fig 5

Extended Data Fig 6

Extended Data Fig 7

Extended Data Fig 8

Extended Data Fig 9

Extended Data Fig 10

DATA AVAILABILITY

The mass spectrometry proteomics data reported in this paper, excluding the PRM data have been deposited in the ProteomeXchange Consortium⁷⁸ via the PRIDE⁷⁹ partner repository with the dataset identifier PXD041152 and 10.6019/PXD041152. The PRM data was uploaded to Panorama⁸⁰ and can be accessed at https://panoramaweb.org/HfwV6S.url. All Tapioca predictions can be downloaded from (and scores ≥ 0.15 viewed) at https://tapioca.princeton.edu/.