Profiling Protein-Protein Interactions in the Human Brain by Refined Co-Fractionation Mass Spectrometry

Abstract

Proteins usually execute their biological functions through interactions with other proteins and by forming macromolecular complexes, but global profiling of protein complexes directly from human tissue samples has been limited. In this study, we utilized co-fractionation mass spectrometry (CF-MS) to map protein complexes within the post-mortem human brain with experimental replicates. First, we used concatenated anion and cation Ion Exchange Chromatography (IEX) to separate native protein complexes in 192 fractions, then proceeded with Data-Independent Acquisition (DIA) mass spectrometry to analyze the proteins in each fraction, quantifying a total of 4,804 proteins with 3,260 overlapping in both replicates. We improved DIA quantitative accuracy by implementing a constant amount of bovine serum albumin (BSA) in each fraction as an internal standard. Next, advanced computational pipelines, which integrate both a database-based complex analysis and an unbiased protein-protein interaction (PPI) search, were applied to identify protein complexes and construct protein-protein interaction networks in the human brain. Our study led to the identification of 486 protein complexes and 10,054 binary protein-protein interactions, which represents the first global profiling of human brain PPIs using CF-MS. Overall, this study offers a resource and tool for a wide range of human brain research, including the identification of disease-specific protein complexes in the future.

Introduction

Protein complexes and dynamic protein-protein interactions (PPIs) play crucial roles in all biological processes in cells^1–3. Disruption or perturbation in these interactions could lead to the deregulation of cellular processes and the onset of human diseases^4–6, as well as therapeutic strategies⁷. Therefore, the comprehensive, large-scale analysis of protein complexes or functional PPIs within an organism/tissue has been a long-standing objective in the post-genomic era. The success of this objective can enhance our understanding of protein function, cellular processes, and ultimately establish the connection among genotype, proteotype and phenotype⁴.

Traditionally, protein complex or PPI studies were carried out through the evaluation of binary interactions using molecular techniques such as yeast two-hybrid assay and protein complementation assay^8–10. While these techniques are still in use¹¹, recent advancement in mass spectrometry (MS) enables diverse approaches for characterizing the protein interactome. MS-based strategies can be divided into two main categories: (i) targeted approaches and (ii) untargeted approaches^{12, 13}. Targeted approaches, like affinity purification coupled to MS (AP-MS), focus on specific interactions through a protein of interest (“bait”) using an antibody or affinity tag. AP-MS is a standard tool for identifying PPIs and has been used for larger-scale studies^14–17. Proximity labeling (PL) methods, such as BioID/TurboID and APEX/APEX2, capture transient interactions by fusing target proteins with promiscuous enzymes that biotinylate nearby biomolecules^18–22. Alternatively, untargeted approaches explore the entire interactome, detecting protein interactions on a global scale. Techniques like cross-linking MS (XL-MS) employ cross-linking reagents to stabilize transient interactions by forming covalent bonds between nearby amino acid residues from different interacting proteins, and to identify specific cross-linked peptides in a proteome scale^23–26. Co-fractionation MS (CF-MS) determines interacting proteins based on their co-elution during a separation phase^27–31. Moreover, computational methods predict PPIs by integrating diverse biological data, such as gene sequences, protein structures, co-expression and evolutionary relationships³². Overall, these untargeted methods aim to offer a comprehensive view of PPIs in cells or tissues.

Among the untargeted approaches to study PPIs, CF-MS has emerged as a promising technique with high throughput capability in identification and quantification of native protein complexes and PPIs at the global scale^27–31. The basic principle of CF-MS is that proteins that interact with each other are likely to co-elute during the separation process, resulting in their co-identification by mass spectrometry. CF-MS does not require genetic manipulation of proteins, and it identifies endogenous, physiologically relevant interactomes³³. Moreover, CF-MS can analyze the distribution of the same protein in different fractions, elucidating the stoichiometries and dynamics of any given protein within distinct co-isolated complexes. Integrating CF-MS with the data-independent acquisition (DIA) method has been shown to improve the sensitivity of CF-MS technique^34–36. Unlike the data-dependent acquisition (DDA) mode which selects peptides based on their intensity for fragmentation³⁷, The DIA selects all peptides within a defined m/z range³⁸. Thus, incorporating DIA in CF-MS expands protein coverage, reducing missing data³⁶.

While global interactome studies with CF-MS have been applied in a wide diversity of organisms and model systems^{28, 29, 33, 39}, this approach has yet to be applied to human clinical samples, such as brain tissues. However, understanding the human brain protein interactome is crucial to decipher human-specific and brain-specific characteristics, as well as related brain functions and diseases, including Alzheimer’s disease (AD), Parkinson’s disease, and schizophrenia. To fill this gap, we applied CF-MS to profile PPIs in postmortem AD human brains. We leveraged detailed biochemical fractionation via Ion Exchange Chromatography (IEX) and optimized high-throughput DIA analysis for protein measurement in each fraction. This approach identified 10,054 potential PPIs, with 2,389 being novel, from the human brain.

Materials and Methods

Sample preparation of human brain tissue

Human brain tissue samples were provided by Banner Sun Health Research Institute. Proteins were extracted from brain tissues as previously described with modifications^{40, 41}. Briefly, 100 mg brain samples were lysed in 1 mL of lysis buffer (tissue:buffer ratio of 1:10 (w/v), 10 mM HEPES, 250 mM sucrose, 5 mM MgCl₂, 0.1% (v/v) dodecyl-β-D-maltopyranoside (DDM), 1 mM DTT, and 1X Protease and Phosphatase Inhibitor Cocktail from Roche). Samples were homogenized using the Dounce homogenizer (45 beats) on ice. After lysis, 1.2 μL of turbonuclease (with the activity of at least 200 units/μL, Millipore Sigma) was added to hydrolyze nucleic acids at 4 °C for 30 min. The lysate was centrifuged at 17,000 × g for 30 min at 4 °C and protein supernatant was collected. Bicinchoninic acid (BCA) assay and an SDS-PAGE short gel⁴² were used to measure protein concentrations in the supernatant.

Biochemical fractionation

As shown previously²⁹, one PolyWAX LP anion-exchange chromatography column (200 × 2.1 mm, 5 μm beads with 1000 Å pore size; PolyLC) was connected to two cation-exchange PolyCAT A columns (each 200 × 2.1 mm, 5 μm, 1000 Å; PolyLC) for separating protein complexes using an Agilent HPLC 1220 system. Buffer A contained 10 mM HEPES, pH 7.6, 3 mM NaN₃; and buffer B included buffer A and 1.5 M NaCl. All the buffers and columns were chilled on ice-water bath. The proteins (500 μg) were fractionated in an elution program (100% buffer A for 30 min loading, 0–67% buffer B until 108 min, and 67–100% buffer B to 130 min, flow rate at 0.1 mL/min). The eluted protein was monitored by UV absorption at 280 nm, collected in ~200 fractions (30 sec per fraction), and dried using SpeedVac concentrator.

Digestion and desalting of co-fractionated proteins

Each dried fraction (assuming ~2.5 μg protein) was dissolved in 10 μL master mix, consisting of 50 mM HEPES, pH 8.5, 8 M urea, 1 mM DTT, 10 ng/μL Bovine serum albumin (BSA), and 2 ng/μL Lys-C with an enzyme:protein ratio of ~1:125 (w/w), using the Mantis automated liquid handling robot (Formulatrix). The BSA (100 ng) was spiked in each fraction as the internal standard. After 3 h incubation, the sample was diluted with 50 mM HEPES, pH 8.5, to reduce urea to 2 M, followed by trypsin digestion overnight (final trypsin concentration at 10 ng/μL in 40 μL, enzyme:protein of 1: 6, w/w). Peptides were further reduced by 1 mM fresh DTT for 30 min and alkylated with 10 mM iodoacetamide for 30 min in the dark at 21 °C. The digested peptides were desalted using C18 StageTip (ThermoFisher) with an enhanced wash step (15 times with 50 μL wash buffer). Desalted peptides were dried using SpeedVac concentrator.

LC-MS/MS analysis

The dried peptides were resuspended into 5% formic acid and centrifuged at 21,000 × g for 5 min. The resulting supernatant was then transferred into an HPLC sample insert tube and centrifuged again at 21,000 × g for 10 min. For each fraction, 10% of the sample (equivalent to ~0.25 μg without considering sample loss) were injected into an Ultimate 3000 RSLCnano system coupled with an Exploris 480 mass spectrometer (ThermoFisher). The peptides were separated in a 30 min gradient with a flow rate of ~0.25 μL/min and a column temperature of 65 °C. The mass spectrometer was operated in data-independent acquisition (DIA) mode, with the following parameters: MS1 scan range of 400–1000 m/z, an isolation window of 20 m/z, window overlap of 1 m/z, default charge state of 2, MS1 and MS2 resolutions of 30K, and 30% normalized collision energy for HCD fragmentation.

DIA data analysis

The DIA data were analyzed using DIA-NN version 1.8.1⁴³ with a spectral library consisting of 20,386 proteins generated from in silico tryptic digestion of the UniProt reviewed human database (March 2023). The spectral library was created through deep neural networks with retention time. The raw files were searched with the following parameters: two missed cleavage, fixed modification of Cys carbamidomethylation, variable modification of Met oxidation, and a 400–1000 m/z mass range. Match between runs (MBR) was enabled for the data search, while all other parameters were kept at default settings. The resulting DIA-NN report file was processed in RStudiov3.3. To get 1% false discovery rates (FDR) of peptides and proteins, DIA-NN report file was filtered at a q-value threshold of 1% for the precursor, library precursor, and protein group. Precursors corresponding to more than one protein group were removed. The retained precursors were then used for label-free quantification using the MaxLFQ algorithm⁴⁴ inside the DIA-NN R package. Given that equal protein loading is not applicable in CF-MS experiments, we bypassed the default normalization of precursors or proteins and instead extracted the original (non-normalized) precursor ions for protein quantification.

Identification of protein complexes

The protein data from DIA-NN quantification⁴³ were normalized using the spiked standard of BSA in individual fractions. The “filter_profiles” and “clean_profiles” functions under “PrInCE” library were used to edit a CF profile^{45, 46}. Briefly, proteins lacking three consecutive abundance values were removed. Proteins that were not present in either replicate were also filtered out. The remaining protein profiles were smoothened with the moving average of 4. Then the resulting data were loaded as a txt file via sample ID key as recommended (https://github.com/anfoss/PCprophet)⁴⁷. The protein complex analysis on PCprophet was performed using Python v3, employing searches against CORUM 4.0⁴⁸ and hu.map.2.0⁴⁹ databases. Default parameters were used, with an FDR at 0.05. The generated ‘ComplexReport.txt’ file was utilized to identify positive complexes, based on the filter criterion ‘Is Complex = Positive’.

Protein-protein interaction identification

PrInCE (Prediction of Interactomes from Co-elution data) was used for CF-MS analysis to identify the high-confidence PPI network^{45, 46}. The CF-MS protein profiles above were scaled (i.e., maximum abundance to 1), and processed by fitting mixtures of one to five Gaussians to each profile. Profiles with an R-squared value of below 0.75 were filtered out. Model selection was then performed using the corrected Akaike information criterion to determine the best Gaussian model for each profile, followed by the “calculate_feature” function. In this step, six measures of distance or similarity were computed for each pair of proteins reflecting the likelihood of a physical interaction based on their mass spectrometric profiles^{45, 46}. These features were computed separately for each replicate, and then the features detected in either replicate were combined using the “concatenate_features” function.

To train a classifier for PPI network inference, a reference set of interactions from the CORUM 4.0 database⁴⁸ was utilized. A 10-fold cross-validation strategy was employed, with a random forest classifier consisting of 100 trees. Pairs of proteins within the same complex were designated as true positives, while pairs of proteins in different complexes were treated as true negatives. The classifier ranked the experimental protein pairs as candidate interactions based on the results from all cross-validation folds. Protein pairs passing the precision of 0.99 (i.e., 1% FDR) were accepted as high-confidence PPI for further evaluation. For novel PPI, additional threshold of correlation of greater than 0.5 in both replicates was required.

Structure prediction of protein-protein interactions

PPI structures were predicted using AlphaFold2-multimer⁵⁰ using a Colabfold v1.5.2⁵¹. In Colabfold, the generation of Multiple Sequence Alignments (MSAs) was performed using “MMSeqs2_uniref_env”, ensuring comprehensive and high-quality data input. During the prediction process, the template search was set to pdb100. To enhance computational efficiency, the number of recycles was set to one. The model type was selected to be “alphafold2_multimer_v3” and three neural network models were utilized when predicting structures. All other parameters were set to default. The structures were visualized in UCSF ChimeraX⁵².

Results and Discussions

Untargeted profiling of protein complexes and PPI in human brain by CF-MS

To profile protein interactions and complexes in the human brain, we performed a high throughput protein co-elution experiment using the CF-MS method (Fig. 1). Our experiment involved several key steps (Fig. 1A): (i) lysing a pooled human brain sample under native condition to prevent the disruption of natural protein complexes, (ii) separating protein complexes largely based on surface charge using IEX with concatenated anion-exchange and cation-exchange columns, (iii) digesting proteins with the addition of an internal standard (BSA protein) at the equal amount, and (iv) analyzing the resulting peptides by LC-MS/MS with the DIA mode. The acquired raw data were analyzed (Fig. 1B) using DIA-NN software⁴³ for protein identification and quantification. The CF-MS results were processed to identify protein complexes and binary PPIs using PCprophet⁴⁷ and PrInCE⁴⁶ programs, respectively.

Fig. 1: Experimental and bioinformatic workflow of Co-fractionation mass spectrometry in this study.

A) Human brain tissue was lysed under a mild condition to release protein complexes with minimal disruption to their composition. The cell lysate was separated by Ion Exchange Chromatography (IEX) in High-Performance Liquid Chromatography (HPLC). The fractions collected at different retention time (RT) were spiked with an internal standard (BSA), digested with Lys-C and trypsin, analyzed using LC-MS/MS in a DIA mode. MS raw files were processed to identify putative protein complexes. B) The DIA-NN software was utilized for protein/peptide identification and quantification. The protein matrix was normalized by the abundance of the internal standard and filtered to improve data quality. For hypothesis-driven analysis, PCprophet was employed with CORUM and hu.MAP 2.0 databases. For discovery-driven analysis, PrInCE was used alongside a random forest classifier.

Following the CF-MS method, we obtained 192 IEX fractions from the human brain lysate and analyzed each fraction by the DIA strategy (Fig. 2A, 2B) and performed a replicated experiment. At the peptide/protein FDR of 1%, we identified 4,607 and 3,650 proteins from the two biological replicates, totaling 4,804 unique proteins (Table S1). These identified proteins were subjected to a filtering process to improve dataset quality (Fig. 2C): (i) the protein intensities were normalized using the BSA internal standard; (ii) stochastic measurements were alleviated by only retaining proteins detected in at least 3 consecutive fractions; (iii) a moving average smoothing window of 4 fractions was applied to the protein profiles; (iv) proteins identified only in one replicate were filtered out. As a result, 3,260 shared proteins by both replicates were accepted for further analysis (Table S1).

Fig. 2: Global assessment of CF-MS data from the human brain.

A) The elution profile from the IEX chromatography, showing the gradient (the red line) and UV absorbance, along with retention time (RT). Fractions were collected every 30 sec. B) The number of proteins identified in each fraction. Fractions before the 30-min RT or after the 126-min RT contained little proteins and were excluded from the analysis, resulting in 192 fractions with identified proteins. C) The workflow to obtain a high-quality dataset. Protein intensities from DIA-NN were normalized, and the protein list was refined to include only those proteins presenting a minimum of 3 consecutive abundance values. Protein profiles were smoothed using a window of 4 fractions. D) The correlation of eluted protein peaks (based on MS results) for 3,260 proteins shared by both replicates. E) A heatmap demonstrating the reproducibility of shared protein profiles.

To assess experimental consistency between the two replicates, we utilized correlation of eluted protein peaks (Fig. 2D) and a heatmap to visualize the 3,260 proteins across the 384 fractions from the two replicates (Fig. 2E). Notably, no substantial differences in protein patterns were observed between the two replicates, suggesting high reproducibility. Through these steps of extensive quality control, we were able to generate a reliable dataset for subsequent protein complex and PPI analyses.

Improvement of quantitative accuracy in CF-MS by the internal standard

Due to variations in sample complexity and protein quantity across all fractions, accurately quantifying the amount of protein in each fraction by label-free DIA analysis presents a challenge. To mitigate this issue, we spiked a consistent amount of BSA protein (100 ng) to each fraction as an internal standard (Fig. 1A). The internal standard was used to normalize for experimental variations during sample processing, such as digestion efficiency, sample loss, and fluctuations in DIA LC-MS/MS analysis. The BSA intensity in each fraction was represented by the mean intensity of its shared peptides among all fractions, excluding any BSA peptides that overlapped with the human proteome. To assess how the equal amount of spiked BSA behaved across the fractions, we examined its intensity profile and found that the BSA intensity varied significantly across the fractions (Fig. 3A, Table S2), reflecting experimental variability. After normalization, we obtained the median intensity of all proteins to indicate the protein abundance in each fraction. This resulted in an MS-based elution profile (i.e., “Median intensity” by DIA) that was correlated with the IEX chromatography pattern (i.e., OD280 by “UV absorbance”) (Fig. 3B and Fig. 2A). We further investigated the correlation between the “Median intensity” and the “UV absorbance” of the eluted fractions before and after BSA normalization. The Pearson correlation coefficient (R) value increased from 0.808 to 0.851 (Fig. 3C), suggesting the enhancement in protein quantification.

Fig. 3: Internal standard in CF-MS improves analysis accuracy and complex identification.

A) Bar plot showing the abundance profile of the BSA internal standard across various fractions, along with retention time. B) Elution profile illustrating the median intensity of all proteins from each fraction pre-normalization (light blue) and post-normalization (brown). C) Correlation between UV absorbance from IEX and median intensity before and after internal standard normalization. A linear trendline was added, along with the Pearson correlation coefficient. D) A Venn diagram illustrating the protein complexes identified before and after BSA normalization.

To further discern the advantages of using an internal standard, we compared the non-normalized and normalized datasets to identify protein complexes in the two repeated experiments. We employed PCprophet as our search tool, referencing the CORUM v4.0⁴⁸ and hu.MAP 2.0⁴⁹ human complex databases. As expected, in the summed protein complexes from two databases were shared by both replicates, the normalized dataset resulted in a higher count of identified complexes, while about 80% of the protein complexes overlapped with the result from unnormalized data (Fig. 3D). Some complexes were only identified from the post-normalization data, such as the alpha-dystrobrevin-ZO1 complex and the HuMAP2–06971 complex, with their individual protein elution profiles shown (Fig. S1). This finding indicates that leveraging an internal standard for calibration improves analytical accuracy, leading to the identification of a greater number of protein complexes.

Evaluation of computationally derived protein complexes by PCprophet

To identify the reported complexes from the CF-MS dataset, we utilized the complex-centric tool PCprophet⁴⁷ along with a reference database, which encompasses both CORUM v4.0⁴⁸ and hu.MAP 2.0⁴⁹ databases. Using a 5% FDR, PCprophet search identified a total of 454 complexes: 111 from CORUM and 375 from hu.MAP (Fig. 4A, Table S3). Interestingly, only 32 complexes (7.0% of 454) were found to overlap between the two databases (Fig. 4B). This indicates the inherited differences in these databases: CORUM exclusively contain 3,684 manually curated complexes based on experimental data⁴⁸, while hu.MAP features 6,965 complexes derived from the integration of mass spectrometric results with physical and proximity interaction data⁴⁹. A similar result was obtained using the PrInCE program (Fig. S2).

Fig. 4: Analysis of protein complexes identified from CF-MS data using PCprophet.

A) Stacked bar plot displaying the protein complexes identified in two databases. B) Venn diagram illustrating the overlapping complexes identified across the two databases, with total complex numbers shown. Due to varied complex naming in these databases, comparisons were made using the names of subunits. C) Examples of protein complexes relevant to brain and/or neuronal development. Their elution profiles are shown with scaled abundance (with the highest peak normalized to 1 for each protein). The PPI network for each subunit is also depicted.

To further assess the reliability of FDR calculations by these computational programs, we created a decoy dataset by randomizing the elution profiles of each protein and subjected it to the same analysis process. The complexes identified from this dataset were labeled as “false discoveries.” The FDR rates obtained were also below 5% (see Fig. S3), supporting the accuracy of the FDR estimations provided by these programs.

Among the protein complexes identified through the PCprophet approach, several notable complexes were actin-related protein-2/3 (ARP2/3) complex, prefoldin complex, LSM complex, and COP9 signalosome (Fig. 4C). The ARP2/3 complex is known to regulate actin polymerization, cell migration, and adhesion⁵³. In the brain, it is essential for neocortical neural development and neuritogenesis⁵⁴. The prefoldin complex acts as a molecular chaperone, preventing protein misfolding and promoting cellular homeostasis⁵⁵. The LSM proteins are assembled into a subcomplex of U6 spliceosome⁵⁶, and recent studies suggest the contribution of splicing dysfunction to Alzheimer’s disease^{42, 57, 58}. The COP9 signalosome (CSN), an evolutionarily conserved complex with eight subunits, plays a vital role in cellular ubiquitylation, DNA-damage response, cell-cycle control, and gene expression^59–61. Recent research highlights its involvement in mitigating neuroinflammatory responses during cerebral injury⁶².

Inference of high confidence PPIs in human brain by a discovery-driven approach

To derive novel interactomes from human brain CF-MS proteome profiles, we utilized a machine-learning pipeline named PrInCE⁴⁶. PrInCE integrates machine learning techniques to identify PPIs from mass spectrometric data, leveraging the elution patterns of known protein complexes as a basis for training a classifier^{27, 63}. We used the CORUM database and a random forest classifier to predict the PPIs and their respective precision (Fig. 5A). Our CF-MS data revealed 10,054 PPIs at a 1% FDR (Fig.5B and Table S4), in which most of the PPIs (n = 7,589) were previously identified as intra-complex PPI in CORUM (Fig. 5B). However, 76 PPIs emerged as inter-complex IDs that could indicate novel protein complexes, and they could also represent possible false discoveries, in line with the 1% FDR. Importantly, a substantial number of new PPIs (n = 2,389) not documented in the reference CORUM database were also identified (Fig. 5B). We further mapped the PPIs on a network using Cytoscape (Fig. 5C)⁶⁴. To refine the network, we used Markov CLustering Algorithm (MCL) from clusterMaker2 to subdivide big clusters into smaller ones⁶⁵.

Fig. 5: Protein-protein interactions identified from CF-MS data using PrInCE.

A) Precision plot created by PrInCE based on the top 500,000 PPIs organized by their FDR ranking. We hypothesize that “true positive” (TP) denotes the intra-complex interactions, while “false positive” (FP) refers to the inter-complex interactions in the CORUM database. B) Bar plot showing the accepted PPIs with an FDR below 0.01. “Intra-complex” indicates PPIs found in the CORUM database; “Inter-complex” shows potential FPs; and “novel” designates PPIs absent in CORUM. C) Interactome visualization of accepted PPIs using Cytoscape. For a rigorous network analysis, the Markov CLustering Algorithm (MCL) from clusterMaker2 was employed. Red edges highlight novel interactions, and grey points to previously identified interactions.

To gain detailed perspective on the protein complexes identified, we conducted an additional analysis of some individual complexes. For example, CAPN1 and CAPNS1 subunits forms the mu-calpain complex, in which CAPN1 functions as the catalytic subunit, while CAPNS1 acts as the regulatory subunit⁶⁶. The two proteins were tightly co-eluted until fraction 122, and then the CAPNS1 regulatory subunit appears to be eluted alone (Fig. 6A), suggesting only a portion of CAPNS1 in the sample is complexed with the CAPN1 subunit. The complex structure is also predicted by Alphafold2 multimeter^{50, 51} (Fig. 6B). Mu-calpain is known to be a ubiquitous calcium-sensitive protease primarily found in the cytoplasm or in proximity to the cell membrane and regulates the normal physiological neuronal function through calcium regulations⁶⁶. Its activation is triggered by calcium concentrations in the micromolar range in vitro⁶⁶. Disturbances in calcium homeostasis can lead to persistent and pathological activation of calpain, which has been associated with various neurodegenerative diseases. Pathological activation of calpain can result in the cleavage of numerous neuronal substrates, negatively impacting neuronal structure and function, and thereby inhibiting essential neuronal survival mechanisms and/or advancing neurodegenerative processes such as Alzheimer’s disease⁶⁷.

Fig. 6: Examples of the known and novel protein-protein interactions identified using PrInCE.

A) Elution profiles of the CAPN1 and CAPNS1 proteins that form the mu-calpain complex. B) Mu-calpain structure predicted by Alphafold2 multimeter. Two different views of structure are shown. C) Elution profiles of the PEPD and PITRM1 proteins. D) Alphafold2 predicted structure of the PEPD/PITRM1 complex. Two different views of structure are shown. The quality of the predicted 3D structures was evaluated by predicted Local Distance Difference Test (pLDDT) and predicted Template Modeling Score (pTM).

We also identified a large number of novel putative PPIs, one of which was between PITRM1 and PEPD. PITRM1 is a peptidasome, a member of pitrilysin oligopeptidase family containing an inverted zinc-binding motif, which is responsible for degrading presequences and unstructured peptides⁶⁸. Recent research has also revealed its involvement in the degradation of amyloid beta^68–70, suggesting of a crucial role in the development of AD⁷¹. Our analysis has shown that the elution profile of PITRM1 is highly correlated with that of PEPD (Fig. 6C), a dipeptidase capable of hydrolyzing di- or tri-peptides containing a prolyl or hydroxyprolyl residue in the C-terminal position, which functions in proline recycling⁷². The alphafold2 multimeter was utilized to predict potential structures of this interaction (Fig. 6D)^{50, 51}. Although the exact biological significance of this interaction is unknown, it can be speculated that they coordinate to cleave cellular peptides based on their respective functions.

Conclusion

In summary, we implemented an improved DIA CF-MS approach, which integrated the power of IEX and DIA LC-MS/MS proteomics, to provide a large-scale analysis of endogenous protein complexes and PPIs within the human brain. The inclusion of an internal standard protein clearly enhanced the precision of protein quantification across all fractions that contain diverse levels of proteins. From the human brain sample, this refined method revealed 10,054 PPIs, with 2,389 being previously unidentified in the database, providing a valuable resource for subsequent research. Moreover, the meticulous methodologies paved the way for the PPI comparison directly from clinical specimens under different physiological or disease conditions.

While IEX chromatography provides excellent resolution in separating protein complexes under native conditions, its major limitation is the lack of molecular weight information, unlike size exclusion chromatography (SEC). If multiple small complexes co-elute together, they might be misinterpreted as a single, larger complex. Integrating IEX with SEC fractionation methods could lead to a more accurate interpretation of protein complexes in complex cell lysates.

Supplementary Material

FigS3

Fig. S3: False positive rate analysis using a decoy dataset.

FigS1

Fig. S1: Examples of protein complexes only identified after internal standard normalization.

FigS2

Fig. S2: Comparison between PrInCE and PCprophet in identifying protein complexes and PPIs.

TableS1

Table S1: Quantification of protein abundance across 192 fractions for two replicates.

TableS2

Table S2: Internal standard abundance from BSA.

TableS3

Table S3: Identification of protein complexes using PCprophet.

TableS4

Table S4: Discovery of protein-protein interactions (PPIs) by PrInCE.

Data availability statement

The mass spectrometry raw data, DIA-NN output as well as meta-data have been deposited to the ProteomeXchange Consortium via the PRIDE repository. These can be accessed using pxd044083 and pxd044084.