Volume 19, Issue 3 2300684
Open Access

Proteomic workflows for deep phenotypic profiling of 3D organotypic liver models

Stefania Koutsilieri

Stefania Koutsilieri

Department of Physiology and Pharmacology, Karolinska Institutet, Stockholm, Sweden

Contribution: Data curation (lead), Formal analysis (lead), ​Investigation (lead), Methodology (equal), Visualization (supporting), Writing - original draft (lead)

Search for more papers by this author
Evgeniya Mickols

Evgeniya Mickols

Department of Physiology and Pharmacology, Karolinska Institutet, Stockholm, Sweden

Department of Pharmacy, Uppsala University, Uppsala, Sweden

Contribution: Formal analysis (supporting), ​Investigation (supporting), Methodology (supporting), Visualization (supporting)

Search for more papers by this author
Ákos Végvári

Ákos Végvári

Division of Chemistry I, Department of Medical Biochemistry and Biophysics, Karolinska Institutet, Stockholm, Sweden

Contribution: Conceptualization (equal), Formal analysis (equal), Funding acquisition (equal), ​Investigation (equal), Methodology (equal), Resources (lead), Validation (lead), Writing - review & editing (supporting)

Search for more papers by this author
Volker M. Lauschke

Corresponding Author

Volker M. Lauschke

Department of Physiology and Pharmacology, Karolinska Institutet, Stockholm, Sweden

Dr Margarete Fischer-Bosch Institute of Clinical Pharmacology, Stuttgart, Germany

University of Tübingen, Tübingen, Germany


Volker M. Lauschke, Dr Margarete Fischer-Bosch Institute of Clinical Pharmacology, Stuttgart, Germany.

Email: [email protected]

Contribution: Conceptualization (lead), Project administration (lead), Supervision (lead), Visualization (equal), Writing - review & editing (lead)

Search for more papers by this author
First published: 21 March 2024

Ákos Végvári and Volker M. Lauschke shared last authorship.


Organotypic human tissue models constitute promising systems to facilitate drug discovery and development. They allow to maintain native cellular phenotypes and functions, which enables long-term pharmacokinetic and toxicity studies, as well as phenotypic screening. To trace relevant phenotypic changes back to specific targets or signaling pathways, comprehensive proteomic profiling is the gold-standard. A multitude of proteomic workflows have been applied on 3D tissue models to quantify their molecular phenotypes; however, their impact on analytical results and biological conclusions in this context has not been evaluated. The performance of twelve mass spectrometry-based global proteomic workflows that differed in the amount of cellular input, lysis protocols and quantification methods was compared for the analysis of primary human liver spheroids. Results differed majorly between protocols in the total number and subcellular compartment bias of identified proteins, which is particularly relevant for the reliable quantification of transporters and drug metabolizing enzymes. Using a model of metabolic dysfunction-associated steatotic liver disease, we furthermore show that critical disease pathways are robustly identified using a standardized high throughput-compatible workflow based on thermal lysis, even using only individual spheroids (1500 cells) as input. The results increase the applicability of proteomic profiling to phenotypic screens in organotypic microtissues and provide a scalable platform for deep phenotyping from limited biological material.


  • ACN
  • acetonitrile
  • ADME
  • absorption, distribution, metabolism and excretion
  • CYP
  • cytochrome P450
  • FA
  • formic acid
  • FDR
  • false discovery rate
  • FT
  • freeze-thaw
  • GO
  • Gene ontology
  • HCD
  • higher-energy collisional dissociation
  • LC
  • liquid chromatography
  • LF
  • label-free
  • metabolic dysfunction-associated steatotic liver disease
  • MS
  • mass spectrometry
  • PBPK
  • physiologically based pharmacokinetic
  • PBS
  • Phosphate buffered saline
  • PHH
  • primary human hepatocytes
  • PMU
  • ProteaseMAX/urea
  • RIPA
  • radioimmunoprecipitation assay
  • RT
  • room temperature
  • SEM
  • standard error of the mean
  • TEAB
  • triethylammonium bicarbonate
  • TMT
  • tandem mass tags

    The recent decades have witnessed major breakthroughs in chemistry, molecular biology, and genetics that have fueled drug discovery and development. Nevertheless, result translation from preclinical to clinical stages remains poor, entailing low success rates of 7%–17% during clinical stages.[1] Multiple reasons have been identified for these stagnating results. One key factor is the incomplete understanding of target biology resulting in inaccurate associations between target and disease states.[2-4] To overcome these limitations, there is a resurgence in phenotypic drug discovery, which has become a prevalent discovery strategy in both academia and the pharmaceutical industry.[5] Modern phenotype-based strategies use cell-based assays and have resulted in many novel drugs and drug candidates for a multitude of indications.[6, 7]

    Phenotypic screens require both a cell model that accurately reflects the phenotype of interest and methods to profile and analyze phenotypic alterations. Organotypic and microphysiological tissue models have emerged as promising methods to culture human cells with relevant phenotypes and functions.[8, 9] Owing to its pivotal role in drug metabolism, the ex vivo culture of liver cells has received substantial attention. For drug development applications, 2D sandwich cultures, micropatterned co-cultures and liver spheroids are widely used.[10] Particularly liver spheroids of primary human hepatocytes (PHH) have been shown to retain hepatic metabolism and functionality for multiple weeks.[11, 12] Hepatic spheroids can moreover reliably predict liver toxicity,[13, 14] CYP induction[15] and hepatic clearance[16, 17] and have been used as relevant models of liver diseases, such as metabolic dysfunction-associated steatotic liver disease (MASLD).[18, 19]

    Comprehensive molecular characterization of hepatic cells and tissues using mass spectrometry (MS)-based protein quantification constitutes a key modality for deep phenotypic assessments that can reveal targets and off-targets, identify mechanisms of action and inform physiologically based pharmacokinetic (PBPK) modeling.[20-22] Over the years, a multitude of diverse proteomic protocols have been developed for liver spheroids that differ in input, cell lysis and quantification strategies. Although different protocols have been benchmarked across simulated data sets,[23] artificial standards,[24] biological fluids,[25] and human tissues,[26] studies evaluating their performance on ex vivo microtissues that comprise a limited number of cells are lacking.

    Herein, we compared the performance of different protocols for the proteomic analysis of organotypic human liver spheroids assessing the variables of cellular input, lysis and quantification method. We found that different input (1500–150,000 cells) and lysis strategies (PMU, RIPA, freeze/thawing) affect the coverage of the liver proteome. Furthermore, using a phenotypic model of MASLD, we demonstrate that the choice of proteomic protocol affects the identification of pathogenic pathways and, ultimately, the biological conclusions. The results shine a light on the importance of proteomic protocols for the profiling of organotypic cultures and can provide guidance for the selection of experimental workflows for deep molecular phenotype analyses.


    2.1 Primary human liver spheroid culture

    Primary human hepatocytes were purchased from BioIVT (Brussels, Belgium). Information about the demographic and medical characteristics of the donors are available in Table S1. Cryopreserved PHH were seeded in 96-well ultra-low attachment plates (Corning) at a cell density of 1500 cells/well in PHH culture medium supplemented with 10% fetal bovine serum (HyClone FBS, Nordic Biolabs) as previously described.[27] After spheroid formation for 7 days, the medium was replaced with serum-free PHH culture medium and replenished every 48 h. Where indicated, MASLD was modeled by inducing hepatocellular steatosis via supplementation with 240 μM oleate and 240 μM palmitate for another 7 days, as previously reported.[18]

    2.2 Proteomic sample preparation

    For proteomic analyses, we pooled 96 spheroids (corresponding to 144,000 cells; donor 1), 48 spheroids (72,000 cells; donor 1), 24 spheroids (36,000 cells; donor 1) or 8 spheroids (12,000 cells; donor 1) or used individual spheroids (1500 cells; donor 2). Individual spheroids or pools were washed twice with PBS and once with 0.1× PBS to remove excess salts. Subsequently, the cells were lysed using three different protocols: chemical lysis with ProteaseMAX/urea (PMU) or RIPA or thermal lysis via freeze-thawing.

    For lysis in PMU, spheroids were suspended in 50 µL of 1M urea (Sigma–Aldrich) and 0.1% ProteaseMAX (Promega) in 50 mM Tris-HCl, pH 8.5 and 10% acetonitrile (ACN) adding 1 µL of 100× protease and phosphatase inhibitors (Pierce). The samples were sonicated in a water bath for 10 min with a VibraCell probe (Sonics & Materials) at 10% amplitude. Lysates were spun down at 13,000 g at 4°C for 10 min.

    For lysis in RIPA buffer, spheroids were centrifuged at 4°C in 1× RIPA at 13,000 g for 10 min before precipitation with four volumes of chilled acetone overnight at −20°C. Acetone was evaporated at room temperature (RT) for 30 min and the protein pellets were dissolved in 6.25 µL of 8M urea. Following water bath sonication for 10 min, 37.5 µL of 50 mM Tris-HCl buffer was added.

    Proteins prepared with PMU or RIPA protocols were reduced with 0.8 µL of 0.5M dithiothreitol (Sigma) at 37°C for 45 min, alkylated with 2.5 µL of 0.5M iodoacetamide (Sigma) for 30 min at RT in the dark and subjected to proteolysis by incubation with 1 µL of 0.1 µg µL−1 trypsin (Promega) over night at 37°C. The digestion was stopped with 3 µL formic acid (FA) to a final concentration of 5% v/v. The samples were cleaned on a C18 HyperSep plate with 5–7 µL bed volume (Thermo Fisher) and dried under vacuum using a Vacufuge concentrator (Eppendorf). Samples were redissolved in 10 µL of 50 mM triethylammonium bicarbonate (TEAB), supplemented with 3 µL ACN and 10 µg of TMT-10plex reagents (Thermo Fisher) in 2 µL of dry can, and incubated at RT for 2 h with shaking at 550 rpm. The labeling reaction was stopped by adding 1.7 µL of 5% hydroxylamine (Sigma). Six individual samples were combined to one analytical sample and dried in vacuum, followed by clean up on a C18 HyperSep plate.

    For thermal lysis, samples were subjected to four cycles of freezing in liquid N2 for 2 min in 5 µL of 100 mM TEAB, pH 8 in a 96-well plate (PCR Plate 96 LoBind, Eppendorf) and thawing at 70°C on a heating block for 2 min. Proteins were then denatured at 90°C for 5 min and cooled to RT before further processing. Proteins prepared by freeze-thawing were digested without alkylation by directly adding 1 µL of 0.1 µg µL−1 trypsin and incubated over night at 37°C. Samples were then supplemented with 2 µL of TMT-10plex reagents (only 126, 127N, 128N, 129N, 120N, and 131 were used and the strategy is thus henceforth referred to as “6plex”) at 5 µg µL−1 concentration and incubated at RT for 2 h with shaking at 350 rpm. Three replicates of treated and three replicates of control samples were prepared simultaneously and combined into a single analytical sample in each TMT-experiment. Labeling was stopped by adding 1.7 µL of 5% hydroxylamine and incubation for 15 min at RT. TMT-labeled peptides were combined directly into a sample vial and dried in vacuum.

    2.3 Data acquisition of LC-MS/MS proteomic analysis

    Peptides were reconstituted in solvent A (2% ACN and 0.1% FA in water), loaded on a 2 cm PepMap nano-trap column and separated on a 50 cm EASY-Spray C18 column (Thermo Fisher) connected to an Ultimate 3000 nanoUPLC system (Thermo Fisher) with a 120 min long gradient: 4%–26% of solvent B (2% water and 0.1% FA in ACN) for 120 min, 26%–95% for 5 min, and 95% of solvent B for 5 min at a flow rate of 300 nL min−1. Mass spectra of TMT-labeled samples were acquired on a Q Exactive HF hybrid quadrupole Orbitrap mass spectrometer (Thermo Fisher) ranging from m/z 375 to 1700 at a resolution of R = 120,000 (at m/z 200) targeting 1 × 106 ions for maximum injection time of 80 ms, followed by data-dependent higher-energy collisional dissociation (HCD) fragmentations of precursor ions with a charge state 2+ to 7+, using 45 s dynamic exclusion. The tandem mass spectra of the top 18 precursor ions were acquired with a resolution of R = 60,000, targeting 2 × 105 ions for maximum injection time of 54 ms, setting quadrupole isolation width to 1.4 Th and normalized collision energy to 34%. The reversed-phase chromatographic separation of label-free samples was identical, but the mass spectrometer was operated with m/z 375 to 1500 scanned in full mass and the top 17 precursors were fragmented with 28% NCE at 30,000 mass resolution.

    2.4 Proteome data analysis

    Raw data files were analyzed using Proteome Discoverer v2.5 (Thermo Fisher) with the MS Amanda 2.0 search engine against the human protein database (SwissProt, 20,306 protein consensus entries downloaded on February 28, 2022). A maximum of two missed cleavage sites were allowed for full tryptic digestion, while setting the precursor and fragment ion mass tolerance to 10 ppm and 0.02 Da, respectively. Carbamidomethylation of cysteine was specified as a fixed modification except for freeze-thaw samples, while TMT6plex on lysine and N-termini (+229.1629 Da), oxidation on methionine as well as deamidation of asparagine and glutamine were set as dynamic modifications. Identical settings were applied for label-free analysis except that no TMT6plex modification was used. Throughout the manuscript we only consider proteins as successfully identified whose abundance could be accurately quantified. Initial search results were filtered with 5% FDR using Percolator node in Proteome Discoverer. In alignment with common guidelines for mass spectrometry-based proteomics,[28, 29] our approach quantified proteins based on one or more unique peptides. The proteomics data files have been deposited to ProteomeXchange via the PRIDE partner repository with the data identifier PXD047462.

    2.5 Analysis of hepatic steatosis

    To image hepatic steatosis, spheroids were fixed with 4% paraformaldehyde, washed 3× with PBS and stained overnight using 2 µM Nile Red. The stained spheroids were imaged on a OperaPhenix High Content Screening System (Perkin Elmer). Intracellular triglyceride levels were quantified using the AdipoRed Adipogenesis Assay Kit (Lonza), as previously reported.[19]

    2.6 Gene ontology and pathway analyses

    Gene ontology (GO) and pathway analyses were conducted with an FDR < 0.05 using WebGestalt.[30] For GO, the parameters of cellular component, biological process and molecular function were examined. Pathway analyses were conducted using KEGG pathways as reference.

    2.7 Statistical analyses

    Statistical analysis was performed using Prism version 10.1.1 (GraphPad Software). All results presented as mean values ± standard error of the mean (SEM). Three biological replicates were employed for each condition. P-values or, where applicable, FDRs < 0.05 were considered significant.


    3.1 The relevance of cellular input rapidly saturates

    To understand the importance of critical variables in proteomic workflows, we evaluated the importance of input, lysis strategy and quantification methods on proteome coverage and biological conclusions (Figure 1). To this end, we used primary human liver spheroids and first assessed the number of quantified proteins as a function of the cellular input. Unlike organoids, liver spheroids are formed from fully mature cells that do not proliferate resulting in microtissues that are morphologically homogeneous and stable for at least 6 weeks.[27] We observed that with higher initial material, the number of quantified proteins increases monotonously with diminishing returns (Figure 2A; Table S2). The number of unique peptides identified per protein as a measure of coverage increased from 1500 to 12,000 cells but did not increase further thereafter (Figure 2B; Table S3). When decreasing the input 12-fold from 144,000 to 12,000 cells, the number of quantified proteins decreased only by 15.4% and 18.3% in PMU and RIPA, respectively. However, when only a single spheroid was analyzed (1500 cells), protein numbers reduced by 3-fold (1629 to 631 proteins in PMU and 1460 to 534 proteins in RIPA). When comparing 144,000 cells (96 spheroids) to 1500 cells (a single spheroid) lysed with the same method, we found that 69% of proteins were uniquely identified in the 144,000-cell sample (Figure 2C; Table S4). Furthermore, protein abundances of common proteins were skewed towards a systematic underprediction of protein levels in single spheroid samples (r = 0.66; Figure 2D).

    Details are in the caption following the image
    Graphical overview of the workflow applied from sample preparation to proteomics analysis. aa, amino acid; FT, thermal lysis using freeze-thaw cycles; PMU, ProteaseMAX/urea; TMT, tandem mass tags.
    Details are in the caption following the image
    The association between protein input and the number of identified proteins rapidly saturates. (A) The number of quantified proteins is shown as a function of the cellular input and the lysis method. * indicates p < 0.05 using heteroscedastic two-tailed t-tests. N = 3 biological replicates per condition are combined and analyzed collectively as explained in the methods section. (B) Number of unique peptides identified per protein across the different lysis methods and input amounts. Averages are shown on top of the violin plots. (C) Venn diagram of the proteins quantified using 144,000 and 1500 cells as input, both lysed with PMU. (D) Scatter plot depicting the log-transformed abundance of proteins identified in both samples using 144,000 and 1500 cells as input. Mean abundances of n = 3 biological replicates per condition are depicted. Note that protein abundances are overall underestimated with lower inputs. (E) Scatter plot depicting the correlation of log-transformed abundances between proteins isolated from 144,000 and 36,000 cells when saturating the loading capacity of the column. (F, G) Ranked protein abundances of uniquely identified proteins in the 144,000- (F) and 36,000-cell samples (G).

    To assess whether differences were due to cell number or overall protein amounts, we lysed 36,000 and 144,000 cells with PMU and compared the number of quantified proteins when saturating the loading capacity of the column (3–4 µg per sample). In total, 1449 proteins were quantified in both samples and their abundances correlated overall well (r = 0.82; Figure 2E). However, 291 and 477 proteins could only be quantified in the 36,000- and 144,000-cell samples, respectively. GO analysis of these proteins showed that they were significantly enriched in ribosomal proteins (Enrichment ratio = 3.8; FDR = 0.0015). These unique proteins were mostly lowly abundant, but, interestingly, also included some highly expressed proteins, such as GSTA2, RPS27A and C4B (Figure 2F,G). Combined, these results suggest that cellular input can majorly impact the number of identified proteins when falling below a certain threshold (between 1500 and 12,000 cells for liver spheroids). While these effects are primarily due to changes in the amount of injected protein, lowly abundant proteins can be variable between samples even when normalizing for protein input.

    3.2 The lysis protocol exerts a significant impact on the number of proteins quantified

    Next, we evaluated the effect of lysis method. To this end, we first quantified protein yields and found that PMU extraction resulted in consistently higher amounts of isolated protein compared to RIPA buffer (Figure 3A). When normalizing for input amount of samples isolated via PMU or RIPA, abundances showed an excellent correlation (r = 0.85) with 80% of proteins being detected in both samples (Figure 3B). Notably, nuclear proteins were significantly overrepresented upon PMU isolation (Enrichment ratio = 1.2; FDR = 0.03). Besides PMU and RIPA, thermal lysis protocols have been used for protein isolations from liver microtissues.[27] Strikingly, spheroid lysis by repeated freeze-thaw cycles resulted in 43%–69% more proteins being quantified compared to the use of chemical lysis, with 305 proteins being uniquely identified in freeze-thaw lysed samples (Figure 3C). It is likely that this improved performance is due to the fact that the freeze-thaw protocol does not require clean-up steps that often entail sample loss and proteins can be directly subjected to tryptic digestion. These results thus demonstrate that thermal lysis based on freeze-thawing constitutes the most effective methodology for protein extraction from primary human liver spheroids when the available input material is low.

    Details are in the caption following the image
    The choice of lysis method has major impacts on protein yield. (A) Protein yield as a function of the cellular input and the lysis buffer. (B) Scatter plot depicting the log-transformed abundance of proteins identified in both, samples lysed with PMU or RIPA when the injected amount was normalized for yield. (C) Venn diagram of the proteins quantified in single spheroids (n = 3) lysed with ProteaseMAX/urea (PMU), RIPA or thermal lysis using freeze-thaw cycles (FT).

    3.3 Comparison of label-free and TMT techniques

    Lastly, we compared the impact of the quantification strategy by lysing 144,000 cells with PMU and analyzing the proteomes using either tandem mass tags or label-free quantification. Using the label-free method, 2523 proteins could be quantified of which 834 were not detected using TMT labeling (Figure 4A). As absolute abundances cannot be meaningfully compared between the methods, we used ranked correlations. Notably, ranked protein abundances showed relatively high concordance with correlation coefficients of 0.79 (Figure 4B). Proteins that were exclusively quantified using the TMT strategy were predominantly found within the low abundance range with the notable exception of KRT9, which was the most abundant protein upon TMT labeling, but not identified in label-free samples (Figure 4C). In contrast, proteins uniquely identified with the label-free strategy encompassed both lowly and moderately abundant proteins, including RPS27A, MCOLN2, and TSPAN14 (Figure 4D).

    Details are in the caption following the image
    Comparison of label-free and TMT-based quantification strategies. (A) Venn diagram showing the overlap of proteins identified upon TMT-labeling or label-free (LF) quantification. Both samples were isolated from 144,000 cells using ProteaseMAX/urea. (B) Scatter plot depicting the correlation between ranked abundances upon TMT-labeling or LF quantification. Protein ranks are sorted in ascending order with the most abundant protein having the highest rank. (C, D) Ranked protein abundances of uniquely identified proteins in TMT-labeled (C) and LF samples (D).

    3.4 Proteomic workflows impact the quantification of proteins critical for hepatic functions

    Proteins involved in the absorption, distribution, metabolism and excretion (ADME) of drugs are essential to emulate hepatic functionality ex vivo. We thus compiled a list of 29 clinically important hepatic transporters and drug metabolizing enzymes identified in FDA labels (Table S5) and compared their expression in isogenic liver spheroids using different proteomic workflows. Overall, the different methods differed substantially in ADME quantification performance (Figure 5A). With high cell numbers as input (≥72,000 cells), TMT labeling quantified at least 69% of important ADME proteins, with the choice of lysis buffer having a relatively minor impact on outcome. Coverage increased even further to 92% of ADME proteins when a label-free quantification strategy was employed. In contrast, drastic differences between lysis protocols were apparent for low inputs (≤36,000 cells) where PMU lysis resulted in approximately 30% more ADME proteins than lysis by RIPA. When using single spheroids, the FT protocol allowed the quantification of 55% of ADME proteins (16/29), whereas PMU and RIPA yielded only 38% (11/29) and 10% (3/29), respectively.

    Details are in the caption following the image
    Proteomic workflows drastically differ in coverage of clinically relevant ADME proteins. (A) Number of the clinically relevant ADME proteins detected with the different proteomic workflows (PMU = ProteaseMAX/urea; LF = label-free). Note that there is an overall increase in the number of identified ADME proteins with increasing cell amount. For low input amounts, lysis via repeated freeze-thaw cycles (FT) results in the highest ADME protein coverage. (B) Heatmap depicting the mean protein abundances (log2 transformed) of clinically relevant ADME proteins after TMT-based quantification. (C) Heatmap depicting the mean protein abundances (log2 transformed) of clinically relevant ADME proteins after LF quantification. (D) Pie charts showing CYP isoform compositions as quantified by the different proteomic workflows. Isoform compositions of representative liver samples are shown as reference.[22, 31, 32] (E) Correlation plot showing the fraction of clinically relevant CYP isoforms in human liver spheroids and human liver samples.

    The different protocols exhibited differential utility for the analysis of membrane-integral transporters and drug metabolizing enzymes (Table S6). The quantification of ABC and SLC transporters was not possible from single spheroids, regardless of the lysis method, but continuously increased with higher cellular input. All clinically relevant transporters were successfully quantified when 144,000 cells were lysed with PMU followed by label-free sample analysis, whereas only 44% of these transporters (4/9) were identified when using TMT labeling. Phase I and phase II drug metabolizing enzymes were overall more readily detected than transporters with 75%–95% being quantified for cellular inputs above 12,000 cells. Importantly, lysis of single spheroids using the freeze-thaw protocol resulted in the quantification of 80% of drug metabolizing enzymes (16/20), whereas PMU and RIPA only identified 55% (11/20) and 15% (3/20), respectively.

    Next, we compared CYP isoform composition and found drastic differences between workflows (Figure 5D). While CYP3A4 was the most abundant CYP when analyzing single spheroids (74.5%–100% of identified CYPs), the use of higher cellular inputs indicated that CYP3A4 and CYP2E1 are present at approximately similar amounts (17.8%–27.8% for CYP3A4 and 23.8%–34.7% for CYP2E1). Other abundant CYPs included CYP2C9 (10.8%–18.2%), CYP2C8 (6.5%–14%) and CYP2D6 (0%–10.4%), whereas levels of CYP2A6 (1.3%–4.7%), CYP2B6 (0.4%–1.6%), CYP2C19 (0%–0.7%) and CYP2J2 (0%–0.4%) were considerably lower. These results align closely with isoform compositions identified in liver resections,[22, 31, 32] where CYP3A4 (13.4%–26.4%), CYP2C9 (12.8%–26.4%), CYP2E1 (6.2%–21.1%) and CYP2C8 (8.4%–17.7%) were most highly expressed, while CYP2A6 (3.8%–21.5%), CYP2B6 (1.1%–3.2%), and CYP2J2 (0.1%–0.9%) were less abundant (Figure 5E). These results demonstrate that the presented workflows can provide reliable quantifications of both highly and lowly abundant hepatic CYPs and show that 3D liver spheroids maintain a physiologically relevant CYP isoform composition.

    3.5 The choice of sample preparation method impacts biological outcomes

    To examine the potential impact of proteomic workflows on molecular phenotype assessments and their biological conclusions, we performed pathway analysis in a pathophysiological model of MASLD in which 3D human liver spheroids were exposed to pathophysiological concentrations of free fatty acids. First, we confirmed that spheroids became indeed steatotic, as evidenced by the appearance of intracellular lipid droplets and increased triglyceride levels (Figure 6A,B). For cellular inputs below 36,000 cells, analysis of samples lysed with PMU or RIPA identified only a small number of differentially regulated pathways (n = 1–7 with FDR < 0.05), including PPAR signaling, differential expression of ribosomes and peroxisomal components, as well as valine, leucine and isoleucine degradation (Table S7). In contrast, when samples were lysed by repeated freeze-thawing, 20 differentially regulated pathways were identified with statistical significance, including gluconeogenesis, the citrate cycle, oxidative phosphorylation, as well as xenobiotic metabolism by CYPs. Increasing the cellular input (72,000 or 144,000 cells) resulted in higher numbers of significant pathways (up to 42 with FDR < 0.05), with PMU lysis identifying over 50% more pathways than RIPA (Table S7). No clear advantage of label-free quantification compared to TMT-labeling was identified.

    Details are in the caption following the image
    Proteomic workflows impact the analytical results in a 3D model of metabolic dysfunction-associated steatotic liver disease. (A) Representative fluorescence microscopy images of steatotic 3D liver spheroids and isogenic controls. Dashed line indicates spheroid outline. Scale bar = 100µm. (B) Quantification of intracellular triglyceride levels in steatotic spheroids and non-steatotic controls. **** indicates p < 0.0001 in a heteroscedastic two-tailed t-test. (C–H) Bubble plots of selected pathways associated with non-alcoholic fatty liver disease show the statistical evidence for differential regulation in steatotic spheroids compared to controls upon analysis with different workflows. Statistical significance expressed as -log(false discovery rate [FDR]) is shown on the ordinate. The dashed line corresponds to FDR = 0.05. Circle size indicates the enrichment ratio.

    Among pathways that are robustly linked to hepatic steatosis in vivo,[33, 34] we identified aberrations in gluconeogenesis, fatty acid and amino acid metabolism as well as the citrate cycle (Figure 6C–H). These results indicate that global proteomics can successfully identify key molecular hallmarks of MASLD, such as dysregulated nutrient fluxes, oxidative stress indicated by altered glutathione levels or decreased mitochondrial respiration. However, while PMU and RIPA lysis required high cell numbers (≥72,000 cells) to identify most pathways, thermal lysis only required single spheroids (1500 cells) to yield overall similar conclusions. These results demonstrate that proteomic workflows based on repeated freeze-thaw cycles can reduce the amount of input material required to arrive at statistically sound biological conclusions.


    Advances in molecular medicine over the last 30 years have resulted in a shift in drug discovery towards target-based approaches. The main advantages of target-based approaches are that they have been easier to implement than phenotypic approaches and that they allow the use of target-oriented toolkits, such as crystallography, mutagenesis and molecular pharmacology.[35] However, this reductionist approach largely uncouples pharmacology from pathophysiology and is thus prone to produce compounds with excellent target selectivity, which, however, might not exert the anticipated physiological responses due to our limited understanding of disease mechanisms. In support of these concerns, an extensive recent review of the origins of all approved drugs revealed that only 123 (10.7%) of all 1144 approved small-molecule drugs have been discovered by target-based discovery, whereas 1021(89.3%) were identified by phenotype-based approaches.[36] The advent of organotypic human tissue models with physiologically relevant phenotypes allows for target-agnostic discovery strategies with moderate-to-high throughput. However, to fully leverage the advances of such emerging phenotypic models, sensitive and scalable technologies are required that can provide accurate results despite the low inputs.

    Our study demonstrates that, as expected, input constitutes a critical factor impacting proteome coverage. GO analysis showed mitochondrial and ribosomal proteins to be underrepresented with low input, likely due to the high loss of these proteins and peptides during the sample preparation procedure.[37] The comparison between the lysis methods, specifically PMU and RIPA, led to a substantial difference in protein yield, with PMU providing almost twice the protein yield compared to RIPA. This discrepancy can likely be ascribed to the precipitation-based methodology utilized by RIPA, which entails higher protein and peptide losses.[38] Interestingly, when the injected amount of protein was normalized, PMU- and RIPA-extractions led to highly concordant protein quantifications whereas some differences between cell numbers remained even when protein input was normalized.

    Besides the amount of input, the choice of quantification strategy had major effects on proteome coverage. We here identified up to 1926 proteins using TMT-labeling, while a label-free approach identified considerably more (n = 2523) in agreement with previous findings.[39, 40] These numbers are overall consistent albeit slightly lower than the number of proteins previously identified in hepatocytes.[41-43] Among ADME proteins, the label-free strategy identified clinically relevant ABC (MDR1, BSEP and MRP2/3) and SLC transporters (OCT1 and OATP1B1); in contrast, SLC transporters were not detected when using TMT labeling. Unlike transporters, phase I and phase II metabolic enzymes were more consistently detected across samples. However, when input was low, only PMU but not RIPA extraction resulted in broad coverage of CYPs and UGTs.

    Liver spheroids have demonstrated their utility in phenotypic screens for compounds stimulating liver regeneration using chemogenomic probes with high target selectivity.[44] However, when screening compounds with less well characterized pharmacology, strategies for molecular target identification must be developed that map disease-relevant phenotypic changes to specific protein targets or signaling pathways, ideally proteome-wide.[45, 46] As an example of a disease-relevant assay, we here used a primary human liver spheroid model of MASLD. This model is steatotic and insulin resistant, thus recapitulating two major functional hallmarks of MASLD, and has been comprehensively characterized at the transcriptomic and lipidomic level.[18, 19] Furthermore, using liver spheroids we functionally dissected the crosstalk between hepatocytes and liver macrophages in MASLD[47, 48] and could identify roles of subpopulations of myeloid in curbing hepatic metabolic stress in obesity.[49] However, comprehensive phenotypic profiling of this model at the proteome level had so far been limited due to the high cell numbers required for conventional proteomics. Importantly, we here found that protein extraction using repeated freeze-thaw cycles was identified to outperform chemical lysis methods by over 40%. This improved performance is likely at least in part due to the lower complexity of the freeze-thaw protocol that does not require clean-up steps, which are prone to sample loss. Pathway enrichment analysis identified 20 differentially regulated pathways in steatotic spheroids, which are in close alignment with in vivo alterations. Thus, while expression levels of individual proteins can fluctuate considerably when using only single spheroids as input, our results demonstrate that meaningful molecular phenotype changes can nevertheless be detected at the pathway level even from individual spheroids.

    In summary, our findings underscore that input material, lysis protocols and quantification strategies can have important impacts on proteomic coverage and biological interpretations. These results have important implications for the interpretation of the existing literature, particularly with regards to the meta-analysis of published findings generated using different proteomic methodologies, and incentivize standardization to increase reproducibility. We present standardized and scalable workflows that allow to draw reliable and biologically meaningful conclusions from the analysis of individual microtissues, thereby extending the applicability of proteomic profiling to phenotypic screens using organotypic cultures and organoids where cell numbers are severely limited.


    Stefania Koutsilieri: Data curation (lead); formal analysis (lead); investigation (lead); methodology (equal); visualization (supporting); writing – original draft (lead). Evgeniya Mickols: Formal analysis (supporting); investigation (supporting); methodology (supporting); visualization (supporting). Akos Vegvari: Conceptualization (equal); Formal analysis (lead); methodology (equal); investigation (supporting). Volker Lauschke: Conceptualization (equal); Funding acquisition (lead); supervision (lead); writing – original draft (supporting); writing – review & editing (lead).


    The work received funding from the Swedish Research Council [grant numbers: 2021-02801 and 2023-03015], by the Ruth och Richard Julins Foundation for Gastroenterology [grant number 2021-00158], by the Knut and Alice Wallenberg Foundation [Grant VC-2021-0026], by the International Foundation for Ethical Research (IFER; Chicago, USA) and by the Robert Bosch Foundation, Stuttgart, Germany. Furthermore, this project has received funding from the Innovative Medicines Initiative 2 Joint Undertaking (JU) under grant agreement No 875510. The JU receives support from the European Union's Horizon 2020 research and innovation programme and EFPIA, Ontario Institute for Cancer Research, Royal Institution for the Advancement of Learning McGill University, Kungliga Tekniska Högskolan and Diamond Light Source Limited. SK is a recipient of a scholarship from the Onassis Foundation.


      V.M.L is CEO and shareholder of HepaPredict AB, as well as co-founder and shareholder of PersoMedix AB. The other authors declare no relevant interests.


      All data is available from the corresponding authors upon reasonable request.