Methods and Protocols

The human hepatocellular carcinoma cell lines Huh7.5 and HepG2‐hNTCP were kindly provided by Ralf Bartenschlager, University of Heidelberg, Heidelberg, Germany and Stephan Urban, University Hospital Heidelberg, Heidelberg, Germany, respectively (Blight et al, 2002; Ni et al, 2014). Huh7.5 cells were cultivated in growth medium containing sodium pyruvate (Dulbecco's Modified Eagle Medium #31053 (Gibco), 10% FCS (Gibco), 1% GlutaMAX (Gibco), 1 mM Sodium pyruvate (Gibco), 100 U/ml penicillin⁄streptomycin (Gibco)), while HepG2‐hNTCP cells were cultivated in the absence of sodium pyruvate and using 1% Glutamin (Gibco) instead of GlutaMAX. Cell line authentication was performed using Multiplex Cell Authentication by Multiplexion (Heidelberg, Germany) as described (Castro et al, 2013). The SNP profiles matched known profiles or were unique. Purity of cell lines was validated using the Multiplex cell Contamination Test by Multiplexion (Heidelberg, Germany) as described (Schmitt & Pawlita, 2009). No Mycoplasma, SMRV, or interspecies contamination was detected.

Primary human hepatocytes were prepared at the University of Leipzig and at the University Hospital Heidelberg. Informed consent of the patients for the use of tissue for research purposes was obtained corresponding to the ethical guidelines of University of Leipzig and University Hospital Heidelberg, respectively. Tissue samples were acquired by partial hepatectomy and originate from healthy sections of resected liver tissue. Hepatocytes were isolated as described recently (Kegel et al, 2016), and hepatocytes were cultivated in adhesion medium (Williams’ Medium E (Biochrom F1115), 10% FCS (Gibco), 0.1 μM dexamethasone, 0.1% insulin, 2 mM l‐glutamine (Gibco), 100 U/ml penicillin⁄streptomycin (Gibco)).

Plasma samples from patients with chronic HBV infection were obtained after informed consent at the University Hospital Freiburg (ethics committee approval number: 227/15). Plasma levels of IFNα were assessed by a custom multiplex assay (Eve Technologies Corp, Canada).

For knockdown studies, cells were transfected with 50 nM siRNA using Lipofectamine RNAiMax (Invitrogen) according to manufacturer's protocol. Cells were incubated with the siRNA complexes for 20 h. The following siRNAs of Dharmacon (GE healthcare) were used ON‐TARGETplus Human USP18 (11274) siRNA—SMARTpool, and ON‐TARGETplus Non‐targeting pool.

USP18 cDNA (https://www.ncbi.nlm.nih.gov/nuccore/{"type":"entrez-nucleotide","attrs":{"text":"AL136690","term_id":"12052903","term_text":"AL136690"}}AL136690) was provided by the Vector and Clone Repository of the Genomics and Proteomics Core Facility of the DKFZ and subsequently cloned in pMOWS‐TreT‐puro (Pfeifer et al, 2010) using MefI/EcoRI and BamHI restriction enzymes (NEB).

For stable transduction, 800,000 phoenix ampho cells were seeded in 6‐well plates (TPP) and the next day co‐transfected with 8 μg pMOWS‐TreT‐USP18‐puro or pMOWS‐TreT and 2 μg pMOWS‐TAM2 using calcium phosphate precipitation in growth medium supplemented with 25 μM chloroquine.

After 8‐h incubation, medium was replaced by growth medium and incubated overnight. The next day supernatant was harvested, filtered, and supplemented with 8 μg/ml polybrene. 1 ml of supernatant was added to 200,000 Huh7.5 cells seeded in 6‐well plates (TPP), which were transduced by centrifugation for 3 h at 340 × g at 37°C. Selection was performed using 0.75 μg/ml puromycin (Sigma), starting 48 h after transduction.

600,000 Huh7.5 cells or 1,000,000 HepG2‐hNTCP were seeded in 6‐well plate format (TPP) 1 day in advance. Prior to stimulation, cells were washed three times with DPBS (Pan Biotech) and growth factor‐depleted in starvation medium (Dulbecco's Modified Eagle Medium #31053 (Gibco), 1% (v/v) GlutaMAX (Gibco), 1 mM Sodium pyruvate (Gibco)) supplemented with 1 mg/ml BSA and 25 mM HEPES (Gibco) for 3 h (Huh7.5 cells). HepG2‐hNTCP cells were growth factor‐depleted overnight (15 h) in starvation medium without sodium puryvate and HEPES. The Huh7.5 cells were kept in 1.5 ml of the indicated medium, and the HepG2‐hNTCP were cultivated in 1 ml of the medium described above. Due to these differences in media volume, the amount of Roferon added per ml was adjusted to ensure that the same total amount of Roferon was added per well. For co‐immunoprecipitation experiments, 1.5 million Huh7.5 cells or 2.5 million HepG2‐hNTCP were seeded on 60 mm tissue culture dishes (TPP) and kept in 3.7 ml of the respective medium. After growth factor depletion, cells were stimulated on a 37°C heating block by addition of interferon alpha 2a (PBL 11000‐1), Roferon (Roche, PZN 08543409), or with recombinant human interferon gamma (R&D, 285‐IF‐100) and harvested at different time points. For sensitization experiments, cells were prestimulated with IFNα and stimulated 24 h later by addition of IFNα.

For primary human hepatocytes, 1 or 1.5 million viable cells were seeded in 6‐well collagen‐coated plates (Bio Coat, Corning) 1 day prior to the experiment. The next day, cells were gently washed twice with DPBS (PAN Biotech) before cells were growth factor‐depleted for 3 h in starvation medium (Williams’ Medium E (Biochrom F1115), 2 mM l‐glutamine (Gibco), 100 U/ml penicillin⁄streptomycin (Gibco)) prior to the experiment.

For model purposes, interferon concentrations given in units/ml and μg/ml were converted to nM based on information supplied by the datasheet (IFNα‐2a) or product information (Roferon). In addition, the activity of Roferon (Roche) and IFNα‐2a (PBL) were compared using the Huh7/LucUbiNeo cell line (Lohmann & Bartenschlager, 2014), a Huh7‐Lunet cell line with a stably replicating HCV genotype 1b (Con1) subgenomic replicon under the selective pressure of G418 (0.5 mg/ml). The replicon contains a neomycin phosphatase as well as a firefly luciferase reporter gene instead of the viral structural genes and harbors replication‐enhancing mutations in the nonstructural genes (Con1‐ET). To perform the titration, 75,000 cells were seeded in 96 well plate format in the absence of G418 and the next day IFNα and Roferon was added in two‐step serial dilutions. After 48 h, cells were lysed in luciferase lysis buffer (1% Triton‐x 100, 25 mM glycyl‐glycine pH 7.8, 15 mM MgSO₄, 4 mM EGTA pH 7.8, 10% glycerol, 1 mM DTT) and stored at −80°C. Luciferase activity was measured on a Mithras2 LB 943 (Berthold Technologies). Signal intensities were normalized to untreated cells and were fitted by four‐parameter Hill kinetics to determine the IC₅₀ concentrations (Fig EV4A). Based on these data, equipotent concentrations for IFNα and Roferon were calculated (Fig EV4B).

Cellular fractionation was performed to obtain cytoplasmic and nuclear protein lysates. Lysis buffers were freshly supplemented with the protease inhibitors Aprotinin and AEBSF (Sigma). Cells were lysed in 250 μl cytoplasmic buffer (10 mM Hepes, 10 mM KCl, 0.1 mM EDTA, 0.1 mM EGTA, 1 mM NaF, 1 mM Na₃VO₄, 0.4% NP40) and gently scraped and transferred to Eppendorf tubes. Lysates were vortexed for 10 s and centrifuged at 1,000 × g, at 4°C for 5 min. Supernatants were transferred (cytoplasmic fraction) and pellets were washed with 250 μl washing buffer (10 mM Hepes, 10 mM KCl, 0.1 mM EDTA, 0.1 mM EGTA, 1 mM NaF, 1 mM Na₃VO₄,) and centrifuged at 1,000 × g, at 4°C for 5 min. Supernatants were discarded and 45 μl nuclear lysis buffer was added (20 mM Hepes, 25% Glycerin, 400 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1 mM NaF, 1 mM Na₃VO₄, 0.4% NP40). Lysates were vortexed 10 s every 2 min for 15 min in total. Nuclear fraction was collected by taking the supernatant after 5 min centrifugation at 20 817 × g at 4°C.

Total cell lysates were prepared by lysing cells in 1 × RIPA lysis buffer (1% NP40, 0.5% DOC, 0.1% SDS, 250 mM NaCl, 2.5 mM EDTA, 50 mM Tris pH 7.2). Cells were lysed in 250 μl lysis buffer, scraped, transferred to Eppendorf tubes, tumbled for 20 min at 4°C, and subjected to sonication (Sonopuls, Bandelin, for 30 s, with 75% amplitude, 0.1 s on 0.5 s off). Whole cell lysates were collected after 10 min centrifugation at 4°C at 20,817 × g.

The concentration of protein lysates was determined by Pierce™ BCA Protein Assay Kit (Thermo‐Fisher) and measured on the InfiniteF200Pro plate reader (Tecan). 10 or 20 μg samples were prepared for quantitative immunoblotting. SOCS3 was immunoprecipitated overnight with SOCS3 antibody (Merck) and protein G sepharose beads (GE Healthcare) supplemented with 0.1 ng SBP‐SOCS3. For co‐immunoprecipitation experiments, 650 μg or 1,200 μg of total cell lysates were incubated overnight with STAT2 antibody (Merck) or IRF9 antibody (Cell Signaling) and protein A sepharose beads (GE Healthcare). Samples were loaded in randomized order to avoid correlated blotting errors (Schilling et al, 2005). Blots were developed using self‐produced ECL solutions (Solution 1: 0.1 M Tris pH 8.5, 2.5 mM Luminol, 0.4 mM p‐coumaric acid; Solution 2: 0.1 M Tris pH 8.5, 0.018% H₂O₂) or using ECL Western Blotting Reagents (GE healthcare) on the CCD camera‐based ImageQuant LAS4000 (GE Healthcare). To remove previous antibody signals, HRP groups were quenched with H₂O₂ as described previously (Sennepin et al, 2009) or antibodies were removed by incubation with stripping buffer (0.063 M Tris pH6.8, 2% SDS, 0.7% β‐mercaptoethanol) at 65°C. Bands were quantified using ImageQuant software (GE healthcare).

For fluorescence‐based detection (Appendix Fig S1), blots were developed using the Odyssey near‐infrared fluorescence scanner (LI‐COR Biosciences) using a secondary antibody coupled to IRDye 800CW near‐infrared dye.

The following antibodies were used Phospho‐Stat1 (Tyr701) (58D6) #9167; Phospho‐Stat2 (Tyr690) #4441; USP18 (D4E7) #4813, all from Cell Signaling. STAT1, CT #06‐501; STAT2, CT #06‐502 both from Merck. IRF9/ISGF3c #610285 (Bectin Dickenson), Calnexin #Adi‐SPA‐860 (Enzo Life Sciences), HDAC1 (10E2) #sc81598 (Santa Cruz), β‐Actin #A5441 (Sigma‐Aldrich), and SOCS3 #ab16030 (Abcam). IP was performed using the following antibodies SOCS3 #04‐004 (1B2), STAT2 CT #06‐502, both from Merck, and IRF9 #76684 (D2T8M) from Cell Signaling. Secondary antibodies include rabbit and mouse specific antibodies raised in goat, coupled to HRP (Dianova) and Goat anti‐Rabbit IRDye 800CW #926‐32211 (LI‐COR Biosciences).

cDNA of human USP18 (https://www.ncbi.nlm.nih.gov/nuccore/{"type":"entrez-nucleotide","attrs":{"text":"AL136690","term_id":"12052903","term_text":"AL136690"}}AL136690) was provided by Genomics and Proteomics Core Facility of the DKFZ and subsequently cloned in pGEX2T vector (GE Healthcare) using BamHI and EcoRI/MfeI (NEB) restriction sites.

Recombinant proteins were produced upon IPTG addition in transformed BL21‐CodonPlus(DE3)‐RIL competent bacteria (Agilent) and were purified using GST or SBP isolation as described previously (Raia et al, 2011).

SBP‐SOCS3 calibrator was kindly provided by Anja Zeilfelder (Klingmüller lab, DKFZ, Heidelberg). SBP‐STAT1ΔN, SBP‐STAT2ΔN, and GST‐IRF9 were established previously in our laboratory (Maiwald et al, 2010).

Concentration of calibrators was determined using BSA protein standard (Pierce) on SDS‐PAGE gel stained with SimplyBlue Safe stain (Invitrogen) according to manufacturer's protocol.

To determine molecules per cell, the cell number of Huh7.5 or HepG2‐hNTCP was counted with the Neubauer improved counting chamber for each treated condition. For primary human hepatocytes from Patients 1–3, a dilution curve was established, which was based on protein concentrations derived from different amounts of cells lysed in 250 μl 1 × RIPA lysis buffer. For primary human hepatocytes from Patient 4–6, cells were fixed with paraformaldehyde and nuclei were stained with DAPI. 16 images per condition were taken using a Motorized Widefield Microscope (Zeiss Cell Observer). Using Fiji software, the respective number of cells per dish was quantified.

Different amounts of calibrators were spiked in 10 μg whole‐cell lysates and quantitative immunoblotting was performed with the indicated antibodies. The linear regions of the calibration curves were fitted with a linear regression model in R, and the amount of endogenous signal was interpolated. Uncertainties were computed as standard error of the mean of the different samples assuming log‐normally distributed signals.

Total RNA was extracted using RNeasy kit (Qiagen) according to manufacturer's instruction. Clearing of the lysates was achieved using QIAshredder spin column (Qiagen). RNA concentrations were determined by absorbance (Nanodrop2000, Thermo Scientific), and reverse transcription was performed with 1 μg of RNA in 20 μl according to manufacturer's instruction (High Capacity cDNA Reverse Transcription Kit from Applied Biosystems). Quantitative PCR was performed on the LigthCycler480 (Roche) using primers and dual hybridization probes in 2× Probes Master (Roche). Cycling protocol consisted of 5 min pre‐incubation at 95°C, 50 amplification cycles (95°C for 10 s, 60°C for 30 sec and acquisition at 72°C for 1 s), and 2 min cooling. Quantification cycles (Cq) were determined by absolute quantification with second derivative maximum method using the software LightCycler480SW1.5.1. Samples for calibration curve were included in each measurement to assess efficiency of primer hybridization.

Data were normalized with the geometric mean of the reference genes hypoxanthine‐guanine phosphoribosyltransferase (HPRT), TATA box‐binding protein (TBP), and glyceraldehyde‐3‐phosphate dehydrogenase (GAPDH).

Primers were designed using the Universal Probe Library Assay Design Center (Roche Applied Biosciences) and manufactured by Eurofins. The utilized UPL probes and primers sequences for human genes are listed in Table 1.

Primers for quantitative RT‐PCR of human genes

Primers are listed in 5′‐ to 3′‐end orientation.

Promoter analysis of the following human genes was performed of a 3,000 bp region consisting of 1,250 bp in front and 1,750 bp inside the corresponding gene: IRF1 (>{"type":"entrez-nucleotide","attrs":{"text":"NC_000005.10","term_id":"568815593","term_text":"NC_000005.10"}}NC_000005.10:c132492039‐132489039 Homo sapiens chromosome 5, GRCh38.p13 Primary Assembly); SOCS3 ({"type":"entrez-nucleotide","attrs":{"text":"NC_000017.11","term_id":"568815581","term_text":"NC_000017.11"}}NC_000017.11:78358329‐78361329 Homo sapiens chromosome 17, GRCh38.p13 Primary Assembly); DDX58 (>{"type":"entrez-nucleotide","attrs":{"text":"NC_000009.12","term_id":"568815589","term_text":"NC_000009.12"}}NC_000009.12:c32527446‐32524446 Homo sapiens chromosome 9, GRCh38.p13 Primary Assembly); HERC5 (>{"type":"entrez-nucleotide","attrs":{"text":"NC_000004.12","term_id":"568815594","term_text":"NC_000004.12"}}NC_000004.12:88455354‐88458354 Homo sapiens chromosome 4, GRCh38.p13 Primary Assembly); and IFI44L (>{"type":"entrez-nucleotide","attrs":{"text":"NC_000001.11","term_id":"568815597","term_text":"NC_000001.11"}}NC_000001.11:78619137‐78622137 Homo sapiens chromosome 1, GRCh38.p13 Primary Assembly). The promoter analysis was performed using Findpatterns of the GCG sequence analysis package using W2H (Senger et al, 1998). The following consensus sequences were tested: GAS (gamma‐activated sequence) TTNCNNNAA (Darnell et al, 1994); and ISRE (interferon‐stimulated response element) TTTCNNTTYY (Dill et al, 2014). Gene annotation is displayed using Geneious v11.1 (Kearse et al, 2012).

Electrophoretic mobility shift assays were performed with nuclear protein lysates obtained from cellular fractionation. Oligonucleotides probes used in the assay were synthesized (Sigma) harboring the GAS‐binding region of the human IRF1 promoter (5′‐CATTTCGGGGAAATCAGGC‐3′). After annealing, oligonucleotides were end‐labeled with [γ‐³²P]‐adenosine triphosphate (3,000 Ci/mmol; PerkinElmer) using the T4 polynucleotide kinase (New England Biolabs) and further purified by gel filtration on Illustra ProbeQuant G‐50 Micro Columns (GE Healthcare). Nuclear lysates (5 μg) were normalized in 19‐μl reaction mixtures containing 20 mM HEPES, pH 7.9; 1 mM DTT; 0.1 mM EDTA; 50 mM KCl; 1 mM MgCl₂; 5% glycerol; 200 μg/ml BSA; and 1 μg poly[dI‐dC] (Sigma). The reaction was placed on ice for 20 min before incubating with ³²P‐labeled oligonucleotide probes (10,000 cpm) for 20 min at room temperature. For the supershift samples, the lysate‐DNA mixture was incubated with 2 μg of antibodies recognizing either STAT1 (Merck, CT #06‐501), STAT2 (Merck, CT #06‐502), or IRF9 (Abcam, ab126940) for 10 min at room temperature. Samples were resolved on a 4.5% native polyacrylamide gel (37.5:1) in a 0.5 × Tris‐borate‐EDTA (TBE) buffer for 6 h at 130 V at 4°C. Gels were dried for 60 min at 80°C and visualized by PhosphorImager.

Immunoblot data were normalized to housekeepers Calnexin, Actin, HDAC1, or recombinant SBP‐SOCS3 using GelInspector (Schilling et al, 2005). For each target, data points were scaled together by means of the R package blotIt (Kaschek, 2011) assuming log‐normally distributed signals. Independent experiments that contained more than three overlapping data points with 1,400 pM IFNα treatment were used as a reference for scaling. Different scaling factors for each gel as well as time course values used for parameter estimation were simultaneously determined by generalized least squares estimation, with data points assumed to be log‐normally distributed. Uncertainties correspond to 68%‐confidence intervals of the estimated data points. For the control model (Huh7.5), 10,945 single data points (Dataset EV1) were scaled by blotIt to obtain 1,902 data points with confidence intervals that served for calibration of the model (Dataset EV2). In addition, the 16 determined amounts of molecules per cell of the feedback proteins were utilized for model calibration. For the validation of the control model with EMSA and protein data (Figs 4B and EV3D), 134 single data points (Dataset EV3) were scaled to obtain 45 data points with confidence intervals (Dataset EV4). For the validation of the control model with qRT‐PCR data (Fig 4C), 315 single data points (Dataset EV3) were averaged to obtain 105 data points with standard error of the mean (Dataset EV4). For the Roferon model (Fig 4D), 891 single data points (Dataset EV7) were scaled to obtain 221 data points with confidence intervals (Dataset EV8) and for the HepG2‐hNTCP model, 1,274 single data points (Dataset EV9) were scaled to 305 data points with confidence intervals (Dataset EV10). For the validation of the HepG2‐hNTCP model with qRT‐PCR data (Appendix Fig S7C), 252 single data points (Dataset EV11) were averaged to obtain 84 data points with standard error of the mean (Dataset EV12). For patient‐derived primary human hepatocytes, no biological replicates were available. Here, experimental errors were estimated from the signal variance of the hepatocytes prestimulated with 1,400 pM IFNα, assuming that the corresponding underlying time course stays constant after stimulation (Dataset EV13).

The modeling process was performed by means of the R package dMod (Kaschek, 2017; Kaschek et al, 2019). In total, the mathematical model consists of 41 species and 75 reactions that were derived by the law of mass‐action and Michaelis–Menten kinetics. The reactions are justified based on published literature (Appendix Table S1). Observables were computed with respect to model states as indicated in Appendix Table S2. Parameter values of the global optimum for the Huh7.5 core model and profile likelihood‐based confidence intervals are shown in Appendix Table S3. Parameters were log‐transformed to ensure positivity and enable optimization over a broad range of magnitudes (Raue et al, 2013). Calculation of analytical steady‐state expressions (Rosenblatt et al, 2016) and application of model reduction for ODE models (Maiwald et al, 2010) was incorporated by a set of parameter transformations (Appendix Table S4). In some cases during the model simplification procedure, parameter values were fixed instead of changing the model structure to keep a biological meaningful model structure. Parameters were estimated by the method of maximum likelihood performing a deterministic multi‐start optimization of 1,000 randomly chosen parameter sets by means of the trust region optimizer trust (Geyer, 2004). Parameter values were not restricted by fixed borders. Instead, in order to prevent the optimizer from finding solutions with very low or high parameter values, we constrained the model parameters with a weak L₂ prior that contributed with one to the likelihood, if the parameter differed by five orders of magnitude from 1. When computing the profile likelihood (Appendix Fig S12), these L₂ priors were substracted in order to ensure an exclusively data‐based identifiability analysis. To show reliability of the optimization, the 200 best optimization runs were displayed as a waterfall plot sorted by their objective values (Appendix Fig S13; Raue et al, 2013). Different local optima were found multiple times, and the global optimum was found in 18 of the 200 cases. To test identifiability of the parameters and to calculate confidence levels for the estimated values, the profile likelihood (Raue et al, 2009) was computed for each parameter. In total, 12 initial values, 17 scaling and offset parameters and 56 dynamical parameters were estimated for the Huh7.5 model. Profile likelihoods (Appendix Fig S12) showed finite confidence intervals for 74 out of 85 parameters. From the remaining parameters, three showed confidence intervals open to minus infinity and eight were open to plus infinity. However, due to the biological significance of the parameters and to ensure that the model selection analysis and the application of the model to different cell systems by L₁ regularization was performed without additional constraints, no further model reduction was applied. Estimated parameter values with corresponding confidence intervals and the resulting model parameters obtained after transformation are summarized in Appendix Table S4. For the receptor model, different structures were evaluated by means of the Bayesian information criterion (Schwarz, 1978).

To analyze experimental measurements performed with the Roferon, the calibrated Huh7.5 model (control model) was utilized with all parameters being fixed except for scaling and offset parameters as well as the binding affinity of the ligand (parameter BindIFN) that were re‐estimated by the method of maximum likelihood.

For analysis of the HepG2‐hNTCP data, the previously estimated binding affinity for Roferon was fixed. Scaling and offset parameters as well as parameters corresponding to the molecules per cell in the system, i.e., totSTAT1, totSTAT2, totIRF9, and synthUSP18, were re‐estimated. The four parameters, synthUSP18mRNAbasal_OE, synthUSP18_inh, synthUSP18mRNAbasal_inh, and synthUSP18mRNA_inh, that were used for incorporation of inhibitor and overexpression conditions were not considered. The remaining 60 parameters of the model were estimated with an additional L₁ constraint as previously described (Merkle et al, 2016). In our case, a combination of L₁ and L₂ prior, i.e., an elastic net (Zou & Hastie, 2005), was applied that penalizes least when parameters take the exact value as in the control model. Penalization strength was chosen such that L₁ and L₂ prior contribute to the same extent if the parameter value differed by one order of magnitude to the control model. Optimization was performed by means of a trust region optimizer which was adapted to L₁ regularization as part of the dynamic modeling framework dMod. For each of 25 different regularization strengths, we performed 200 optimization runs starting from randomly chosen parameter sets. The 5,000 resulting fits were evaluated by means of a combination of goodness‐of‐fit and number of parameters being different from the control model. Based on Bayesian information criterion, we defined the objective value 4 × ln(n) × k − 2 × ln(L), where n is the number of data points, k the number of parameters different from the control model and L the value of the likelihood function. Compared to classical BIC, our definition favors models with low amounts of necessary L₁ parameters. The resulting parameter estimates are summarized in Appendix Tables S5 and S6. For the HepG2‐hNTCP model, we obtained two out of the 60 L₁ parameters (indicated by ratio_parameter) to be different from Huh7.5.

For the analysis of primary human hepatocyte data, we performed L₁ regularization similar to the case of HepG2‐hNTCP. For each of six patients, parameters defining the number of molecules per cell of the proteins as well as scaling and offset parameters were re‐estimated. The remaining primary human hepatocyte‐specific parameters were assumed to be the same for all patients and were estimated by means of L₁ regularization (see Appendix Table S7 for parameter estimates). For the primary human hepatocyte model, we obtained five out of 60 L₁ parameters to be different from Huh7.5 cells. The gene‐specific parameters that were estimated to link the occupied binding sites predicted by the mathematical model to the respective target genes (Fig 4C) are listed in Table S8.