Assessing drug-induced liver injury liabilities in human and animal hepatocytes under static and perfused spheroid culture conditions

Abstract

Drug-induced liver injury (DILI) remains a major challenge in drug development, highlighting the need for reliable in vitro tools to assess hepatotoxicity and improve translation from animal to human studies. Models using hepatocytes from preclinical species are also needed to evaluate species-specific toxicity. In this study, we evaluated the basal function and DILI sensitivity of primary hepatocytes from human, monkey, rat, and dog cultured for up to 14 d in static monolayers, static spheroids, and spheroids cultured in a microfluidics-based microphysiological system (MPS). Hepatocyte function and injury responses were assessed using albumin, urea, and liver enzymes. Cells were exposed to species-specific DILI compounds chlorpromazine (CPZ), bosentan (BOS), and fialuridine (FIAU). Across platforms, human and monkey hepatocytes exhibited greater functional stability and sensitivity to DILI compounds than rat and dog hepatocytes. CPZ and BOS induced cytotoxicity primarily in human and monkey hepatocytes, while FIAU produced species-dependent effects consistent with known in vivo outcomes. The microfluidics-based MPS exhibited modestly improved hepatocyte spheroid function relative to static models, although limited MPS throughput constrained our ability for testing drugs beyond FIAU. Overall, these results demonstrate that integrating multi-species hepatocyte spheroids across static and microfluidic platforms enables comparative DILI assessment and supports improved preclinical-to-clinical translation. This study demonstrates that static primary hepatocyte spheroids provide a robust, scalable, and physiologically relevant platform for detecting species-specific drug-induced liver injury, with gene expression profiling offering stronger mechanistic insight than traditional functional biomarkers. By showing that perfused microphysiological systems may not substantially improve performance over static spheroids for the tested hepatocyte preparations, the work supports a pragmatic tiered testing strategy that balances throughput, cost, and biological complexity in preclinical safety assessments.

Preclinical drug safety evaluations require experimental systems that faithfully capture biological differences across species while providing mechanistic insight into human-relevant adverse effects, and this is especially noteworthy for drug-induced liver injury (DILI) (Olson et al. 2000; Ribeiro et al. 2019; Baudy et al. 2020; Marx et al. 2025). Unexpected liver toxicity findings during human clinical trials are costly and potentially preventable causes of late-stage candidate failures (Negi et al. 2022; Loewa et al. 2023; Mehta et al. 2025). Consequently, improving the translational power of preclinical models, including cell-based assays, remains an active area of investigation and an acute regulatory need (Avila et al. 2023). Cell-based liver models have long been used widely for studies of pharmacokinetics and liver injury (Soldatow et al. 2013). However, these models struggle to preserve hepatocyte function in culture and have many other limitations that include: Loss of liver lobule-like polarization, diminished metabolic capacity, and dedifferentiation occurring within hours to days of culture (Dunn et al. 1989; Berthiaume et al. 1996; Wang et al. 2004; Smith et al. 2012). These limitations restrict the utility of traditional in vitro liver models for studies of chronic or delayed-onset DILI. As a result, there is growing interest in advanced in vitro liver platforms such as spheroids and microfluidics-based microphysiological systems (MPS) that can recapitulate and maintain hepatocellular function, enable longer-term and repeated exposures, and reveal more accurate comparison of human-relevant drug responses (Bell et al. 2016; Luckert et al. 2017).

Liver MPS are becoming more widely accepted in drug development because they are designed to recapitulate physiologically relevant microenvironments. Liver MPS typically employ microfluidic perfusion to deliver nutrients and oxygen, remove waste, and support tissue-like gradients of shear stress factors that are believed to be essential for maintaining stable hepatocyte functionality and morphology (Domansky et al. 2010; Ewart et al. 2017; Ribeiro et al. 2019). Over the past decade, numerous liver-specific and multi-organ MPS platforms have been developed and commercialized, many of which use primary hepatocytes from human or preclinical species, or induced pluripotent stem cell (iPSC)-derived hepatocyte-like cells (Ewart et al. 2017; Baudy et al. 2020; Low et al. 2021). Integration of MPS into early safety assessment also presents an opportunity to enhance translation from preclinical species to humans while improving the mechanistic understanding of drug-induced injury responses (Jang et al. 2019; Negi et al. 2025).

One area of special interest in liver MPS is the integration of spheroids within microfluidic devices, combining the longevity and physiological relevance of microtissues with enhanced nutrient exchange and dynamic flow. Multiple studies have shown that hepatocyte spheroids preserve physiological CYP450 activity, albumin production, urea synthesis, and transporter function better than 2D monolayers (Ramaiahgari et al. 2014, 2017; Bell et al. 2017, 2018). Spheroids also have enhanced longevity (Scholten et al. 2026) which enables assessment of complex DILI phenotypes such as mitochondrial dysfunction, oxidative stress, cholestatic injury, and adaptive versus adverse responses—all key characteristics of hepatotoxic compounds (Rusyn et al. 2021)—over extended exposure periods. In addition to drug safety testing, liver spheroids are increasingly adopted in chemical toxicity studies, where long-term, low-dose exposures are critical for evaluating chronic effects of environmental pollutants and mixtures (Ramaiahgari et al. 2017; Forsythe et al. 2018; Seo et al. 2022; Negi et al. 2023). Comparisons of microfluidic MPS and static spheroids showed that both models exhibited broadly similar sensitivity to acetaminophen and fialuridine at relevant clinical concentrations, demonstrating their utility for improved prediction of hepatotoxicity in preclinical studies (Foster et al. 2019).

Several experimental platforms have been engineered to incorporate spheroids and microfluidics where tissue constructs can be housed in dedicated tissue chambers connected to media reservoirs via microchannels, enabling continuous recirculation (Cox et al. 2022). Such flow-based systems have shown improved long-term functionality compared with static cultures. Overall, the integration of advanced in vitro models, particularly spheroids and microfluidic MPS, into cross-species DILI assessment strategies has the potential to significantly improve early detection of species-specific toxic liabilities (Jang et al. 2019; Negi et al. 2025; Yen et al. 2026). Accordingly, this study aimed to characterize the phenotypic, functional, and transcriptional biomarkers of both basal and drug-treated cultures of primary hepatocytes derived from human, monkey, rat, and dog across multiple in vitro platforms. These platforms included (i) traditional static monolayer cultures, (ii) static spheroids generated using the ultralow attachment plates—Elplasia, and (iii) perfused spheroids cultured within the HUMIMIC Chip2 (TissUse, GmbH) MPS. We evaluated performance of each platform for up to 14 d by assessing viability and functional stability, as well as the platform-dependent differences in drug responses. In addition, we investigated species-specific DILI outcomes by simulating repeated exposures of the cultures with well-characterized hepatotoxicants, including fialuridine (FIAU), bosentan (BOS), and chlorpromazine (CPZ). Through this comparative, multi-platform translational approach, we sought to determine whether culture format and perfusion may influence hepatocyte phenotype and drug-induced injury profiles, thereby providing a comprehensive framework for cross-species evaluation in preclinical DILI assessment.

Materials and methods

Overall experimental design

The study was designed to investigate hepatocyte viability, function, and liver injury responses across multiple mammalian species using a range of cell culture platforms. The experimental study is summarized in Fig. 1. Experiments were performed using primary hepatocytes, all obtained from BioIVT (Baltimore, Maryland, United States), from human (Lot No. IIN), cynomolgus monkey (Lot No. VRX), Wistar rat (Lot No. JJU), and beagle dog (Lot No. JEO). The cell sources and vendors are listed in Table S1. Hepatocytes were cultured under 3 main in vitro conditions: (i) static 2D monolayer on collagen I-coated 96-well plates at a density of 70,000 cells/well, (ii) static spheroid (2.5D) model formed using ultra-low attachment platform Elplasia (Corning, Corning, New York) at 40,000 cells/well (∼500 cell/spheroid), and (iii) microfluidics-based (3D) spheroid-containing MPS using the HUMIMIC (TissUse GmbH, Berlin, Germany) platform, with cells seeded at 50,000 cells/chip (∼550 cells/spheroid). All cultures were maintained at 37 °C in a humidified atmosphere with 5% CO₂, and media was refreshed every 1 to 2 d. At each media change, aliquots were collected for subsequent biochemical assays. Culture initiation was defined as day 0, and experiments were conducted over 14 d, depending on the platform. Treatments commenced on day 6 of culture and were continued at each media replenishment every other day. In 2D and static spheroid platforms, all compounds were applied across a concentration range of 1 to 100 µM; however, cells cultured in MPS were treated only at 100 µM of FIAU due to lower throughput. Throughout the culture period, hepatocyte viability and functional capacity were monitored by quantifying albumin and urea secretion, in addition to liver injury biomarkers lactate dehydrogenase (LDH), and aspartate aminotransferase (AST) in the conditioned media.

Fig. 1.

Overall study design. A) Experimental design for comparative hepatotoxicity assessment in static monolayer (2D) cultures, static spheroids (2.5D Elplasia), and perfused spheroids in the TissUse MPS (3D-Humimic Chip2) platform. B) Primary hepatocytes from human, monkey, rat, and dog were tested in each platform. C) Experimental timelines for studies in different platforms. The day of hepatocyte seeding into each platform was designated as day 0. Drug treatment with chlorpromazine (CPZ), bosentan (BOS), and fialuridine (FIAU) began on day 6, and medium containing freshly prepared drugs (in designated concentrations) was replaced every other day. Samples (medium and cells) were collected at the indicated time points for further analyses.

Cell culture methods and drug treatments

Static monolayer model

Primary human hepatocytes were thawed Spheroid Spin Medium (Z990032) and initially cultured in INVITROGRO CP Spheroid Plating Medium (Z990033) supplemented with 0.4% (v/v) Spheroid Media Supplement A (Z990035) and 2.2% (v/v) TORPEDO Antibiotic Mix (Z990000). Cells were maintained in this plating medium for the first 6 d post-thaw. On day 6, the medium was replaced with complete INVITROGRO Spheroid Maintenance Medium (Z990034) containing 0.4% (v/v) Spheroid Media Supplement A (Z990035) and 2.2% (v/v) TORPEDO Antibiotic Mix (Z990000). Primary monkey, rat, and dog hepatocytes were thawed directly into their respective species-specific INVITROGRO CP media (monkey: Z990003; rat: Z990028; dog: Z990025). The composition of these media was proprietary and has been optimized by the manufacturer for each species (Table S2). For the 2D static culture model, hepatocytes were seeded at a density of 70,000 cells per well in collagen I-coated 96-well plates in the medium volume of 200 μl/well. Plates were pre-coated with 200 µg/ml collagen I (CB354249, Corning) according to the manufacturer’s instructions prior to cell seeding. For all culture conditions, the first medium change was performed 48 h after seeding, followed by 50% medium replacement every other day thereafter. For monkey and dog hepatocyte cultures, subsequent medium changes were performed using the corresponding species-specific INVITROGRO CP media supplemented with 2.2% (v/v) TORPEDO Antibiotic Mix (Z990000). For rat hepatocyte cultures, medium replenishment was performed using species-specific medium supplemented with 2.2% (v/v) TORPEDO Rodent Antibiotic Mix (Z990027).

Static spheroid model

For in-house spheroid generation, primary hepatocytes were seeded into ultra-low attachment Corning Elplasia 96-well plates (catalog no. 4442). Each well of the Elplasia plate contains 79 microwells; hepatocytes were seeded at a density of 500 cells per microwell, corresponding to approximately 40,000 cells per well in the medium volume of 250 μl/well. Following cell seeding, plates were centrifuged at 200× g for 2 min to facilitate rapid cell aggregation and uniform spheroid formation. Cultures were maintained at 37 °C in a humidified incubator with 5% CO₂. The first complete medium change was performed 48 h after seeding to remove residual debris and non-aggregated cells. Thereafter, 50% of the culture medium was replaced every other day using hepatocyte maintenance medium for the duration of the experiment.

Perfused spheroid model

HUMIMIC Chip2 devices were obtained from TissUse GmbH (Berlin, Germany). Each device has 2 independent chips, each consisting of 2 separate culture compartments and 3 integrated micropumps enabling dynamic fluid flow between compartments. Prior to cell seeding, both compartments were filled with 300 µl of hepatocyte-specific medium and pre-equilibrated in a humidified incubator (37 °C, 5% CO₂) for at least 3 d. The Chip2 circuits were connected to the HUMIMIC Starter control unit, which operated the pumps at an initial frequency of 0.5 Hz (30 beats per minute) with a pressure of +350 mbar and a vacuum of −300 mbar. On day 0, the HUMIMIC system was paused to allow for cell inoculation. Chips were transferred to a laminar flow hood, and the liver compartment was visually inspected under a brightfield microscope to confirm the correct positioning of the microwell insert. A total of 300 µl of pre-warmed medium was carefully added to each compartment, starting with the liver chamber, while keeping the pipette tip centered and close to, but not in contact with the microwell insert. After medium addition, the compartments were sequentially sealed with provided screw-on caps, beginning with the liver compartment, and the chips were returned to the incubator (37 °C, 5% CO₂) without connection to the pump system. Subsequently, 100 µl of medium was gently removed from the liver compartment, and 100 µl of a hepatocyte suspension (50,000 viable cells per chip) was added. The total medium volume in this model is 600 μl/device. Following cell seeding, the caps on each compartment were tightly sealed using a 7 mm hexagonal socket wrench, and the chips were centrifuged at 100× g for 2 min at room temperature to facilitate cell settling. Chips were then mounted onto the HUMIMIC holder, connected to the tube adapters, and placed into the incubator. The HUMIMIC Starter was programmed to resume flow at 0.5 Hz with a pressure/vacuum setting of ±200 mbar. From day 1 onward, daily media changes were performed by transferring each chip into a biological safety cabinet, unscrewing the reservoir compartment and replacing 300 µl from with fresh hepatocyte-specific medium. The pump frequency was increased to 0.75 Hz (45 BPM), and on day 2, pressure and vacuum were adjusted to ±300 mbar. These conditions were maintained for the remainder of the experiment. Medium changes continued daily thereafter for functional endpoint assessments.

On day 6 of culture, cells were exposed to test compounds CPZ, BOS, or FIAU or vehicle control (0.2% DMSO). For static spheroids, all 3 compounds were tested, whereas in the 3D microfluidic spheroid MPS platform only FIAU (100 µM) was evaluated due to the limitation of throughput—8 individual tests could be performed on one HUMIMIC Starter controller with 2 controllers operating simultaneously. Treatments were administered every other day through day 14. Stock solutions of each compound were prepared in DMSO, and working concentrations were generated by serial dilution, maintaining a final DMSO concentration of ≤0.2% (v/v); vehicle control cultures received 0.2% DMSO. For all culture formats, medium was replaced with drug-containing medium at 2× concentration to achieve a final 1× dose in the wells or chips, and cultures were re-dosed every other day until day 14. During each medium exchange, spent culture media samples were collected, centrifuged to remove cellular debris, and stored at −80 °C until further analysis.

Biochemical assays

Due to small-volume sample limitations of in vitro assays, selected biomarkers were examined based on our previous studies with complex in vitro models of the liver (Sakolish et al. 2021; Kato et al. 2022; Lim et al. 2023; Negi et al. 2025). Albumin was measured using the E88-129 kit from Bethyl Laboratories (Montgomery, Texas). Urea secretion was quantified using Stanbio Urea Nitrogen (BUN) Test (Cat No. SB-0580-250; Fisher Scientific, Waltham, Massachusetts). Lactate dehydrogenase (LDH) release and aspartate transaminase (AST) activity were measured using colorimetric methods. LDH assay kit (Ab102526) was purchased from Abcam (Waltham, Massachusetts). AST activity kit (Cat No. 701640) was obtained from Cayman Chemical Company (Ann Arbor, Michigan). All assays were performed according to the protocols provided by the respective vendors.

RNA isolation

Total RNA was isolated from hepatocytes cultured in monolayers, and spheroids cultured in static condition or in microfluidic MPS. For static monocultures and static spheroid cultures, RNA extraction was performed in the microwells after removal of the medium using TRIzol Reagent (Invitrogen, Carlsbad, California) followed by purification with the Direct-zol RNA Miniprep Kit (Zymo Research, Irvine, California) according to the manufacturer’s instructions. For perfused spheroid culture experiments, the sealed compartment was gently opened and both apical and basal media were aspirated. Then, the chip’s microwell inserts were carefully transferred into sterile 1.5 ml microcentrifuge tubes. TRIzol reagent (300 µl) was added directly to each insert and incubated at room temperature for 10 min. During incubation, samples were vortexed briefly every 5 min to ensure complete cell lysis. Inserts were then removed from the tubes, and an equal volume of 100% molecular-grade ethanol was added to the lysates. The mixture was transferred to Zymo-Spin IICR columns, and RNA purification was completed according to the Direct-zol manufacturer’s protocol. For all samples, RNA was eluted in 50 µl of DNase/RNase-free water, and concentrations were measured using the NanoDrop 8000 spectrophotometer (Fisher Scientific). All RNA samples were stored at −80 °C until library preparation and RNA sequencing.

Gene expression library preparation and sequencing

Total RNA samples were submitted for library preparation and sequencing to the Texas A&M AgriLife Genomics and Bioinformatics. Quality control was performed on a representative subset of samples (n = 11 per plate, randomly selected). RNA purity was determined by measuring absorbance at 230 to 280 nm using a microvolume DropletQuant spectrophotometer (PerkinElmer, Waltham, Massachusetts). RNA integrity was assessed via capillary electrophoresis on an Advanced Analytical 5200 Fragment Analyzer System (Agilent, Santa Clara, California). Total RNA concentration was quantified using absorbance at 260 nm on a DropletQuant Spectrophotometer. mRNA was isolated from 150 ng of total RNA using a Nextflex Poly-A Selection kit. Unique dual-indexed cDNA libraries were constructed using the Nextflex Rapid Directional RNA 2.0 kit on a Sciclone NGSx automated liquid handling platform, with reactions scaled to 2/5 of the manufacturer’s standard volume. PCR cycle optimization was performed on a pooled library sample using real-time PCR, with 14 cycles determined as optimal to reach plateau phase. The library concentration was determined by using a fluorometric quantification assay using Molecular Devices Spectramax M2 fluorescent plate reader and Thermo Fisher Scientific Picogreen Assay. Libraries were then pooled by equal mass using a Janus Mini liquid handler. The pooled library was sequenced on an Illumina NovaSeq X + sequencer (1.5B flowcell) using a 2-channel configuration with paired-end reads of 2 × 150 bp. Sequence cluster identification, quality prefiltering, base calling, and uncertainty assessment were done in real time using Illumina’s NCS 1.0.2 and RFV 1.0.2 software with default parameter settings. Sequencer base call files (BCL) were demultiplexed and formatted into FASTQ files using bcl2fastq 2 2.19.0, yielding 1.9 billion demultiplexed clusters, with an average of 20 million clusters per sample. Data quality was evaluated using FastQC 0.11.9 and MultiQC v1.9. The FASTQ files were uploaded to Gene Expression Omnibus (GEO; accession number GSE333420).

Transcriptomic data analysis

The data analysis workflow was adapted from previous study (Negi et al. 2025). Specifically, single-end FASTQ files (R1) were processed using fastp (version 0.23.4) (Chen et al. 2018). Preprocessing steps included trimming of homopolymer sequences and quality-based filtering using a Phred score threshold of 20. Reads from dog, monkey, human, and rat were aligned to the ROS_Cfam_1.0, Macaca_fascicularis_6.0, GRCh38.p14, and mRatBN7.2 genome assemblies, respectively, using STAR (version 2.7.11b) (Dobin et al. 2013). The resulting alignments were sorted using SAMtools (version 1.19.2) (Danecek et al. 2021) and gene-level counts were quantified using HTSeq (version 2.0.5) (Anders et al. 2015). Quality control filtering was applied to remove low-quality samples and transcripts. Samples with fewer than 1 million total reads or with less than 50% uniquely mapped reads were excluded. This resulted in the removal of 13 samples, comprising 12 rat samples (across FIAU treatment and control groups in both static monolayers and static spheroid platforms; n = 3 per group) and one dog sample (day 4, static monolayers). The remaining samples were examined for outliers using principal component analysis (PCA), and no further exclusions were deemed necessary. Additionally, transcripts not detected across all samples within each species were excluded prior to downstream analysis. Normalization of count data within each species was carried out using the Bioconductor package DESeq2 (version 1.46.0) (Love et al. 2014).

For cross-species comparisons, human Ensembl gene identifiers were converted to HUGO Gene Nomenclature Committee (HGNC) gene symbols, and nonhuman Ensembl identifiers were mapped to corresponding human orthologs using the Bioconductor package biomaRt (version 2.62.1) (Durinck et al. 2009). Temporal patterns in gene expression were explored by generating a heatmap of normalized counts for all genes with identified human orthologs (n = 12,283) using the pheatmap package (version 1.0.13) with default clustering and distance metrics. Trends over time were further assessed for selected Phase I, II, and III enzyme genes using linear modeling (lm() in R). Trend directionality was recorded, and statistical significance was determined using false discovery rate (FDR)-adjusted P-values with a cutoff of <0.05. To characterize treatment effects, gene expression in FIAU-treated samples was compared with matched controls within each species and platforms using the DESeq2 package (Love et al. 2014). Log2 fold change estimates were shrunk using the lfcShrink() function (type = “apeglm”). Differentially expressed genes (DEGs) were defined as those met thresholds of FDR-adjusted q-value <0.05 and absolute log2 fold change > 1.5. Pathway enrichment analysis was performed on DEGs to investigate biological mechanisms underlying treatment effects. Enrichment of Reactome and WikiPathways pathways was conducted using the ReactomePA (version 1.50.0) (Yu and He 2016) and clusterProfiler (version 4.14.6) (Yu et al. 2012) R packages, respectively. To further support hazard characterization, a gene set enrichment workflow based on the Key Characteristics framework (Tsai et al. 2025) was applied. For all pathway analyses, the background gene set consisted of the normalized genes retained after low-count filtering within each species.

Statistical analysis and data availability

Statistical analyses were performed using GraphPad Prism version 9.2 (GraphPad Software, San Diego, California, United States). Comparisons among groups were conducted using one-way analysis of variance (ANOVA) followed by Tukey’s multiple comparisons test. Where applicable, 2-way ANOVA followed by Šidák’s multiple comparisons test was used. Data are presented as mean ± SD or as heatmaps, as indicated in the corresponding figures. A P-value < 0.05 was considered statistically significant. Significant differences are indicated in the figures by a single asterisk (*), regardless of the exact P-value threshold below 0.05.

All raw data for the experiments presented herein have been uploaded to EveAnalytics website (see Table S3 for the web link). Gene expression data have been uploaded to Gene Expression Omnibus (GEO; accession number GSE333420).

Results

Characterization of species-specific hepatocyte function in static monolayers, static spheroids, and perfused spheroids

Primary hepatocytes from human, monkey, rat, and dog livers were cultured (Fig. 1C) in static monolayers, static spheroids (Elplasia), and perfused spheroids (TissUse MPS). Representative photomicrographs illustrating morphology of the spheroids in static and perfused cultures are shown in Fig. 2. Human hepatocytes used in these studies were “spheroid-qualified” by the vendor; other species’ hepatocytes were not assessed for spheroid formation by the vendor. Distinct species-dependent differences in microtissue formation were observed. Human hepatocytes consistently formed compact, well-defined spheroids in both models. These tissue constructs displayed uniform size distribution and tight cellular organization by day 6, with sustained structural integrity through day 14. It is worth noting that due to their high degree of compaction, primary human hepatocyte spheroids occasionally dislodged from the microcavities during media exchanges in static cultures and when the hepatocyte chamber in microfluidic MPS was opened at the end of the experiment on day 14. In contrast to primary human hepatocytes, hepatocytes from monkey, rat, and dog did not form compact spheroids in static culture conditions. The cells formed loose cellular aggregates that remained contained within the microcavities throughout the 14-d culture period. Notably, both monkey and rat hepatocytes formed more compact, but not spheroid-shaped aggregates in microfluidic MPS, suggesting that perfusion and microenvironment may influence microtissue organization. Dog hepatocyte aggregates in microfluidic MPS were more organized than in static cultures but were not as compact as cells from other species. These findings demonstrate that hepatocyte microtissue formation is highly species- and platform-dependent, highlighting the importance of careful model selection when designing physiologically relevant in vitro liver systems for preclinical drug testing.

Fig. 2.

Representative bright-field images of hepatocytes in the static (2.5D-Elplasia) culture and microfluidic (3D-Humimic Chip2) MPS from day 6 A) and day 14 B). In 2.5D-Elplasia, there are 79 microcavities in each well; ∼500 cells are expected to sediment into each microcavity. In each Humimic Chip2 MPS device, there are 91 microcavities; ∼550 cells are expected to sediment into each microcavity.

To evaluate hepatocyte functionality across the 3 culture modalities, liver function (secreted albumin and urea) and liver injury (LDH and AST) markers were measured every other day for up to 14 d. Figure 3 shows the time course dynamics in these biomarkers normalized to the initial cell seeding density. Overall, species-dependent differences were observed in both initial activity and the extent to which hepatocyte function was maintained over time. The most robust and stable albumin secretion was detected in primary human hepatocytes across all platforms. Among the tested platforms, the highest albumin levels were observed in microfluidic MPS and the lowest in static monolayer cultures; however, the differences among platforms were minor. Notably, albumin secretion by primary human hepatocytes increased or stabilized over the course of the 14-d culture period, indicating effective adaptation and maintenance of hepatocyte function under the long-term culture conditions. Monkey, rat, and dog hepatocytes displayed slightly higher albumin production in the microfluidic MPS as compared with static spheroids or monolayer cultures. Rat hepatocytes showed the most pronounced early platform-dependent differences with the highest albumin levels in the microfluidic MPS, and the lowest in static monolayer cultures. In contrast to human hepatocytes, albumin secretion in animal hepatocytes cultures gradually declined over time in all platforms, although the rate of decline varied by species and culture system. Urea production generally followed the trends observed for albumin (Fig. 3). All tested hepatocytes showed modest declines in urea levels over time. Despite these temporal changes, urea levels across species were relatively comparable, suggesting that basic metabolic activity was preserved in all systems. Markers of cellular injury, including AST and LDH, showed generally similar trends across species and platforms (Fig. 3). They subsided over time in culture with the largest drop found in microfluidic MPS and the rise in LDH release observed in primary human hepatocytes past day 6 in monolayer cultures. These results demonstrate that human hepatocytes exhibit superior and sustained liver function across all culture platforms, while cells from non-human species show progressive functional decline, collectively highlighting both species- and platform-dependent differences in hepatocyte performance that must be considered when selecting in vitro models for preclinical drug evaluation.

Fig. 3.

Comparative analysis of the time course in the liver function markers (albumin and urea) and cell injury-related markers (LDH and AST) across species and culture platforms. Data is organized by species (columns) and biomarker (rows). Each plot contains data from 3 culture platforms (see inset in the top right corner)—squares are monolayer cultures, triangles are static spheroids, and circles are perfused spheroids in MPS.

Temporal dynamics of species-specific hepatocyte gene expression profiles in static monolayers, static spheroids, and perfused spheroids

Gene expression profiling is valuable for determining whether hepatocytes maintain their liver-like phenotype in culture, especially across culture platforms (Baudy et al. 2020; Homan 2023). Accordingly, we evaluated the temporal patterns in gene expression in primary hepatocytes from 4 species cultured in 3 different platforms over 14 d (Fig. 4). All expressed genes in primary human hepatocytes (n = 12,283) were clustered and arranged by days in culture (day 0 represents cells immediately post thawing, and days 4 and 14 were collected from untreated samples) and platform (monolayers, static spheroids, and perfused spheroids in MPS) (Fig. 4A). Transcripts in other species were matched to their human orthologs and then ordered in the same way as the genes selected from primary human hepatocyte data. The most prominent difference among samples was between day 0 (cells before plating) and subsequent culture days as expression of many genes was either up- or down-regulated after plating in a consistent manner in all species. The only exception to this was primary human hepatocyte gene expression data from day 4 cultures in static spheroids—most of the gene expression was similar to that on day 0. However, by day 14 the expression levels across the entire transcriptome were similar to those in monolayer cultures at either days 4 or 14. It should be noted that because of the low throughput, we did not examine day 4 cultures in perfused spheroids. Still, day 14 gene expression was similar in all culture platforms.

Fig. 4.

Gene expression analysis of hepatocytes cultured as monolayers (96-well plate, 2D), static spheroids in Elplasia (2.5D), and perfused spheroids in MPS (3D). A) Heatmap of normalized gene expression counts for all overlapping genes with human orthologs from day 0 (pre-culture baseline) through post-plating time points (days 4 and 14). B) Temporal expression patterns of ADME-related genes, including cytochrome P450 monooxygenases (Phase I metabolism), phase II conjugation enzymes, and drug transporters. Genes with statistically significant temporal trends are color-coded—dark purple indicates upregulation (positive slope), orange indicates downregulation (negative slope). Gray denotes genes for which no human ortholog was identified in non-human species; white indicates genes with no statistically significant differential expression over time.

Next, we examined more closely the expression of representative ADME genes (Fig. 4B). Similar patterns to the whole transcriptome were observed—many of these genes changed in expression by day 4 in monolayer cultures. The trends for each transcript are shown in the right-hand panel of Fig. 4B. The expression of most ADME genes trended downward over time, particularly in rat and dog hepatocytes. Several genes maintained their expression levels or were upregulated, suggesting the preservation of some metabolic capacity in culture over 14 d. Among Phase I metabolism transcripts, expression of CYP2A6, CYP2B6, CYP2C8, and CYP2E1 was significantly downregulated over 14 d while CYP1A2, CYP1A2, CYP1B1, and ALDH3A1 showed increased or maintained expression. Among Phase II enzymes that demonstrated variable temporal trends, GSTM3 and GSTP1 showed increase in expression, while SULT family members exhibited predominantly decreasing expression over time. Among hepatic drug transporters, ATP-binding cassette subfamily B members (ABCB) demonstrated increased expression, while solute carrier organic anion transporter (SLCO) family members exhibited predominantly downregulated expression. Beyond the clear difference in expression patterns in monolayer cultures, few platform-specific differences were observed between static and fluidic spheroid cultures. Collectively, these results show that while culture platform and time in culture influence hepatocyte gene expression, the most pronounced transcriptomic shift occurs upon plating regardless of species origin, with selective preservation of key ADME genes highlighting both the utility and limitations of in vitro systems.

Assessment of species-specific DILI across different culture platforms

We evaluated the hepatotoxic effects of several model compounds across 3 culture platforms employed in this study. To test the ability of each model to demonstrate known in vivo drug toxicity in humans and demonstrate known species similarities and differences in drug-specific effects, we used CPZ, BOS, and FIAU. CPZ is a commonly prescribed antipsychotic that has known risk of hepatotoxicity, adverse event that may arise through combined cytotoxic and cholestatic mechanisms and has been observed in both humans and preclinical species (Larrey and Ripault 2013; Mueller et al. 2015). BOS is an endothelin receptor antagonist that is known to induce liver injury in humans via cholestasis because of inhibition of the bile salt export pump (Fattinger et al. 2001; Leslie et al. 2007). However, comparable toxicity was not observed in rodent studies even under conditions of transporter inhibition, and only variable effects were reported in dogs at exposures approximately 10× the C_max (US FDA 2001). FIAU is an investigational nucleoside analog that showed no hepatic toxicity in preclinical studies but caused fatal liver injury during clinical trials (Institute of Medicine 1995; McKenzie et al. 1995). FIAU-associated hepatotoxicity in the clinical trial was characterized as severe, delayed, and often fatal—15 patients were enrolled of which 7 developed severe toxicity after taking FIAU for at least 9 wk (5 died and 2 survived only after liver transplantation) (Institute of Medicine 1995). Hepatotoxicity of FIAU did not follow typical liver injury pattern: Patients developed hepatic failure with only mild jaundice and minimal changes in serum enzymes, plus rapid lactic acidosis, pancreatitis, microvesicular steatosis, and signs of myopathy and/or peripheral neuropathy. Hepatotoxic potential of FIAU is believed to be linked to mitochondrial DNA depletion, steatosis, and cholestasis (Honkoop et al. 1997; Kleiner et al. 1997), and this compound is often used to test species differences.

Accordingly, to assess species specificity and compare model performance, we used chlorpromazine (CPZ; human C_max= 0.5 μM), bosentan (BOS; human C_max= 7.4 μM), and fialuridine (FIAU; human C_max= 1 μM) as case studies. In static monolayer and static spheroid culture platforms, cells were exposed to these drugs in concentration-response: CPZ (1, 10, 30 μM), BOS (1, 10, 30, 200 μM), and FIAU (1, 10, 30, 100 μM). Due to limited throughput in the microfluidic perfused spheroid MPS platform—only 16 total conditions could be examined—experiments were conducted only with vehicle and FIAU at 100 μM. In all platforms, cells were dosed repeatedly—every other day starting on day 6 of culture and continuing until day 12 for a total of 4 doses (Fig. 1C). A broad set of endpoints was evaluated, including time-course measurements of liver injury biomarkers (AST and LDH) and hepatocyte functional marker (albumin).

Figure 5 shows the dose–response, species-, and time-dependent effects of CPZ, BOS, and FIAU in static monolayer cultures. CPZ is known to be hepatotoxic in both humans and rats by a mixed cytotoxic and cholestatic mechanisms (Mueller et al. 2015). CPZ induced clear toxicity across all biomarkers in primary human hepatocytes. Liver injury markers (AST and LDH) were elevated even at the lowest tested concentration (10 μM, 20× C_max), and albumin secretion was reduced at all concentrations. In primary monkey hepatocytes, treatment with CPZ resulted in an early release of AST and LDH, which declined with repeated dosing, consistent with cytotoxicity. Albumin levels decreased significantly after the second dose across all concentrations. In primary rat hepatocytes, responses were more variable. LDH showed no significant changes, whereas AST displayed a declining trend with consecutive doses. Albumin also declined after the third dose, though these changes were not statistically significant. In primary dog hepatocytes, no significant changes in LDH were observed. AST increased during the initial days of treatment, when compared with controls, but declined with repeated dosing, consistent with cytotoxic effects. Albumin levels were initially elevated at the highest concentration (100 μM) but leveled-off with continued treatment. Overall, these data highlight toxic effects of CPZ in all tested species, with primary human hepatocytes showing the most consistent and pronounced responses across all biomarkers.

Fig. 5.

Assessment of known human DILI-causing drugs on hepatocyte function and injury markers in monolayer cultures of cross-species hepatocytes. Hepatocyte monolayers were exposed for up to 8 d to a range of concentrations of 3 reference DILI compounds: CPZ, BOS, and FIAU. Cellular injury was assessed by quantifying the release of clinical liver injury biomarkers AST and LDH in the culture medium. Hepatic function was evaluated by measuring secreted albumin levels. Data are presented as heatmaps representing the log2 fold change in biomarker levels relative to day-matched vehicle-treated controls. Columns are arranged by compound, day, and concentration. Rows represent different species. Red shading indicates an increase, and blue indicates a decrease relative to vehicle levels. Each condition was tested in triplicate (n = 3) and the average values are shown. Significance relative to vehicle-treated controls was assessed using one-way ANOVA and the asterisks (*) denote statistically significant (P < 0.05) differences.

BOS causes cholestatic liver injury in humans but not in rats (Leslie et al. 2007), and some liver toxic effects have been reported in the dog study (US FDA 2001). We found that with BOS treatment, LDH was significantly increased in human and monkey hepatocytes, whereas AST was elevated only in human hepatocytes. Albumin levels were significantly decreased in both human and monkey hepatocytes. In rat hepatocytes, no significant changes were observed in LDH or AST. Albumin levels showed a significant increase on day 6, although the effects were variable over time. In dog hepatocytes, LDH was not affected, while AST increased during the initial days of treatment but returned to baseline with continued dosing. Albumin levels were unaffected, although biphasic and variable effects were observed over the course of the experiment in both rat and dog hepatocytes. Overall, BOS exhibited pronounced hepatotoxicity effects in human and monkey hepatocytes, while rat and dog hepatocytes exhibited more variable responses, particularly in functional markers.

FIAU was without liver toxicity findings in preclinical studies but fatal hepatotoxicity was observed in 5 subjects during phase II clinical trial (Institute of Medicine 1995). Its mechanisms of hepatotoxicity are thought to be dependent on mitochondrial dysfunction followed by steatosis and cholestasis (Kleiner et al. 1997). In studies with FIAU, LDH, and AST were increased in human hepatocytes in a time- and concentration-dependent manner. Concomitantly, albumin levels were significantly decreased across the same conditions. In primary hepatocytes from all preclinical species tested (monkey, rat, and dog), no consistent effects on LDH or AST were observed; however, variable and occasionally biphasic responses were noted. Albumin in monkey hepatocytes was significantly increased, while in rat hepatocytes, albumin tended to increase at higher concentrations. In dog hepatocytes, no significant changes in albumin were observed. Overall, these results in traditional 2D monolayer cultures demonstrate that FIAU-induced hepatotoxicity was most pronounced in human hepatocytes, with animal hepatocytes showed variable responses, particularly in functional markers.

Figure 6 shows the data for static spheroids cultured in Elplasia platform. CPZ treatment increased LDH and AST levels in human hepatocytes; however, with repeated dosing, these effects became less pronounced, and by day 8 only the lowest concentration (10 μM) maintained an increase in LDH and AST. Albumin secretion decreased in a concentration-dependent manner, consistent with the observed toxicity. Monkey hepatocytes exhibited a similar pattern, with increases in LDH and AST and a concentration-dependent decrease in albumin. In rat hepatocytes, LDH and AST levels decreased rather than increased, which likely reflects overt cytotoxicity and loss of viable cells rather than a lack of toxicity. Albumin secretion showed a declining trend but was not statistically significant. Dog hepatocytes demonstrated a similar profile, with decreases in LDH and AST consistent with overt cytotoxicity. Albumin showed an initial decrease early in treatment but no clear effects with continued dosing.

Fig. 6.

Assessment of known human DILI-causing drugs on hepatocyte function and injury markers in static spheroid cultures (Elplasia). Hepatocyte spheroids were exposed for 8 d to a range of concentrations of 3 reference DILI compounds: CPZ, BOS, and FIAU. Cellular injury was assessed by quantifying the release of clinical liver injury biomarkers—AST, and LDH in the culture medium. Hepatic function was evaluated by measuring secreted albumin levels. Data are presented as heatmaps representing the log2 fold change in biomarker levels relative to day-matched vehicle-treated controls. Columns are arranged by compound, day, and concentration. Rows represent different species. Red shading indicates an increase, and blue indicates a decrease relative to vehicle levels. Each condition was tested in triplicate (n = 3) and the average values are shown. Significance relative to vehicle-treated controls was assessed using one-way ANOVA and the asterisks (*) denote statistically significant (P < 0.05) differences.

For BOS treatment, LDH increased significantly in human hepatocytes only at the highest dose on day 8. Early increases were observed but did not reach statistical significance. Albumin was significantly decreased only at the highest concentration. Monkey hepatocytes showed similar effects, with a concentration-dependent increase in LDH, no change in AST, and a significant concentration-dependent decrease in albumin. Rat hepatocytes showed no major changes in LDH or AST under BOS exposure, although slight decreases in these toxicity markers were observed, again suggesting potential overt cytotoxicity rather than lack of response. Albumin levels increased at lower concentrations but showed no change at higher concentrations. In dog hepatocytes, LDH and AST decreased rather than increased, with a slight increase in LDH at the lowest BOS concentration (10 μM). Albumin decreased significantly early in treatment (day 2), followed by a biphasic pattern at later time points.

For FIAU, LDH and AST showed no significant changes in human hepatocytes; however, albumin levels were significantly decreased in a dose-dependent manner, with effects becoming more pronounced with repeated dosing. Monkey hepatocytes also showed no changes in LDH or AST, but albumin increased on days 6 and 8, which may indicate a stress response. Rat hepatocytes showed no notable changes in LDH, AST, or albumin. Dog hepatocytes exhibited variable LDH responses, generally showing decreases rather than increases, consistent with loss of viable cells. Albumin levels showed no major changes, except for an increase at the highest concentration (200 μM) on day 8. Collectively, the data in Elplasia platform demonstrates species-dependent differences in hepatotoxic responses, with human hepatocytes showing the most robust and interpretable toxicity profiles across all endpoints.

Figure 7 shows the results of the experiments with FIAU in perfused spheroids in MPS. FIAU was tested only at 100 µM due to the platform’s limited throughput. Across cells from all tested species, most biomarkers showed relative downward change as compared with their day-matched vehicle controls, which likely reflects reduced cellular activity and potential early cytotoxicity; however, these trends could not be confirmed statistically because of the low sample size and high variability within the platform. Human hepatocytes exhibited a modest declining trend after FIAU treatment, but this did not reach significance for the same reasons. In the other species, no clear or consistent effects were observed for any biomarkers. These data in perfused MPS platform show the current limitations in sensitivity and chip-to-chip variability of complex in vitro models.

Fig. 7.

Assessment of FIAU 100 µM on perfused hepatocyte spheroid function and injury markers. Hepatocyte spheroids were cultured in Humimic Chip2 and exposed for 8 d to FIAU (100 µM). Cellular injury was assessed by quantifying the release of clinical liver injury biomarkers—AST, and LDH in the culture medium. Hepatic function was evaluated by measuring secreted albumin levels. Data are presented as heatmaps representing the log2 fold change in biomarker levels relative to day-matched vehicle (0.2% DMSO)-treated controls. Columns are arranged by compound, day, and concentration. Rows represent different species. Red shading indicates an increase, and blue indicates a decrease relative to vehicle levels. Each condition was tested in duplicate (n = 2) and the average values are shown.

Assessment of species-specific gene expression changes in response to FIAU across different culture platforms

We also examined the effects of FIAU (100 μM treatments over 4 consecutive rounds from days 6 to 12, Fig. 1C) on gene expression profiles in primary hepatocytes across species and platforms (Fig. 8). First, we used PCA to plot individual samples within each species (Fig. 8A). Excellent consistency was observed between biological replicates for all species and conditions, indicated by close overlap of the individual samples in all 4 PCA plots. In primary human hepatocytes, we found that after 4 d in culture (no treatment), gene expression profiles in cells cultured in monolayers were more distinct from those in the pre-plating cells (day 0) than in cells cultured in static spheroids. Because microfluidic MPS could not be opened until the end of the experiment, no day 4 samples were available. However, by day 14, gene expression profiles in untreated cells cultured in different culture systems have largely converged—separating in the first principal component. Gene expression profiles of cells treated with FIAU (day 14) in static monolayers and static spheroids showed the largest difference from vehicle-treated cells; the difference in microfluidic spheroids in MPS platform was relatively minor. In primary monkey hepatocytes, the largest difference was observed between days in culture with no effect of FIAU and only small differences between culture conditions. The same was also observed in primary rat hepatocytes, albeit FIAU effects could be examined only in microfluidic spheroids in MPS platform due to sample loss from other experiments. In primary dog hepatocytes, in addition to the separation by day of culture, there was also a notable difference by platform, but little effect of FIAU treatment.

Fig. 8.

Treatment effects on hepatocyte gene expression induced by FIAU across species. A) Principal component (PC) analysis of the individual replicate samples is displayed separately for each species. Platforms are shown in colors and days of culture and treatment conditions are indicated next to each cluster. B) Venn diagrams illustrating the overlap of differentially expressed orthologs for each platform (left) and species (right). A total number of genes is shown in brackets and the numbers inside the Venn diagram show overlap or unique genes. C) Summary of pathway analysis of species-specific gene expression changes in response to FIAU in static monolayers (2D), static spheroids (2.5D), and perfused spheroids in MPS (3D). Pathway analysis was performed as detailed in Methods using Key Characteristic Gene Sets (ORA = over-representation analysis, GSEA = gene set enrichment analysis), Reactome, and Wiki Pathways. A dot indicates that the pathway was significantly enriched (q-value <0.05). Gray shading indicates conditions that were not examined using gene expression because no samples were available for analysis. Full listing of affected pathways is included in Table S4.

We then examined differentially expressed genes (DEGs; absolute log₂ fold change >1.5 and false discovery rate (FDR) < 0.05) in each species comparing the effects of FIAU (100 μM) to vehicle-treated samples on day 14 across species and tested platforms (Fig. 8B). When we compared species effects within each platform (left column), we found that the largest effect of FIAU on gene expression was observed in primary human hepatocytes cultured in static monolayers (1,705 DEGs) and static spheroids (1,724 DEGs); in the perfused spheroids in 3D MPS, the effect in human cells was actually far smaller than in rat or dog hepatocytes. Overlap in DEGs among species was modest—about 1/3 of the DEGs affected by FIAU in monkey and dog hepatocytes were common to those in human hepatocytes. When we examined platform effects in each species (right column), we found that 2/3 of the DEGs were conserved between static monolayer and static spheroid cultures; all but one DEGs from perfused spheroids in MPS were also in common with 2 other platforms. For monkey and rat hepatocytes, the overlap was much smaller; however, the overall transcriptional response to FIAU was muted as compared with that in human hepatocytes. These findings demonstrate that FIAU elicits species-specific transcriptional responses, with human hepatocytes showing the greatest sensitivity across species. The high concordance between static monolayer and static spheroid platforms suggests that static culture conditions adequately capture FIAU-induced gene expression changes, while the attenuated response in perfused spheroids in MPS may reflect either protective effects of media flow, altered drug exposure kinetics, or technical variability due and low throughput. The lack of robust transcriptional responses in non-human species aligns with lower effects on the function and injury biomarkers (Figs 5 to 7), supporting the conclusion that FIAU toxicity is predominantly human-specific phenomenon at both functional and gene expression levels.

Lastly, we investigated the biological pathways underlying FIAU-induced transcriptional changes. Figure 8C shows the results from pathway enrichment analysis using 3 complementary approaches: Key Characteristics (Tsai et al. 2025), Reactome (Jassal et al. 2020), and WikiPathways (Kelder et al. 2012). Data are organized by principal toxicity pathway, then species (rows) and by pathway analysis approach and culture platform (columns) and significant enrichment (q-value <0.05) is indicated by a black dot. Overall, FIAU effects were detected on pathways or key characteristics that include xenobiotic metabolism and transport, lipid-related metabolism, fibrosis, and inflammation. The significant effects were largely concordant among pathway enrichment approaches insofar the significant effects were largely constrained to the effects of FIAU on primary human hepatocytes. In addition, the pathway-level effects were concordant between static monolayer and static spheroid platforms. In contrast, the perfused spheroids in MPS platform yielded no significantly enriched pathways except for fibrosis and lipid/bile acid metabolism, potentially reflecting the reduced transcriptional response observed in this system and/or insufficient statistical power due to limited sample size. Collectively, gene expression results showed that FIAU elicits a predominantly human-specific transcriptional response involving effects on xenobiotic metabolism, lipid metabolism, and inflammatory pathways, with strong concordance between static culture platforms (both monolayers and spheroids) but an attenuated response in perfused spheroids.

Discussion

Legislative, ethical, scientific, financial, and other drivers (Han 2023; US FDA 2025, 2026) are facilitating rapid development of complex in vitro models and MPS—experimental platforms that aim to augment drug and chemical safety evaluations, and complement or replace animal tests where they lack translational relevance (i.e. development of monoclonal antibody-based therapeutics) (Kleinschmidt-Doerr et al. 2026; Wu et al. 2026). Among the organs that are of primary focus for such advances is the liver (Baudy et al. 2020; Kang et al. 2024), and specifically DILI which remains one of the leading causes of late-stage attrition in drug development and post-market withdrawals. Because of the attention that in vitro liver models have received over the decades (Soldatow et al. 2013), there are many options that include different culture platforms, cell types, and species. The aspirations to evolve these types of models into more physiologically relevant assays often entails addition of microfluidic conditions, yet such advanced models are more challenging to operate and scientific justification is necessary for adding complexity to existing investigative and pre-clinical safety pipelines (Devine et al. 2025). This study directly addressed whether continuous perfusion enhances hepatocyte spheroid performance beyond static 3D culture. Specifically, our study aimed to address several key points: (i) functional performance and longevity across platforms, (ii) species-specific responses to known DILI compounds (CPZ, BOS, and FIAU) using hepatocytes from rat, dog, monkey, and human donors, and (iii) gene expression changes in different species and models over time and in response to a mitochondrion toxicant FIAU.

Comparative platform performance and the role of perfusion

This study directly addressed whether continuous perfusion enhances hepatocyte spheroid performance beyond static spheroid cultures. Several previous studies have shown that perfusion does improve the long-term maintenance of the hepatocytes in culture (Rubiano et al. 2021; Lim et al. 2023) and also improves prediction of DILI (Ewart et al. 2022). Surprisingly, we found that perfusion provided minimal functional improvement over static spheroids made from primary hepatocytes, even though previous studies in this platform showed advantages of culturing HepaRG cells in a spheroid format under perfusion (Materne et al. 2015). While albumin and urea production were sustained in the microfluidic MPS platform, levels were comparable to static spheroids, suggesting that well-formed spheroids already achieve sufficient cell-cell contact and matrix interactions to maintain differentiated function without dynamic flow, results that are consistent with findings in static liver spheroids (Cox et al. 2020). While our result may be donor-specific as we have previously showed that primary human hepatocytes from another donor perform well in different MPS models than in static monolayers (Negi et al. 2025), this finding has practical implications. Specifically, perfused systems introduce substantial technical complexity, reduce throughput, and increase costs. Therefore, their advantages over simpler static spheroid culture modalities may need to be tested for each lot of cells to evaluate proportional gains in hepatocyte functionality. Indeed, the human hepatocytes used in our study also showed higher albumin levels compared with previous studies, reflecting donor-specific variability, a persistent challenge when using primary cells. Across all platforms, both biomarkers declined over time, consistent with gradual loss of function during extended culture. Similarly, hepatocyte monolayers of this particular donor lot remained surprisingly functional for 14 d, though with greater declining function as compared with both spheroid platforms.

Species-specific toxicity detection

In certain instances, compounds associated with DILI in humans, either in clinical trials or in post-marketing surveillance, produce no clear signs of toxicity in preclinical animal testing, illustrating the possible mismatch between preclinical outcomes and clinical observations (Olson et al. 2000). Complex liver in vitro models have been proposed as tools to guide the selection of animal safety models, as they can mimic species-specific differences using liver cells from various preclinical animal species (Zhou et al. 2019). They can be used to select appropriate compound-specific pre-clinical species for general repeat-dose toxicity tests that must be conducted in 2 mammalian species, typically one rodent (e.g. rat) and one non-rodent (e.g. dog or non-human primate) (Prior et al. 2020). Accordingly, 3 hepatotoxic compounds: CPZ, BOS, and FIAU were tested in this study across species and culture platforms. Concentration-response relationships were evaluated in static monolayers and static spheroids; however, in the perfused MPS platform, only FIAU was tested due to the limited throughput.

Interestingly, static monolayers were more sensitive in detecting drug-induced cytotoxicity when compared with static spheroids. This increased sensitivity may result from higher drug exposure per cell due to the larger surface area-to-volume ratio in monolayers. However, this hypothesis was not evaluated in our study. We also found that static spheroids most accurately captured expected species-specific toxicity profiles (Table 1). CPZ showed effects across all species and endpoints, as expected (Greer et al. 2010; Broeders et al. 2015). BOS was least hepatotoxic in the rat, also similar to the result in previous reports (Leslie et al. 2007). For FIAU, which causes severe human mitochondrial toxicity but is less toxic in rodents due to differences in mitochondrial targeting transporter (Lee et al. 2006), static spheroids reflected this species difference more accurately than monolayer cultures. Thus is likely due to better-preserved metabolic competence. Still, static monolayers showed heightened sensitivity to all compounds, potentially generating false positives from non-physiological stress responses despite their larger surface area-to-volume ratio enhancing drug exposure. In the perfused spheroids, FIAU treatment effects did not reach statistical significance, primarily due to limited sample size, a direct consequence of low MPS throughput. This constraint considerably limits the number of compounds, concentrations, and replicates that can be tested, restricting statistical power for routine screening applications in MPS. Another possibility to explain the difference among models could be that Humimic Chip2 devices have polydimethylsiloxane (PDMS) surfaces known to bind hydrophobic molecules (Auner et al. 2019); however, FIAU is a hydrophilic molecule with low octanol-water partition coefficient meaning it is not expected to partition into or bind to PDMS surfaces as confirmed in another PDMS-based liver chip (Foster et al. 2019).

Table 1.

Summary of in vitro (this study) and in vivo (see References) observations on species-specific hepatotoxic effects of CPZ, BOS, and FIAU.

		In vitro observations			Summary of in vivo observations
		ALB	LDH	AST
Human
CPZ	2D	↓	↑	↑	Well-established cause of idiosyncratic cholestatic hepatitis/jaundice in humans (Plaa and Priestly 1976; Padda et al. 2011)
	2.5D	↓	↑	↑
	3D	–	–	–
BOS	2D	↓	↑	↑	Dose-related ALT/AST elevations and clinically recognized hepatotoxicity; liver monitoring required (Fattinger et al. 2001; Rubin et al. 2002)
	2.5D	↓	↑	–
	3D	–	–	–
FIAU	2D	↓	↑	↑	Severe delayed mitochondrial hepatotoxicity, lactic acidosis, liver failure, deaths/transplantation in hepatitis B trial (McKenzie et al. 1995)
	2.5D	↓	↕	↕
	3D	–	–	–
Monkey
CPZ	2D	↓	↑	↕	Monkey-specific in vivo evidence for chlorpromazine-induced cholestatic hepatitis is limited (Plaa and Priestly 1976; Padda et al. 2011)
	2.5D	↓	↑	↕
	3D	–	–	–
BOS	2D	↓	↑	–	Published monkey-specific in vivo hepatotoxicity data are limited (Fattinger et al. 2001; Rubin et al. 2002)
	2.5D	↓	↑	↕
	3D	–	–	–
FIAU	2D	↑	↑	–	Studies in monkeys did not clearly reproduce the fatal delayed mitochondrial liver failure observed clinically (Lewis and Dalakas 1995; McKenzie et al. 1995)
	2.5D	↑	↑	–
	3D	–	–	–
Rat
CPZ	2D	↓	–	↓	In vivo studies report impaired bile flow and cholestatic injury (Plaa and Priestly 1976; Padda et al. 2011)
	2.5D	–	↓	↓
	3D	–	–	–
BOS	2D	↕	–	↓	Rat in vivo data do not consistently reproduce human hepato-toxicity or cholestasis through BSEP/Bsep inhibition and altered bile acid handling (Fattinger et al. 2001; Rubin et al. 2002)
	2.5D	↕	↓/–	–
	3D	–	–	–
FIAU	2D	↕	↓	–	Studies in rats did not reproduce severe delayed mitochondrial hepatic failure observed in humans (Lewis and Dalakas 1995; McKenzie et al. 1995)
	2.5D	↕/–	↕	–/↕
	3D	–	–	–
Dog
CPZ	2D	↕	↓	↓	Dog-specific in vivo evidence for chlorpromazine cholestatic hepatitis is limited (Plaa and Priestly 1976; Padda et al. 2011)
	2.5D	↓	↓	↕
	3D	–	–	–
BOS	2D	↑/–	–	↑	Dog studies did not establish the same clinically important ALT/AST monitoring signal seen in humans (Fattinger et al. 2001; Rubin et al. 2002)
	2.5D	↕	↓	↓
	3D	–	–	–
FIAU	2D	–	–	↑/–	Standard dog toxicology did not clearly reproduce FIAU’s severe delayed human hepatic failure phenotype (Institute of Medicine 1995; Lewis and Dalakas 1995; McKenzie et al. 1995)
	2.5D	↑	↕	–/↓
	3D	–	–	–

Symbols used in the table: [↑] increase, [↓] decrease, [↕] mixed/inconsistent observations, [–] no significant effect.

Transcriptional stability across culture formats and advantages of transcriptomics for both hazard and mechanism studies in pre-clinical safety

It has been suggested that -omics datasets, such as gene expression profiling, are critical to matching MPS or other complex in vitro models to a fit-for-purpose biological question (Homan 2023). In our study, gene expression profiling revealed remarkable consistency across platforms in all species tested. The most pronounced differences occurred between day 0 (post-thaw) and subsequent timepoints, reflecting cell adaptation to culture conditions, a well-documented phenomenon (Dall et al. 2025). Critically, differences between static monolayer, static spheroid, and perfused spheroid platforms were minimal, indicating that core hepatocyte transcriptional programs are preserved across these experimental conditions, at least in the donor preparations used herein. Analysis of ADME genes showed time-dependent downregulation in all platforms, with only select genes maintaining stable expression over 14 d. Such progressive loss of gene expression, documented in previous studies (Bell et al. 2016), underscores the challenge of maintaining fully differentiated hepatocyte phenotype in vitro. Species-specific differences were also apparent, particularly for genes absent or differentially expressed in rat, dog, and monkey hepatocytes compared with human orthologs. These data indicate that spheroid platforms need further refinement and characterization at the functional activity level because while gene expression is certainly an important readout, we did not test if decreases in mRNA levels diminish our ability to identify DILI compounds in these experimental conditions.

We also found that transcriptional analysis more accurately reflected species-specific FIAU hepatotoxicity than functional biomarkers; the highest concordance was observed in static spheroids. Substantial overlap in differentially expressed genes between static monolayer and static spheroid cultures indicated similar response patterns, albeit the magnitude of change in terms of the number of genes and pathways was attenuated in perfused MPS. This dampened response may reflect more physiological conditions where dynamic flow reduces acute stress, or alternatively, lower effective drug concentrations due to increased clearance, requiring detailed pharmacokinetic characterization. We also note that FIAU-induced gene signatures were remarkably consistent with known effects of this human-specific hepatotoxic agent, with significant enrichment of pathways related to lipid metabolism, bile acid homeostasis, and cholestasis (Kleiner et al. 1997). While fibrosis has not been reported in limited human studies with FIAU, the molecular signatures pointing to fibrosis are informative as mechanistic hallmarks of the severity of micro-steatosis observed in humans. The enrichment of fatty acid and lipid metabolism pathways in FIAU-treated human hepatocytes suggests a broad disruption of mitochondrial function, bile acid homeostasis, and metabolic signaling changes consistent with the known clinical manifestations of FIAU toxicity, including hepatic steatosis and lactic acidosis resulting from mitochondrial dysfunction (Borgne-Sanchez and Fromenty 2025). Notably, in the hepatocytes from pre-clinical species, these metabolic pathway alterations were only identifiable in rat hepatocytes cultured in microfluidic-based MPS. The human-specific enrichment of pathways related to mitochondrial dysfunction, lipid dysregulation, and cholestasis provides not only the translational relevance of both static monolayer and static spheroid human hepatocyte models for detecting early manifestation of mitochondria disfunction-related DILI signatures but also offers additional mechanistic insights into adverse effects of FIAU in human hepatocytes. Notably, these mechanistic signatures were concordant between monolayer and spheroid platforms in human hepatocytes but largely absent in the perfused MPS, except for fibrosis and lipid/bile acid metabolism pathways, potentially reflecting reduced transcriptional response and/or insufficient statistical power.

Practical implications and implementation strategy

While functional biomarkers provide useful indices of hepatocyte health, gene expression analysis offered superior mechanistic accuracy and may serve as a primary endpoint in the future screening pipelines. In addition, perfusion-based MPS requires characterization with larger number of compounds and broader compound panels before routine adoption, because throughput and cost remain practical barriers (Rusyn et al. 2022b; Kang et al. 2024). One limitation of this study is that in order to compare different culture modalities and species, we were not able to test a large number of compounds or donors; therefore, only limited conclusions can be drawn and the assessment of a larger number of compounds would be needed to be able to determine advantages and limitations of various models. This is a common challenge with complex in vitro models. For example, recent studies in a different MPS showed encouraging results in terms of human DILI prediction (Ewart et al. 2022); however, the “large” set of compounds for an experiment in MPS measures in 10 to 20 substances, while simpler test models (e.g. traditional or ultra-low attachment 96-well plates) can accommodate greater number of compounds and concentrations. Previous scale-up attempts for the liver MPS in terms of the number of compounds that could be tested have not resulted in improvements in accuracy of DILI prediction, at least in the case of some of the higher-throughput liver MPS platforms (Kato et al. 2022). Therefore, our current findings and previous reports (Rusyn et al. 2022b) support a tiered testing strategy for MPS integration into drug and chemical safety evaluation pipelines. In Tier 1 (High-Throughput Screening): 2D monolayers can be used for rapid hazard identification during discovery, leveraging cost-effectiveness and scalability despite functional limitations. In Tier 2 (Mechanistic Assessment): static spheroid cultures may be most suitable for refined toxicity assessment and mechanistic investigations, offering improved physiological relevance, sustained function, and extended culture capability for concentration-response studies. In Tier 3 (Advanced Mechanistic Studies): perfused and multi-cellular MPS may be used for specific applications that have been shown to benefit from dynamic medium flow conditions (Lauschke et al. 2019; Zhou et al. 2019): repeat-dose exposures, drug metabolism, or cell-cell interactions, as well as for derisking unexpected clinical findings requiring mechanistic elucidation in human-relevant contexts (Ewart and Roth 2021; Ewart et al. 2022; Marx et al. 2025). This strategy, already in place at most end-users of cell-based models for drug and chemical safety assessment (Beilmann et al. 2025; Kleinschmidt-Doerr et al. 2026), optimizes resource allocation, balances throughput with physiological complexity, and maximizes hazard identification throughout drug development.

The tiered evaluation strategy could also help in overcoming another limitation of our study—lack of population variability. Indeed, population variability in pharmacodynamics and pharmacokinetics is a well-established challenge in drug and chemical safety evaluations (Rusyn et al. 2022a; Varshavsky et al. 2023) and experiments with primary human hepatocytes from different donors have been proposed as the means to both quantify inter-individual variability (Tsamandouras et al. 2017) and discover potential genetic and genomic determinants of such differences (Schadt et al. 2008; Harrill et al. 2009; Niemeijer et al. 2024). The use of pooled lots of primary hepatocytes from different subjects is a common strategy to minimize donor-to-donor variability; however, while this approach is widely used and offers some advantages (Goel et al. 2024), it still does not alleviate the need for quantitative characterization of population variability (Dorne et al. 2012). Thus, while our results should be qualified by the fact that the data we provide herein is from a single donor for human, monkey, and dog hepatocytes and a pooled lot for the rat, and that the human lot was selected for enhanced spheroid formation, future studies should address the need for evaluating inter-individual differences in studies of hepatotoxicity using cells from both humans and pre-clinical species.

In summary, this study supports previous observations that static spheroid cultures provide a robust, scalable, and physiologically relevant platform for detecting species-specific hepatotoxicity (Bell et al. 2016; Proctor et al. 2017; Vorrink et al. 2018; Li et al. 2026). At the same time, the added value of perfusion in a liver in vitro model was more nuanced. On the one hand, perfusion resulted in enhancement of the functional performance of hepatocyte spheroids. On the other hand, the low throughput of Humimic Chip2 model makes drug testing difficult and additional studies are needed to fully establish the value of this model for DILI assessment. We also found that gene expression profiling offers superior mechanistic insight compared with traditional functional biomarkers and should be integrated into toxicity screening pipelines. Advanced MPS platforms such as the Humimic Chip2 system hold promise for addressing specific mechanistic questions but require further optimization to justify their routine use. A tiered testing strategy that strategically combines static monolayers, static spheroids, and perfused MPS offers the most pragmatic and scientifically sound approach to improving DILI prediction and reducing late-stage attrition in drug development.

Author contributions

Chander K. Negi and Ivan Rusyn wrote the manuscript; all authors read and edited the manuscript; Chander K. Negi, Philip Hewitt, Stephen S. Ferguson, and Remi Villenave designed the research; Chander K. Negi, Chi Lam T. Chan, Han-Hsuan D. Tsai, Ibrahim Alshammari, and Courtney Sakolish performed the research; Chander K. Negi, Chi Lam T. Chan, Han-Hsuan D. Tsai, and Ivan Rusyn analyzed the data.

Supplementary material

Supplementary material is available at Toxicological Sciences online.

Funding

None declared.

Conflicts of interest. None declared.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files.