Matrix viscoelasticity promotes liver cancer progression in the pre-cirrhotic liver

1Type 2 diabetes mellitus (T2DM) and obesity are important risks for HCC, and it is estimated that up to 30% of HCCs in non-alcoholic steatohepatitis (NASH) occur at a precirrhotic stage when matrix stiffness is still low, and these patients often have poor glycaemic control,. AGEs are produced by the non-enzymatic glycation of serum or tissue proteins during T2DM or can be ingested through the consumption of food prepared at high temperatures. Over time, AGEs accumulate in the matrix due to decreased clearance and metabolism– and can biochemically modify collagen and ECM proteins. Although AGEs in NASH do not appear to increase stiffness greatly, they could affect ECM viscoelasticity. Tissues and ECMs are generally viscoelastic, exhibiting viscous energy dissipation in response to mechanical perturbations and a time-dependent mechanical response, such as stress relaxation in response to deformation. Recent research has shown that changes in ECM viscoelasticity, independent of stiffness, have impacted cell behaviours, including proliferation and migration of breast cancer cells–. Viscoelasticity can modulate tissue growth dynamics, symmetry and the growth of cancer cells. Here we investigated the role of AGE-mediated changes in ECM mechanical properties on NASH and HCC progression.

2To evaluate the potential role of AGE-mediated changes on the ECM, we studied the mechanical properties of liver samples from patients with T2DM, and in individuals with NASH with or without T2DM (average disease activity score, 4; fibrosis stage, 0–1; Fig. 1a). Liver AGEs were significantly higher in patients with T2DM, or with NASH and T2DM compared with those without T2DM (Fig. 1b). Using atomic-force microscopy (AFM), we found that patients with T2DM had similar stiffness (Fig. 1c,d) but higher hysteresis areas under loading–unloading cycles, corresponding to viscous energy dissipation or loss, indicating higher viscoelasticity (Fig. 1e,f). Furthermore, using rheometry, we confirmed that the livers of patients with NASH and T2DM without cirrhosis had similar storage moduli but a higher loss tangent and faster stress relaxation under a constant deformation (Fig. 1g–k).

3To investigate the link between AGEs and viscoelasticity, we studied a mouse model fed high-AGE diet (HiAD). These mice exhibit steatosis, hepatocyte ballooning, insulin resistance and higher liver AGEs compared with those on regular (chow) or fast-food (FFD) diets (Fig. 1l,m). On the basis of AFM (Fig. 1n–p) and rheometry (Fig. 1q–t) analyses, the stiffness was similar (Fig. 1n,q), but the viscoelasticity increased (Fig. 1o,p,r–t), in mice fed a HiAD compared with those fed chow or FFD. Notably, inhibiting AGE production with pyridoxamine (PM) or preventing the formation of AGE–collagen cross-links with alagebrium (ALT-711) reversed the changes in viscoelasticity (Fig. 1o,p,r–t). Together, these data indicate that AGEs in the precirrhotic ECM cause increased viscoelasticity.

4The above results prompted us to investigate the potential causal connections between ECM viscoelasticity and HCC progression. Studies on NASH-related HCC commonly used a high-fat diet combined with the carcinogen diethylnitrosamine, genetically modified models or a CDAA diet. However, it is important to follow the typical pathogenesis characterized by cell injury and inflammation preceding HCC. To experimentally imitate the conditions with increasing ECM stress relaxation before HCC onset, we used a more relevant model to modulate the metabolic/matrix milieu before tumour seeding. We hydrodynamically delivered human MET (hMET) with mutant β-catenin or control vectors to chow-, FFD- or HiAD-fed mice. Earlier appearance and faster growth of transformed foci (glutamine synthetase (GS) and MYC-tag positive) were observed in HiAD-fed mice compared with in mice on the chow or FFD diet (Fig. 2a–c and Extended Data Fig. 1a–c). Importantly, tumour growth diminished, and the survival rate improved after AGE inhibition with PM or after breaking collagen–AGE cross-links with ALT (Fig. 2d–g and Extended Data Fig. 1d–f).

5We next modulated AGE receptors using two additional approaches. We first addressed AGER1, the clearance receptor. As individuals with T2DM plus NASH and mice on a HiAD have significantly downregulated AGER1 (also known as DDOST), we sought to reverse this by AAV8-mediated AGER1 delivery before the hydrodynamic injection (HDI) (Fig. 2h). AGER1-reconstituted mice had higher levels of AGER1 (Fig. 2i) and lower levels of AGE receptors (Fig. 2j) owing to improved uptake, similar stiffness (Fig. 2k) but lower stress relaxation (Fig. 2l,m and Extended Data Fig. 2a), and reduced growth of transformed foci (Fig. 2n,o and Extended Data Fig. 2b), compared with mice that were injected with the control AAV8 construct.

6Second, we studied mice with hepatocyte-specific deletion of RAGE (also known as Ager; proinflammatory AGE receptor; RAGEHepKO) that were fed a HiAD, and hydrodynamically injected these mice with hMET with mutant β-catenin (Extended Data Fig. 2c). These mice had reduced growth of tumour foci (Extended Data Fig. 2d,e), decreased liver AGEs (Extended Data Fig. 2f), similar stiffness (Extended Data Fig. 2g) and reduced viscoelasticity (Extended Data Fig. 2h–j) compared with RAGEfl/fl mice fed the HiAD.

7Together, these studies with four animal models all converge on the conclusion that AGEs in the ECM create a more viscoelastic and tumorigenic environment that promotes HCC.

8The data presented so far suggest that AGE accumulation is critical to increased ECM viscoelasticity and HCC progression. We next sought to elucidate how ECM viscoelasticity is modulated. Type-1 collagen networks are thought to be critical regulators of tissue mechanics. AGEs can directly cross-link collagen and affect matrix mechanics by altering its helical structure,. Collagen networks were imaged using second harmonic generation (SHG) with two-photon microscopy in liver samples from humans (healthy control, NASH, NASH + T2DM; Extended Data Fig. 3a), as well as in mice (chow, HiAD, HiAD + PM, HiAD + ALT-711 and in AAV8-AGER1-injected livers; Extended Data Fig. 3b). The collagen network exhibited bundling in patients with T2DM + NASH and in mice on the HiAD, whereas the network appeared to be more organized with thinner fibres in healthy humans, and in mice on the chow or HiAD diets after PM and ALT treatment or AGER1 reconstitution. Inhibiting AGE production or disrupting AGE–collagen cross-links decreased the amount of highly cross-linked insoluble collagen (Fig. 3a). We analysed decellularized native ECM from chow- or HiAD-fed mice using SHG to evaluate collagen fibre length, spatial orientation and interconnectivity. Compared with the well-organized collagen network in mice on the chow diet, in HiAD-fed mice, the network exhibited less connectivity and had collagen bundles with shorter fibres and lower fibre–fibre angles (Fig. 3b). Similar structures with shorter fibres and angles were seen in neutralized reconstituted collagen hydrogels with AGEs (Fig. 3c). The interactions between AGEs and collagen telopeptide lysine residues can interfere with fibre elongation and the formation of a structurally cohesive collagen network. To evaluate the properties of AGE-cross-linked collagen hydrogels and the relationship between low-connectivity networks and matrix characteristics, we performed rheological measurements. AGE-cross-linked collagen hydrogels had similar stiffness (Fig. 3d) but faster stress relaxation compared with non-cross-linked collagen (Fig. 3e). We next measured the distribution of Young’s moduli by AFM mapping of hydrogels. We saw that, while there was heterogeneity in local moduli in both gels, the ranges of elastic moduli were slightly higher in the collagen-only case (Extended Data Fig. 3c,d). The distribution of hysteresis areas (viscoelasticity) demonstrated higher frequencies in AGE-cross-linked collagen hydrogels (Extended Data Fig. 3e,f). To further study the collagen-rich matrix with stiffness comparable to that of the liver, interpenetrating polymer network (IPN) 3D hydrogels made from AGE-modified collagen and alginate were tested (Fig. 3f). Similar changes in architecture and connectivity were seen after AGE exposure (Fig. 3g). Breaking the AGE–collagen cross-links with ALT-711 decreased the amount of insoluble collagen (Fig. 3h) and increased the fibre length, fibre–fibre angles (Fig. 3i,j) and network connectivity. ALT-711 itself did not change the collagen matrix (Extended Data Fig. 3g) or its mechanical properties (Extended Data Fig. 3h–j). Indentation tests showed restored viscoelasticity after using ALT-711 (Fig. 3k). IPN hydrogels mixed with AGE-modified collagen exhibited similar stiffness (Fig. 3l), but faster stress relaxation compared with non-modified collagen (Fig. 3m–o) based on rheological measurements.

9To further study how the matrix architecture impacts viscoelasticity, we simulated collagen matrices with a variation in two structural parameters to model the activity of AGEs: average collagen fibre bundle length and bundling angle, as a proxy for heterogeneity (Fig. 3p,q). The collagen matrices are connected by a combination of weak bonds, known to act within and between collagen fibres (Extended Data Fig. 4a–g) and could underlie stress relaxation and viscoelasticity in type-1 collagen matrices, and cross-links due to AGEs– (Extended Data Fig. 4h,j). Bundled networks are more heterogeneous and exhibit faster stress relaxation compared with unbundled networks (Fig. 3r). Reducing the fibre length (Fig. 3s and Extended Data Fig. 5a–e) or bundle angle (Fig. 3t and Extended Data Fig. 5f–i) increases the heterogeneity of the network leading to faster stress relaxation.

10Together, these computational results show that the changes in architecture that we observed due to AGE activity in vivo and in vitro, namely enhanced heterogeneity and shorter fibre length, are predicted to lead to faster stress relaxation and enhanced viscoelasticity.

11We next focused on mechanotransductive pathways leading to tumour growth and invasion by faster stress relaxation. RNA-sequencing (RNA-seq) and bioinformatics analyses were performed on mouse liver samples (in the chow, FFD, HiAD, HiAD + PM and RAGEHepKO groups), to identify potential pathways. Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses were performed on differentially expressed genes between the HiAD and FFD groups to categorize the enriched functional classification (Fig. 4a and Extended Data Fig. 6a). In the Hippo signalling pathway group, YAP/TAZ-regulated genes were enriched in HiAD-fed mice and decreased after PM treatment and in RAGEHepKO mice (Fig. 4a). YAP activation, as indicated by a nuclear signal of active non-phosphorylated YAP (Extended Data Fig. 6b,c,g) and the induction of targets CTGF (also known as CCN2) and CYR61 (CCN1), was observed in mice on the HiAD, but decreased after PM treatment or in the RAGEHepKO mice on the HiAD (Extended Data Fig. 6d). YAP-positive hepatocytes were not observed in close proximity to collagen bundles in the HiAD group (Extended Data Fig. 6e). To substantiate that HCC in faster-stress-relaxing ECM was promoted by YAP and TAZ, we co-injected dominant-negative TEA domain transcription factor 2 (dn-TEAD2) or control plasmids by HDI with hMET and mutant β-catenin (Fig. 4b). The TEAD family of transcription factors are important binding partners for YAP and TAZ, and TEAD2 in particular was linked to HCC progression. We found that the number of GS/MYC double-positive foci significantly decreased after dn-TEAD2/HDI co-injection compared with in mice that were injected with the control plasmid (Fig. 4c,d and Extended Data Fig. 6f (haematoxylin and eosin images)), and these foci were negative for active, nuclear YAP (Fig. 4c (bottom, inset)). In mice treated with HDI + control vector co-injection, active YAP localized to the nuclei of GS-positive cells. Downstream targets CTGF and CYR61 were downregulated after injection with dn-TEAD2 (Fig. 4e). Together, these data indicate that HCC in faster-stress-relaxing ECM was promoted by YAP.

12Given these results, our next question was how the viscoelastic ECM conveys proliferative and invasive signals. To directly test the effects of viscoelasticity, we constructed IPNs of alginate and reconstituted basement membrane (rBM) matrix hydrogels for 3D culture studies (rBM-IPNs). HCC cells (Fig. 5 and Extended Data Fig. 7a–e) were encapsulated in rBM-IPN hydrogels, with the same stiffness (Young’s modulus, 2 kPa) but different levels of viscoelasticity (Fig. 5a). In high-viscoelasticity gels, cells exhibited increased proliferation (Fig. 5b,e), YAP activation (Fig. 5c,f) and the formation of TKS5-positive (also known as SH3PXD2A) and MT1-MMP-positive invadopodia-like structures as well as a more elongated shape (Fig. 5d,g). When the Young’s modulus was gradually increased to 5 kPa, we did not observe an increase in cell proliferation, YAP activation or induction of its target genes (Extended Data Fig. 7a–e). Invadopodia are thin linear protrusions that are linked to cancer cell invasion and migration. In high-viscoelasticity gels, we detected canonical markers such as TKS5 and MT1-MMP (Fig. 5d and Extended Data Fig. 7f) as well as active integrin β1, and phosphorylated myosin light chain (p-MLC) (Extended Data Fig. 7g,h, respectively). Antibody controls are shown in Extended Data Fig. 7i. As integrin-β1–matrix interactions modulating mechanotransduction were reported in breast cancer cells in high-viscoelasticity gels,, we used integrin-β1-specific blocking antibodies (Fig. 5h–j). These reduced proliferation (Fig. 5h) and YAP target CTGF (Fig. 5i) and increased cell circularity (Fig. 5j).

13Notably, on the basis of our RNA-seq data, TNS1 was significantly induced in an AGE- and viscoelasticity-dependent manner (Figs. 4a and 5k). TNS1 was described as part of a molecular clutch that mediates the mechanical linkage between ECM-bound integrins and the actomyosin cytoskeleton to generate traction force–. To study the role in mechanosignalling, proximity ligation assays (PLA) were performed in HCC cells in low- versus high-viscoelasticity hydrogels to capture the potential interaction between TNS1 and integrin β1. High-viscoelasticity conditions promoted their interaction (Fig. 5l,m and Extended Data Fig. 7j (antibody controls)). We next studied the effects of TNS1 or integrin β1 knockdown (CRISPR–Cas9; Fig. 5n–p). This reduced cell proliferation (Fig. 5n and Extended Data Fig. 8a) as well as the activation of YAP and target genes (Fig. 5o and Extended Data Fig. 8b,e) in high-viscoelasticity ECM. Invadopodia formation was reduced, and cell circularity increased after TNS1 or integrin β1 knockdown (Fig. 5p and Extended Data Fig. 8c,d). Hep3B cells exhibited similar results (Extended Data Fig. 9). To further analyse the importance of the TNS1–integrin β1 interaction in invadopodia formation and proliferation, we transfected cells with a vector containing a mutant TNS1 (with a deletion of the PTB domain preventing its binding to integrin β1, hereafter, dd-TNS1) or full-length TNS1 (Extended Data Fig. 10a) and embedded them in low- or high-viscoelasticity hydrogels. We assessed whether dd-TNS1 is capable of binding to integrin β1 using a PLA. In dd-TNS1-transfected cells, the PLA signal was very low, whereas, in the full-TNS1–tdTomato-positive cells, the signals co-localized (Extended Data Fig. 10b,c). dd-TNS1-transfected cells in high-viscoelasticity hydrogels had lower proliferation (Extended Data Fig. 10d,h), and reduced active YAP and its target genes (Extended Data Fig. 10e,i,k), while their circularity increased (Extended Data Fig. 10f,j) and they did not form protrusions (Extended Data Fig. 10g). As RhoA is involved in regulating cell shape and migration, and it was described to be modulated by TNS1, we next studied its activity. Cells in high-viscoelasticity hydrogels exhibited increased RhoA-GTPase activity in a TNS1- and integrin-β1-dependent manner (Extended Data Fig. 11a,b). LATS1 activity decreased and YAP was more active in cells in high-viscoelasticity gels, whereas the reverse was seen after TNS1 or integrin β1 knockdown. Pharmacological inhibition of ROCK-GTPase and myosin 2 resulted in an increase in cell circularity, a lower proliferation rate and reduced expression of YAP targets (Extended Data Fig. 11c–e).

14To study the in vivo role of TNS1, we used a CRISPR–Cas9-directed approach whereby TNS1 was targeted at the time of HDI. Mice on chow or HiAD diets were hydrodynamically injected as before, including either single-guide RNA for Tns1 (sgTNS1) or control sgRNA plasmids, and were euthanized 7 weeks after the injections (Extended Data Fig. 12a). In the sgTNS1-treated mice, Tns1 expression and the number of transformed MYC+GS+ foci decreased compared with in the control-sgRNA-injected mice (Extended Data Fig. 12b–d, respectively), and YAP targets Ctgf (also known as Ccn2) and Cyr61 (Ccn1) had lower expression (Extended Data Fig. 12e). Together, these data suggest that TNS1 is part of a molecular clutch that mediates the mechanical linkage between ECM-bound integrins and the actomyosin cytoskeleton responding to high viscoelasticity (Fig. 5q).

15Here we show that AGEs in the liver ECM create a faster stress-relaxing viscoelastic niche leading to the activation of mechanosignalling promoting HCC. Although matrix stiffness in advanced fibrosis/cirrhosis and its effects on cancer progression have been extensively studied, our data are the first, to our knowledge, to demonstrate how changes in matrix viscoelasticity, independent of stiffness, affect HCC growth. This is clinically very relevant as increasing viscoelasticity could be a risk factor foretelling more invasive features of HCC in NASH/T2DM. Current guidelines exclude precirrhotic patients from HCC screening paradigms; therefore, new viscoelasticity-based imaging approaches will need to be developed to identify the population at risk. T2DM is a major risk factor not only for liver cancer progression but also in breast, colon and pancreatic cancer–, and this may point to the crucial role of a more viscoelastic matrix in patients with diabetes.

16Collagen cross-linking mediated by AGEs has been thought to contribute to increasing stiffness similarly to LOXl2 or TTG, modifying the helical structure of collagen–. However, our studies identified that, in certain contexts, increased collagen cross-linking leads to weakly bound shorter fibres with lower interconnectivity, together promoting a viscoelastic niche. This in turn facilitates changes in cellular shape, cytoskeletal reorganization and the formation of invadopodia-like structures,,,. This could be relevant to not only the liver peritumoral matrix, but also other cancers in which changes in viscoelasticity have been observed, but are not yet implicated functionally for progression,.

17We also identified the major role of the TNS1–integrin β1–RhoA–YAP mechanotransduction pathway responding to increasing viscoelasticity. While the role of YAP activation in various cancers is known, as is its role in responding to viscoelasticity in vitro–,,,, the in vivo role of YAP as an inducer of cancer progression in response to viscoelasticity changes was previously unclear. Furthermore, another element of interest relates to the role of TNS1 as a key mechanosensory in this integrin–YAP signalling pathway. TNS1 regulating and sustaining the activity of the molecular clutch could be essential for protrusion formation and migratory activity of cells,, and these can be further explored in future studies.

18In conclusion, we identified the central role of increasing viscoelasticity in the liver tumour niche. Viscoelasticity-activated mechanocellular pathways promise to be a diagnostic and/or therapeutic areas in NASH/T2DM-related HCC and beyond.

19All human samples were de-identified and exempted (exemption 4). Human liver samples were obtained from the Stanford Diabetes Research Center (SDRC), Donor Network West (DNW), Stanford Tissue Bank and Clinical Biospecimen Repository and Processing Core (CBRPC) of the Pittsburgh Liver Research Center (PLRC). This was approved by Stanford University Institutional Review Board (IRB, 67378) and the Pittsburgh Liver Research Center Review Board. Histology was evaluated for necroinflammation, hepatocellular ballooning, ductular reaction and fibrosis by a hepatopathologist in a blinded manner, and non-alcoholic fatty liver disease activity score (NAS) scores were provided. Detailed information about the donors, including sex, age, diagnosis and NAS score, are summarized in Supplementary Table 1.

20All animal experiments were conducted according to the experimental procedures approved by the Institutional Animal Care and Use Committee at Stanford University and Palo Alto VA. Mice were maintained at temperatures and humidity ranges of 20 to 26 °C and 30% to 70%, respectively. Mice were housed in standard cages under 12 h–12 h light/dark cycles with ad libitum access to water and food unless otherwise indicated. Mice were placed on control chow or a FFD (17.4% protein, 20% fat and 49.9% carbohydrate, AIN-76A) supplemented with high-fructose corn syrup in the drinking water at a final concentration of 42 g l−1 for up to 14 weeks, or HiAD (prepared by cooking FFD at 120 °C for 20 min) as previously described supplemented with high-fructose water. Wild-type C57BL/6J male mice (aged 8 to 10 weeks) were purchased from the Jackson Laboratory. RAGEfl/fl mice on the C57B6 background were gifted by B. Arnold. RAGEHepKO mice were generated by crossing RAGEfl/fl mice with Albumin-cre mice (Jackson Laboratory) for several generations. Male RAGEfl/fl mice and RAGEHepKO mice aged 8 to 10 weeks were used. To generate hepatocyte transgenic AGER1 (Ddost) mice, wild-type mice were injected with adeno-associated virus 8 (AAV8)-control green fluorescent protein (AAV8-control) or AAV8-thyroxine-binding globulin-AGER1 recombinase (AAV8-AGER1) (5 × 1011 genome copies, Vector BioLabs) at the end of week 5 of feeding (the rationale of the specific targeting in hepatocyte has been previously reported). To study the inhibition of AGE formation or AGE–protein cross-linking, wild-type mice fed with HiAD were injected intraperitoneally daily with pyridoxamine hydrochloride (Sigma-Aldrich) (60 mg per kg, daily), alagebrium (ALT-711) (Sigma-Aldrich, 1 mg per kg), or vehicle (Tris-HCl), from the fifth week of feeding. At the time of euthanasia, photographs of the livers were taken, and the number and the size of lesions on the surface were recorded.

21Oncogenic plasmids were delivered to the mouse liver by hydrodynamic tail vein injection as previously described. In brief, pT3-EF1α-h-hMET, pT3-EF1αh-WT-β-catenin-MYC-tag, pT3-EF1α-h-S45Y-mutant-β-catenin-MYC-tag and pCMV/SB transposase were from S. P. S. Monga’s laboratory. Plasmids used for in vivo experiments were purified using the Invitrogen Endotoxin-free Maxi prep kit (Sigma-Aldrich). A combination of 20 µg of pT3-EF5-hMET and 20 µg of pT3-EF5α-WT-β-catenin-MYC or pT3-EF5α-S45Y-β-catenin-MYC, along with 1.6 µg of SB (25:1) was diluted in 1 ml of endotoxin-free saline (AdipoGen), filtered through 0.22 mm filter (Millipore), and injected into the tail vein of mice in 5 to 7 s at the at the end of week 7 of chow, FFD or HiAD feeding. To inhibit YAP activity, the mice were injected with dn-TEAD2 (pT3-EF1α-h-dnTEAD2, 60 µg) or negative control vector (empty pT3-EF1α, 60 µg), (from S. P. S. Monga’s laboratory) by hydrodynamic tail vein injection at the end of week 7 of chow or HiAD feeding with hMET, S45Y-mutant-β-catenin and SB. To knockdown TNS1, mice were injected with a CRISPR–Cas9-based vector linking two sgRNAs targeting Tns1 exon 1 and exon 7 (pX333-TNS1, 50 µg) or negative control vector (empty pX333, 50 µg) by hydrodynamic tail vein injection at the end of seventh week of chow or HiAD feeding at the same time as with hMET, S45Y-mutant-β-catenin and SB. The sequences of the sgRNAs used in this study are provided in Supplementary Table 2.

22Mice were monitored 3 days per week for any sign of pain or distress, and the body weight and body condition score were evaluated after hydrodynamic tail vein injection. The mice were immediately euthanized if they lost more than 10% of their body weight or if they had a body condition score of 2 or less during that time. The tumour size (macroscopic) never reached more than 3 mm.

23Paraffin-embedded tissue samples were cut into 5 µm sections, deparaffinized and rehydrated. For antigen retrieval, slides were boiled in citrate buffer (0.01 M, pH 6.0) using a microwave oven on high power for 5 min and cooled down to room temperature. After incubation in 3% aqueous H2O2 to quench the endogenous peroxidase, the sections were washed in PBST (PBS with 0.1% Tween-20, v/v) washing buffer, blocked with 5% goat serum (EMD Millipore) diluted in PBST at room temperature for 1 h and incubated with primary antibodies diluted in 2% goat serum in PBST at 4 °C overnight.

24For immunohistochemistry, slides were incubated with appropriate biotinylated secondary antibodies (Supplementary Table 3) for 1 h, and then processed according to the ABC Peroxidase Standard Staining Kit (Thermo Fisher Scientific) for 30 min. The slides were stained with 3,3′-diaminobenzidine (Abcam) for 5 s to 5 min and counterstained with haematoxylin (Thermo Fisher Scientific) for 45 s. The images were scanned using the Leica Aperio AT2 system at Stanford Human Pathology/Histology Service Center. Serial sections were incubated with GS and MYC-tag primary antibodies, and a cluster of cells (at least 20 cells) positive both for GS and MYC-tag were counted as early foci. The antibodies used in this study are shown in Supplementary Table 3.

25AGEs were measured using the OxiSelect Advanced Glycation End Product Competitive ELISA Kit (Cell Biolabs) in the serum and the liver homogenate according to the manufacturer’s instructions. In brief, 10 mg liver samples were homogenized in PBS. After measuring the protein concentration, 300 µg protein was added to a 96-well ELISA plate and incubated for 1 h at room temperature. After incubation with the secondary HRP-conjugated anti-AGE antibodies, the reaction was halted with a stop solution, and the plates were read at 450 nm.

26According to the manufacturer’s recommendations, total RNA was extracted from snap-frozen liver tissues and cells using the RNeasy mini kit (Qiagen). Complementary cDNA was created from an identical amount of RNA using the iScript cDNA synthesis kit (Bio-Rad). The PowerUp SYBR Green PCR Master Mix (Applied Biosystems) was used for quantitative PCR with reverse transcription (RT–qPCR) on the 7900HT machine (Applied Biosystems), and the data were evaluated using the 2−Ct technique. As an endogenous control, Arbp (also known as Rplp0) and GAPDH were used to standardize the data in mouse and human, respectively. A list of the primer sequences used in this study is provided in Supplementary Table 4.

27Measurements on frozen liver tissues were performed on livers embedded in OCT compound (Sakura), snap frozen by direct immersion into liquid nitrogen and cut into 100 μm sections on the Leica CM1900-13 cryostat. The samples were kept in a protease inhibitor cocktail during the AFM analyses (Roche Diagnostic). For the hydrogel mapping collagen or AGE-modified collagen was neutralized with 5× DMEM and 1× NaOH and the hydrogels were kept overnight on 37 °C. The Bruker Resolve BioAFM system was used to take measurements (Bruker). For the indentation, Novascan Tech modified silicon nitride cantilevers (k = 0.01 N m−1, k = 0.06 N m−1) with a borosilicate glass spherical tip (diameter 10 and 5 μm, respectively) were used.

28For each session, cantilevers were calibrated using the thermal oscillation method. AFM force maps were performed on 94.7 μm × 94.7 μm fields. Each experimental group included at least four different human or mouse samples, with two sections from each, and three different areas generated per section. For hydrogel mapping, each group included at least four different gels, and five different areas generated per gel. Data analyses were performed using the Hertz model in NanoScope Analysis v.1.9. and Mountains SPIP v.9.

29Measurements were optimized to assess storage modulus, loss modulus, loss tangent and stress relaxation based on methods previously described. In brief, liver samples were prepared using an 8 mm diameter punch (Integra Miltex). The height of the slices ranged from 3 to 5 mm in the uncompressed state. The samples were kept hydrated during all of the experiments with DMEM. Parallel plate shear rheometry was performed on the ARES-G2 rheometer (TA instruments) at room temperature using TA TRIOS software v.5.1.1 (TA instruments). For all of the measurements, the upper plate was initially lowered to touch the sample, and 0.01 N of nominal initial force (~300 Pa) was applied to ensure adhesive contact of the sample with the plates. Measurements were taken first with a dynamic time sweep test (2% constant strain, oscillation frequency 1 radian s−1, measurements taken for 600 s), then stress relaxation (10% initial strain, measurements taken for 600 s).

30The hydroxyproline assay for total liver collagen was performed as previously described. In brief, liver samples were homogenized and denatured in 6 N HCl. Hydrolysed samples were then dried and washed three times with deionized water, followed by incubation in 50 mM chloramine T oxidation buffer for 20 min at room temperature. The samples were incubated with 3.15 M perchloric acid (Sigma-Aldrich) for 5 min, then with p-dimethylaminobenzaldehyde (Sigma-Aldrich). The absorbance of each sample was measured.

31For insoluble collagen, the liver or collagen gel was first homogenized in 0.5 M acetic acid at a 1:4 ratio (for example, 800 μl for 200 mg liver) to make 20% liver homogenate. Next, 500 μl of 20% liver homogenate was added onto 1 ml of 0.5 M acetic acid and the tubes rotated at 4 °C overnight. The samples were centrifuged at 20,000g for 30 min to collect the pellet, resuspended in pepsin (2 mg ml−1 in 0.5 M acetic acid) and incubated at 4 °C overnight. The next day, the samples were centrifuged and the pellets were collected. The samples were then analysed using the hydroxyproline assay, described previously.

32AGE-BSA was prepared as previously described,. In brief, glycolaldehyde (Sigma-Aldrich) was dissolved in 10 mg ml−1 BSA/PBS to a final concentration of 33 mM. The solutions were incubated at 37 °C for 72 h, followed by dialysis against PBS. The dialysed solutions were sterilized with 2 µM filters, and aliquots were stored at −80 °C.

33Collagen type I (telopeptide intact, Corning, 354236, Corning collagen I, rat tail, used in all experiments) was incubated with 2.5 mg ml−1 in AGE-BSA 0.1% acetic acid (Merck) to obtain 3 mg ml−1 collagen solution at 4 °C for 4 weeks. BSA was mixed with collagen as the control. Alagebrium chloride (ALT-711, 20 mg ml−1) was added as the AGE cross-linking inhibitor. Collagen gels were polymerized by mixing 3 mg ml−1 AGE-modified or non-modified collagen solution with 10× PBS and neutralized with 1× NaOH (Merck) and incubated at 37 °C for 90 min leading to the formation of gels. For most gels, 1.6 mg ml−1 collagen was used.

34According to the manufacturer, low-molecular-mass, ultrapure sodium alginate (Provona UP VLVG, NovaMatrix) with a molecular mass of <75 kDa was used for fast-relaxing substrates. For slow-relaxing substrates, sodium alginate (high molecular mass) was used (FMC Biopolymer, Protanal LF 20/40, high molecular mass, 280 kDa). Alginate was treated with activated charcoal, dialysed against deionized water for 3–4 days (molecular mass cut-off, 3,500 Da), sterile-filtered, lyophilized and then reconstituted to 3.5 wt% in serum-free DMEM (Gibco). The use of low/high-molecular-mass alginate resulted in high/low viscoelasticity IPNs.

35Mouse livers were decellularized in situ by detergent (0.5% (w/v), sodium deoxycholate, 250 ml per mouse) and water (50 ml per mouse) perfusion at a pump speed of 0.2 ml min−1. After the final perfusion, the livers were removed and washed overnight in PBS. For AGE-modified collagen gels, the samples were prepared as previously described. Gels were imaged 1 day after formation.

36For SHG imaging, all of the samples were imaged using the Leica TCS SP5 multiphoton confocal microscope or the Leica Stellaris 8 DIVE upright confocal microscope. The excitation wavelength was tuned to 840 nm, and a 420 ± 5 nm narrow band-pass emission controlled by a slit was used for detecting the SHG signal of collagen. The images were recorded using an inverted confocal laser-scanning microscope (Leica TCS SP8) equipped with a ×20 water-immersion objective for confocal reflection imaging. An Ar+ laser at 488 nm was used to illuminate the sample, and the reflected light was detected with photomultiplier tube (PMT) detectors. Scans were at 1,024 × 1,024 pixels, and all of the images were taken 80–100 μm into the samples. Collagen measurements were performed using CT-Fire software (v.2.0 beta) (https://loci.wisc.edu/software/ctfire) and ImageJ v.1.53t (https://imagej-nih-gov.stanford.idm.oclc.org/ij/).

37Alginate was transferred to a 1.5 ml Eppendorf tube (a polymer tube) and kept on ice for each viscoelastic gel. For rBM-IPNs, alginate was mixed 30 times before adding rBM (Corning) at 4 °C. Collagen-IPNs were created by diluting and neutralizing AGE-modified or AGE-unmodified collagen gels with 10× DMEM and 1 M NaOH (Merck) at 4 °C. All substrates had a final concentration of 10 mg ml−1 alginate, 4.4 mg ml−1 rBM or 1.6 mg ml−1 collagen after additional DMEM was added. This was pipette-mixed, and the resulting mixture was kept on ice. Calcium sulfate was added to 1 ml Luer lock syringes (Cole-Parmer) and stored on ice to maintain the constant Young’s moduli of the substrates with high and low viscoelasticity. The polymer mixes were divided into individual 1 ml Luer lock syringes (polymer syringes) and placed onto ice as well. The polymer syringe was linked to the calcium sulfate syringe to create gels. The two solutions were quickly combined using 30 pumps on the syringe handles, and the resulting mixture was placed into a well of an eight-well Lab-Tek plate (Thermo Fisher Scientific) that had been precoated with rBM. After moving the Lab-Tek dish to a 37 °C incubator, the gel was allowed to form for 1 h before a full medium was added.

38Rheology experiments were performed using a stress controlled AR2000EX rheometer (TA Instruments). IPNs were directly deposited onto the lower Peltier plate for rheology testing. The gel was then slowly contacted by a 25 mm flat plate, creating a 25 mm disk gel. To stop dehydration, mineral oil (Sigma-Aldrich) was applied to the gel’s edges. The storage and loss moduli had equilibrated by the time the time sweep was done, which was at 1 rad s−1, 37 °C and 1% strain.

39For the stress relaxation experiments, after the time sweep, a constant strain of 5% was applied to the gel at 37 °C, and the resulting stress was recorded over the course of 4 h.

40Human HCC cell lines Huh7 (gift from P. Sarnow) and Hep3B (purchased from ATCC) were cultured in high-glucose DMEM (Gibco) with 10% fetal bovine serum (FBS, Gibco) with 1% penicillin–streptomycin (Life Technologies). All cells were cultured at 37 °C in 5% CO2.

41TNS1 was knocked down in Huh7 and Hep3B cells using TNS1 sgRNAs or integrin β1 sgRNAs. A control sgRNA sequence was used as the negative control. sgRNAs were cloned into pMCB306 (Addgen plasmid 89360, sgRNA expression vector with GFP, puromycin resistance), then co-transfected into cells with lentiCas9-Blast (Addgen plasmid 52962, expresses human codon-optimized Streptococcus pyogenes Cas9 protein and blasticidin resistance from the EFS promoter). Transfected cells were selected by puromycin and tested for TNS1 or integrin β1 expression 2 days after transfection. A list of the sequences of the sgRNAs used in this study is provided in Supplementary Table 2.

42For PTB-domain-deleted TNS1 (dd-TNS1) assays, dd-TNS1 or a full-length TNS1 was cloned into a tdTomato-N1 vector (gift from S.-H. Lo). Huh7 cells were transfected with these plasmids and cultured for 24 h before encapsulation in IPNs hydrogels.

43Huh7 or Hep3B cells were transfected with plasmids containing the human TKS5 (also known as SH3PXD2A) with either mNeonGreen or mScarlet tag (gift from L. Hodgson) for all invadopodia analyses.

44For analyses of YAP activation, invadopodia formation and proliferation, Huh7 cells were serum-starved overnight and encapsulated in IPNs. In brief, cells were washed with PBS, trypsinized using 0.05% trypsin/EDTA, washed once, centrifuged and resuspended in serum-free medium. The concentration of cells was determined using the Vi-Cell Coulter counter (Beckman Coulter). After Matrigel was mixed with the alginate, cells were added to this polymer mixture and deposited into a cooled syringe. The solution was then vigorously mixed with a solution containing CaSO4 and deposited into wells of a chambered cover glass (LabTek). The final concentration of Matrigel and alginate was 4.4 mg ml−1 and 10 mg ml−1, respectively. 5 mM, 15 mM and 50 mM CaSO4 were used to generate varied stiffness of IPNs with 0.8 kPa, 2 kPa and 5 kPa, respectively. The final concentration of cells was 3 × 106 cells per ml of IPN. The cell-laden hydrogels were gelled in an incubator at 37 °C and 5% CO2 for 60 min and then were cultured in a DMEM medium containing 10% FBS. After 1 day, cells were collected for RT–qPCR, western blotting and immunostaining analysis.

45Inhibitors in the 3D cell culture were used at the following concentrations: 10 μM Y-27632 to inhibit ROCK (Abcam); 50 μM blebbistatin (Abcam) to inhibit myosin II; and 1 μg ml−1 monoclonal integrin-β1-blocking antibody (Abcam, P5D2). Vehicle-alone controls for these inhibitors were as follows: DMSO for blebbistatin, and latrunculin-a; deionized water for Y-27632; and IgG nonspecific antibody (Sigma-Aldrich, I5381) for integrin-β1-blocking antibody. Y-27632 and blebbistatin were added to the culture medium directly. Integrin β1-blocking antibody was incubated with Huh7 cells on ice for 1 h before encapsulation in IPNs and added to the culture medium directly.

46RAGEfl/fl (wild-type) mice were fed chow, FFD or HiAD for 14 weeks. A group of HiAD-fed mice was injected intraperitoneally daily with PM. A cohort of RAGEHepKO mice was placed on a HiAD for 14 weeks as described previously. RNA was prepared from 2–3 mice per group, and RNA-seq was performed at Novogene with paired-end 150 bp reads (NovaSeq 6000 Sequencing System, Illumina). Gencode gene annotations version M18 and the mouse reference genome major release GRCm38 was derived from https://www.gencodegenes.org/. Dropseq tools v.1.1249 were used for mapping the raw sequencing data to the reference genome. The resulting UMI-filtered count matrix was imported into R v.3.4.4. Before differential expression analysis using Limma v.3.40.650, sample-specific weights were estimated and used as coefficients alongside the experimental groups as a covariate during model fitting with Voom. t-tests were used for determining differentially (P < 0.05) regulated genes between all possible experimental groups. GSEA was conducted using the preranked GSEA method within the KEGG databases with the online tool g:Profiler (https://biit.cs.ut.ee/gprofiler/gost). RNA-seq heat maps and unsupervised hierarchical clustering was performed with g:Profiler. RNA-seq data are available under Gene Expression Omnibus accession number GSE245016.

47Cells were washed with PBS and lysed with RIPA buffer. The homogenate was centrifuged, and the supernatant was collected. Protein concentrations were determined using the Pierce BCA Protein Assay Kit (Thermo Fisher Scientific). Protease inhibitor (Roche) and phosphatase inhibitor (Roche) were added to all the lysis procedures mentioned above, and 10–50 µg of the protein samples were loaded onto the SDS–polyacrylamide gel. The proteins were transferred to a polyvinylidene difluoride membrane or nitrocellulose membrane, which was blocked with 5% BSA in TBST and then incubated with primary antibodies at 4 °C overnight. The blots were washed with TBST and further incubated with horseradish-peroxidase-conjugated secondary antibodies. The signal was detected by adding Western-Bright enhanced chemiluminescence substrate (Advansta) or SuperSignal West Pico PLUS Chemiluminescent Substrate (Thermo Fisher Scientific) and imaged with film or the iBright CL1500 imaging system (Thermo Fisher Scientific). The images were processed and analysed using NIH ImageJ (v.1.53t) and iBright Analysis software v.5.2.1 (Thermo Fisher Scientific). The antibodies used in this study were shown in Supplementary Table 3.

48Frozen sections of livers or gel-embedded cells were washed twice with PBS and fixed with 4% paraformaldehyde at 4 °C overnight. Sections were permeabilized in PBS with 0.4% (v/v) Triton X-100 for 10 min. After blocking with 5% goat serum in PBST at room temperature for 1 h, cells were incubated with primary antibodies (Supplementary Table 3) diluted in 2% goat serum in PBST at 4 °C overnight. The slides were washed and then incubated with secondary antibodies at room temperature for 1 h. Coverslips were washed with PBST between incubations and mounted with an anti-fade mounting medium with DAPI. Fluorescence images were taken using the Leica TCS SP8, multi-photon Leica Stellaris 8 DIVE upright Confocal, and Zeiss Airyscan2 LSM980 systems. Images were processed using NIH ImageJ (v.1.53t). To quantify cell circularity and cell area, the confocal images of cells were analysed in ImageJ v.1.53t (https://imagej-nih-gov.stanford.idm.oclc.org/ij/). Circularity, mathematically calculated as 4π × area × (perimeter)−2, ranges from 0 to 1, with a value of 1 being a perfect circle.

49The Duolink proximity ligation assay kit (Sigma-Aldrich) was used to determine the interaction between TNS1 and integrin β1 in Huh7 cells. Reagents were used according to the manufacturer’s instructions, and the steps were optimized. In brief, anti-TSN1 and anti-integrin β1 were used as the primary antibodies. The primary antibodies bound to a pair of oligonucleotide-labelled secondary antibodies (PLA probes), the hybridizing connector oligos joined the PLA probes if they were close by and the ligase created the DNA template needed for rolling-circle amplification (RCA). Labelled oligos hybridized to the complementary sequences in the amplicon and produced discrete red fluorescent signals that could be seen by confocal microscopy (Leica Microsystems). NIH Image J (v.1.53t) software was used to count the signal, and the average counts were used for the plot.

50In this study, we used an agent-based model for simulating the discrete structure of collagen matrices. Details of the model and all of the parameter values used in the model are provided in the Supplementary Methods and Supplementary Table 5. The computational domain is rectangular with 20 × 20 × 5 µm in the x, y and z directions. A periodic boundary condition exists only in the x and z directions. In simulations, the motions of the cylindrical elements are updated at each time step through the Langevin equation and the Euler integration scheme. Three types of matrices were used in this study: fibrillar matrix; long, tight-bundle matrix; and short, loose-bundle matrix. Fibril assembly is initiated by the nucleation of seed fibrils through the appearance of one cylindrical segment in random positions, followed by elongation up to either 3 µm or 5 µm through the addition of segments without consideration of depolymerization. Effective collagen concentration calculated using the specific volume of collagen (0.73 ml g−1) is 3.65 mg ml−1. The fibrillar matrix is constructed with cross-linkers in the absence of bundlers (Extended Data Fig. 4d,h). During the assembly of the fibrillar matrix, individual fibrils are interconnected by cross-linkers of which two binding sites bind to any part of two fibrils. There is no preference for a cross-linking angle in this process. The long, tight bundle matrix is assembled with cross-linkers in the presence of bundlers that connect individual fibrils in a parallel, staggering manner into tight bundles (Extended Data Fig. 4g,i). Cross-linkers connect fibrils within each bundle or fibrils belonging to different bundles. The short, loose-bundle matrix is created by bundlers that connect fibrils at their ends with a specific angle (Extended Data Fig. 4e,f,j). While the bundlers are permanently bound to fibrils, the cross-linkers can unbind from fibrils at a rate that exponentially increases with an increasing force, following Bell’s law. Fibrils are permanently bound to two boundaries normal to the y direction (that is, +y boundary located at y = 20 µm and −y boundary located at y = 0 µm). After completion of matrix assembly, 20% strain is applied to the +y boundary in the x direction, whereas the −y boundary is fixed. After reaching the 20% strain, the strain is held at a constant level to measure stress relaxation.

51At least three biological replicates were performed for all in vivo experiments, and in vitro experiments were repeated at least three times. Data are presented as mean ± s.e.m. Statistical analyses were performed using GraphPad Prism v.10.0.3 (GraphPad Software). Normality distribution was assessed using the Kolmogorov–Smirnov test. Two-tailed unpaired t-tests and one-way ANOVA with Tukey tests were used to analyse data with a normal distribution. Data with a non-normal distribution were analysed using Wilcoxon rank-sum tests and Kruskal–Wallis tests with Dunn’s test. P < 0.05 was considered to be statistically significant.

52Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

53Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-023-06991-9.