Modular Micro-Physiological Human Tumor/Tissue Models Based on Decellularized Tissue for Improved Preclinical Testing

High attrition rates associated with drug testing in 2D cell culture and animal models stress the need for improved modeling of human tumor tissues. In previous studies, our 3D models on a decellularized tissue matrix have shown better predictivity and higher chemoresistance. A single porcine intestine yields material for 150 3D models of breast, lung, colorectal cancer (CRC) or leukemia. The uniquely preserved structure of the basement membrane enables physiological anchorage of endothelial cells and epithelial-derived carcinoma cells. The matrix provides different niches for cell growth: on top as monolayer, in crypts as aggregates, and within deeper layers. Dynamic culture in bioreactors enhances cell growth. Comparing gene expression between 2D and 3D cultures, we observed changes related to proliferation, apoptosis and stemness. For drug target predictions, we utilize tumor-specific sequencing data in our in silico model, finding an additive effect of metformin and gefitinib treatment for lung cancer in silico, validated in vitro. To analyze mode-of-action, immune therapies such as trispecific T-cell engagers in leukemia or toxicity on non-cancer cells, the model can be modularly enriched with human endothelial cells (hECs), immune cells and fibroblasts. Upon addition of hECs, transmigration of immune cells through the endothelial barrier can be investigated. In an allogenic CRC model, we observe a lower basic apoptosis rate after applying PBMCs in 3D compared to 2D, which offers new options to mirror antigen-specific immunotherapies in vitro. In conclusion, we present modular human 3D tumor models with tissue-like features for preclinical testing to reduce animal experiments. This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 International license (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution and reproduction in any medium, provided the original work is appropriately cited. # contributed equally; § shared senior authors

Medicine (Jannasch et al., 2015;Linke et al., 2007;Schanz et al., 2010) (Ref, DE 30 2014 008, Int.: 1 255 634, US: 79168704)), which is derived from an intestinal porcine tissue matrix. When seeded with tumor cells, it is referred to as OncoVaSc-TERM ® . The BioVaSc-TERM ® -technology and its derivatives are already being used for vascularized skin equivalents as well as for intestinal and airway models (Groeber et al., 2016;Schweinlin et al., 2017). As the decellularized and sterilized intestine of one single pig is sufficient to generate about 150 3D tumor models, the approach minimizes the use of animals for research.
The extremely versatile biological matrix "Small Intestinal Submucosa with mucosa" (SISmuc) is produced by removing all vessel structures and offers different tissue niches: i) on the surface of villi, ii) within former crypts, and iii) within deeper layers of the collagen matrix. As a unique feature of the matrix, structures of the basement membrane are preserved after decellularization, which enables physiological anchorage-dependent growth of epithelial cells (Stratmann et al., 2014). Furthermore, this structure enables the testing of substances that are intended to eliminate invasive tumor cells once they cross this natural border. Early steps in metastasis formation can also be studied in order to understand underlying mechanisms and to identify possible points of action. Our matrix-based lung cancer models were already shown to accurately reflect drug responses upon targeted therapies observed in the clinic -not just better than 2D, but partly also more reliably than animal models (Göttlich et al., 2018;Stratmann et al., 2014).

Applications of 3D tumor tissue models
Building on these previous findings, we now characterized our modular micro-physiological 3D tissue models using microarrays and immunohistochemical staining for stem cell markers to analyze differences between 2D and 3D conditions in more detail. Supported by in silico analyses, our 3D test system offers a time-efficient way to test effects of drug combinations on complex tissues, to analyze effects on signaling pathways, and to identify interdependencies. As an example, we present a new combination therapy for lung cancer treatment combining metformin and gefitinib. Our 3D tissue culture test system not only validates the prediction. It also allows functional tests to identify requirements for a synergistic response.
The modular approach further enables the addition of hECs to 3D tissue models. These hECs were seeded onto the former serosal side of the decellularized intestine so that toxic effects on the vasculature and the diapedesis behavior of immune cells can be addressed. To include the stromal compartment, we established different co-culture models with primary fibroblasts and found that dynamic cell culture within flow bioreactors overcomes problems with poor stroma development in vitro (Nietzer et al., 2016). In the context of immune therapeutic approaches, we observed an enhanced efficacy compared to 2D culture of MHCI independently acting trispecific T-cell engagers in a 3D leukemia model with artificial endothelium. Moreover, CRC 3D models show a drastically reduced induction of tumor cell apoptosis by allogeneic peripheral blood mononuclear cells (PBMCs) in monocultures and also in fibroblast-enriched co-culture mod-Innovative three-dimensional (3D) approaches reflecting relevant aspects of the in vivo situation reach from transwell cultures, spheroids, synthetic or biological matrices, 3D printed models and organoids to whole organ-on-a-chip models (Albritton and Miller, 2017;Katt et al., 2016;Nath and Devi, 2016;Santo et al., 2017;van Duinen et al., 2015;Weeber et al., 2017;Xu et al., 2014). As so far no "magic bullet" exists for all cancer types, stages and patients, no single test system can be expected to meet the testing requirements of all different therapeutic approaches. Depending on the respective questions and the therapeutic strategies to be assessed, the required test systems have to be specifically adapted. More advanced therapeutic options, including targeted and immune therapies and their combinations with more conventional chemotherapeutics or radiation approaches, require suitable complex models. Optimized in vitro models could also reduce the large number of animals required to analyze different combinations of compounds.
There are, however, defined features that are commonly advantageous for reliable models. A fully human setting avoids extrapolations across species boundaries. A gradual 3D growth of cancer cells is more likely to result in structures similar to those found in tumors in vivo. Due to their unphysiologically high proliferation index (often 80 to 90%), 2D models generally lead to an overestimation of cytostatic effects (Cree et al., 2010). General tissue architecture, which includes response-modulating accessory cells as found in the tumor microenvironment, can only be replicated in complex 3D models. Moreover, a loss of tumor stroma over time, which is frequently observed in transplantable in vivo tumor models as well as in organoids (Yada et al., 2018;Kim et al., 2019), may render these test systems insensitive to important contributions from the stromal compartment. Serious limitations also apply for the testing of immunotherapeutic agents in humanized animal models: Apart from the enormous cost and effort, species-specific differences of cytokines typically result in an impaired maturation and differentiation of certain human immune cell subsets in a murine microenvironment (Rongvaux et al., 2013(Rongvaux et al., , 2014Song et al., 2019;Walsh et al., 2017;Anselmi et al., 2020).
Being able to monitor immune-mediated tumor control in a predictive, fully human in vitro system will thus significantly advance preclinical research.

3D tumor tissue models
The present study aims to establish high-output modular 3D tissue drug test systems that model tissue niches sufficiently well to enable a reliable testing of drug responses. While investigations of homophilic cell-cell-interactions, cell migration, invasion, and changes of signaling pathways may also yield reasonably reliable results in 2D tumor cell culture models, interactions of tumor cells with different other cell types such as fibroblasts, hECs and immune cells are challenging to explore in such 2D systems.
In this study we present modular human 3D tumor models for lung, colorectal, and breast cancer as well as leukemia. Their complexity can be adapted according to the therapeutic question. All models are based on the patented BioVaSc-TERM ® (Biological Vascularized Scaffold -Tissue Engineering Regenerative els. This indicates that complex 3D models with allogeneic cells offer a therapeutic window for testing antigen-independent immune-stimulatory or -inhibitory agents.
These promising advantages of our modular 3D culture approach should reduce attrition rates and save animal experiments on account of its more physiologic tissue features, which are examined in detail in the following.

Animals
The matrices SISmuc and SISser (used for the leukemia model with an endothelial barrier) were prepared from porcine jejunal segments of young pigs (German landrace, local supplier). The animals used for organ removal were not bred specifically for research purposes but obtained from a local supplier who breeds animals for food production. SISmuc consists of small intestine submucosa (SIS) and mucosa (muc). SISser consists of SIS and serosa (ser). All explantations were in compliance with the German Animal Protection Laws ( §4 Abs. 3), and all animals received humane care in compliance with the guidelines of the FELASA, WHO and FDA (WHO-TRS978 Annex3 und FDA-OCTGT Preclinical Guidance) after approval by our institutional animal protection board (registration reference number #2532-2-12, Ethics Committee of the District of Unterfranken, Würzburg, Germany). Pigs were anesthetized and then euthanized prior to surgery, therefore not suffering from pain or distress. To keep the number of sacrificed animals low, many different organs were removed for various research projects.

Matrix preparation
Chemical decellularization of jejunal segments was performed according to standardized protocols (Jannasch et al., 2015;Linke et al., 2007;Schanz et al., 2010). The jejunum was explanted, rinsed and chemically decellularized with a sodium deoxycholate monohydrate solution. Afterwards the vascular tree was manually removed, and the product was sterilized (γ-sterilization).

Primary cells and cell lines
Primary human cells were obtained with informed consent according to ethical approval granted by the Institutional Ethics Committee of the University Hospital Würzburg (approval number 182/10).
All cells were cultured under standard culture conditions (37°C, 5% CO 2 ). We used HROC24 and HROC87 cells (kindly provided by Dr Michael Linnebacher, University Hospital Rostock) (Mullins et al., 2019) cultured in DMEM/F12 medium with Glu-taMAX™ (Gibco/Life Technologies, USA) supplemented with 10% FCS. HCC827, MCF7 (both purchased from DSMZ), MDA-MB-231 cells (purchased from ATCC), THP-1 (purchased from DSMZ) and SW480 (purchased from DSMZ) were cultured in RP-MI 1640 medium with GlutaMAX™ (Gibco/Life Technologies, USA) supplemented with 10% FCS, except for HCC827 cells, which were cultured with 20% FCS. Human dermal fibroblasts were isolated from the dermis of foreskin biopsies and cultured in DMEM with GlutaMAX™ (Gibco/Life Technologies, USA) supplemented with 10% FCS and 100 mM sodium pyruvate. Primary microvascular hECs were isolated from the dermis of foreskin biopsies and cultured in VascuLife Media (Lifeline Cell Technology, USA). Endothelial cell cultures were morphologically confirmed to contain no fibroblasts after isolation from a skin biopsy. If fibroblasts were found in the cultures, cells were sorted using magnetic associated cell sorting beads directed against the endothelial marker CD31 (Miltenyi, Germany). Fibroblasts and hECs were cultured in the presence of 1% penicillin/streptomycin (P/S) after isolation until the first passaging was performed. PBMC for trispecific antibody treatment (delivered by PD G. Stuhler, HLA-A2 negative, CMV positive) were frozen at -80°C and thawed overnight prior to use in RPMI with 10% FCS and 1% P/S. PBMC were generated from buffy coats via Ficoll gradient isolation and cultured in RPMI 1640 medium with GlutaMAX™ (Gibco/Life Technologies, USA) with 10% FCS.
We applied the culture conditions recommended by the supplier to the respective cell lines and used the lowest quantity of FCS that is feasible for the different cell types to reduce FCS consumption. So far, the used cells did not grow with defined FCS-free media. In case of co-culture models, equal parts of cell-specific media were mixed. Cell lines were used in passages up to 20, while primary cells were only used until passage 6. Tumor cell lines were passaged twice a week, and medium was changed every 2 to 3 days. Cells were checked for mycoplasma contaminations every 8 to 10 weeks by PCR (Minerva Biolabs, Germany).

Preparation of tumor models
Single pieces of SISmuc or SISser matrix were fixed between two metal rings (cell crowns) and seeded with 100,000 tumor cells on the mucosal side or with 300,000 hECs on the submucosal side. To create a tight endothelial barrier, hECs were seeded 3 to 5 days prior to seeding with tumor cells. Tumor models were placed in 12-well plates with 1 mL medium in the inner compartment and 1.5 mL in the outer compartment of the cell crown. For static cell culture, the models were cultured for 11 days before treatment with the test substance. During static cell culture, medium was changed every 2 to 3 days. For dynamic culture, the models were placed in bioreactor chambers after 3 days of static pre-culture. The customized bioreactors used, described in (Nietzer et al., 2016), were connected to a peristaltic pump and a medium reservoir containing 45 mL cell-specific medium. The pump ensured a constant medium flow of 3-4 mL/min, avoiding additional sheer stress to the cells. After 14 days of dynamic culture, the models were treated for 3-5 days with the test substance. During dynamic cell culture, medium was changed every 7 days. All test substances were administered via the cell culture medium.

Treatment of cells in 2D and 3D
For 2D cell culture, cells were cultured in 12-well plates. Tipifarnib (Selleck Chemicals, USA) in DMSO was applied in different concentrations to the cell culture medium for 5 days with a medium change on the third day. For 3D cell culture, hECs were seeded on SISser and cultured for 14 days prior to adding THP-9000, Keyence, Japan). DAPI was counted automatically using the software BZ Analyzer (Keyence, Japan) to determine the total cell number. Ki-67-positive cells were counted manually. For each image, the number of Ki-67-positive cells in percent of total cell number was calculated, and the mean of each sample was used for statistical analysis.

Statistical analysis
All pairwise comparisons were assessed using a Student's t-test for unpaired samples. Results were corrected for multiple testing using Bonferroni's adjustment. P values < 0.05 were considered significant. Presentation of the flow cytometry data in the coordinate system was realized by SigmaPlot Version 10.0 (Systat Software, Inc.). M30 and Ki-67 quantifications were analyzed and presented with R 1 .

Histological and immunohistochemical staining
As an endpoint measurement, the tumor models were fixed in a 4% paraformaldehyde solution for 2 h, embedded in paraffin, and cut in a microtome at 3-5 µm. Hematoxylin and eosin (H&E) staining (Morphisto, Germany) was performed according to the manufacturer's protocol. Non-immunofluorescent immunohistochemical staining was performed using the 3,3'-diaminobenzidine (DAB) system (DCS Innovative Diagnostik-Systeme, Germany) according to the manufacturer's protocol. Here, DAB is oxidized in the presence of a horseradish-peroxidase and hydrogen peroxide, resulting in the formation of a brown dye at areas of antibody binding.
Primary antibodies were diluted 1:100 in antibody diluent (DCS Innovative Diagnostik-Systeme, Germany) and incubated over night at 4°C. As secondary antibodies, donkey anti-mouse IgG (Thermo Fisher Scientific Cat# A-31571, RRID: AB_162542) conjugated to Alexa-647 or donkey anti-rabbit IgG (Thermo Fisher Scientific Cat# A-31572, RRID: AB_162543) conjugated to Alexa-555 were diluted 1:400 in antibody diluent and incubated for 1 hour at room temperature (RT). Cell nuclei were counterstained using 4′,6-diamidino-2-phenylindole (DAPI), which binds to DNA. DAPI was diluted in the embedding medium Fluoromount-G (Thermofisher Scientific, USA). Images were taken using a digital microscope (BZ-9000, Keyence, Japan). 1 cells to the medium on day 14, with or without tipifarnib, and cultured for the next 5 days (with a complete medium change on day 17). In case of the trispecific antibody treatment (AB1: VLαCD3-scFvαHLA-A2; AB2: VHαCD3-scFvαCD45; antibodies provided by PD G. Stuhler), antibody 1 and/or 2 (each 30 nM) and 200,000 PBMC as apoptosis inductors in THP-1 cells via trispecific antibodies were added to 200,000 THP-1 cells in each well of the 12-well plate (in Vasculife ® medium with 1% P/S) and incubated for 24 h. The setting in the 3D model was the same as in the 2D model with the difference that, where indicated, the cell + DMSO/trispecific antibodies mixture was added to each cell crown onto the surface of the 3D endothelial monolayer. Treatment with 10 mM metformin (Merck, Germany) and 1 µM gefitinib (Absource Diagnostics, Germany) of 3D models started on day 11 for 3 days until day 14 of 3D cell culture.

Flow cytometry
Cells in the supernatant were resuspended and transferred to a reaction tube. For cell death quantification, cells were centrifuged (4°C, 300 x g, 5 min) and washed in sterile PBS with 0.5% BSA. Cells were resuspended in 500 µL PBS with 0.5% BSA, and 5 µL 7-aminoactinomycin D (7-AAD) (Biozol, Germany) was added per sample. 7-AAD intercalates into the DNA and is used to distinguish apoptotic cells (7-AAD+) from viable cells (7-AAD-). After incubation in the dark on ice for 10 min, FACS measurements were performed (FACSCalibur, BD, Germany). Data analysis was performed with the software FlowJo 6.0. For treatment with trispecific antibodies, we confirmed that the primary hECs we used for the model reacted with the anti-HLA-A2 antibody (BD Pharmingen, Germany) in flow cytometry prior to reseeding the scaffold.

TUNEL Assay
To detect DNA damage in hECs, the in situ Cell Death Detection Kit, POD (Roche, Switzerland) was used according to the manufacturer's instructions. Staining was visualized using a BZ-9000 microscope (Keyence, Japan).

Apoptosis measurements using M30 ELISA
For quantification of epithelial apoptosis, the M30 CytoDeath™ (PEVIVA ® , TECOmedical, Germany) assay was used to measure the caspase 9-cleaved product of cytokeratin 18 in the cell culture supernatant every 24 h during treatment. This neo-epitope is specifically expressed by apoptotic epithelial cells. The M30 CytoDeath™ assay was performed according to the manufacturer's protocol. Supernatants of static cell cultures were obtained from the inside of the cell crowns. Supernatants of dynamic cell cultures were obtained via a sampling port integrated into the tubing of a flow bioreactor.

Quantification of proliferation
Ki-67 staining was quantified as follows: At least 6 images of each sample were taken with a fluorescence microscope (BZ-network reconstruction was combined with dynamic simulations of cellular pathways: First, the network topology was created and edited in CellDesigner (version 3.5.1) (Matsuoka et al., 2014) and exported as xml-file. Importantly, the Boolean logic of the network was considered, i.e., activating and inhibitory interactions between receptors, proteins and protein cascades. Modifying crosstalk was also implemented. Next, dynamic simulations were conducted applying the SQUAD tool (Di Cara et al., 2007). SQUAD reads the topology of the network using the xml file generated from CellDesigner. The ON-and OFF-states of the network were also considered for dynamic modeling. SQUAD automatically interpolates between the different network states using exponential functions. To model signal propagation in the network, SQUAD further considers the Boolean network topology, concatenating the exponential functions accordingly. The trajectories of full or partial activation down to no inhibition were calculated for the whole network and for all included proteins. Different mutational profiles and treatment were integrated into the dynamic simulation via the SQUAD perturbation function. Protocols were saved in prt file format. Outcomes were assessed by readouts proliferation and apoptosis as basic markers and considering all available data (in particular phosphorylation status of involved proteins and protein cascades).
Module omics data analysis: Data from microarray experiments was preprocessed in the R v3.2.1 environment with Bioconductor packages affy and vsn; VSN normalization and RMA probeset summary were performed with the vsnrma command using default parameter settings. Fold change values of treated (test dataset) vs untreated (reference dataset) were calculated and log2-transformed for 2D and 3D samples individually. Cut-off values for significance were set to ± 1.0 log2 fold change. Gene transcripts that showed differential expression upon treatment in HCC827 but not in A549 cells were selected for further analysis. Differentially expressed genes (DEGS) that showed up-or downregulation in HCC827 in 3D but not in 2D in response to gefitinib were clustered into gene lists, respectively. The functional enrichment analysis using the DEGS was performed using g:Profiler web tool (Raudvere et al., 2019), and relevant enriched terms (adj. p value < 0.05) were shown as table and plotted using ggplot2 package version 3.3.2 (Wickham, 2009) in R version 4.0.2 (according to (Fuchs et al., 2020)).

Generation of micro-physiological human tumor models on a porcine tissue matrix
In order to generate human tumor models representing the in vivo situation more closely than 2D cell cultures or simple spheroids, we used pieces of the porcine intestinal jejunum deriving from the BioVaSc-TERM ® . These comprise an intestinal loop ( Fig. 1A) with an arterial entrance (red arrow) and a venous outward flow (blue arrow) and enable reseeding of the whole vascular tree with hECs. For the 3D tumor tissue models, only the former gut wall without the vessel system is used after decellularization. The porcine matrix consists of three different main

Ultrastructural analysis
Samples were washed with pre-warmed PBS with calcium and magnesium, and cell-free edges were removed with a scalpel prior to fixation in a 6.25% or 2.5% solution of glutaraldehyde overnight at 4°C. Further sample preparation was performed at the Imaging Core Facility, Biocenter, University of Würzburg. For ultrastructural analysis of the unseeded scaffold as well as the 3D tumor models, scanning (SEM) electron microscopy images were also taken at the Imaging Core Facility, Biocenter, University of Würzburg. To analyze changes in scaffold structure after cell culture, the tumor models were decellularized after 14 days of cell culture. To this aim, the samples were incubated twice in sodium deoxycholate (1 h at RT), washed twice for 10 min in PBS, incubated in DNase I solution for 2 h at 37°C and rinsed in PBS prior to fixation.

Microarray data
For RNA isolation with the RNeasy Micro Kit (Qiagen, Germany) from cells in 3D experiments, frozen SISmuc stored at -80°C was used. Isolation was carried out according to the manufacturer's protocol. While still frozen, tissue was cut into small pieces with a scalpel and transferred into 350 µL Buffer RLT in a 2 mL reaction tube. Heat-sterilized steel beads were added to the tubes, and the tissue was lysed at 50 Hz for 5 min with a tissue lyser device. Samples were centrifuged at full speed for 3 min, and supernatants were transferred to a new 1.5 mL tube. One volume of RNAsefree 70% ethanol was added, and samples were mixed by pipetting. Following transfer, complete contents were centrifuged in RNeasy MinElute spin columns for 15 s at 8,000 x g. Further steps were carried out according to the manufacturer's protocol.
Isolation of RNA from cells from 2D experiments was done according to the RNeasy Mini Kit protocol. RNA concentrations were determined by absorbance measurements at 260 and 280 nm using the NanoQuant Plate for the Infinite M200 plate reader (Tecan, Switzerland). Quality control of RNA samples using the Agilent 2100 Bioanalyzer and related reagents as well as all microarray experiments and data analysis were performed by Claus-Jürgen Scholz and colleagues from the Core Unit Systems Medicine at the Interdisciplinary Centre for Clinical Research (IZKF), University of Würzburg. In brief, 100 ng of RNA was first reverse-transcribed to cDNA, before cDNA was re-transcribed into labeled cRNA using the Affymetrix IVT Plus Kit. Subsequently, 130 µL hybridization mix prepared from labeled cRNA fragments was hybridized for 16 h at 45°C and 60 rpm to GeneChip PrimeView arrays. Microarrays were washed and stained in the Affymetrix Fluidics Station 450. Microarrays were scanned using a GeneChip Scanner 3000 7G (Affymetrix, USA). The data were deposited with GEO database under accession number GSE79078.
Bioinformatics and in silico modeling Module 3D tissue simulations: Signaling network reconstruction was based on available literature and on biochemical and human interactome database sources such as KEGG and IMEx. To model individual drug actions as well as effects of combinations, dynamic simulations of cellular pathways also considered known impacts of different drugs. For data-driven modeling, signaling Cells grown in static cell culture reach a state similar to tissue homeostasis after about 11 days of culture. In the absence of major changes in proliferation and apoptosis, cells can then be used for drug tests over a period of approximately 7 days (Fig. S1 2 ). While dynamic models in bioreactors have a lower through-put, they benefit from better nutrient supply and waste disposal via the medium flow, thus mimicking in vivo conditions more closely. As dynamic culture improves cell growth and maintenance, the duration of drug application can be extended for up to several weeks.

The tissue matrix
After the decellularization process as described before (Jannasch et al., 2015;Linke et al., 2007;Schanz et al., 2010), the main part of the SISmuc matrix is composed of collagen fibers covered with the structurally preserved basal lamina on the mucosal side of the matrix ( Fig. 2A). The villi structures are intact (Fig. 2B). With-layers ( Fig. 1B): (i) the mucosa, (ii) the small intestinal submucosa (SIS) and (iii) the serosa. Subsequently, two different tissue matrices can be derived: the SISmuc is composed of the former SIS and the mucosa, while the SISser contains the former SIS and the serosa. As a unique feature, structures of the basement membrane are preserved on top of the mucosa after the chemical decellularization process.
Both matrices can be fixed between two cell crowns, resulting in inserts for 12-well plates similar to a transwell system (Fig.  1C,D,E). They can also be fixed in flow bioreactors for dynamic cell culture (Fig. 1E). Different designs of cell crowns are possible: either they can be used upside down for two-sided cell culture (white cell crown in Fig. 1C) or upright for one-sided cell culture (brown cell crown in Fig. 1C). Dynamic culture in bioreactors enhances cell growth on the matrix (blue in Fig. 1F) compared to static cell culture (Fig. 1F). muc. Different stages of cancer cell invasion can be mimicked depending on culture condition and cell line. MCF7 breast cancer cells clearly grow on top of the matrix and inside the former crypt structures under static (Fig. 3A) and dynamic culture conditions (Fig. 3C). In contrast, MDA-MB-231 breast cancer cells invade into the whole matrix under static (Fig. 3B) and dynamic culture conditions (Fig. 3D) as shown before (Wallstabe et al., 2019). Immunohistochemical staining of pan-cytokeratin (PCK, green in Fig. 3E,F) and vimentin (Vim, red in Fig. 3E,F) reveals their different epithelial-to-mesenchymal transition (EMT) status. Dynamic conditions promote the invasive character of MDA-MB-231 cells, leading to an increased tumor cell mass within the mucosal part of the SISmuc (Fig. 3D) as well as in the basal membrane, pores ranging from 1 to 2 µm are visible (Fig. 2C). These pores in the basement membrane have a sharp border and overall rounded shape, suggesting that they are naturally occurring pores rather than damage deriving from the decellularization process. In contrast, the former serosal side of the matrix shows no intact basement membrane structure: Only single collagen fibers can be observed clearly at higher magnification (Fig. 2D,E).

Differentiation between invasive and non-invasive tumor cells
The invasive behavior of cells can be classified by investigating cell growth morphology when seeded on the surface of the SIS-  cer models based on the tissue matrix technology (Göttlich et al., 2018;Stratmann et al., 2014). In order to reveal involved pathways and biological functions, we now performed microarray-based gene expression analyses of the two lung cancer cell lines A549 and HCC827 in 2D and 3D and compared the respective differences. To adjust for the qualitative difference in RNA isolated from either 2D or 3D culture and to focus on genes relevant for drug response, we compared the ratios of samples treated or untreated with gefitinib, an EGFR tyrosine kinase inhibitor (TKI), under 2D or 3D culture conditions. As HCC827 cells harbor an activating mutation in the EGFR gene, while A549 cells do not, only HCC827 were expected to respond to this drug. Gene transcripts with differential expression upon gefitinib treatment in HCC827 but not in A549 cells were selected for further gene ontology (GO) analysis (Tab. 1; increased Vim expression (red in Fig. 3F), whereas non-invasive MCF7 cells only show strong PCK expression (green in Fig. 3E).

Enhanced tumor stem cell marker expression under 3D conditions
Cancer stem cells are discussed to present the key subpopulation that should be targeted efficiently for long-term therapeutic success. For CRC model generation, we used two early passage cell lines recently isolated from patients with different tumor stages (HROC24: low grade, HROC87: high grade). HROC24 cells grow mostly as monolayers on top of the matrix and within crypt structures, whereas the more malignant HROC87 form aggregates on top and in former crypt structures of the intestinal matrix. In order to compare stem cell characteristics between 2D and 3D culture conditions, we investigated the expression of the tumor stem cell markers CD44, CD133, CD166 and EpCAM (Willis et al., 2008). In both cell lines, expression of all stem cell markers was only found in 3D conditions, whereas in 2D conditions CD44 and CD166 were not detected in either cell line (Fig. 4).

Differences in gene expression upon targeted therapy under 2D conditions and in 3D tissue models
In previous studies, we have shown an improved correlation of drug responses with clinical observations in 3D lung can-

Fig. 4: 3D cell culture of colorectal cancer (CRC) cells on the SISmuc matrix increases the expression of cancer stem cell (CSC) markers
Glass slides from 2D and paraffin-embedded sections from 3D cultures were immunohistochemically stained for the stem cell markers CD44 (A to D), CD133 (E to F), CD166 (I to L) und EpCAM (M to P). While the CRC cell lines HROC24 and HROC87 are negative for CD44 (A, B) and CD166 (I, J) under conventional 2D conditions, both cell lines express these markers when cultured on SISmuc (C, D, K, L). Scale bars given in M to P: 100 μm (same scale for A to P), representative images of n = 3.

Cell proliferation
Negative regulation of cell proliferation 9 5.72•10 -3 a The table shows selected biological processes and pathways that are deregulated in 3D but not in 2D. Proliferative processes and associated metabolism are down-regulated under 3D conditions. The number of deregulated genes that are involved in each process is shown. Adjusted (multiple testing corrected) p-values are given on the right. Analyses are based on microarray data from 1 experiment in each condition (n = 1).
metabolism and ADP metabolism, which are all linked to the Warburg effect) and downregulation of DNA replication (replication initiation, strand elongation, mitotic spindle organization, chromosome condensation). These highly significant observations are in line with the previously observed reduced cell proliferation under 3D conditions. Upregulated genes are, in contrast, linked to positive regulation of apoptotic processes, to the extracellular matrix (ECM) (epithelium development, ECM organization, tissue morphogenesis), to cell differentiation, migration and signal transduction. An increase of tran-pathway enrichment with bubble-plots for pathways upregulated in 3D compared to 2D in Fig. S3 2 and downregulated pathways in Fig. S4 2 ). Gene lists that showed down-and upregulation in 3D models compared to 2D conditions were subjected to an enrichment analysis. A high number of downregulated genes in 3D culture suggests more stringent cell cycle control (cell division, cell cycle checkpoint, G1/S phase transition, cell proliferation), fewer cancer-like metabolic processes (less pronounced expression of pathways such as canonical glycolysis, pyruvate tions is of considerable interest. We thus generated models including cell types from the tumor microenvironment such as hECs, immune cells, and fibroblasts. This enables the analysis of new substances and immune therapies. Toxic side effects on non-cancerous cells also can be investigated. In order to test recently described trispecific hemibodies which target CD45 positive tumor cells in HLA-A2 positive leukemia patients, we seeded HLA-A2 positive hECs on the serosal side of our biological matrix and co-cultured this artificial endothelium with the human HLA-A2 and CD45 dual-positive leukemia cell line THP-1 (Bersi, 2017). In order to assay efficacy and specificity of these hemibodies as an antigen-specific immunotherapeutic approach (Banaszek et al., 2019) in a tissue-like context, we analyzed the induction of apoptosis in THP-1 cells by flow cytometry and compared results from 3D to conventional 2D cell culture. In addition, we compared the efficacy to treatment with the clinically investigated farnesyltransferase inhibitor tipifarnib (Stieglitz et al., 2015). Hemibodies can only recruit T-cells towards dual-positive CD45 and HLA-A2 positive leukemic cells when both antibody parts (AB1 and AB2) bind their respective targets in close proximity. They then assemble to form a functional antibody that also binds to CD3 on T-cells. After addition of HLA-A2-negative PBMCs as effector cells, tumor cell lysis was induced (Fig. 6E,F). Importantly, after 5 days of tipifarnib treatment, we observed the induction of apoptosis in a dose-dependent manner in THP-1 cells in the 3D model and severe damage of the endothelial monolayer confirmed by CD31 staining (Fig. 6D), which, in contrast, was not affected by trispecific hemibody treatment (Fig. 6H). Furthermore, the hemibody treatment yielded a higher efficacy in the 3D model (60% apoptotic THP-1 cells after only 24 hours of treatment compared to 40% apoptotic THP-1 cells in 2D) as shown in Figure 6E and F. Moreover, the endothelium can be investigated as a barrier for immune cells about to enter the tumor, which is an important issue in immune therapies (Galon et al., 2006). While the endothelial monolayer is shown in green by CD31 staining, tumor cells are stained only with DAPI (Fig. 6I ). With the help of flow bioreactors, the diapedesis of immune cells, in this case PBMCs (shown in green), was investigated under dynamic medium flow conditions (Fig. 6J).
Finally, by adding human dermal fibroblasts to different tumor cell lines cultured in our 3D models, we observed different effects on tissue morphology. HCC827 lung tumor cells and A549 cells (Fig. 7A,B in green) showed largely unaltered cell growth patterns compared to monocultured tumor cells. The added fibroblasts simply migrated to deeper matrix areas, while the tumor cells stayed on top of the matrix (Fig. 7C,D in red). Astonishingly, in co-culture with colorectal SW480 cancer cells, non-cancerous fibroblasts induced tumor cell in-scripts related to inhibition of cell proliferation was observed. All differentially expressed genes with fold-change values are listed in the supplementary file (Tab. S1 3 , raw data, and Tab. S2 4 , GO-term analysis results).

Simulation module: in silico synergistic effect of gefitinib and metformin validated including mutation constellation in 3D tissue lung cancer models
To demonstrate the potential of our 3D model to support exploratory pre-clinical research, bioinformatic simulations were used to discover new targets and promising treatment combinations for CRC and NSCLC (Baur et al., 2019;Göttlich et al., 2018). The work of Göttlich et al. (2018) pointed towards 5' AMP-activated protein kinase (AMPK) activation as a potential treatment strategy for lung cancer. As metformin acts (among other effects) also as an AMPK activator, we tested the combination of gefitinib and metformin on HCC827 cells. In line with our bioinformatic prediction, we observed a significantly increased effect on apoptosis induction for the combination (Fig. 5A). In contrast, proliferation was only affected by gefitinib, but not by metformin, and no enhanced effect on the inhibition of proliferation could be measured for the combination (Fig. 5B).
These data were integrated into a signaling network topology of HCC827 cells (Fig. 5C). We then simulated the dynamics of the signaling network (Fig. 5D). All proteins and their activities change over time, and effector nodes for key cellular responses are calculated in this simulation. The change over time for key cellular responses (proliferation, apoptosis) and central proteins (p53, mTOR, AMPK, etc.) is shown in Figure 5D. This allows an estimation of the cellular responses, for instance enhanced induction of the node "apoptosis" by the combination therapy compared to gefitinib monotherapy, whereas proliferation shows a similar profile in both treatment scenarios (Fig. 5D).
Thus, comparing the pharmacological effects in our in vitro 3D cell culture model (Fig. 5A,B) with the in silico prediction shows concordant results. The combination of in vitro and in silico modeling allows to accurately relate the tumor mutation constellation with the most appropriate drug combination. Moreover, as the modular in vitro test system as well as the in silico simulation module are generic, the approach can be used for testing many other defined drug targets, drugs or drug combinations beyond the example shown in this manuscript.

Modular refinement of the 3D tissue models by adding different cell types of the tumor microenvironment
Further cellular components such as the tumor stroma impact drug responses. Given the strong immunosuppressive effects exerted by the tumor stroma, a targeting of tumor-stroma interac-in mono-and co-cultures with the allogenic immune cell populations. Under 3D culture conditions, the PBMC-mediated apoptosis in SW480 mono-cultures was reduced in all samples compared to 2D cultures (Fig. 7G), where variation was very high compared to 3D data. The same effects could be seen in SW480 co-cultures with fibroblasts (Fig. S2 2 ). vasion into deeper matrix layers, which was accompanied by formation of tumor cell aggregates and by a partial destruction of former crypt structures, indicating the generation of a desmoplastic reaction (Fig. 7E,F). To explore whether this model can be applied for testing of immunotherapies, we added immune cells (PBMCs) and measured the induction of apoptosis leukemia tissue model the matrix SISser in which the small intestinal submucosa and the serosa are preserved.
Mina Bissel and others have stressed the huge importance of the microenvironment for physiological cell growth: isolated cells lose their functionality as cues from the microenvironment critically determine the cell fate (Bissell et al., 2003;Bissell, 1981). Advances in 3D technologies, particularly organ-

Discussion
We report here a modular 3D tissue culture system to generate improved human tumor models with an in vivo-like microphysiology. To this aim, we use a decellularized tissue matrix from porcine gut termed SISmuc in which the mucosa is preserved on top of the small intestinal submucosa. In addition, we used for a  A and B), which show a well-differentiated cellular morphology, co-culture (B) with dermal fibroblasts (red in B) has no morphological effect on the growth pattern of these tumor cells. (C,D) Also, in A549 cells (green in C and D), the co-culture with dermal fibroblasts (red in D) does not influence the growth pattern in 3D cell culture. (E,F) In contrast, when highly malignant colorectal SW480 tumor cells (green in E and F) are co-cultured with dermal fibroblasts (red in F), the formation of tumor clusters (white arrows in F) that overgrow former crypt structures is observed (n = 3). G: When allogenic PBMCs were added to SW480 tumor cells, allogenic (and thus largely antigen-independent) tumor cell killing was assessed in order to evaluate the potential usefulness of the models as test systems for MHCI-dependent immune-mediated tumor cell killing. In 2D (n = 6), the apoptosis induction was much higher (with a higher variation) than in 3D (n = 6) in mono-as well as in co-culture with fibroblasts (see Fig. S2 2 ). A to F: immunohistochemical staining of PCK (green) and vimentin (Vim). Nuclei are counterstained with DAPI (blue). Scale bar in F: 50 μm for A to F. All PBMCs derived from the same donor: n = 2 (5 technical replicates/experiment) for 2D and n = 3 for 3D.
Further aspects of the models should improve predictivity and mode-of-action analysis: (i) The 3D tissue architecture, its composition of ECM components, and the structure of the basement membrane and bioreactors lead to a more physiological growth of carcinoma cells according to their invasive traits, gene expression/pathway activity (apoptosis, proliferation) and stem cell markers. (ii) Semi-quantitative in silico simulations based on these gene expression data allow focused predictions of drug targets and drug combinations. (iii) Addition of cells from the tumor microenvironment enables testing of antigen-specific and MHCIindependent immune therapy and toxicity, and transmigration studies across the endothelium, including assessment of the impact of activated fibroblasts on invasion. Moreover, allogenic immune reactions are reduced in 3D tumor models.
For model generation, tumor cells are seeded on the luminal side -the mucosal side -of the decellularized intestine. Similar to epithelium-derived carcinomas, they are thus physiologically anchored to the basement membrane. We could show by ultrastructural analysis that only on the mucosal, but not on the serosal side, structures of the basement membrane are preserved as a closed layer, showing small, supposedly natural holes at the bottom of the former crypts. These holes are too small for undirected cell migration through them and were proposed to have functions in immune cell trafficking (McClugage and Low, 1984;Takahashi-Iwanaga et al., 1999;Takeuchi and Gonda, 2004). Holes in the basement membrane are also reported in the human bronchus after decellularization (Howat et al., 2002). In our experiments, we observed that the basement membrane represents a strong barrier for tumor cell growth in all investigated cancer entities when the tumor cells are non-invasive. Interestingly, the breast cancer cell line MDA-MB-231, which was formerly described to be highly invasive (Neve et al., 2006), overcomes this barrier and scatters over the whole mucosa in form of single cells as reported before (Wallstabe et al., 2019). In line with hypotheses on EMT and invasion (Zhang and Weinberg, 2018), these MDA-MB-231 cells show a more mesenchymal marker expression, indicating a higher EMT status than the non-invasively growing MCF7 breast cancer cells, as shown here for the first time on our matrix in a 3D tumor model.

Exploiting the modular micro-physiological 3D tissue models for testing immune therapies, fibroblasts and immune cells
Different cell types can also be used to study primary stromal components of tumors. As hECs do not only contribute to the tumor stroma by building a physical barrier but also by secreting soluble signaling molecules, they play an important role when it comes to drug responses in tumor cells (Gubbiotti et al., 2020;Meurette and Mehlen, 2018). Endothelial cells can be seeded on the basolateral side of the matrix to investigate diapedesis of immune cells. For this purpose, specific cell crowns were developed to enable a 3-day upside-down pre-incubation with hECs. These cell crowns can also be manufactured by 3D print-oids, provide new options in the field of more physiological and personalized models for drug testing and cancer research (Drost and Clevers, 2018;Simian and Bissell, 2017). Nevertheless, organoids have significant disadvantages like their inside-out topology that hinders practical analyses, the very expensive culture conditions due to the need of Matrigel ® and growth factors, and high donor variation 5 . Moreover, during organoid formation, dedifferentiation processes are inevitable. Stroma components are often not implemented into the organoids (Baker et al., 2016;Fan et al., 2019) or even get lost (Kim et al., 2019). Immune components can, in contrast, easily be added from peripheral blood (Dijkstra et al., 2018). Animal models such as patient-derived xenografts (PDX) present other challenges: Apart from ethical concerns, they are time-consuming and generally show high variability, which calls for high numbers of animals to achieve statistically meaningful results. While these practical limitations may be overcome, other problems caused by the species barrier are intrinsic to these models. For example, host factors from mice or rats in some cases fail to activate their corresponding receptors on human cells as we have, next to others, demonstrated before (Baur et al., 2019;Francone et al., 2007;Jeffers et al., 1996;Anselmi et al., 2020;Rongvaux et al., 2014).
Decellularized matrices provide complex cues that enhance cell growth, differentiation (Berger et al., 2020) and -as shown here -also stemness potential. The main advantage of our 3D tissue model system compared to organoids is its modular and standardizable set-up with a distinct orientation in a transwell culture plate. Thereby, transmigration of tumor and immune cells across the preserved basement membrane or the added endothelial barrier can be easily evaluated, as the monolayer integrity is visualized by CD31 staining.
To generate in vitro models close to the in vivo situation suitable for standardization and large-scale production, we use one single pig for the generation of about 150 tumor models out of one single intestine and generate tumor models based on a standardized operating protocol (Göttlich et al., 2016). We utilize the jejunum as in this part one intestinal loop is supplied with blood via one artery and one vein. This enables complete cell removal by thoroughly flushing the vessels with detergent as published before (Jannasch et al., 2015;Linke et al., 2007;Schanz et al., 2010). All investigated cell lines showed a stable proliferation index as well as apoptosis levels between day 11 and about day 18 of a static 3D cell culture, which enables one week of drug testing in a kind of tissue homeostasis. The proliferation index is generally reduced in 3D models in comparison to 2D models, which is more representative of in vivo conditions. In investigated cell lines, we find different indices by counting the percentage of Ki-67 positive cells as done in the clinic. Most cell lines range between 20 and 40 percent, which reflects observations from the clinic (Warth et al., 2014;Inwald et al., 2013). Thus, these models should reduce the risk for false-positive results in cytostatic drug testing (Cree et al., 2010). ported by the tumor stroma. This provides new options for drug testing in different tissue-like contexts. For example, it can be tested if substances interfere with the interaction of tumor cells with surrounding stroma cells and can thereby hinder tumor cell invasion.
An adjustment of in vitro to patient conditions is, for instance, suggested by the fact that our 3D model -similar to patientsdoes not imply a correlation between HSP90 inhibitor efficacy and a KRAS mutation in lung tumors (Göttlich et al., 2018). While such correlation had been suggested by 2D and animal models (Sos et al., 2009), clinical studies revealed this to be most likely an artifact (Socinski et al., 2013).

In which aspects are modular 3D tissue models an improvement?
To analyze differences between tumor cells cultured in 2D versus 3D in more detail, we first looked at stemness. CD44 functions as a receptor that binds to hyaluronic acid and other components of the ECM (Nusgens, 2010). It is widely recognized as a stem cell marker in solid tumors, acts as a co-receptor and regulates cell behavior by integrating external signals (Wang et al., 2018). Other markers expressed by healthy stem cells, such as, e.g., CD133, also are found on cancer stem cells (CSC) (Sahlberg et al., 2014;Wang et al., 2012). EpCAM is important for epithelial cell-to-cell contacts but also has oncogenic potential in cell signaling by releasing its intracellular domain (Munz et al., 2009). Together with CD44, EpCAM significantly affects the clinical course in CRC (Liu et al., 2014). CD166, which is found in the intestinal stem cell niche, also is discussed as a cancer stem cell marker (Levin et al., 2010). Staining of these markers revealed a higher expression of all investigated markers under 3D conditions compared to 2D conditions. As tumor cells can acquire mesenchymal features and stem cell characteristics during invasion and development of resistance (Shibue and Weinberg, 2017), it remains open whether CSCs are more likely than non-CSCs to grow on the OncoVaSc-TERM ® scaffold or whether stem cell features develop in non-CSCs due to the 3D culture conditions. Still, considering that the postulated CSC population comprises the cells that are most likely to escape from conventional cytostatic chemotherapy or radiotherapy, their targeting is thought to be essential for long-lasting therapeutic success (Shibue and Weinberg, 2017). The enhanced stemness of tumor cells cultured on the matrix in 3D may explain their generally higher chemoresistance compared to 2D cultured cells, which should be advantageous for the identification of efficient drugs for cancer therapy. Nevertheless, inhibitors that are well known to work in the clinic, such as gefitinib, do reliable work in our 3D tissue test systems (Stratmann et al., 2014;Göttlich et al., 2018). To determine more general differences between 2D and 3D expression in molecular detail, we compared the ratio of treated to non-treated gefitinib-responsive and non-responsive tumor cells carrying an EGFR biomarker or not. Gene ontology (GO) enrichment analysis of microarrays of 2D-and 3D-cultured cells normalized to untreated cells of the respective culture condition reveals highly significant differ-ing. Even though no closed basement membrane structure is preserved on this serosal side of the matrix, hECs line up to form a closed endothelium-like structure.
To explore further applications of our model, we also generated a leukemia model in which the acute myeloid leukemia (AML) cell line THP-1 was co-cultured with hECs. As depicted in Figure 6A and B, we could show a dose-dependent efficacy of the drug tipifarnib on top of a 3D endothelium with values similar to 2D. Furthermore, tipifarnib treatment resulted in side effects indicated by damage to the endothelium. In clinical studies, tipifarnib did not succeed in AML therapy (Thomas and Elhamri, 2007).
We could observe a higher efficacy and specificity in our leukemia 3D model when we tested trispecific antibodies (Banaszek et al., 2019) as an expansion of the bispecific T-cell engaging (BiTE) therapy concept (Goebeler and Bargou, 2020). Here, the CD3-binding site of a split trispecific antibody, referred to as a hemibody, is reconstituted after binding to another hemibody. For reconstitution, both hemibodies have to bind specifically to an antigen expressed on leukemic tumor cells (CD45, marker of hematopoietic cells) and to an antigen expressed in all nucleated cells of the patient (HLA-A2). Thereby HLA-A2-negative but CD45-positive immune cell populations from donors and HLA-A2-positive but CD45-negative patient cells as hECs should not be affected. Accordingly, we could not observe toxic effects in only HLA-A2-positive endothelial cells, demonstrating next to its efficacy also the specificity of this immunotherapeutic strategy. This has allowed us to show that our models are suitable to test and explore antigen-specific but MHCI-independent immune therapies in vitro. In contrast to other frequently used endothelial 3D cell culture models, like transwell plates, our model allows direct and close contact of hECs to the ECM. This is important as ECM cues change signaling in endothelial cells (Fukuhara et al., 2009). Furthermore, integrating the endothelium in our model enables us to study the transmigration of immune cells over the endothelial barrier. This is an important feature for the improvement of immune therapies, as the presence of immune cells within the tumor is currently the best predictor for a treatment response to immune checkpoint inhibition (Lu et al., 2019;Tumeh et al., 2014).
Fibroblasts are another major part of the tumor stroma, contributing to physical and immunological barriers (Kalluri, 2016). These stromal cells can also be included in the 3D models. Notably, with or without tumor cells, fibroblasts cross the basement membrane of our matrix and grow inside the whole matrix. Their relevance to tumor tissue morphology is demonstrated in our 3D CRC model, where SW480 CRC cells show signs of collective invasion in co-cultures with primary dermal fibroblasts (Nietzer et al., 2016). This results in the destruction of former crypt structures in the SISmuc (Nietzer et al., 2016). The role of fibroblasts in tumor metastasis has also been confirmed in animal models (Gao et al., 2019), and they are important determinants of cancer therapy (Valkenburg et al., 2018). Our new and refined in vitro models thus reflect the behavior of non-invasive tumor cells as well as single cell invasion, but also collective invasion sup-Finally, we show that a 3D CRC model with or without fibroblasts is much more resistant towards induction of apoptosis by allogenic PBMCs than a 2D model. This represents a first step towards establishing immunotherapeutic 3D test systems with material from different donors. Thus, the model provides a good starting point for testing MHCI-dependent immunotherapeutics such as immune checkpoint-inhibitors. In addition to antigen-specific T-cell engaging testing, we have already demonstrated that our 3D tumor models can be used to test CAR T-cell efficacy in solid tumors as a further antigen-specific approach and to develop new engineering strategies (Stuber et al., 2020;Wallstabe et al., 2019).
To summarize, benefits of our platform include a reduction of animal experiments, improved screening modalities for anticancer drug development, the possibility to include immune cells and stromal components, and hence a more reliable testing of immune therapies. We are thus confident that our modular and standardizable test platform, which closely mimics cellular in vivo conditions, has a high potential for basic and applied research. ences in gene clusters for proliferation, cell cycle and apoptosis. Given the more physiological environment and growth pattern of 3D cell culture, there is reason to believe that our 3D models reflect clinical responses more accurately than conventional 2D models.
To understand the systems biology of targeted treatments as applied in the clinic and to connect our in vivo-like 3D models to the in silico prediction module, we integrated further analysis parameters going beyond viability, such as proliferation, apoptosis and signaling (Stratmann et al., 2014). Thus, central signaling cascades, known from literature and databases, are set up in silico with the software CellDesigner as a topology to reflect interdependencies by feedback loops and redundancies. The interconnectivity of the network that includes parameters such as kinase cascades, mutations, in vitro measured drug responses of proliferation, apoptosis, and EMT, subsequently provides a basis for optimized simulations of drug effects. The in silico module starts from a Boolean decision network, which then is dynamically simulated and investigated by the software SQUAD using differential equations (Di Cara et al., 2007). This enables approximations of systemic drug responses even without detailed kinetic information. As a generic example, we tested the effects of the AMPK activator metformin on HCC827 cells, as AMPK activation was predicted to be a promising target before (Göttlich et al., 2018). Metformin is discussed to reduce the risk to die from cancer in diabetic patients (Chae et al., 2016;Sun et al., 2015). Increased apoptosis induction by metformin and the EGFR inhibitor gefitinib was predicted in silico and demonstrated in vitro. Starting from these experimental observations and combined with in silico modeling, hypotheses can be generated for mode-of-actions that are exemplified in the following: Looking at the topology, the induction of proliferation by metformin monotherapy could be explained systemically by cellular pathway switching: Upon mTOR inhibition, a compensatory activation of the MEK-ERK pathway is likely to occur. Though there is no direct connection between these pathways, the cellular system change is instead mediated by a concerted reaction of the complete network (connection via PI3K-AKT, and both drugs have an enhancing effect via p53 on apoptosis). However, this MEK-ERK pathway activation could then be targeted by the combination with gefitinib. Hence, there is no direct interaction between the pathways. Instead, a whole system state switch occurs, which would have been difficult to hypothesize by just looking at the data obtained from our 3D culture system. Thus, "wet lab" research performed in this culture system can benefit from additional in silico modeling. Vice versa, in silico modeling can only be as good as the data the algorithms are fed with. While integrated mutational signatures cannot yet predict the response to a treatment, in silico analyses can identify promising candidates that can then be pretested in our 3D tumor tissue models. This could strongly reduce the number of animals required to develop combination therapies, as only the most promising combinations would have to be tested in animals. In addition, testing of individually tailored therapies could be greatly accelerated, which may become critical for precision medicine in oncology.