Personalized models of breast cancer desmoplasia reveal biomechanical determinants of drug penetration

Breast cancer desmoplasia heterogeneity contributes to high disease mortality due to discrepancies in treatment efficacy between patients. Personalized in vitro breast cancer models can be used for high throughput testing and ranking of therapeutic strategies to normalize the aberrant microenvironment in a patient-specific manner. Here, tumoroids assembled from patient-derived cells cultured in microphysiological systems including perfusable microvasculature reproduce key aspects of stromal and vascular dysfunction. Increased hyaluronic acid and collagen deposition, loss of vascular glycocalyx and reduced perfusion, and elevated interstitial fluid pressure in the models result in impaired drug distribution to tumor cells. We demonstrate the application of these personalized models as tools to rank molecular therapies for the normalization of the tumoroid microenvironment and to discover new therapeutic targets such as IL8 and CD44, which may ultimately improve drug efficacy in breast cancer patients.


INTRODUCTION
Tissue desmoplasia is a key process underlying disease progression in breast cancer. The induced physical heterogeneity in the breast tissue presents a major obstacle to treatment 1,2 . Breast tumor cells (TCs) are largely responsible for desmoplastic remodeling of the tumor microenvironment, both directly through aberrant deposition of matrix proteins and indirectly through activation of stromal cells [3][4][5] .
Increased ECM density, reduced vascular perfusion and barrier function, and reduced lymphatic drainage all contribute to increased interstitial fluid pressure (IFP) in the breast cancer microenvironment, resulting in reduced immune cell infiltration 6 and impaired drug distribution from blood to TCs [7][8][9] . Therapeutic strategies to normalize the desmoplastic breast cancer microenvironment are currently in clinical trials 10 . However, the patient-to-patient heterogeneity of breast cancer, hence its response to different treatments, makes it critical to identify personalized strategies with the greatest therapeutic potential.
Patient-derived TCs formed into structures with increasing cellular complexitysingle-cell type spheroids 11 , multi-cell type tumoroids 12,13 , and stem cell-derived organoids 14 -have recently attracted attention as laboratory models to rapidly evaluate response to molecular therapies 15 . Incorporating various "-oid" structures into microphysiological models presents the opportunity to broaden their scope and capture key morphological and functional aspects of the tumor microenvironment [16][17][18][19] . An additional attractive feature of microfluidic-based models is the extremely fine control over mechanical and biochemical stimuli imparted on the cells within 20,21 . However, current models have so far been unsuccessful in recapitulating the complexity of the desmoplastic breast cancer microenvironment 22 , particularly the aberrant ECM and vasculature, which jointly determine drug distribution into the tumor.
Harnessing the capabilities of microphysiological models to recapitulate the pathophysiological complexity of the aberrant TC microenvironment in breast cancer desmoplasia could provide new therapeutic strategies to improve clinical care for breast cancer patients.
We have recently demonstrated culture of cancer spheroids assembled from ovarian and lung TC lines in microphysiological models containing human microvascular networks (MVNs) 23 . These MVNs can 4 be perfused with therapeutic molecules to assess their permeability across the vascular endothelium and resulting TC death. In this work, we expand on this methodology to culture tumoroids containing primary breast TCs or breast TC lines, evaluating their microenvironments to demonstrate differential desmoplastic stromal and vascular remodeling with resulting responsiveness to normalizing treatments.
Herein, we show that these vascularized tumor models can both provide quantitative metrics to rank the efficacy of different therapeutic strategies, and empower the discovery of new molecular targets to normalize the desmoplastic breast cancer microenvironment.

Vascularized tumoroids on-chip reproduce variations in breast cancer desmoplasia
We first made use of breast TC lines representative of three main disease molecular sub-types 24 to establish the model's capabilities: MCF7 (estrogen receptor, ER, and progesterone receptor, PR, positive), SKBR3 (human epidermal growth factor receptor 2, HER2, positive) and MDA-MB-468 ("triple-negative", epidermal growth factor receptor, EGFR, positive). Tumoroids approximately 500 µm in diameter were formed by co-culture of TCs with human fibroblasts (FBs) in non-adherent well plates for 4 days, which allowed the formation of stable cell aggregates even for TCs that would not aggregate in mono-culture (SKBR3 and MDA-MB-468). We employed our recently described 23 methodology to co-culture the tumoroids in 3D gels within microfluidic devices containing human microvascular networks (MVNs, Figure 1a), which self-assemble over 7 days from human endothelial cells (ECs) and the same FBs used to form the tumoroids. Gene expression analysis of therapeutic target receptors on TCs in tumoroid MVN devices confirmed high ER and PR in MCF7, high HER2 in SKBR3, and high EGFR in MDA-MB-468 in the MVN devices ( Figure 1b). EGFR gene expression significantly increases for all TCs isolated from the MVN devices compared to the same TCs cultured in 2D well plates, possibly due to signaling from additional cell types in the MVN devices. HER2 expression also increases approximately 10-fold for the MCF7 cells. As a result, measurable levels of EGFR and HER2 proteins are found in the microenvironment of all three tumoroids (Figure 1c). 5 Expression of those receptors is principally co-localized with the TCs (Figure 1d), yet sectioning of tumoroids in MVN devices revealed a cytokeratin-rich core of dead TCs and, possibly, FBs where receptor expression is lost (Figure 1d, Supplementary Figure 1a). The tumoroid dead cores may be the result of hypoxia-induced necrosis, as previously observed in tumors in vivo and TC aggregates larger than 500 µm 25   Overall, these results show that different tumoroids can alter their surrounding microenvironment in drastically different ways. In addition, despite the relative simplicity of TC lines compared to primary TCs, the tumoroids tested are capable of building complex microenvironments in the MVN devices that mimic key aspects of the progression of breast cancer desmoplasia 4 . Specifically, the MCF7 tumoroids appear to be representative of an early disease stage, where TCs still possess a primarily epithelial phenotype and a fibrotic ECM 'sheath' forms around the tumoroids that is associated with restricted TC migration and invasion; the MDA-MB-468 tumoroids, instead, may be representative of a later disease stage, whereby TCs aggressively invade the tumoroid microenvironment through a denser ECM deposited by both TCs and CAFs, with SKBR3 tumoroids somewhat in between these two extremes ( Figure 1j). MDA-MB-468 tumoroids also recapitulate an additional feature of disease progression in that they often become vascularized by the surrounding MVNs (Supplementary Figure 3). We next assessed whether these vascular and ECM changes result in differences between drug distribution to the tumoroids.

Loss of vascular hyaluronic acid contributes to increased tumoroid interstitial fluid pressure
We have previously shown that MVNs in the vicinity of TC aggregates can partially lose vascular barrier function, resulting in higher permeability across the endothelium 23   resulting in increased filtration and hydraulic conductivity; n = 3. Significance assessed by one-way ANOVA after confirming a normal distribution of the data; p < 0.05 *, p < 0.01 **, p < 0.001 ***. 10 The loss of vascular barrier function appears to be incompatible with decreased drug distribution to the TME, whereby higher therapeutic molecule concentrations would be expected to reach the TCs across the leaky endothelium. However, our model allows the assessment of additional factors contributing to the impaired trans-vascular drug distribution in the tumor microenvironment to better understand this phenomenon. First, the morphology of the MVNs is altered in the vicinity of the tumoroids, as evidenced by a loss of vessel density and vessel specific surface area available for drug transport in the The elevated IFP in breast cancer desmoplasia hinders drug penetration and negatively correlates with cancer survival 9,31,36 . The vascularized tumoroid models present an opportunity to better understand this 11 phenomenon and help identify therapeutic strategies that can improve drug delivery. It is well understood that increased ECM density in the desmoplastic stroma contributes to increased IFP by resisting the passage of interstitial flow via low matrix permeability, k 9,31 . HA is a key determinant of this contribution through its capability to bind large quantities of fluid 37  These results confirm that vascular HA plays an important role in maintaining a low Lp in control MVNs, and that loss of vascular HA in tumoroid MVNs is associated with a concurrent increase in Lp, therefore in IFP. Despite its only partial role in elevated IFP, loss of vascular HA may be used as a marker for cancer-associated vascular dysfunction. For this reason, we next assessed possible causes for this change in the tumoroid MVNs.  We next set out to establish whether these cytokines directly contribute to vascular HA degradation.

Inhibition of IL8 restores endothelial barrier function and enhances drug penetration
Recombinant versions of the five cytokines were perfused through control MVNs at a concentration of 5 ng mL -1 to assess changes in permeability after short (< 15 min) and long (12 hours) exposure times.
TNFα and CCL2 increased MVN permeability to dextran at short times, while significant long-term changes were produced by IL8, IL12, TNFα, and CCL2, with IL8 showing the largest increase in permeability (1.9 x 10 -8 cm s -1 to 5.7 x 10 -8 cm s -1 , Figure 3b). End-point analysis of vascular HA concentration by immunofluorescence showed significant degradation by CCL2 at short times and IL8 at long times (Figure 3c, Supplementary Figure 7c). Exposure of control MVNs to IL8 confirmed loss of vascular HA at concentrations as small as 1 ng mL -1 (Figure 3d). Interestingly, IL8 gene expression increases for all TCs in the MVN devices (Figure 3e, Supplementary Figure 7d). IL8 has angiogenic effects 50 , and its expression in TCs may be enhanced by paracrine signaling from ECs in the TME.
Overall, these results point to IL8 as a candidate target cytokine to prevent vascular HA degradation, and subsequent loss of vascular barrier function. 15 We tested this hypothesis by treating MDA-MB-468 tumoroid MVNs with molecules directed against IL8: an IL8-blocking mAB, and the broad anti-inflammatory small molecule Dexamethasone. All treatments were administered in the MVN devices over 4 days. We first assessed changes in vascular HA  These results identify IL8 as an effective target to aid recovery of the vascular glycocalyx as a way of increasing drug penetration and distribution in breast tumors. Importantly, the results also suggest that targeting pathophysiological mechanisms affecting the vascular ECM can provide a different strategy to normalize the desmoplastic TME other than targeting the stromal ECM. To compare the effects of these two therapeutic strategies, we next assessed potential treatments to degrade stromal HA.

Desmoplastic stroma normalization strategies can be ranked for different tumoroid models
We treated, again, MDA-MB-468 tumoroids in MVNs for four days, using HA-ase, a CD44-blocking mAB, and a TGFβ-blocking mAB. While HA-ase and the CD44-blocking mAB target stromal HA directly by degrading it or blocking TCs and FBs from binding to it, respectively, the TGFβ-blocking mAB indirectly targets HA by depriving FBs in the tumoroids of TGFβ, a key stimulant of stromal HA production 51 . Similar to the treatments directed against IL8, blocking of CD44 and TGFβ in the tumoroid stroma decreased MVN permeability (2.9 x 10 -8 cm s -1 for CD44, 3.8 x 10 -8 cm s -1 for TGFβ, Figure 4a) and IFP, as seen by an increase in effective MVN permeability with applied intravascular 16 pressure. This effect was particularly pronounced for the CD44-blocking treatment, whereby effective permeability values were only slightly lower than for control healthy MVNs. HA-ase, instead, did not lower permeability (5.9 x 10 -8 cm s -1 ), but rather increased effective permeability under intravascular pressure to levels higher than control (Supplementary Figure 7e) We further hypothesized that degradation of stromal HA can lead to increased drug distribution to TCs.
To test this, the tumoroid MVNs were treated for four days with Trastuzumab and Cetuximab, a model drug combination with expected cytotoxic effects 32,33 to compare drug effectiveness in the different tumoroids, at a concentration of 20 µg mL -1 to match expected levels in circulation 34 . Preliminary data showed significant TC death in tumoroids in well plates at concentrations as low as 2 µg mL -1 when assessed by fluorescence imaging of a cell death marker (Supplementary Figure 8a, b). We used the same method to evaluate cell death as a result of co-perfusion with the molecules targeting stromal HA in the MVN devices. Sections of the fixed tumoroids post-treatment revealed increased drug penetration compared to controls (Figure 4f), validating our hypothesis that stromal HA degradation increases interstitial drug transport. Despite the relatively high mAB concentration perfused, significant cell death was observed only in the MCF7 tumoroids (Figure 4g Overall, these results confirm our hypothesis that stromal HA degradation in the tumoroid microenvironments correlates with increased drug efficacy due to improved penetration. The results also showcase the capability of the models to assess relative impact for different therapeutic strategies in different tumoroid models. We next applied this capability to rank therapeutic strategies in vascularized tumoroids assembled from patient-derived TCs. assessed by one-way ANOVA after confirming a normal distribution of the data; p < 0.05 *, p < 0.01 **, p < 0.001 ***, p < 0.0001 ****.

Personalized tumoroid models pinpoint effective desmoplasia normalization strategies
We formed tumoroids with TCs from two breast cancer patients and cultured them within MVN devices and a lack of a clear trend with applied intravascular pressure, confirming an increase in IFP even for these tumoroids (Figure 5b). These results further confirm that our models can display a wide range of breast cancer heterogeneity between different disease types and patients. Significance assessed by one-way ANOVA after confirming a normal distribution of the data; p < 0.05 *, p < 0.01 **, p < 0.001 ***, p < 0.0001 ****.
Both patient-derived tumoroids show higher stromal HA concentrations than control MVNs (1410 ng mg -1 and 1854 ng mg -1 for Patient 1 and Patient 2, respectively, compared to 994 ng mg -1 ). We tested the treatment regimens described above to attempt normalization of the tumoroid microenvironments, including those treatments rebuilding vascular HA and those degrading stromal HA (Figure 5c). All treatments produce an overall decrease in stromal HA concentration, although not significant for the IL8-blocking mAB in Patient 1 tumoroids and the TGFβ-blocking mAB in Patient 2 tumoroids. Aside from HA-ase, which virtually obliterates HA in the tumoroids, the decrease in HA concentration is greatest when blocking TGFβ in Patient 1 tumoroids and blocking CD44 in Patient 2 tumoroids.
Remarkably, while the IL8 blocking mAB and Dexamethasone decrease overall HA concentrations less than the other treatments, they produce levels comparable to MVN controls, possibly indicating a greater extent of ECM normalization by targeting both the vascular and stromal HA. Most importantly, these results show that even when assembled from primary TCs, the vascularized tumoroid models are capable of capturing variations in the effectiveness of different treatments in normalizing the desmoplastic microenvironments.
Finally, we assessed the efficacy of the different treatments in producing changes in TC death when used in combination with Trastuzumab and Cetuximab, as previously done with the TC line tumoroids.
We found that Patient 1 tumoroids are more susceptible to treatment with the therapeutic mABs compared to Patient 2 tumoroids (Figure 5d), likely the result of differential expression of target receptors between the two TC types (HER2 expression confirmed by the supplier only for Patient 1 TCs). For Patient 1, cell death increases significantly with combination treatment with HA-ase and the TGFβ-blocking mAB, for which HA degradation was greatest, as similarly seen in MCF7 tumoroids.
Despite the overall inefficacy of Trastuzumab and Cetuximab in targeting Patient 2 tumoroids, we observed increased cell death with combination treatment of the mABs with Dexamethasone (and, although not significant, with the IL8-blocking mAB). A different treatment approach, for example by standard small molecule chemotherapy or different mAB target, may prove more effective in producing significant TC death for Patient 2. Taken together, these results confirm that patient-derived 22 vascularized tumoroid models can help rank therapeutic strategies by effectiveness in normalizing the tumoroid microenvironment and improving delivery and efficacy of clinically relevant therapies.

DISCUSSION
Normalization of desmoplastic cancers is currently being approached therapeutically by targeting cell contractility of TCs and CAFs or their aberrant ECM deposition 10 , both with the purpose of preventing blood vessel constriction and improving drug penetration. Of these treatments, TGFβ-targeting appears particularly attractive due to the key role uncovered for this factor in stimulating desmoplastic collagen and HA deposition 55 . Our results using vascularized tumoroid models confirm that targeting TGFβ can be successful in normalizing the tumoroid microenvironments for certain patients 55 , as seen in models The heterogeneity of breast cancer desmoplasia makes patient-specific in vitro models uniquely positioned to pinpoint treatment strategies that are most effective in achieving a therapeutic response 15 .
The MVN models allow for culture of patient-derived tumoroids in functional microvascular beds that allow perfusion of relevant therapeutic molecules under physiological vascular flow conditions. In addition, tumoroids from different parts of an excised tumor may also be tested individually to address intratumor heterogeneity. MVNs were formed here with primary human umbilical vein ECs (HUVECs) and lung FBs, which were chosen for their vasculogenic potential and robust formation of perfusable MVNs even when exposed to pro-inflammatory factors 23,47 . Future modifications of the models may include immortalized MVN cells sources with improved reproducibility 60   (#48404-097, VWR). ECs (6 million mL -1 ), FBs (2 million mL -1 ), and tumoroids were co-injected with fibrin gel solution within the central channel of the microfluidic device, and cultured over 7 days through daily Vasculife medium changes in the side channels. A monolayer of ECs was seeded on the gel surfaces in the side channels on day 4 of culture, as previously described 46 .

MVN permeability and fluid flows
Permeability of fluorescent molecules in the MVNs was measured by confocal microscopy and image analysis with the software ImageJ as previously described 46  Pa 63 . Pressurization of the MVNs through the microfluidic device side channels was done as shown previously 34 using a FlowEZ pressure regulator (Fluigent), and the effective MVN permeability for intravascular pressures up to 1500 Pa was measured as above. The increase in effective permeability of FITC was used to calculate the hydraulic conductivity of the MVNs, Lp, using 64 : where P is the MVN permeability at intravascular pressure Δp = 0 Pa and Peff is the effective permeability. No endothelial reflection of the small solute and no difference in osmotic pressure across the endothelium were assumed, as demonstrated before for this system 34 . The interstitial flow resulting from MVN pressurization in the vicinity of the tumoroids was measured by fluorescence tracking of a 30 µm bleached spot, as previously described 34 .

Interstitial fluid pressure modeling
Computational modeling of IFP in the MVN devices was performed using the software COMSOL where p is the fluid pressure, u is the fluid velocity, µ is the fluid viscosity, k is the matrix permeability, AS/V is the ratio of the MVN lateral surface and the matrix volume, Lp,M is the hydraulic conductivity of 26 the side monolayer, assumed one order of magnitude higher than that of the MVNs based on our previous assessments 34

Gene and protein expression
Gene expression was assessed by quantitative polymerase chain reaction (qPCR). Tumoroids from 4 microfluidic devices, 4 tumoroids per device, were collected by separating PDMS and glass using a scalpel to access the gel, as previously described 65 . Gel biopsy punches (1 mm diameter, #15110-10, Ted Pella) centered at each tumoroid were collected and the matrix between cells lysed with Liberase

MVN and tumoroid treatments
Treatments were administered in the microfluidic devices through perfusion across the MVNs. HA-ase

Data interpretation and statistics
The number of technical and biological repeats is reported in each figure caption. When possible without compromising clarity, single biological repeat data points are provided in the graphs. Statistical analysis of the data was performed with the software Prism (GraphPad, version 9), and the specific statistical test used is reported in each figure caption. Mean differences with p value < 0.05 were taken as significant. 29