Quantification of the accumulation of nanomedicine in tumours

We first determined nanomedicine tumour accumulation in three mouse models with differing degrees of vascularization, stroma composition and target-site localization (Fig. 1a). The tumour models were A431 human epidermoid carcinoma, MLS human ovarian carcinoma and CT26 murine colon cancer. As a nanocarrier, we employed a 67kDa-sized poly(N-(2-hydroxypropyl) methacrylamide) (PHPMA) polymer, as this prototypic albumin-sized macromolecule has consistently provided us with high levels of tumour accumulation in a variety of models15,16,17. We used fluorescence reflectance imaging (FRI) and hybrid CTFMT to visualize and quantify the biodistribution and tumour accumulation of DY750-labelled PHPMA (Fig. 1b,c and Supplementary Fig. 1). When normalized to average tumour volume at the timepoint of analysis (250 mm3), at 72h post intravenous (i.v.) injection, we found average levels of target-site localization of 5.01.7, 8.51.6 and 10.21.7 percent of the injected dose (%ID) for A431, MLS and CT26 tumours, respectively, exemplifying sustained localization to tumours over time, as well as different accumulation patterns in the three models (P=0.0024, one-way analysis of variance (ANOVA); Fig. 1d and Supplementary Fig. 1). The tumours were then excised, and DY750-labelled PHPMA accumulation patterns were validated ex vivo using FRI (Supplementary Fig. 2). The collected tumours were fixed, sectioned and stained for biomarker assessment.

a, A schematic of the experimental protocol aimed at identifying tumour-tissue biomarkers that correlate with nanomedicine accumulation in tumours. The tumour accumulation of the prototypic polymeric nanocarrier, PHPMA, was assessed using CTFMT in three distinct mouse models with varying degrees of tumour targeting. Subsequently, correlation analyses were conducted using 23 tumour-tissue microenvironment features associated with tumour-targeted drug delivery, focusing on aspects related to the vasculature (red), stroma (green), macrophages (blue) and cellular density (grey). The dashed lines indicate double stained features. For further details, please refer to Supplementary Table 1. The illustration was created with BioRender.com. b, FRI-based, longitudinal optical imaging of DY750-labelled PHPMA accumulation in the tumours of mice with A431, MLS and CT26 tumours representing low, medium and high levels of target-site accumulation, respectively (the white dashed circles indicate tumour location, and one mouse per tumour model is shown). c,d, Longitudinal CTFMT visualization (c) and quantification of DY750-labelled PHPMA tumour accumulation (d) in percent of the injected dose (100% is equal to 2nmol of dye) normalized to 250mm tumour volume. The statistical significance between the two models was assessed via individual Students t-tests (A431 versus MLS, *P=0.0168; A431 versus CT26, **P=0.0025) and between all models via one-way ANOVA (#P=0.0024). Each data point represents a CTFMT scan of one animal.

We analysed 23 tumour microenvironment features associated with tumour-targeted drug delivery (Supplementary Table 1). These included vascular features, such as vessel density (CD31), perfusion (lectin) and angiogenesis (VEGFR2); lymph vessels (LYVE-1); extracellular matrix components, such as SMA, collagen I and collagen IV; tumour-associated macrophages (TAM; F4/80); and tumour cell density (4,6-diamidino-2-phenylindole). In addition, we analysed combinations of the above, via immunofluorescent double-stainings, to, for example, assess vessel support (SMA+/CD31+), vessel function (lectin+/CD31+) and the fraction of angiogenic vessels (VEGFR2+/CD31+).

The tumour-tissue biomarkers were captured and quantified via fluorescence microscopy and correlated with nanocarrier accumulation in A431, MLS and CT26 tumours (Fig. 2). Regarding blood vessel density and perfusion, we observed an overall good agreement between the number of (perfused) vessels and DY750-labelled PHPMA accumulation. The CT26 tumours had the highest number of total and functional blood vessels (89.035.9 and 48.018.8, respectively; Fig. 2a,b,g,h), and this was in line with their high level of polymer accumulation (10.21.7%ID per 250mm3; Fig. 1d). Conversely, A431 tumours had low levels of total and functional blood vessels (28.515.1 and 25.615.5, repectively; Fig. 2a,b,g,h), aligning with their low accumulation of DY750-labelled PHPMA (5.01.7%ID per 250mm3; Fig. 1d). Interestingly, while CT26 tumours had the highest absolute numbers of total and functional blood vessels, A431 tumours presented with the highest relative level of perfused vessels (91.3%, as compared with 62.7% for MLS and 54.9% for CT26; Supplementary Fig. 3j). This indicates that the absolute number of (functional) blood vessels is a more important factor determining nanomedicine tumour targeting than the relative fraction of vascular perfusion. In good agreement with this, also the absolute numbers of SMA+, Col I+, Col IV+ and VEGFR2+ blood vessels (Fig. 2c,d,i,j) correlated better with DY750-labelled PHPMA tumour accumulation than the relative fractions of SMA+, Col I+, Col IV+ and VEGFR2+ vessels (Supplementary Fig. 3jn).

af, Immunofluorescence stainings for all blood vessels (CD31) (a), actively perfused vessels (lectin) (b), pericyte-supported vessels (SMA) (c), angiogenic vessels (VEGFR2) (d), lymphatic vessels (LYVE-1) (e) and TAM (F4/80) (f) in A431, MLS and CT26 tumours. Scale bar, 50m. gl, Quantification of the immunofluorescence images for CD31+ vessels (g), lectin+ vessels (h), SMA+ vessels (i), VEGFR2+ vessels (j), LYVE-1+ vessels (k) and F4/80 (l) (no., number). The black bars indicate means. *P<0.05, **P<0.01 (Students t-test). Note that the analysis in gi is based on 10 magnification images, while the analysis in jl is based on 20 magnification. mr, Correlation of PHPMA tumour accumulation at 72h post injection (in percent of the injected dose (100% represents 2nmol of dye) normalized to 250mm tumour volume) with the respective tumour-tissue biomarker features (CD31+ vessels (m), lectin+ vessels (n), SMA+ vessels (o), VEGFR2+ vessels (p), LYVE-1+ vessels (q) and F4/80 (r)). The trendlines are shown per tumour model (colour-coded) and for all tumours together (black). The R2 values indicate the coefficient of determination and reflect the goodness of fit. Each data point represents one animal.

Regarding the retention component of nanomedicine tumour targeting, we particularly looked at LYVE-1+ lymphatic vessels and F4/80+ TAM. Interestingly, we observed that the tumour model with the highest level of PHPMA accumulation, that is, CT26, had almost double the number of LYVE-1+ lymphatic vessels as A431 and MLS (Fig. 2e,k). This indicates that the absence of effective lymphatics as a mediator of nanomedicine retention in tumours may be less important than originally anticipated18. It actually even suggests the opposite, which is that a certain degree of functional lymphatics in tumours may be needed to assist in attenuating the high interstitial fluid pressure that is typical of tumours19. A very good correlation was found between the density of TAM and nanomedicine accumulation (Fig. 2f,l,r). The area fraction of TAM increased from 2.2% to 5.1% to 7.7% for A431, MLS and CT26 tumours, respectively, correlating almost linearly with the increased tumour accumulation in these models (Fig. 1d) and resulting in good R2 values both within and across the three models (Fig. 2r). This finding corroborates an increasing number of notions that TAM act as a key reservoir for nanomedicine retention in tumours8,20. It furthermore implies that TAM density seems to be a suitable tumour-tissue biomarker to predict nanomedicine tumour accumulation.

Feature importance was assessed using gradient tree boosting (GTB). GTB is a machine learning technique for building predictive regression models based on a set of yes/no decision trees21,22,23. The trained GTB model considered all 23 features analysed as a regression model and was applied to predict polymeric nanomedicine tumour accumulation (Fig. 3a). Given the relatively small dataset, the leave-one-out method was employed to avoid the mixing of training and testing datasets. Ten decision trees, with a depth of up to eight questions, were found to be able to properly predict nanocarrier tumour accumulation based on histopathological features (R=0.70; Fig. 3b). As exemplified in Fig. 3c, GTB-based importance assessment identified the percentage of lectin+ (that is, functional vessels percentage) and angiogenic (that is, VEGFR2 vessels percentage) blood vessels, the density of TAM (that is, F4/80 area fraction (AF)) and the total, SMA+ and Col I+ number of blood vessels (that is, CD31 number, SMA number and Col I vessels number, respectively) as predictive features.

a, Schematic workflow. Tumour-tissue biomarkers were stained, quantified and correlated with the tumour accumulation of PHPMA nanocarriers. GTB-based machine learning was employed to rank feature importance using predicted versus measured PHPMA tumour accumulation values (Y, yes; N, no; B14, biomarker 14). b, N-fold cross-validation of predicted versus measured PHPMA tumour accumulation patterns illustrates the accuracy of the employed GTB method for predicting nanomedicine tumour targeting (in percent of the injected dose (100% represents 2nmol of dye) normalized to 250mm tumour volume). c, Ranking of the importance of the identified tumour-tissue biomarker features based on their assignment in the GTB decision trees (%, biomarker positive vessels of the number total vessels; no., number). The error bars indicate the standard deviaitoin (n=14).

When aiming to establish a biomarker for patient stratification, the practicality of the approach and the presence of a proper dynamic range are crucial. This implies that in the features identified via GTB, the functionality of tumour blood vessels needs to be excluded, because lectin cannot be injected in patients. For the fraction of VEGFR2+ blood vessels, the dynamic range is small (Supplementary Fig. 3l), making it unlikely to serve as a good biomarker. Moreover, as for the number of SMA+ and Col1+ blood vessels, double-staining would be required. This can be done preclinically with immunofluorescence, but is not typically performed in histopathological protocols in routine clinical practice. In follow-up studies with additional tumour models, we therefore focused on blood vessel and TAM density as tissue biomarkers.

The feature importance and biomarker potential of tumour blood vessels and TAM were confirmed in a panel of ten tumour models. This panel was selected to encompass models with very different tumour microenvironment architectures (thereby reflecting the heterogeneity observed in human tumours24) and consisted of six PDX and four CDX xenograft models. To ensure broad applicability of blood vessel and TAM density as biomarkers for predicting nanomedicine accumulation, we decided to employ a second drug-delivery system in these ten models, replacing the prototypic polymeric nanocarrier PHPMA with a PEGylated liposome formulation similar to Doxil/Caelyx25. Initially, fluorescent DiI-labelled liposomes were used to visualize the accumulation and distribution of liposomes in tumours. The highest levels of liposome accumulation were observed in E35CR and Calu-3 tumours, and the lowest levels were found in A549 and Calu-6 tumours (Fig. 4a).

a, Fluorescence microscopy analysis of Dil-labelled PEGylated liposomes (in red) in ten tumour models at 24h after i.v. administration Scale bar, 200m. The blood vessels are stained in green and the cell nuclei in blue. b, Tumour accumulation of PEGylated liposomal DXR in six PDX (green dots) and four CDX (red dots) tumour models. Individual and mean (black bars) tumour concentrations of DXR are shown for 20 mice per group and 5 mice per timepoint. c, Total tumour accumulation over time of PEGylated liposomal DXR (that is, AUC0120h). Values represent meanstandard error of the mean. d, Histopathological DAB staining of tumour blood vessels (CD31) and TAM (F4/80) for the ten models. Scale bars, 100m. eh, Quantification of blood vessel (e) and TAM (g) density based on DAB staining and correlation of blood vessel (f) and TAM (h) density with total liposomal DXR tumour accumulation (no., number of vessels or TAM per field of view).

We next used doxorubicin (DXR)-loaded liposomes and determined drug accumulation in tumours using high-performance liquid chromatography. For each of the ten models, this was done for four timepoints, with five tumours per timepoint (Fig. 4b). Total DXR concentrations over time were quantified and expressed as the area under the curve (AUC). In good agreement with the DiI-liposome fluorescence data (Fig. 4a), AUC determination demonstrated that tumour DXR concentrations were highest in E35CR and Calu-3, making these the highest drug-accumulating models, with drug levels three to five times higher than those of the majority of other models (Fig. 4c). A549 and Calu-6 were again found to accumulate the lowest amounts of liposomes, with DXR concentrations five to ten times lower than most other models. Interestingly, when comparing all AUC values together, it was furthermore found that PDX models presented with higher overall levels of liposomal DXR accumulation than CDX models (Fig. 4c).

In clinical practice, pathology protocols involve light (and not fluorescence) microscopy. Accordingly, we switched to 3,3-diaminobenzidine (DAB) staining and studied blood vessel and TAM density via standard histopathology in the ten PDX and CDX models. As shown in Fig. 4dh, we found that the three models with the lowest accumulation levels upon administration of liposomal DXR, that is, SW620, A549 and Calu-6 models (Fig. 4c), also presented with the lowest levels of CD31 and F4/80 staining. Across the ten different tumour models, there was a good correlation between tumour blood vessel and TAM density and nanomedicine accumulation (Fig. 4f,h). It should be noted in this regard, however, that the E35CR model was identified as a clear outlier, as it presented with the highest levels of Dil- and DXR-loaded liposome accumulation (Fig. 4ac), while its levels of CD31+ blood vessels were intermediate (Fig. 4f) and those of F4/80+ TAM were very low (Fig. 4g). When determining the area fraction of CD31 and F4/80 instead of the number of CD31+ and F4/80+ cells, observations were identical for all of the above notions, confirming the robustness of the tumour-tissue biomarkers identified (Supplementary Fig. 4). Altogether, these results demonstrate that there is a good correlation between the levels of the tumour blood vessels and TAM and the level of nanomedicine tumour accumulation.

Having identified tumour blood vessels and TAM as key features correlating with nanomedicine tumour accumulation, we next explored the robustness, validity and potential clinical applicability of combined tumour blood vessel and macrophage scoring, with the aim of developing a simple and straightforward biomarker protocol for patient stratification. This protocol is primarily designed to help predict which individuals from a heterogeneous patient population should be excluded in clinical trials, because their tumours are likely to show low nanomedicine accumulation and poor therapeutic efficacy (Fig. 5a).

a, Schematic workflow demonstrating the concept of patient stratification in cancer nanomedicine clinical translation based on tumour-tissue biopsies, created with BioRender.com. b, DAB staining illustrating the density of tumour blood vessels (CD31) and TAM (F4/80) in tumours, reaching from lowest (score 1) to highest (score 4) levels of blood vessel and macrophage density. Biomarker scores indicate 1 for absent, 2 for low, 3 for intermediate and 4 for high. Scale bars, 100m. c, Colour-coded heatmap, representing the distribution of CD31 and F4/80 product scores in the ten PDX and CDX tumour models with differing degrees of PEGylated liposomal DXR tumour accumulation. Tumours are ranked from high to low AUC, from top to bottom. Tumour-tissue biomarkers were scored by ten blinded observers, who each analysed three tissue sections per tumour model (n=30 in total). The colour intensity reflects the number of product scores. d, Schematic displaying the distribution of true and false positives and negatives in the tumour-tissue biomarker product score heatmap. e, Receiver operating characteristic (ROC) curve, generated on the basis of the tumour-tissue biomarker product scores, exemplifying very high diagnostic accuracy differentiating between low and high nanomedicine tumour accumulation (ROC curve is based on the scores in c; the red dashed line represents randomness and the units of the axis are in %).

We conceived a DAB-based histopathological scoring setup in which we considered 1 for absent, 2 for low, 3 for intermediate and 4 for high for the expression of both tumour-tissue biomarkers (Fig. 5b). Ten blinded observers, including three board-certified pathologists, were asked to score 60 tumour sections (30 for CD31 and 30 for F4/80; 6 for each tumour model). As shown in Fig. 5c, the colour-coded scoring intensities demonstrate that for tumour models with low CD31 and F4/80 product scores, the levels of liposomal DXR accumulation were also low. With a cut-off score of 6 to differentiate between tumours with low versus high nanomedicine accumulation, the blinded observers product scores correctly identified SW620, A549 and Calu-6 as true negatives (Figs. 4ac and 5c,d). Conversely, six out of seven models with good nanomedicine accumulation were correctly identified as true positives (Fig. 5c, d). The E35CR model turned out to be false negative, as its low CD31 and F4/80 product score incorrectly indicated that it would not accumulate liposomes well, which it clearly did do (Fig. 4ac). No false positives were detected (Fig. 5c,d). Altogether, nine out of ten tumour models could be correctly associated with low versus high nanomedicine accumulation on the basis of our tumour blood vessel and TAM biomarker product score.

To quantify the biomarker performance of our product score, we determined the area under the receiver operating characteristics (AUROC) curve. The AUROC curve represents a probability assessment, with a value of 0.5 resulting in a straight 45-line reflecting randomness (represented by the dashed red line in Fig. 5e). The AUROC curve represents the capability of a biomarker to distinguish between different classes, in this case between low versus high nanomedicine tumour accumulation. We obtained an AUROC value of 0.91 for our blood vessel and TAM product score (Fig. 5e), which is generally considered excellent for predicting nanomedicine tumour targeting, following the published criteria26.

The robustness and translatability of our biomarker product score were assessed in immunocompetent mouse models and in patient samples. The former were included to rule out the possibility that the presence of T cells plays an important role in determining nanomedicine delivery to tumours. To this end, we analysed PHPMA accumulation in orthotopic 4T1 triple-negative breast cancer tumours in BALB/c mice and PEGylated liposome accumulation in subcutaneous and orthotopic Hep55.1C liver tumours in C57BL/6J mice. As shown in Supplementary Fig. 5, good correlations between blood vessel and TAM product scores and nanomedicine tumour targeting were observed, as exemplified by R2 values of 0.51, 0.86 and 0.63, respectively. This confirms that our biomarker product score remains valid in syngeneic and orthotopic tumours in immunocompetent mice.

Next, we aligned our biomarker product score with the most comprehensive clinical dataset available on nanomedicine tumour targeting in patients27. In this study, the researchers used 111In-labelled PEGylated liposomes and quantitative SPECT imaging to assess nanomedicine tumour accumulation in 17 patients with different type of tumour27. For the most prevalent tumour types included, that is, ductal breast cancer, squamous cell carcinoma of the lung and squamous cell head and neck cancer, we collected matching tumour resection samples as well as primary tumour biopsies from the Biobank archive of the Institute of Pathology at RWTH Aachen University Hospital (Supplementary Table 5). Blood vessel (CD31+) and TAM (CD68+) density were analysed in ten different patient samples for each of the three cancer types, always in five different microarray sections for each individual tumour specimen. The expression levels and patterns of F4/80 and CD68 on TAM were demonstrated to be similar (Supplementary Fig. 6). Representative CD31 and CD68 stainings for breast, lung and head and neck cancer lesions are shown in Fig. 6a,b. Using QuPath software28, we quantified blood vessel and TAM density in these tumours and found that breast cancer typically presents with much lower levels of both tumour-tissue biomarkers as compared with lung and head and neck cancer (P<0.001 and P<0.0001 for blood vessels and P<0.05 for TAM; Fig. 6c,d).

a,b, Representative DAB stainings of blood vessels (a) and TAM (b) in tumour tissues obtained from patients with breast, lung and head and neck (H&N) cancer (all data in this figurre are based on tumour resections, and the data based on biopsies are shown in Supplementary Fig. 7). c,d, Quantification of blood vessels (c) and TAM (d) in ten patient samples for each tumour type (no., number per field of view; significance is indicated in P values based on Students t-test). e, Tumour accumulation of 111In-labelled PEGylated liposomes in patients with breast, lung and head and neck (H&N) cancer (in percentage of the injected dose per kilogram tumour). The data are replotted based on the work in ref. 27 (significance is indicated in P values based on Students t-test). f, Means of blood vessel and TAM product scores plotted against means of liposome tumour targeting, showing that biomarker product scoring correctly identifies breast cancers as poorly nanomedicine accumulating lesions. The error bars indicate the distribution of %ID and product score values (standard deviations on the x-axis and minima and maxima on the y-axis; n=310 as it is based on the means of c, d and e).

The liposome tumour targeting data from ref. 27 is replotted in Fig. 6e. In line with our rationale and reasoning, it can be seen that ductal breast cancer lesions in patients (5.33.0%IDkg1) accumulate radiolabelled PEGylated liposomes significantly less well than lung (18.26.6%IDkg1; P<0.05) and head and neck (33.017.6%IDkg1; P<0.05) squamous cell carcinomas. When generating tumour-tissue biomarker product scores based on the number of blood vessels and TAM per tumour type and when plotting these product scores against the average level of liposome accumulation per tumour type, we found that breast cancers clustered in the lower left corner, thereby pinpointing them as true negatives (Fig. 6f). For the majority of lung and head and neck cancer lesions, the product scores were much higher than for breast cancer, thereby classifying them as true positives. In a final validation study, we also employed the original primary tumour biopsies for biomarker assessment. For the 30 patients samples initially included, 28 primary biopsies were available. As exemplified by Figure S7, the results obtained in biopsies are very similar to those obtained in resected tumour tissues, again clearly identifying ductal breast cancers as poorly accumulating lesions. Thereby, they not only confirm the robustness of our approach but also showcase its clinical translatability. Altogether, these findings provide compelling proof-of-concept for the use of tumour blood vessels and TAM as tissue biomarkers for predicting nanomedicine tumour targeting.

Read more here:
Histopathological biomarkers for predicting the tumour accumulation of nanomedicines - Nature.com