Supervised learning on impedance cytometry data for label-free biophysical distinction of pancreatic cancer cells versus their associated fibroblasts under gemcitabine treatment

Abstract

Chemotherapy failure in pancreatic cancer patients is widely attributed to cancer cell reprogramming towards drug resistance by cancer associated fibroblasts (CAFs), which are the abundant cell type in the tumor microenvironment. Association of drug resistance to specific cancer cell phenotypes within multicellular tumors can advance isolation protocols for enabling cell-type specific gene expression markers to identify drug resistance. This requires the distinction of drug resistant cancer cells versus CAFs, which is challenging since permeabilization of CAF cells during drug treatment can cause non-specific uptake of cancer cell-specific stains. Cellular biophysical metrics, on the other hand, can provide multiparametric information to assess the gradual alteration of target cancer cells towards drug resistance, but these phenotypes need to be distinguished versus CAFs. Using pancreatic cancer cells and CAFs from a metastatic patient-derived tumor that exhibits cancer cell drug resistance under CAF co-culture, the biophysical metrics from multifrequency single-cell impedance cytometry are utilized for distinction of the subpopulation of viable cancer cells versus CAFs, before and after gemcitabine treatment. This is accomplished through supervised machine learning after training the model using key impedance metrics for cancer cells and CAFs from transwell co-cultures, so that an optimized classifier model can recognize each cell type and predict their respective proportions in multicellular tumor samples, before and after gemcitabine treatment, as validated by their confusion matrix and flow cytometry assays. In this manner, an aggregate of the distinguishing biophysical metrics of viable cancer cells after gemcitabine treatment in co-cultures with CAFs can be used in longitudinal studies, to classify and isolate the drug resistant subpopulation for identifying markers.

Keywords: Chemotherapy; Impedance cytometry; Machine learning; Pancreatic cancer; Tumor microenvironment.

Figure 1:

A. Pancreatic cancer cells and cancer associated fibroblasts (CAFs) collected from liver metastasis of PDAC patient (T608) are propagated in mice as xenografts, so that reprogramming of the cancer cells interacting with CAFs under gemcitabine treatment can be followed by comparing the viable cancer cells within: B. (i) monocultures vs. (ii) transwell CAF co-cultures, using single-cell impedance and flow cytometry (latter after fluorescent staining). C. Normalized viable cancer cell proportions in the AV⁻ ZNIR⁻ gate for the untreated control and gemcitabine-treated samples show a rise in drug resistance under transwell co-culture with CAFs vs. under monoculture. The gating of viable cancer cells after drug treatment in transwell co-culture with CAFs based on AV⁻ ZNIR⁻ events in flow cytometry (D) can be achieved in a label-free manner with similar accuracy using $\varphi Z_{18MHz}$ vs. $\varphi Z_{0.5MHz}$ plots (E). The viable cancer cells after drug treatment (i.e., the drug resistant subpopulation) exhibit: F. higher opacity, G. lower $\varphi Z_{18MHz}$, and higher electrical size ($\sqrt[3]{\mid Z{\mid }_{0.5\text{MHz}}}$), as validated by FSC data (ESI Fig. S4B-C).

Figure 2.

Multiple impedance metrics are used to compare the biophysical properties of pancreatic cancer cells and CAFs after 48 h of transwell co-culture to simulate the multicellular tumor sample prior to drug treatment. The metrics of: A. $\varphi Z_{0.5MHz}$, B. opacity ($\mid Z_{2MHz}\mid ∕\mid Z_{0.5MHz}\mid$), C. $\varphi Z_{18\mathit{MHz}}$, D. electrical diameter ($\sqrt[3]{\mid Z{\mid }_{0.5\text{MHz}}}$ normalized vs. beads), and E. FSC (flow cytometry) are shown as normalized histograms for ~10,000 cell events. For each impedance metric, the proportion of cells on either side of the indicated 1D gate is used to separate respective cell types. F-I. Significance plots for the respective metrics from triplicate samples. G. The multifrequency impedance phase dispersions of cancer cells versus CAFs.

Figure 3:

Viable cancer cells and CAFs in a homogeneous sample can be identified based on 2D gates from plots of $\varphi Z_{30MHz}$ versus $\varphi Z_{0.5MHz}$ to separate CAFs (A. pink shaded area) and cancer cells (B. blue shaded area), but a significant fraction of each cell type is mis-classified. Upon application to impedance data from a heterogeneous tumor sample that is generated based on 50% cancer cells and 50% CAFs (C), the same 2D gate can separate the respective cell types (D.i: blue shaded area and pink shaded area), but 2186 cancer cells are misclassified as CAFs (False Negatives) and 1098 CAFs are misclassified as cancer cells (False Positives) (D.ii), thereby giving an accuracy of 72.6% for this 2D gating approach. E-G. Flow cytometry after EpCAM staining of cancer cells shows near-perfect classification, as 50.7% cancer cells and 49.3% CAFs.

Figure 4:

A.i. Following transwell co-culture of cancer cells and CAFs for 48 h, in absence of drug treatment (later to be done with gemcitabine treatment), the respective cell types are separated to: B. analyze their impedance data clusters over 12 impedance metrics (C). This data is used as the training set to construct a supervised machine learning model (D), as confirmed by analyzing its confusion matrix (E). The trained SVM model, with classification accuracy of 93.7% is applied to classify impedance data from multicellular tumor samples (A.ii).

Figure 5:

Utilization of supervised learning to assess multiparametric classification of impedance data clusters using heterogeneous tumor samples with varying proportions of cancer cells (70% to 5%) to CAFs. A. Scatter plots of $\varphi Z_{30MHz}$ versus $\varphi Z_{0.5MHz}$ are classified by: B. Flow cytometry analysis after EpCAM staining, and C. SVM predicted classes.

Figure 6.

A. Impedance cytometry data from the remaining viable cells after gemcitabine treatment (1 μg/mL for 48 h) of cancer cells and CAFs in a trans-well co-culture system is utilized over 12 impedance metrics to construct a supervised learning model. Scatter plots of impedance phase at 30 MHz ($\varphi {\mathrm{Z}}_{30\text{MHz}}$) vs at 0.5 MHz ($\varphi {\mathrm{Z}}_{0.5\text{MHz}}$) for CAFs (i), and cancer cells (ii) and the confusion matrix for the optimal SVM model (iii). This model is applied to data from heterogeneous samples to predict the respective cell types: B. 68% cancer cells and 32% CAFs. C. 85% cancer cells and 15% CAFs. B.i & C.i: scatter plots of heterogeneous samples. B.ii & C.ii: scatter plots of predicted classes by the SVM model, with the respective accuracies assessed based on a confusion matrix (B.iii & C.iii).

Figure 6.

A. Impedance cytometry data from the remaining viable cells after gemcitabine treatment (1 μg/mL for 48 h) of cancer cells and CAFs in a trans-well co-culture system is utilized over 12 impedance metrics to construct a supervised learning model. Scatter plots of impedance phase at 30 MHz ($\varphi {\mathrm{Z}}_{30\text{MHz}}$) vs at 0.5 MHz ($\varphi {\mathrm{Z}}_{0.5\text{MHz}}$) for CAFs (i), and cancer cells (ii) and the confusion matrix for the optimal SVM model (iii). This model is applied to data from heterogeneous samples to predict the respective cell types: B. 68% cancer cells and 32% CAFs. C. 85% cancer cells and 15% CAFs. B.i & C.i: scatter plots of heterogeneous samples. B.ii & C.ii: scatter plots of predicted classes by the SVM model, with the respective accuracies assessed based on a confusion matrix (B.iii & C.iii).