Dissecting Response to Cancer Immunotherapy by Applying Bayesian Network Analysis to Flow Cytometry Data

Abstract

Cancer immunotherapy, specifically immune checkpoint blockade, has been found to be effective in the treatment of metastatic cancers. However, only a subset of patients achieve clinical responses. Elucidating pretreatment biomarkers predictive of sustained clinical response is a major research priority. Another research priority is evaluating changes in the immune system before and after treatment in responders vs. nonresponders. Our group has been studying immune networks as an accurate reflection of the global immune state. Flow cytometry (FACS, fluorescence-activated cell sorting) data characterizing immune cell panels in peripheral blood mononuclear cells (PBMC) from gastroesophageal adenocarcinoma (GEA) patients were used to analyze changes in immune networks in this setting. Here, we describe a novel computational pipeline to perform secondary analyses of FACS data using systems biology/machine learning techniques and concepts. The pipeline is centered around comparative Bayesian network analyses of immune networks and is capable of detecting strong signals that conventional methods (such as FlowJo manual gating) might miss. Future studies are planned to validate and follow up the immune biomarkers (and combinations/interactions thereof) associated with clinical responses identified with this computational pipeline.

Keywords: Bayesian networks; FACS; flow cytometry; gating; immune networks; immuno-oncology; machine learning.

Figure 1

A full Bayesian network (BN) constructed from fluorescence-activated cell sorting (FACS) data (Checkpoint panel) from 13 patients preimmunotherapy treatment (day 1). The number next to the edge and edge thickness indicate dependency strength; arrows/edge directionalities distinguish parent/offspring relationships during the BN reconstruction, and do not necessarily imply causation flow (see Section 4 Materials and Methods: Bayesian networks modeling for details). “Tsize” node, highlighted in red, is the response status variable (4 responders, 9 nonresponders). Remaining nodes are immune marker variables. Response status node is directly strongly linked with five markers (TIGIT, CD4, CD8, CD160, 4-1BB, designated by dark grey nodes in the network) and, less strongly, with six other markers (TIM3, CD45RA, OX40, CXCR5, KLRG1, LAG3, light grey nodes in the network). See text for further details.

Figure 2

A full BN constructed from FACS data (innate panel) from 13 patients preimmunotherapy treatment (day 1). See Figure 1 legend for further details.

Figure 3

A full BN constructed from FACS data (Adaptive panel) from 13 patients preimmunotherapy treatment (day 1). See Figure 1 legend for further details.

Figure 4

BNs constructed from FACS data (Checkpoint panel, Non-naïve CD4+ T cells subdataset) from 13 patients. (A) day 1, nonresponders only, (B) day 1, responders, (C) day 21, nonresponders, (D) day 21, responders. See Figure 1 legend for further details.

Figure 5

A BN constructed from FACS data (Checkpoint panel, Non-naïve CD4+ T cells subdataset) from 13 patients. The “Contrast” node is the indicator variable with four states, (day 1/nonresponse, day 1/response, day 21/nonresponse and day 21/response). See Figure 1 legend for further details.

Figure 6

Contrast BNs constructed from FACS data (Checkpoint panel, Non-naïve CD4+ T cells subdataset) from 13 patients. (A) at day 1, responders vs. nonresponders, (B) at day 21, responders vs. nonresponders, (C) nonresponders, at day 1 vs. day 21, and (D) responders, at day 1 vs. day 21. “Day” node is the indicator variable with two states (day 1, day 21). “Response” node is the indicator variable with two states (response, nonresponse). See Figure 1 legend for further details.

Figure 7

Contrast BN constructed from FACS data (Adaptive panel, Naïve CD4+ T cells) from 13 patients preimmunotherapy treatment (day 1). “Response” (red node in the graph) is the indicator variable with two states (response, nonresponse). See Figure 1 legend for further details.

Figure 8

“Traditional” flow cytometry assessment of CXCR3 expression. Naïve CD45RA+ CD4+ CD3+ T cells were manually gated using FlowJo flow cytometry analysis software. The fraction (%) of naïve CD4+ T cells manually judged to express CXCR3 across patient responders (R) and nonresponders (NR) are shown. Unpaired two-tailed t-test resulted in p-value of 0.5496, indicating no significant difference between responders and nonresponders.

Figure 9

Effective cumulative distribution functions (ECDFs) for CXCR3 and other nine markers’ compensated fluorescent intensities, compared in responders and nonresponders. (Adaptive panel, Naïve CD4+ T cells, day 1). (A) CXCR3; (B) CCR10; (C) CD25; (D) CCR4; (E) CCR6; (F) PD-1; (G) CD127; (H) CD73; (I) CXCR5; (J) ICOS.

Figure 10

Probability distribution functions (PDFs) for CXCR3 compensated fluorescence intensities, compared in responders and nonresponders. (Adaptive panel, Naïve CD4+ T cells, day 1).

Figure 11

SHAP (Shapley Additive exPlanations) output of the RF model (Adaptive panel, Naïve CD4 cells, responders vs. nonresponders at day 1). The left panel shows a set of “beeswarm” plots reflecting individual variable importance (variables are ranked in descending order) and the impact on the model output (probability of response) depending on the value of the feature (marker). For instance, high values of CXCR3 (red color) act as a positive predictor (right side of the central axis) of response. The right panel shows an example of the interplay between two features, here CXCR3 and CCR10 (low values of CCR10 increase the predictive power of CXCR3). CCR10 was found to be the strongest modulator of CXCR3.