Prediction of Lung Cancer Immunotherapy Response via Machine Learning Analysis of Immune Cell Lineage and Surface Markers

Cancer Biomark 2022 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Overview: Machine Learning to Predict Immunotherapy Response from Blood Immune Profiles

Clinical Need Immune checkpoint inhibitors targeting PD-1 and PD-L1 (pembrolizumab, nivolumab, atezolizumab, durvalumab) benefit only about 20% of non-small cell lung cancer (NSCLC) patients. Identifying responders before treatment would improve outcomes and avoid exposing non-responders to side effects.

Study Design This pilot study applied four machine learning algorithms to mass cytometry (CyTOF) data measuring 30-plus immune cell surface markers from blood samples of 13 NSCLC patients at baseline and after 12 weeks of immunotherapy.

Four Comparisons The analysis evaluated four patient group comparisons: healthy controls versus patients at baseline, responders versus non-responders at baseline, healthy controls versus 12-week responders, and responders versus non-responders at 12 weeks.

Top Performance The Regularized Random Forest (RRF-RF) model achieved AUC 0.918 for distinguishing responders from non-responders at baseline -- before treatment begins. This pretreatment discriminability is the most clinically valuable finding.

TL;DR: Machine learning analysis of blood immune cell profiles from mass cytometry achieved AUC 0.918 for predicting which NSCLC patients will respond to checkpoint inhibitor immunotherapy before treatment starts.
Pages 2-3
Mass Cytometry: Measuring the Immune System in Detail

CyTOF Technology Time-of-flight mass cytometry (CyTOF) simultaneously measures 30-plus markers on thousands of individual immune cells from a single blood sample. Unlike traditional flow cytometry limited to 10-15 markers, CyTOF uses heavy metal-tagged antibodies detectable by mass spectrometry.

Maxpar Direct Kit The Maxpar Direct Immune Profiling kit (Fluidigm) provided a standardized panel of 30 markers covering all major immune cell populations including T cells, B cells, NK cells, monocytes, and dendritic cells.

Data Dimensionality Each patient sample generated a high-dimensional dataset with hundreds of features representing the abundance of each immune cell type and the expression level of each surface marker on each cell population -- creating complex data ideal for machine learning.

Sample Timing Samples were collected at two time points: before treatment (baseline) and after 12 weeks of immunotherapy. This longitudinal design allows both pretreatment prediction and post-treatment immune profile changes to be characterized.

TL;DR: Mass cytometry simultaneously measured 30+ immune markers on blood cells from 13 patients at two time points, generating rich high-dimensional data that machine learning algorithms could mine for response predictors.
Pages 3-4
Machine Learning Methods and Feature Selection

Four Algorithms Applied Random Forest (RF), Partial Least Squares Discriminant Analysis (PLS-DA), Multi-Layer Perceptron (MLP), and Elastic Net (EN) were tested for each of the four patient comparisons, with feature selection integrated into the RF and MLP models.

Regularized Random Forest RRF-RF (Regularized Random Forest for feature selection followed by Random Forest prediction) was the best model for responder-vs-non-responder comparisons. Regularization prevents overfitting by penalizing unnecessary variable inclusion.

0.632 Bootstrapping Validation Because the cohort was too small for standard k-fold cross-validation, 0.632 bootstrapping with 10,000 iterations was used. This validated approach corrects for the optimistic bias of in-bag training estimates while accommodating small sample sizes.

Feature Importance Analysis Variable importance was computed for each algorithm. Features consistently ranked highly across multiple models and also statistically significant by Wilcoxon rank-sum test were designated as key predictive features.

TL;DR: Four ML algorithms with 10,000-iteration bootstrapping validation were applied; RRF-RF best separated responders from non-responders, with feature importance analysis identifying key immune markers.
Pages 5-7
B-Cell Markers as Key Immunotherapy Response Predictors

Unexpected B-Cell Dominance Contrary to the traditional T cell-centric view of anti-tumor immunity, B-cell markers dominated the feature importance rankings across models. CD20+ B-cell subpopulations and their surface markers (IgD, IL-7Ra, CD27, CXCR5) were repeatedly identified as top predictors.

Responders vs Non-Responders at Baseline Patients who later responded to immunotherapy had more CD20+ CD27+ B-cells, more natural killer (NK) cells, and more CD4+ T-cells at baseline. This suggests that pre-existing immune activity predicts treatment responsiveness.

B-Cell Proportions as Novel Features Ratios of surface markers between different B-cell subpopulations (e.g., IgD on CD27+ cells divided by IgD on CD38+ cells) were found predictive -- a novel feature type not previously explored for immunotherapy response prediction.

12-Week Changes After 12 weeks of treatment, responders showed lower CXCR5 on CD20+ CD27+ B-cells and lower CD27 on CD20+ CD38+ B-cells compared to healthy controls, suggesting immunotherapy-specific immune remodeling in responding patients.

TL;DR: B-cell populations and surface markers -- not the expected T cells -- were the dominant predictors of immunotherapy response, with novel marker ratio features providing additional discriminative power.
Pages 7-8
NK Cells and CD4+ T Cells Support Response Prediction

NK Cell Abundance Higher abundance of natural killer (NK) cells at baseline was a top feature distinguishing responders from non-responders. NK cells kill tumor cells through innate mechanisms that operate independently of tumor antigen recognition, suggesting baseline immune competence matters.

CD4+ T Cell Role Higher abundance of CD4+ helper T cells at baseline was also predictive of response. CD4+ T cells coordinate adaptive immune responses and are increasingly recognized as important mediators of checkpoint inhibitor activity.

Immunosuppression Context Non-responders tended to have lower NK cell counts, possibly reflecting stronger tumor-mediated immunosuppression. Tumors that have already suppressed NK cells at diagnosis may be harder to re-activate with checkpoint blockade.

IL-7Ra Expression Diminished IL-7 receptor alpha (IL-7Ra) expression on B-cells was observed in NSCLC patients compared to healthy controls. IL-7 supports B-cell proliferation, and reduced IL-7Ra may indicate a less capable humoral immune response in cancer patients.

TL;DR: Higher NK cell and CD4+ T cell counts at baseline characterized future responders, suggesting that patients with less tumor-mediated immune suppression are better positioned to benefit from checkpoint inhibitors.
Pages 9-10
Limitations and Future Research Directions

Very Small Cohort Thirteen patients is a pilot-scale sample insufficient for definitive conclusions. The high AUC values observed may partially reflect overfitting despite bootstrapping; validation in a larger independent cohort is essential before clinical translation.

Mixed Immunotherapy Agents Patients received four different checkpoint inhibitors (pembrolizumab, nivolumab, atezolizumab, durvalumab). While these target the same PD-1/PD-L1 axis, they differ in molecular binding and side effect profiles. Pooling them assumes interchangeability which may not be fully valid.

Potential Confounders Age, ethnicity, smoking history, and prior treatment status vary across patients and could confound results. The small sample size makes adjusting for these variables impossible while maintaining statistical power.

Path to Clinical Translation The approach requires validation in 100+ patients across multiple institutions, with standardized CyTOF protocols. If validated, the blood-based nature of the test -- avoiding tissue biopsy -- would be a major advantage for repeated monitoring during treatment.

TL;DR: The 13-patient pilot cohort is insufficient for definitive conclusions, but the blood-based approach and novel B-cell findings justify a larger multi-center validation study to determine clinical utility.
Citation: Open Access, 2022. Available at: PMC12364200.