Network-based machine learning approach to predict immunotherapy response in cancer patients

Abstract

Immune checkpoint inhibitors (ICIs) have substantially improved the survival of cancer patients over the past several years. However, only a minority of patients respond to ICI treatment (~30% in solid tumors), and current ICI-response-associated biomarkers often fail to predict the ICI treatment response. Here, we present a machine learning (ML) framework that leverages network-based analyses to identify ICI treatment biomarkers (NetBio) that can make robust predictions. We curate more than 700 ICI-treated patient samples with clinical outcomes and transcriptomic data, and observe that NetBio-based predictions accurately predict ICI treatment responses in three different cancer types-melanoma, gastric cancer, and bladder cancer. Moreover, the NetBio-based prediction is superior to predictions based on other conventional ICI treatment biomarkers, such as ICI targets or tumor microenvironment-associated markers. This work presents a network-based method to effectively select immunotherapy-response-associated biomarkers that can make robust ML-based predictions for precision oncology.

Fig. 1. A network-based machine-learning (ML) approach to identify immunotherapy-associated biomarkers.

a Network visualization to identify genes proximal to immunotherapy targets in a protein-protein interaction (PPI) network. Immunotherapy targets (e.g., PD-1 for nivolumab) are displayed in blue and projected onto a PPI network, followed by network propagation using drug targets as seed genes. Network propagation is depicted as blue arrows. After propagation, drug target-proximal genes were selected by choosing nodes with high propagation scores (high-influence scores). b Identifying Network-based biomarkers (NetBio). Biological pathways (Reactome) enriched with high-influence score genes were selected via the hypergeometric test. c Input features used for machine learning to predict immunotherapy responders and non-responders. d Overview to measure predictive performances. For prediction objectives, we conducted predictions of the drug response and overall survival. For the training and test datasets, we conducted within-study predictions and across-study predictions.

Fig. 2. Predictions of drug response and overall survival for immunotherapy-treated patients.

a–d Immunotherapy-response prediction using the expression levels of drug targets (PD-1, PD-L1, or CTLA4) or network-based biomarkers (NetBio). Leave-one-out cross-validation (LOOCV) predictions for the (a) Gide, (b) Liu, (c) Kim, and (d) IMvigor210 datasets are plotted. Predicted responders (Pred R) and non-responders (Pred NR) are plotted against observed responders (teal) and non-responders (orange). The two-sided Fisher’s exact test was used to compute statistical significance. e–g Overall survival of predicted responders and non-responders based on LOOCV. The predicted responders and non-responders are depicted in red and blue, respectively. The log-rank test was used to measure statistical significance. The light-colored areas indicated 95% confidence interval of each percent survival. h–o LOOCV performance based on NetBio markers; gene-based markers, including PD-1, PD-L1, and CTLA4; and tumor microenvironment (TME)-based markers, including CD8 T cells, T-cell exhaustion, cancer-associated fibroblasts (CAFs), and tumor-associated macrophages (TAMs). GeneBio and TME-Bio include all of the target genes of each category. To quantify performance, we used (h–k) accuracy and (l–o) F1 score. Source data are provided as a Source Data file.

Fig. 3. Predictive performance in three independent melanoma datasets.

a Overall scheme of immunotherapy-response predictions in three independent datasets. The datasets and transcriptomic features used to train and test the machine-learning models and the number of samples for each dataset are displayed. b–d The area under the receiver operating characteristic curve (AUC) for the (b) Auslander, (c) Prat, and (d) Riaz datasets is shown. The random expectation, equaling an AUC of 0.5, is displayed as dotted lines. Expression profiles of cancer-associated fibroblast (CAF) marker genes were not available in the Prat dataset. N.D. not detected. Source data are provided as a Source Data file.

Fig. 4. Comparison of predictive performance using machine learning-based feature selection.

a Overall scheme for comparisons. b Overall predictive performance using NetBio-based or machine learning-based feature selection for 11 independent tests. Statistical significance was measured using the two-sided paired-sample t test. Boxplot shows median value, interquartile range (IQR) as bounds of the box and whiskers that extends from the box to upper/lower quartile ± IQR × 1.5. c Bar plots of predictive performances in 11 different tests, using accuracy, F1 score, or AUC as a metric to quantify performance. Source data are provided as a Source Data file.

Fig. 5. NetBio-based predictions recapitulate the immune microenvironment.

a Research scheme to compute the correlation between NetBio-based predictions and immunogenic features in the TCGA dataset. b Correlation between the predicted drug response using NetBio and immunogenic features in the TCGA cohort. Correlation was measured using Pearson’s correlation coefficient (PCC). c, d NetBio pathways identified from (b) Gide et al. and (c) Liu et al. are shown. The scatterplot displays the correlation between pathway expression and immunogenic features. The light-colored areas indicated 95% confidence interval of linear regression line. The PCC and correlation P values are shown. Source data are provided as a Source Data file.

Fig. 6. The expression levels of NetBio pathways are consistent with immunohistochemistry-based immune phenotypes in bladder cancer.

a Categorization of immune phenotypes using immunohistochemistry. b, c Expression levels of NetBio pathways in various immune phenotypes. For NetBio pathways, chemokine receptors bind chemokines (b) and FcgR activation (c) are shown. The two-sided Mann–Whitney U test was used to compute statistical significance for differential pathway expression levels across different immune phenotype patient groups. Boxplot shows median value, interquartile range (IQR) as bounds of the box and whiskers that extends from the box to upper/lower quartile ± IQR × 1.5. Source data are provided as a Source Data file.

Fig. 7. Combining network-based transcriptome features and the tumor mutation burden (TMB) improves the prediction of the overall survival in PD-L1 inhibitor (atezolizumab)-treated bladder cancer patients.

a Overall scheme of the network-based transcriptome and TMB combined predictions. b, c Prediction of the overall survival using (b) TMB-only and (c) a combined ML model. Statistical significance was measured using the log-rank test. The light-colored areas indicated 95% confidence interval of each percent survival. d Differential expression of the Raf activation pathway between predicted responders from TMB-only predictions and the reclassified subgroup from combined ML model predictions. Boxplot shows median value, interquartile range (IQR) as bounds of the box and whiskers that extends from the box to upper/lower quartile ± IQR × 1.5. The two-sided Student’s t test was used for statistical significance. e Network representation of the atezolizumab target (PD-L1) and Raf activation pathway. f Association between PD-L1 expression, the TMB levels and the expression levels of the Raf activation pathway with overall survival in bladder cancer patients (TCGA BLCA dataset). The light-colored areas indicated 95% confidence interval of each percent survival. Source data are provided as a Source Data file.