scRNA-seq data used in AI prediction models

scRNA-seq and cell-type annotation data (accession number GEO: GSE120575) of melanoma samples from Sade-Feldman et al. [24] were obtained and analyzed. The scRNA-seq dataset includes 48 tumor samples from 32 patients with metastatic melanoma treated with ICIs. This dataset was generated by sorting CD45 ⁺ single cells and profiling gene expression using the Smart-seq2 protocol. The dataset had already been preprocessed and filtered, and we analyzed these filtered data without applying any additional ad hoc filtering. Briefly, as described by Sade-Feldman et al. [24], FASTQ files were aligned to the human reference genome (GRCh37/hg19) using STAR. Gene expression was quantified in transcripts per million (TPM) using RSEM. Quality control excluded (i) cells with zero expression of both CD45 and CD3E, (ii) cells expressing fewer than 1,000 genes, and (iii) cells whose mean log₂(TPM + 1) expression of housekeeping genes was < 2.5.

Our study was conducted using publicly accessible data that are openly available to anyone and do not contain any personally identifiable information. As such, our study was exempt from IRB review.

scRNA-seq data visualization

The data (GSE120575) from a previous scRNA-seq study by Sade-Feldman et al. [24] of melanoma ICI immunotherapy were analyzed using the Seurat R package (version 4.4.0). We downloaded log2(TPM + 1) transformed expression values and used them for downstream analyses. As part of our quality control on the downloaded dataset, we confirmed that mitochondrial percentage was < 5% across cells. ICI treatment response data were available for each patient. The scRNA-seq data for each response group were as follows: responders (17 samples, 5,564 cells) and non-responders (31 samples, 10,726 cells). To identify high-variance genes, we selected genes with a variance > 6, following the same criterion used in a previous study [24], resulting in a total of 4,374 genes. These genes were used for clustering analysis for visualization. Briefly, principal component analysis (PCA) was used for dimensionality reduction, and the number of significant principal components was calculated using built ‘RunPCA’ function. The cells underwent unsupervised clustering, according to the shared nearest neighbor graph (‘FindNeighbors’ function, using the top 10 principal components (PCs); ‘FindClusters’ function, resolution = 0.5) and were visualized by uniform manifold approximation and projection (UMAP) using the top 10 PCs. We initially computed the first 10 principal components and used them in subsequent analyses. Cell type annotation information was obtained from the Sade-Feldman scRNA-seq study (GSE120575) [24].

Differentially expressed gene profiling and pathway analysis in scRNA-seq dataset

Differentially expressed genes (DEGs) from all single cells (n = 16,290) were identified using the Wilcoxon rank-sum test, with an adjusted P value<0.05 according to the Benjamini–Hochberg (BH) method and a log2-fold change > 1. Specifically, DEGs were identified from the downloaded log2-transformed TPM values using the Seurat R package, using the FindAllMarkers and FindMarkers functions to derive DEGs for each group according to a previous study [25]. To identify the meaningful biological pathways related to DEGs, clusterProfilter (v4.6.2) R package was used for Reactome enrichment analyses. clusterProfiler was also used for Signatures, with a false discovery rate (FDR) or BH-adjusted P-value of <0.05 (hypergeometric test) were considered significantly enriched. The immune exhaustion score was calculated using the AddModuleScore function in the Seurat, based on a gene set associated with T cell exhaustion, including PDCD1, CTLA4, LAG3, HAVCR2, and TIGIT [26]. DEGs between CCR7-positive and CCR7-negative B cells were identified using the Seurat FindMarkers function with parameters min.pct = 0.1 and logfc.threshold = 0.25. Multiple testing correction was performed on the raw P values using the FDR method.

Predictive model based on single-gene expression from scRNA-seq data

Six classification models were constructed to predict responders and non-responders based on gene expression data as follows: Extreme Gradient Boosting (XGBoost), Random Forest, Logistic Regression, Support Vector Machine (SVM), Feedforward Neural Network (FNN), and 1D Convolutional Neural Network (1D CNN). All six models use the same input format (cell by DEG expression matrix) of scRNA-seq data (S1A Fig). Feature scaling was performed using “StandardScaler.” The dataset was split into 80% training and 20% test sets, preserving the class distribution. For XGBoost, Random Forest, SVM, and FNN, hyperparameters were optimized using RandomizedSearchCV with stratified 5-fold cross-validation within the training set. The best models were selected based on receiver operating curve-area under the curve (ROC-AUC) score. The optimal models were then retrained on the full training set and evaluated on the test set. In contrast, the 1D-CNN architecture consisted of two Conv1D layers with ReLU activation, batch normalization, max pooling, and dropout regularization to reduce overfitting. The model was trained using the Adam optimizer with binary cross-entropy loss for up to 100 epochs, with early stopping (patient = 5). Model performance was evaluated using ROC-AUC for all classifiers, which was calculated on the held-out 20% test set, whereas accuracy and loss were specifically assessed for the 1D-CNN model. Feature importance was extracted from tree-based models (XGBoost and Random Forest). All analyses were performed in Python (version 3.10.12) using scikit-learn (version 1.6.1) for classical machine learning models and TensorFlow (v2.8.0) with Keras (v2.8.0) for deep learning models. For reproducibility, all random seeds were fixed to 42.

Predictive model based on pathway level from scRNA-seq data

To classify individual cells as originating from responder or non-responder samples, we implemented two deep learning architectures using Keras: a 1D-CNN and a two-dimensional convolutional neural network (2D-CNN). For the pathway-based model, we downloaded KEGG pathways (186 gene sets) from MSigDB (https://www.gsea-msigdb.org/gsea/msigdb). To generate the input file for the 1D-CNN model, we linked individual genes to their corresponding pathways. After removing duplicate gene–pathway combinations, we constructed expression matrices for a total of 12,413 unique gene–pathway pairs across 16,290 single cells (S1B Fig). Meanwhile, in the pathway-level modeling, the two types of information (gene names and pathway names) enabled a 2D data representation, allowing us to apply a 2D-CNN; accordingly, a 2D-CNN model was constructed. For the 2D-CNN model, we generated two-by-two matrix files for each of the 16,290 single cells, where each matrix consisted of gene expression values with 186 pathways represented as columns and 383 genes in the pathways as rows (S2 Fig in S1 File). Given that the input data for the pathway-based model were limited to 383 genes, the model was constructed without an additional feature selection step.

The 1D-CNN model received a flattened vector of gene-pathway features per cell with a shape of (12,413, 1). The flattened vector comprised three sequential convolutional blocks: the first block included 128 filters with a kernel size of 15, followed by ReLU activation, dropout (rate = 0.3), and max pooling (pool size = 2). The second block used 64 filters with the same kernel size and operations, and the third block used 32 filters with a kernel size of 5. The resulting feature maps were flattened and passed through a fully connected dense layer with 64 units and ReLU activation, followed by dropout and a final sigmoid-activated output layer for binary classification.

In parallel, the 2D-CNN model leveraged the spatial structure of the gene-pathway matrix. Each cell was represented as a grayscale image of size 383 × 186 (genes × pathways), with input shape (383, 186, and 1). The architecture consisted of two convolutional blocks: the first with 128 filters (3 × 3 kernel, ‘same’ padding), and the second with 64 filters (3 × 3 kernel), both followed by ReLU activation, dropout (rate = 0.3), and 2 × 2 max pooling. A global max pooling layer was then applied, followed by a dense layer with 64 units, ReLU activation, and dropout, culminating in a sigmoid-activated output layer.

Both models were trained using the Adam optimizer (learning rate = 1e-4) and binary cross-entropy loss. Early stopping (patience = 3) and model checkpointing based on validation accuracy were employed to mitigate overfitting. The 1D-CNN was trained with a batch size of 64 for up to 30 epochs using normalized input features, while the 2D-CNN followed the same training protocol using the reshaped 2D inputs. All models were trained using identical train/validation splits to ensure a fair comparison.

All implementations were developed using Keras (v2.8.0) with TensorFlow (v2.8.0) as the backend. To interpret the predictions of the 2D-CNN model, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the most influential gene-pathway regions. Grad-CAM was implemented using GradientTape from TensorFlow for automatic differentiation. Grad-CAM derives class-discriminative importance scores from the activation maps of the final convolutional layer, weighted by the gradients of the predicted class score. Accordingly, the resulting heatmaps represent activation-based contribution patterns learned by the 2D-CNN, rather than absolute gene expression levels. Color intensity reflects the relative importance of convolutional features associated with specific gene–pathway regions for the model’s prediction.

Specifically, we constructed a sub-model that outputs both the feature maps from the final convolutional layer and the model prediction. For a given input cell, the gradient of the predicted class score with respect to the final convolutional feature maps was computed using GradientTape. These gradients were spatially average-pooled to obtain importance weights for each feature map channel, which were then linearly combined with the feature maps to generate a class-discriminative heatmap. The resulting map highlights gene-pathway regions that most strongly contributed to the decision of the model. The performance of the proposed deep learning models (1D-CNN and 2D-CNN) was evaluated based on their ability to classify single cells according to patient response to ICI therapy (responder vs. non-responder). The ROC-AUC was used as the primary evaluation metric, given its robustness in binary classification tasks, especially under potential class imbalance.

To contextualize the performance of the CNN-based models, a series of conventional machine learning classifiers was also implemented as baselines. These included logistic regression, SVM, Random Forest, extreme gradient boosting (XGBoost), and feedforward neural network (FNN). All baseline models were trained on the same input data as the 1D-CNN (flattened gene-pathway feature vectors). Hyperparameters for each model were optimized using grid search or default configurations provided by standard machine learning libraries such as scikit-learn and XGBoost. The CNN and baseline models were trained and validated on identical data splits to ensure a fair comparison.

Bulk RNA-seq data analysis for AI predictive model validation

For validation purposes, bulk-RNA dataset of melanoma (Liu dataset) was also downloaded from cBioPortal (https://www.cbioportal.org) [27,28]. This dataset includes tumor tissue samples from 144 patients with advanced melanoma treated with anti-PD1 immune checkpoint blockage monotherapy, among whom RNA-seq data were available for 121 patients. Among them, 47 patients were classified as responders, including those with a complete response (CR) or partial response (PR) to ICI, whereas 72 non-responders exhibited progressive disease (PD) or stable disease (SD). Two patients exhibited heterogeneous treatment responses.

In addition, we downloaded raw RNA-seq FASTQ files from melanoma patient tissues treated with either anti-PD-1 monotherapy or combined anti-PD-1 and anti-CTLA-4 therapy (https://www.ebi.ac.uk/ena/browser/home; accession number PRJEB23709) (Gide dataset) [29]. Gene expression values were quantified using the Cancer Genome Atlas Program (TCGA) RSEM RNA sequencing analysis pipeline and normalized using the upper quartile normalization method [30]. Gene expression level was further transformed by log2(expression value +1) prior to downstream analyses. Because MTRNR2L2 was not included in the hg19 based reference gene annotation used by this pipeline, expression values were unavailable. The function of this gene has been elucidated only recently, and therefore this gene was not included in the annotation of RNA-seq analysis pipelines based on early versions of the hg19 reference genome. Data regarding the transcriptome-based molecular subtypes [31] for this dataset were retrieved from a study conducted by Bagaev et al. [32].

We also downloaded 108 bulk RNA-seq datasets (57 on-treatment and 51 pre-treatment) from 65 patients with melanoma, along with the corresponding clinical data, from the GEO database (GSE91061, Riaz dataset) [33]. The clinical characteristics and RNA sequencing platform information for the three independent datasets are summarized in S1 and S2 Tables in S1 File, respectively.

Signature scoring for gene set using bulk RNA-seq data

Signature scores were calculated using the singscore package (v1.19.1) [34] based on gene lists derived from a 2D-CNN prediction model using the KEGG pathway. These signature scores were applied to three independent melanoma bulk RNA-seq datasets from ICI-treated patients.

Statistical analysis

Continuous variables, such as gene expression and immune exhaustion score, were compared using the Wilcoxon rank sum test. Survival analysis was performed using the log-rank test, with visualization using Kaplan–Meier plots. Additionally, univariate or multivariate Cox proportional hazards regression analyses were performed using R (v4.1.2). Forest plots were visualized using the survminer R package (v0.5.0). Survival was analyzed with the survival R package (v3.2-13). All P values reported are two-sided. All statistical analyses were performed with R (v4.1.2) or Python (v3.10.12).

Malignant melanoma TME landscape and immuno-therapy response marker identification

Using the metastatic melanoma scRNA-seq dataset (GSE120575), we obtained a raw UMI matrix of melanoma tissues. After filtration, the expression profiles of 16,290 cells were processed using Seurat R package. UMAP and previously published cell type information revealed distinct cell types, including T lymphocytes and myeloid, B, and plasma cells (Fig 1A). These cell types were identified based on the expression of canonical markers, such as CD3D and CD3E, in T lymphocytes [24]. Additionally, when classified by immunotherapy responsiveness, differences in cell composition and the proportion of specific cell types were observed (Fig 1B and S3 Fig in S1 File). To investigate the difference in transcriptional profiles between responders and non-responders independent of cell types, DEGs were identified between the two groups. This analysis identified 210 DEGs between responders and non-responders, including 86 genes upregulated in responders and 124 genes upregulated in non-responders (Fig 1C). In the non-responder group, gene expression was enriched for pathways related to neutrophil degradation, interferon signaling, interferon-gamma signaling, and MHC class II antigen presentation. In contrast, the responder group exhibited activation of pathways associated with B cell response and B cell receptor signaling (Fig 1D).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. scRNA-seq characteristics of metastatic melanoma by immunotherapy response.

(A) UMAP visualization of cell types in melanoma tissues. (B) UMAP plot illustrating the distribution of immunotherapy responders and non-responders. Responders (blue) and non-responders (gray) exhibit distinct clustering patterns. (C) Volcano plot showing differentially expressed genes (DEGs) between responders (n = 86) and non-responders (n = 124) identified across all cells regardless of cell type from scRNA-seq data (|avg log2FC| > 1). Top 10 genes based on log2FC in responders and non-responders are labeled. (D) Reactome pathway analysis of highly expressed genes identified from scRNA-seq data in responders (left) and non-responders (right). scRNA-seq, single cell RNA sequencing; FC, fold change.

https://doi.org/10.1371/journal.pone.0343633.g001

Key gene contributors identified by feature selection from gene-based prediction model using identified DEGs of scRNA-seq data

Six classification models were built based on the 210 DEGs selected through feature selection to predict responders and non-responders (Fig 2A and S1A Fig in S1 File). In this study, XGBoost, Random Forest, logistic regression, SVM, FNN, and 1D CNN were trained and evaluated. Among these models, XGBoost achieved the highest performance with an area under the ROC-AUC of 0.87, followed by Random Forest (AUC: 0.86) and SVM (AUC: 0.86) (Fig 2B). In contrast, logistic regression (AUC: 0.82) and 1D-CNN (AUC:0.84) showed the lowest performance among the models. To identify the most influential genes with high feature importance in tree-based models with the highest performance, we compared feature importance rankings derived from XGBoost and Random Forest. A significant correlation (R = 0.47, P = 6.6e-13) was observed between the rankings derived from both tree-based methods, highlighting a subset of 29 genes as consistently important for prediction (Fig 2C). As a supplementary assessment of reproducibility, we performed 100 independent random splits of the dataset and repeatedly computed the AUC to derive confidence intervals (Fig 2D). The resampled estimates showed a strong correlation with the original test results (R = 0.93, P = 0.0077), and XGBoost consistently achieved the highest predictive performance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Gene-based classification model using scRNA-seq data for immunotherapy response.

(A) Schematic overview of classification model building using differentially expressed genes (DEGs). Dataset (n = 16,290; baseline and post-treatment) was split into 80% training and 20% test sets. Various models, including XGBoost, Random Forest, Logistic Regression, SVM, FNN, and 1D CNN, were trained using the training set. Hyperparameters were optimized through cross-validation and the best parameters were used to retrain the models. The retrained models were then evaluated on testing set to obtain final performance metrics. (B) ROC curves of classification models. XGBoost achieved the highest area under the curve (AUC = 0.87), followed by Random Forest (AUC = 0.86) and SVM (AUC = 0.86). (C) Scatter plot comparing feature importance rankings from XGBoost and Random Forest. A subset of 29 genes was consistently identified as important for classification. (D) To assess the reproducibility of the predictive model, we permuted the dataset 100 times and measured AUC of the model; the results were similar to the original test (Spearman correlation test). (E) Dot plot showing expression levels of 29 genes in responders and non-responders (related to Fig 2C). (F) Differences in CCR7 and MTRNR2L2 expression between responders and non-responders in baseline and post-treatment samples, respectively (Wilcoxon rank-sum test). %Exp indicates the percentage of cells expressing the gene.

https://doi.org/10.1371/journal.pone.0343633.g002

Expression of these key 29 genes showed distinct patterns between responder and non-responder groups (Fig 2E, S4 Fig, and S3 Table in S1 File). Among these genes, analysis of their association with post-treatment survival using the Liu dataset revealed that CCR7 and MTRNR2L2 were significantly correlated with patient survival. Accordingly, we focused our subsequent analyses on these two genes. Notably, when baseline and post-treatment scRNA-seq samples were analyzed separately, these genes remained significant regardless of treatment time point (Fig 2F).

CCR7 and MTRNR2L2 expression patterns across immune cell types derived from scRNA-seq data

Using cell type annotation information from the scRNA-seq data, we examined the expression patterns of the CCR7 and MTRNR2L2 genes. CCR7 showed high expression in the responder group, consistent with its identification in the prediction model, and was particularly enriched in B cells and memory T cells (Fig 3A). In contrast, MTRNR2L2 exhibited overall high expression in the non-responder group, with elevated expression observed across various T cell populations, including exhausted CD8 ⁺ T cells and cytotoxic lymphocytes (Fig 3B). The remaining 27 genes are presented in S5 Fig in S1 File. Functionally, pathway analysis revealed that CCR7-positive B cells showed enrichment of immune activation-related pathways, such as NF-κB signaling, compared to CCR7-negative B cells (Fig 3C). MTRNR2L2 expression displayed a distinct bimodal distribution pattern (Fig 3D), with the high-expression group exhibiting a significantly elevated immune exhaustion score (Fig 3E). Cells with high MTRNR2L2 expression exhibited a higher proportion of exhausted CD8 ⁺ T cells than low-expression cells (Fig 3F).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Expression patterns of CCR7 and MTRNR2L2 in scRNA-seq data.

(A) CCR7 is predominantly expressed in B cells and memory T cells within the responder group (Wilcoxon rank-sum test). (B) MTRNR2L2 was broadly expressed across multiple cell types; however, its expression was particularly elevated in exhausted CD8 ⁺ T cells within the non-responder group (Wilcoxon rank-sum test). (C) Pathway enrichment analysis of differentially expressed genes between CCR7-positive and CCR7-negative cells within the G1-B cell cluster revealed enrichment of B cell activation-related pathways, including NF-κB signaling, in CCR7-positive B cells. (D) Biomodal distribution pattern of MTRNR2L2 gene expression. (E) High MTRNR2L2 expression was associated with elevated immune exhaustion scores (Wilcoxon rank-sum test). (F) Proportion of exhausted CD8 ⁺ T cells increased in high-MTRNR2L2 expression group.

https://doi.org/10.1371/journal.pone.0343633.g003

The functions of other key genes identified in our analysis (IFITM3, MT2, MTRNR2L8, IFI6, HSPA1A, HSPA1B, RPL9P9, and RPL9P8) in the immune system and cancer are summarized in S4 Table in S1 File [35–47]. Several of these genes participate in inflammatory responses within the immune microenvironment and are implicated in tumor growth and oncogenesis. For example, IFITM3 exhibits dual functionality, contributing to antiviral defense and promoting tumorigenesis [35–37,47]. HSPA1A and HSPA1B, members of the heat-shock protein family, can augment immune activation; however, polymorphisms in these genes have been associated with increased melanoma risk [46]. MT2A is overexpressed in melanoma and has been implicated in downregulating adaptive immunity [38,39].

CCR7 and MTRNR2L2 expression validation using independent bulk RNA-seq cohorts

We examined the association between CCR7 and MTRNR2L2 expression and immunotherapy response and prognosis in three independent cohorts of melanoma patients treated with immune checkpoint inhibitors (ICIs) (Fig 4). In the Liu dataset, CCR7 expression was associated with favorable survival, whereas MTRNR2L2 expression was associated with unfavorable survival. These findings were further validated using two independent datasets (the Gide and Riaz datasets). Although some associations did not reach statistical significance, consistent trends were observed across all datasets. In summary, across all three datasets, CCR7 tended to be highly expressed in patients with favorable treatment responses, and higher CCR7 expression showed a tendency to be associated with improved survival. In contrast, MTRNR2L2 exhibited the opposite trend. Similar patterns were observed in both pre-treatment and on-treatment samples. The expression patterns of the remaining key genes identified by the predictive model, stratified by treatment response, are presented in S6 Fig in S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Clinical significance of CCR7 and MTRNR2L2 across three independent melanoma cohorts with bulk RNA-seq data.

Box plots displaying the expression of genes (CCR7 and MTRNR2L2) in three independent bulk RNA-seq data from melanoma patients with CR/PR (complete response/partial response) and PD/SD (progressive disease/stable disease). Kaplan–Meier survival curves demonstrating associations of genes with overall survival (OS) and progression-free survival (PFS) (log-rank test). Hazard ratio (HR) and 95% confidence intervals from the univariate Cox regression analysis. High CCR7 expression is associated with improved survival, whereas high MTRNR2L2 expression is associated with worse prognosis.

https://doi.org/10.1371/journal.pone.0343633.g004

Prognostic significance of CCR7 and MTRNR2L2 according to histopathologic and molecular subtypes

The three melanoma cohorts used for validation consisted of melanomas from various primary sites. Using these cohorts, we investigated the prognostic significance of CCR7 and MTRNR2L2 genes across histopathological and molecular subtypes. The prognostic significance of CCR7 and MTRNR2L2 expression was consistent regardless of the primary site (Fig 5A) or histologic subtype (Fig 5B) of melanoma. In contrast, when examining the three molecular subtypes of cutaneous melanoma, the immune subtype was associated with the most favorable prognosis, whereas the MITF-low subtype showed the poorest prognosis (Fig 5C). CCR7 gene expression was significantly associated with molecular subtype, showing markedly higher expression in the immune subtype (Fig 5D). The prognostic significance of CCR7 expression was lost in multivariate analysis that included molecular subtype (Fig 5E). When we examined the mutational subtypes using the Riaz dataset [33], CCR7 and MTRNR2L2 expression showed no association with mutational subtype (Fig 5F).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. CCR7 and MTRNR2L2 expression according to the melanoma subtypes.

(A, B) CCR7 and MTRNR2L2 expression showed prognostic significance for overall survival, independent of the primary site (A) and histological subtype of melanoma (B). (C) Prognostic differences according to molecular subtypes, including microphthalmia-associated transcription factor (MITF)-low, immune, and keratin subtypes (log-rank test). (D) CCR7 expression patterns across molecular subtypes (Kruskal–Wallis test). (E) Multivariate Cox regression analysis of CCR7 expression in relation to molecular subtype. ALM, acral lentiginous melanoma; LMM, lentigo malignant melanoma; NM, nodular melanoma; SSM, superficial spreading melanoma. (F) Gene expression patterns across mutational subtypes (Kruskal–Wallis test).

https://doi.org/10.1371/journal.pone.0343633.g005

Pathway-based deep learning approach for predicting immunotherapy response

Although individual genes—such as CCR7 and MTRNR2L2—showed associations with ICI response and prognosis of patients, we examined whether prediction was possible at the pathway level. KEGG pathway-level analysis showed varying predictive performance across different classification models (Fig 6A). In the pathway-level approach, XGBoost (AUC: 0.882) and Random Forest (AUC: 0.842) showed the highest predictive performance, followed by 1D-CNN (AUC: 0.810), 2D-CNN (AUC:0.806), FNN (AUC: 0.804), Logistic Regression (AUC: 0.767), and SVM (AUC: 0.761). To identify and visualize the contribution of individual gene–pathway combinations to ICI response prediction in the 2D-CNN model, we applied the Grad-CAM technique. For both responder and non-responder groups, we generated heatmaps with genes on the x-axis and pathways on the y-axis, where higher values indicate gene–pathway combinations that contributed most to model prediction within each response category (S7 Fig in S1 File). This visualization enabled us to distinguish gene–pathway associations specific to either responders or non-responders. Accordingly, the feature importance analysis of the 2D-CNN model revealed key gene–pathway features that contributed to the prediction of response and non-response (S7 Fig in S1 File). When the differences in gene–pathway feature values between responders and non-responders were visualized using a heatmap (Fig 6B), the two spots showing the largest positive and negative differences each encompassed 64 genes across eight pathways, suggesting that these features represent the most critical elements characterizing responders and non-responders, respectively. This region included a substantial number of immune-related genes, such as immune checkpoint gene CD274, HLA genes, and interferon-associated genes. Conversely, the corresponding region in non-responders primarily involved nine genes belonging to the cell cycle pathway (Fig 6B). The CCR7 gene, which was identified at the single-gene level, was not ranked among the top features in the pathway-level prediction model. Meanwhile, when responder-specific signature scores derived from genes in these pathways from the highest score spot (Fig 6B, red box) were calculated across three independent bulk RNA-seq datasets, no significant differences in prognosis were observed, although a trend toward improved survival was noted (Fig 6C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Pathway-based classification models based on the KEGG pathway.

(A) Single-cell gene expression mapped to KEGG pathways enabled predictive models showing relatively high performance across multiple AI algorithms. (B) Heatmap (left) showing differences in 2D-CNN feature-importance between responders and non-responders; numbers on x- and y-axes denote distinct genes and pathways, respectively. List of pathways and their constituent genes showing the largest differences between responders and non-responders (right panel). (C) Prognostic analysis across three bulk RNA-seq datasets using responder-specific signature scores derived from genes within the eight responder-associated pathways identified in the highest score spot in Fig 6B (log-rank test).

https://doi.org/10.1371/journal.pone.0343633.g006