LUADnet: a deep learning model for prediction of clinical outcomes in lung adenocarcinoma based on gene expression signatures

Introduction

Lung cancer remains the leading cause of global cancer mortalities (1). Lung adenocarcinoma (LUAD) is the most prevalent non-small cell lung cancer (NSCLC), a subtype accounting for 85% of all lung cancer (2). Currently, immuno-chemotherapy combinations are widely available for advanced or metastatic NSCLC without harboring an epidermal growth factor receptor (EGFR), ALK or other driver gene variations (3). However, there are several pitfalls: (I) primary and secondary immune escape develop during drug exposure (4); (II) immune-related adverse events (irAEs) were associated with survival outcome of nivolumab treatment in advanced or recurrent NSCLC (5); (III) Immune checkpoint inhibitor (ICI) agents for treatment of advanced NSCLC are involved in a substantial economic burden (6); (IV) the objective response rate was only 35.9% in NSCLC patients receiving nivolumab plus ipilimumab (7). Therefore, it is crucial to stratify patients into response and non-response groups before immuno-chemotherapy. Therefore, identifying and differentiating patient subgroups that are potentially suitable or unsuitable for immunotherapy—particularly understanding the response differences among various populations at the molecular and microenvironmental levels—has become a crucial prerequisite for optimizing clinical decisions. Accurately categorizing patients into potential responders and non-responders before treatment is essential for achieving individualized, stratified therapy, improving overall efficacy, reducing adverse reactions, and rationally allocating medical resources. The Cancer Genome Atlas (TCGA) database contains patients who have not received immunochemotherapy, making it suitable for selecting patients for consideration of immunechemotherapy. Luadnet can identify responsive patients who are unable to receive immunochemotherapy, thereby saving patients money and conserving medical resources. Meanwhile, this model can identify non-responsive patients, making immunochemotherapy a potential choice.

Deep learning is a vital method for discovering new things about patients by analyzing biological data with mathematical modeling and applying it to clinical and scientific research (8). The classification of gene expression level has always been a key research direction in the field of medical disease research. Quang et al. (9) proposed the domain-adversarial neural network (DANN), which was the first to apply deep learning to genomic sequence analysis. This method can capture the nonlinear relationships between features and is more suitable than support vector machine (SVM) for handling problems with large amounts of data and features. Fiannaca et al. (10) proposed a deep learning method for metagenomic data classification, which uses k-mer representation to map sequences as vectors into numerical space, and trains convolutional neural networks (CNNs) and deep belief networks (DBNs). Li et al. (11) proposed a hybrid category attention neural network called DeepATT for identifying the functional effects of DNA sequences. This method designs appropriate feature extraction and selection methods for specific functional effects, enabling a better understanding of their internal correlations. Yang et al. (12) proposed DNASimCLR, a method that uses CNNs and a contrastive learning-based SimCLR framework to extract complex features from different gene sequences. Lilhore et al. (13) proposed a fusion model for protein sequence classification called ProtICNN-BiLSTM. They seamlessly integrated an attention-based improved convolutional neural network (ICNN) with bidirectional long short-term memory (BiLSTM) units, improving the accuracy of protein sequence classification. However, CNNs have limitations in capturing long-distance dependencies among features in sequences. While some algorithms use recurrent neural network (RNN), long short-term memory (LSTM), and other methods to try to capture long- and medium-range features in DNA/RNA sequences, they have not achieved ideal results. Recently, prognosis models of LUAD based on gene signatures have primarily utilized machine learning algorithms such as least absolute shrinkage and selection operator (LASSO), random survival forest, CoxBoost, generalized boosted regression modeling, and support vector machines (14). However, these approaches may overlook the complex relationships among gene expression levels. To solve these issues, we propose a LUADnet classification model based on multi-scale CNN and self-attention mechanisms. Specifically, we first use the ResNet50 network to extract deep features from the RNA sequencing data. Then, we employ two parallel modules: a global feature extraction module (GFEM) and a multi-scale local feature extraction module (LFEM). In the GFEM, we use a self-attention mechanism to establish the interrelationships between each element of the expression level, enhancing the network’s ability to perceive the global characteristics of the expression level. In the multi-scale LFEM, we primarily focus on the relationship between an element and its neighboring elements, enabling the model to simultaneously learn and extract local features at multiple scales. After obtaining both global and local features, we propose a channel importance selection module, which can filter the global and local features and increase the weight of those features that play a more significant role in subsequent feature extraction and final classification. Finally, we summarize the extracted features from each layer and input them into the classifier as the network’s final output.

In this study, we first depicted the transcriptional profiling of a LUAD patient. Then, we built the LUADnet classification model to predict the response of LUAD patients to non-immunotherapy treatments. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1391/rc).

Methods

Dataset collection and processing

Patient clinical data, survival data, and RNA-seq data were obtained from the UCSC Xena database (https://xena.ucsc.edu/). Genomic alterations were collected from TCGA database (https://cancergenome.nih.gov/). Level 3 log₂(count + 1) data were converted to count data using R. Incomplete and duplicate patient information were excluded. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Survival analysis

Statistical analysis of survival data and prognostic value was visualized using the Survival and Survminer packages (version 0.4.9). Kaplan-Meier curves were conducted to evaluate overall survival (OS). For survival analysis, patients were stratified into response and non-response groups based on clinical information. Survival curves were estimated using the Kaplan-Meier method.

Differentially expressed genes (DEGs) analysis

Differential expression analysis of sequencing data was conducted using the DESeq2 package (15) in R, based on the criteria of |log₂fold change| >0.5 and P value <0.01. The analysis was conducted using default parameters for library size normalization and dispersion estimation. The top 10 DEG were illustrated in a heat heatmap.

Functional enrichment analysis of DEGs

The clusterProfiler R package (version 4.4.3) was used to perform Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses (16) following the criteria of adjusted P value <0.05. To identify biological processes associated with the prognostic signature, we separately analyzed upregulated and downregulated DEGs. The background gene set was defined as the human genome-wide gene set provided by the org.Hs.eg.db database.

Hardware and software for model construction

The experiments in this study are conducted on a high-performance computing platform featuring an Intel^® Xeon^® Platinum 8336C processor (2.30 GHz) and eight NVIDIA GeForce RTX 4090 GPUs, running Ubuntu 20.04.6. This hardware setup provides the computational power necessary for efficient model training. On the software side, our implementation is built using the latest PyTorch version 2.7.0. The ResNet50 architecture is instantiated via the torchvision.models module. Core components of the model—including multi-layer perceptrons, convolutional layers, and attention mechanisms—are constructed using PyTorch’s torch.nn package, specifically leveraging nn.Linear, nn.Conv1d, and nn.MultiheadAttention.

The model is implemented based on PyTorch and trained on GPUs with CUDA support. Cross-entropy loss is used for binary classification learning, with the Adam optimizer employed for parameter optimization—learning rate is set to 1×10⁻⁴, and weight decay to 1×10⁻³.

Training utilizes mini-batch gradient updates with a training batch size of 16 and a validation batch size of 64. The maximum number of training epochs is set to 200. Regarding data partitioning, a fixed test set of 63 samples is used, including 40 response and 23 non-response cases. All remaining samples are allocated to the training set to maximize the data available during training.

To address class imbalance in the training set, we apply SMOTE only during the training phase to perform synthetic oversampling of the minority class, thereby mitigating model bias toward the majority class. The test set maintains its original distribution without any resampling.

Given that the input consists of one-dimensional (1D) sequential data, we introduce random perturbation-based data augmentation during training to enhance robustness. Specifically, Gaussian noise disturbance is applied to the standardized input vector, where each dimension of the noise follows, with noise intensity determined based on the scale of the training data and generated online during training. No perturbation is applied during testing. To improve statistical reliability, all model experiments are independently repeated across 10 different random seeds, and the average of the test set metrics is taken as the final result.

Classification via novel deep learning technique

We use the gene expression data as the input for our deep learning structure, namely LUADnet. LUADnet consists of multiple layers of feature extractors and classifiers. We use the ResNet50 architecture as the backbone network for the feature extractor and modify the two-dimensional (2D) convolutions to 1D convolutions to fit the structure of the data. Additionally, due to the large volume of data, using a 1×3 kernel size may not effectively capture larger feature patterns. Therefore, we modify the convolution kernel size for feature extraction to 1×255 to accommodate the length of the mRNA count data.

The structure of the proposed LUADnet is shown in Figure 1A. We integrate the proposed GFEM, multi-scale LFEM, and feature importance selection module into each feature extraction block, enabling the feature extraction block to capture more comprehensive features. In the GFEM, we use a self-attention mechanism to establish the interrelationships between each gene expression data, enhancing the network’s ability to perceive the global structure of the data. In the multi-scale LFEM, we mainly focus on the relationships between an element and its neighboring elements, allowing the model to learn and extract local features at multiple scales simultaneously. The channel importance selection module is used to filter the global and local features, increasing the weight of features that are more important for subsequent feature extraction and final classification. Finally, we aggregate the features extracted from each layer and input them into the classifier as the network’s final output.

Figure 1 Diagram of the construction of LUADnet. (A) Classification via novel deep learning technique. (B) Dilated convolutions with four different dilation rates to extract multi-scale features. (C) The standard convolution kernel and dilated convolution kernel. (D) The schematic of the GFEM. (E) The schematic diagram of the CSM. BN, batch normalization; CSM, channel selection module; GFEM, global feature extractionmodule; K, key; LFEM, local feature extraction module; MLP, multilayer perceptron; Q, query; ReLU, rectified linear unit; SA, Self-Attention; V, value.

The network replaces all 2D convolution and pooling operations with their 1D counterparts while preserving the residual addition mechanism. After receiving a single-channel 1D sequence as input, the stem layer first applies a 1D convolution with 1 input channel and 64 output channels, a kernel length of 511, a stride of 2, and padding of 255. This is followed by a 1D max pooling operation with a kernel length of 255, a stride of 2, and padding of 127. Consequently, the feature length undergoes two stride-2 downsampling steps in the stem stage.

The backbone consists of four stages with channel sizes of 64, 128, 256, and 512, respectively. In stages 2 to 4, the first residual block employs a stride of 2 for further downsampling, while stage 1 does not downsample. Residual connections are strictly preserved in the code: when the stride is not 1 or when the input and output channel numbers differ, the shortcut branch explicitly introduces a downsampling projection using a 1D convolution with a kernel length of 1 and the same stride as the main branch, followed by batch normalization. This ensures that the outputs from the main and shortcut branches are perfectly aligned in both length and channel dimensions before element-wise addition.

The second 1D convolution within each residual block uniformly uses a kernel length of 255 and padding of 127. When the stride is 1, the feature length remains unchanged, and only the receptive field is enlarged; downsampling occurs only when this convolution is set with a stride of 2, and the dimensional alignment is simultaneously handled by the projection shortcut described above. Therefore, changes in feature resolution are entirely determined by explicit stride settings in the code, and increasing the kernel length itself does not further reduce resolution when the stride is 1.

Moreover, the backbone network still follows the ResNet50 paradigm, preserving the element-wise addition mechanism for cross-layer skip connections and achieving dimensional matching at the entry of each stage through projection shortcuts. This ensures that the key design characteristics of ResNet50 are retained after adaptation to 1D.

LFEM

Local features of mRNA phenotype refer to specific gene expression patterns that influence its biological function and phenotype. By extracting local features, the model can capture short-range dependencies and specific biological patterns, which are particularly important for improving classification performance. Numerous studies have demonstrated that multi-scale features can enhance the performance of classification models. By employing a multi-scale approach, the model can simultaneously learn and extract local features at multiple scales, thereby capturing various gene expression patterns.

As shown in Figure 1B, we use dilated convolutions with four different dilation rates to extract multi-scale features from the gene expression data. Dilated convolution is a convolution technique that expands the receptive field by inserting gaps between elements in the convolution kernel. It can extract multi-scale features without increasing the number of parameters or computational complexity, which is especially advantageous when processing higher-dimensional data. For the 1D dilated convolution, we adopt the standard notation to explicitly specify the summation index and its upper bound, expressing the convolutional output as:

y[i]=∑k=1K−1w[k]⋅x[i+k⋅r][1]

where y is the convolutional output, w denotes the convolutional kernel weights, K is the kernel size, is the kernel index (with being the upper bound of the index), x is the input sequence, and r is the dilation rate of the convolution. After applying a dilated convolution with a dilation rate r, the effective kernel size increases from K to R=K−1⋅r+1. When r=1, the dilated convolution reduces to a standard convolution with kernel size K. Figure 1C below shows the schematic of dilated convolution: a depicts a standard 1×3 convolution kernel, with a receptive field of 1×3; b shows a 1×3 dilated convolution kernel with a dilation rate r=2, resulting in a receptive field of 1×5. The receptive field of the convolution layer for standard convolutions depends on the size and stride of all previous convolution kernels. While the receptive field grows linearly in standard convolutions, dilated convolutions can achieve exponential growth of the receptive field.

After extracting multi-scale features, we add spatial attention layers and residual blocks as supplementary information to enhance the final multi-modal fused features.

Fisa=SA(Fi)[2]

Fire=Conv1×1(Fi)[3]

FiL=FiM×Fisa+Fire[4]

Where Fi represents the features extracted by the i-th feature extractor, Fisa represents the output of the spatial attention module, Conv1×1 represents the output of the 1×1 residual convolution operation, Fire represents the output of the residual block, and FiM represents the output of the dilated convolution. FiL represents the multi-scale local features extracted. Finally, through the multi-scale LFEM, we are able to capture a more comprehensive range of patterns and biological information from the gene expression data, thereby enhancing the feature extraction capability.

GFEM

The global features of mRNA phenotype refer to the broad, large-scale characteristics and patterns in mRNA that can influence the phenotype. Therefore, extracting global features helps capture these long-range dependencies. In the GFEM, we use the self-attention mechanism to capture global information from the gene expression data. Self-attention is a technique used to model relationships between arbitrary gene expression within transcription profiling, enabling it to capture global dependencies within the transcription profiling, rather than being limited to local relationships between adjacent gene expression. In particular, self-attention helps our model scan across the gene expression data and identify which gene types matter most for classification, improving overall performance. The proposed GFEM is illustrated in Figure 1D.

We use linear layers to map the gene expression features into three distinct components: the query Q, key K, and value V. Then, by transposing K and performing a matrix dot product with Q, we obtain the attention weight matrix for each gene. The Softmax function is applied to normalize the attention weight matrix into a probability distribution. The final attention weights determine the importance of each gene relative to other positions. The formula for generating the self-attention matrix is shown below:

Q,K,V=[WQ,WK,WV]·Fi[5]

SA=Softmax(Q∗KTdk)[6]

where, WQ, WK, and WV are the weights of the fully connected layers, and is the scaling factor used to ensure numerical stability and prevent the dot product values from becoming too large. After obtaining the attention weight matrix, it is multiplied by the value V to extract the final global features. The computation for the attention output, clearly indicating that the global features are given by FiG=SA⋅V, i.e., FiG=Softmax​(QK⊤/dk)V, thereby eliminating the previous omission and ensuring that the symbolic derivation is complete and the expression is clear. The extracted global features can effectively model interactions between any genes, capturing long-range dependencies over distances, thus enhancing the feature representation ability of the network.

Channel selection module (CSM)

After extracting the multi-scale local features and global features of the gene expression data, we concatenate the two features along the channel dimension to achieve fusion. Since local features and global features contribute differently to the classification task, in the channel importance selection module, we intentionally assign different weights to the local and global features, aiming to obtain fused features that are more favorable for classification. The schematic diagram of the channel importance selection module is shown in Figure 1E.

After concatenating the local and global features along the channel dimension, we embed attention vectors along the spatial axis. Unlike typical channel attention methods (17,18), we simultaneously apply global max pooling and global average pooling along the channel dimension to retain more information. These pooled features are then sent to the multilayer perceptron (MLP) to learn the importance relationships between the channels. Finally, we use the Sigmoid function to obtain two-channel importance weight vectors. The calculation process is shown below:

FiCat=[FiL;FiG][7]

WcA=Sigmoid(MLP(GAP(Ficat)))[8]

WcM=Sigmoid(MLP(GMP(Ficat)))[9]

where FiCat represents the concatenated local and global features along the channel dimension, GAP represents the global average pooling operation, and GMP represents the global max pooling operation. Finally, we add the two-channel importance weight vectors and multiply them with the concatenated features, resulting in the fusion features where important channels are enhanced, enabling the network to have higher discriminative ability.

Classifier

Current classification methods often only use the features from the final layer of the encoder for classification (19,20), without fully considering that the features from different layers of the encoder contain information with varying resolutions. Therefore, the classifier in this paper aims to comprehensively integrate these features for classification, which can enhance the network’s expressive ability and improve classification performance. The calculation process of the classifier is as follows:

x3=F^3+MP(conv1(F^4))[10]

xi=F^i+MP(conv1(xi+1)),i=1,2[11]

pred=Linear(GAP(x1))[12]

Here, the convolution operation with a kernel size of 1×1 conv1 is used for channel number transformation, MP represents max pooling, which is used for resolution size transformation to enable element-wise addition, and pred represents the final prediction output from the fully connected layer.

Statistical analysis

The statistical significance of survival data between the two groups was assessed using the two-sided log-rank test, as implemented in the survival package in R.

Results

The transcriptional landscape between the response and non-response LUAD patients

Clinical information, survival data, RNA-seq data and genomic alteration data for tumor tissue samples of LUAD were downloaded from TCGA and Xena databases. The clinical response to non-immunotherapeutic treatment was classified into four categories: complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). CR, PR, and SD were grouped as the response category, while PD was classified as the non-response category based on the disease control rate (DCR). The response group consisted of 264 patients with CR, 4 with PR, and 32 with SD, while the non-response group included 63 patients. The OS was 3,597 days. The response group was associated with a significantly longer mean OS of 4,142 days compared to 1,333 days for the non-response group (P<0.0001, Figure 2A).

Figure 2 Overview of the transcriptomic landscape of LUAD patients. (A) Survival curves for overall survival in response and non-response LUAD patients. (B) The volcano plot illustrates the DEGs between response and non-response LUAD patients. The red dots indicate upregulated genes, while the blue dots represent downregulated genes. Genes that do not exhibit statistically significant differences in expression levels are depicted as gray dots. (C) Heatmap depicting the expression of top 10 DEGs between response and non-response LUAD patients. (D) GO enrichment analysis results of up-regulated DEGs in response group. (E) GO enrichment analysis results of up-regulated DEGs in non-response group. (F) KEGG enrichment analysis results of up-regulated DEGs in non-response group. (G) TMB values between response and non-response in LUAD patients. DEGs, differentially expressed genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; LUAD, lung adenocarcinoma; TMB, tumor mutational burden.

A total of 1,306 DEGs were identified between the response group and the non-response group, including 814 upregulated and 492 downregulated genes (Figure 2B; supplementary table 1 available at https://cdn.amegroups.cn/static/public/tlcr-2025-1-1391-1.xlsx). The top 50 DEGs consisted of 17 upregulated and 33 downregulated genes, respectively (Figure 2C).

Subsequently, GO and KEGG enrichment analyses were conducted to summarize the potential functions of the 1,306 DEGs. The GO analysis confirmed that the DEGs upregulated in the response group were involved in various biological processes, including immune activation-related pathways (such as leukocyte migration, myeloid leukocyte activation, and the B cell receptor signaling pathway), transmembrane transport (specifically the regulation of monovalent ion transmembrane transport, sodium-ion transmembrane transport), G protein-coupled receptor (GPCR) signaling (including the adenylate cyclase-modulating GPCR signaling pathway, phospholipase C-activating GPCR signaling pathway, and adenylate cyclase-activating GPCR protein-coupled receptor signaling pathway) (Figure 2D). Immune activation can enhance antitumor immunity; for example, B cell receptor signaling plays a crucial role in promoting this immune response (21).

GO enrichment analysis of DEGs that were upregulated in the non-response group revealed significant enrichment in processes related to cell division. This includes mitotic nuclear division, nuclear division, and nuclear chromosome segregation (Figure 2E). Additionally, the KEGG analysis of the DEGs upregulated in the non-response group primarily indicated enrichment in the cell cycle pathway (Figure 2F). Cell cycle dysregulation can promote oncogenesis and malignant progression, potentially explaining the lack of response (22).

Tumor mutational burden (TMB) between the response and non-response LUAD patients

It is reported that TMB serves as a major predictive biomarker for immunotherapy (23). Therefore, we explored whether there are significant differences in TMB between response and non-response groups in non-immunotherapy treatments, including radiotherapy, chemotherapy, and surgical resection. TMB between the response group and the non-response group was not significantly different (Figure 2G).

LUADnet outperformed classic models

Using GFEM, LFEM, and CSM, 20 genes were selected as features to build the model. However, none of these genes were present in the list of DEGs. First, from a methodological standpoint, our deep learning model integrates both local and GFEMs. The global component is specifically designed to capture long-range dependencies and interactions across the entire feature space. Restricting inputs to a pre-filtered list of DEGs would truncate this global context and undermine the purpose of the architecture. Second, from a biological perspective, prognosis in LUAD is often driven by complex pathway perturbations (24), involving both core drivers (often highly expressed) and numerous co-regulators exhibiting subtle changes. These subtle signals, which are frequently filtered out by traditional cut-offs, can collectively provide valuable information for a deep learning model.

To ensure a robust evaluation despite the limited cohort size, the dataset was randomly partitioned into 10 folds while preserving the proportion of responder and non-responder classes in each fold through stratified sampling. The model was trained on nine folds and tested on the remaining fold, and this process was repeated 10 times. Firstly, as shown in Table 1, our proposed model outperforms all other state-of-the-art models in terms of both precision and recall, while achieving the highest overall F1-score of 0.9193 and accuracy of 0.9206. This validates our model’s effectiveness in handling the classification task.

Model refinement

Secondly, the ablation experiments suggest that the combination of LFEM, GFEM, and CSM enables improving the overall classification performance (Table 2). Meanwhile, excluding LFEM or GFEM shows reductions in accuracy and F1-score, highlighting the limitations of relying solely on local or global feature extraction. The exclusion of CSM also leads to a drop in performance (0.0158 in accuracy and 0.0184 in F1-score), suggesting that channel selection techniques can help refine the model.

Discussion

This study offers a comprehensive understanding of the transcriptional profiling of non-immunotherapy NSCLC patients, categorized into response and non-response groups. The response group demonstrated an activated immune microenvironment, whereas the non-response group exhibited an aberrant cell cycle. Our results suggest that the LUADnet model serves as a robust predictor of DCR in LUAD patients undergoing non-immunotherapy.

Studies have demonstrated that the immune microenvironment of lung cancer patients is closely linked to prognosis (25). In our study, we observed an enrichment of immune activation-related pathways in the response group, including leukocyte migration, myeloid leukocyte activation, and the B cell receptor signaling pathway. This suggests that an active immune state before treatment is associated with better clinical outcomes.

Aberrant cell cycle progression is one of the fundamental mechanisms leading to tumorigenesis and cancer development, making the regulatory factors of the cell cycle machinery viable targets for anti-cancer therapy (12). As a targeted therapeutic agent for LUAD, osimertinib inhibits cell cycle progression by promoting the phosphorylation of p53 and p21 while simultaneously reducing cyclin D1 expression, independent of EGFR activity (26). Apatinib induced cell cycle arrest at the G1 phase and inhibited the expression of Cyclin D1 and CDK4 in NSCLC (27). In our study, the DEGs that were upregulated in the non-response group were found to be enriched in processes related to the cell cycle and cell division. This suggests that effective non-immunotherapy may inhibit tumor growth by arresting the cell cycle. For patients who do not respond to treatment, drugs that inhibit the cell cycle may be potential candidates for therapy.

Multiple studies have demonstrated that TMB may serve as a surrogate marker for overall neoantigen load (28), making highly mutated tumors more susceptible to targeting by activated immune cells (29). It has been reported that TMB is associated with improved survival in cancer patients receiving ICIs across a wide variety of cancer types (30). In this study, TMB is not associated with the DCR, suggesting that there is no correlation with improved clinical outcomes.

The superior performance of our LUADnet model across all evaluation metrics highlights its effectiveness in learning discriminative patterns relevant to the classification task. The high precision and recall scores indicate its capacity to maintain a reliable balance between sensitivity and specificity. Moreover, the consistently strong F1-score and overall accuracy suggest that the model not only generalizes well but also maintains robustness across varying conditions. These results demonstrate the practical value of our approach in real-world classification scenarios, outperforming existing state-of-the-art methods in both accuracy and consistency. Furthermore, the integrated use of LFEM, GFEM, and CSM collectively enhances classification effectiveness. Ablation studies show that removing either LFEM or GFEM results in noticeable declines in both accuracy and F1-score, underscoring the limitations of relying solely on local or global feature extraction. Additionally, excluding the CSM also leads to a performance drop, suggesting that channel selection plays a critical role in refining feature representation and improving model precision

Despite its strong performance, LUADnet does have certain limitations that offer directions for future research. First, the current model operates under supervised learning and relies on annotated training data, which may be limited in quantity and subject to labeling bias. Future work could explore semi-supervised (31) or self-supervised learning (32,33) strategies to reduce reliance on labeled data and improve generalizability to new datasets. Second, while our model is designed to handle complex spatial and channel-wise features, it does not explicitly incorporate temporal or longitudinal data (34), which could be relevant in evolving biological contexts such as disease progression. Integrating temporal modeling components may further enhance the model’s predictive capacity. Third, receiver operating characteristic (ROC) curve analysis with area under the curve (AUC) values, which is crucial for comprehensively evaluating model discrimination, was not performed in the current study. Future work will address these limitations by validating the model in larger, multi-center cohorts and conducting thorough ROC analyses with confidence intervals to better establish the model’s predictive utility. In addition, although cross-validation offers an estimate of performance stability, we did not perform bootstrapping to generate confidence intervals or conduct calibration analysis. Future studies with larger cohorts should incorporate these metrics to more accurately assess model calibration and clinical utility.

Additionally, the interpretability of deep learning models remains a challenge. While the use of attention and channel selection offers some insight into important features, future extensions could incorporate explainable AI (XAI) techniques (35,36) to provide greater transparency and support clinical trust. Finally, while LUADnet has demonstrated superior performance on benchmark datasets, its deployment in clinical environments will require rigorous validation on multi-institutional, heterogeneous datasets to ensure robustness, fairness, and scalability in diverse real-world settings. Finally, we will incorporate multi-omic data to enhance the performance of LUADnet and investigate noninvasive biological data to broaden the model’s applicability.

Conclusions

This study explored the tumor microenvironment of responding and non-responding LUAD patients following non-immunotherapy and developed a robust deep learning model to stratify these patients, which will facilitate therapeutic strategies and alleviate economic burdens.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1391/rc

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1391/prf

Funding: This work was supported by Noncommunicable Chronic Diseases-National Science and Technology Major Project (Nos. 2023ZD0506101 and 2023ZD0506100); Sichuan Science and Technology Program (No. 2025JDDJ0005); the Fundamental Research Funds for the Central Universities (Nos. SCU2026D017 and 2026qzsyxf06); grant (No. RHM25101)from 1·3·5 Project of State Key Laboratory of Respiratory Health and Multimorbidity, West China Hospital, Sichuan University; State Key Laboratory of Respiratory Health and Multimorbidity, State Key Laboratory Special Fund (No. 2060204).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1391/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Miller KD, Fuchs HE, et al. Cancer statistics, 2022. CA Cancer J Clin 2022;72:7-33. [Crossref] [PubMed]
2. Molina JR, Yang P, Cassivi SD, et al. Non-small cell lung cancer: epidemiology, risk factors, treatment, and survivorship. Mayo Clin Proc 2008;83:584-94. [Crossref] [PubMed]
3. Wang M, Herbst RS, Boshoff C. Toward personalized treatment approaches for non-small-cell lung cancer. Nat Med 2021;27:1345-56. [Crossref] [PubMed]
4. Kim JM, Chen DS. Immune escape to PD-L1/PD-1 blockade: seven steps to success (or failure). Ann Oncol 2016;27:1492-504. [Crossref] [PubMed]
5. Haratani K, Hayashi H, Chiba Y, et al. Association of Immune-Related Adverse Events With Nivolumab Efficacy in Non-Small-Cell Lung Cancer. JAMA Oncol 2018;4:374-8. [Crossref] [PubMed]
6. Xia L, Liu Y, Wang Y. PD-1/PD-L1 Blockade Therapy in Advanced Non-Small-Cell Lung Cancer: Current Status and Future Directions. Oncologist 2019;24:S31-41. [Crossref] [PubMed]
7. Hellmann MD, Paz-Ares L, Bernabe Caro R, et al. Nivolumab plus Ipilimumab in Advanced Non-Small-Cell Lung Cancer. N Engl J Med 2019;381:2020-31. [Crossref] [PubMed]
8. Gao Q, Yang L, Lu M, et al. The artificial intelligence and machine learning in lung cancer immunotherapy. J Hematol Oncol 2023;16:55. [Crossref] [PubMed]
9. Quang D, Chen Y, Xie X. DANN: a deep learning approach for annotating the pathogenicity of genetic variants. Bioinformatics 2015;31:761-3. [Crossref] [PubMed]
10. Fiannaca A, La Paglia L, La Rosa M, et al. Deep learning models for bacteria taxonomic classification of metagenomic data. BMC Bioinformatics 2018;19:198. [Crossref] [PubMed]
11. Li J, Pu Y, Tang J, et al. DeepATT: a hybrid category attention neural network for identifying functional effects of DNA sequences. Brief Bioinform 2021;22:bbaa159. [Crossref] [PubMed]
12. Yang M, Wang Z, Yan Z, et al. DNASimCLR: a contrastive learning-based deep learning approach for gene sequence data classification. BMC Bioinformatics 2024;25:328. [Crossref] [PubMed]
13. Lilhore UK, Simiaya S, Alhussein M, et al. Optimizing protein sequence classification: integrating deep learning models with Bayesian optimization for enhanced biological analysis. BMC Med Inform Decis Mak 2024;24:236. [Crossref] [PubMed]
14. Song Z, Wang Y, Zhu M, et al. Exploring ribosome biogenesis in lung adenocarcinoma to advance prognostic methods and immunotherapy strategies. J Transl Med 2025;23:503. [Crossref] [PubMed]
15. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550. [Crossref] [PubMed]
16. Yu G, Wang LG, Han Y, et al. clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS 2012;16:284-7. [Crossref] [PubMed]
17. Wang Y, Guo S, Guo J, et al. Towards performance-maximizing neural network pruning via global channel attention. Neural Netw 2024;171:104-13. [Crossref] [PubMed]
18. Huang H, Chen Z, Zou Y, et al. Channel prior convolutional attention for medical image segmentation. Comput Biol Med 2024;178:108784. [Crossref] [PubMed]
19. Vivek S, Faul J, Thyagarajan B, et al. Explainable variational autoencoder (E-VAE) model using genome-wide SNPs to predict dementia. J Biomed Inform 2023;148:104536. [Crossref] [PubMed]
20. Schlusser N, González A, Pandey M, et al. Current limitations in predicting mRNA translation with deep learning models. Genome Biol 2024;25:227. [Crossref] [PubMed]
21. Zhang Y, Yu Y, Gu X, et al. Exosomes derived from colorectal cancer cells suppress B-cell mediated anti-tumor immunity. Int Immunopharmacol 2025;148:114176. [Crossref] [PubMed]
22. Jiang P, Yu J, Zheng Y, et al. Prognostic significance and immune infiltration analysis of HMGA2 in endometrial cancer. Front Immunol 2025;16:1559278. [Crossref] [PubMed]
23. Shi R, Sun J, Zhou Z, et al. Integration of multiple machine learning approaches develops a gene mutation-based classifier for accurate immunotherapy outcomes. NPJ Precis Oncol 2025;9:54. [Crossref] [PubMed]
24. Li Z, Qiao W, Yu S, et al. Integrating computational pathology and multi-transcriptomics to characterize lung adenocarcinoma heterogeneity and prognostic modeling. Int J Surg 2025;111:5162-81. [Crossref] [PubMed]
25. Yan Y, Sun D, Hu J, et al. Multi-omic profiling highlights factors associated with resistance to immuno-chemotherapy in non-small-cell lung cancer. Nat Genet 2025;57:126-39. [Crossref] [PubMed]
26. Nanamiya R, Saito-Koyama R, Miki Y, et al. EphB4 as a Novel Target for the EGFR-Independent Suppressive Effects of Osimertinib on Cell Cycle Progression in Non-Small Cell Lung Cancer. Int J Mol Sci 2021;22:8522. [Crossref] [PubMed]
27. Xie C, Zhou X, Liang C, et al. Apatinib triggers autophagic and apoptotic cell death via VEGFR2/STAT3/PD-L1 and ROS/Nrf2/p62 signaling in lung cancer. J Exp Clin Cancer Res 2021;40:266. [Crossref] [PubMed]
28. Rizvi NA, Hellmann MD, Snyder A, et al. Cancer immunology. Mutational landscape determines sensitivity to PD-1 blockade in non-small cell lung cancer. Science 2015;348:124-8.
29. Chalmers ZR, Connelly CF, Fabrizio D, et al. Analysis of 100,000 human cancer genomes reveals the landscape of tumor mutational burden. Genome Med 2017;9:34. [Crossref] [PubMed]
30. Samstein RM, Lee CH, Shoushtari AN, et al. Tumor mutational load predicts survival after immunotherapy across multiple cancer types. Nat Genet 2019;51:202-6. [Crossref] [PubMed]
31. Wang J, Liao N, Du X, et al. A semi-supervised approach for the integration of multi-omics data based on transformer multi-head self-attention mechanism and graph convolutional networks. BMC Genomics 2024;25:86. [Crossref] [PubMed]
32. Richter T, Bahrami M, Xia Y, et al. Delineating the effective use of self-supervised learning in single-cell genomics. Nat Mach Intell 2025;7:68-78.
33. Wei Q, Islam MT, Zhou Y, et al. Self-supervised deep learning of gene-gene interactions for improved gene expression recovery. Brief Bioinform 2024;25:bbae031. [Crossref] [PubMed]
34. Ranek JS, Stanley N, Purvis JE. Integrating temporal single-cell gene expression modalities for trajectory inference and disease prediction. Genome Biol 2022;23:186. [Crossref] [PubMed]
35. Ali S, Abuhmed T, El-Sappagh S, et al. Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence. Information Fusion 2023;99:101805.
36. Schwalbe G, Finzel B. A comprehensive taxonomy for explainable artificial intelligence: a systematic survey of surveys on methods and concepts. Data Min Knowl Disc 2024;38:3043-101.