We evaluated our model using four publicly available single-cell RNA sequencing (scRNA-seq) datasets, each representing a distinct drug treatment and cancer context:

For each dataset, scRNA-seq was conducted on cancer cells before the administration of drug treatment, enabling the capture of pre-treatment transcriptomic states. After drug treatment, each cell was assigned a binary response label (0 = resistant, 1 = sensitive) based on post-treatment viability assays. These labels, derived from experimental annotations in the original publications, serve as the ground truth for model training and evaluation.

To address class imbalance in the labeled datasets, we applied different sampling strategies: SMOTE was used for DATA1, and oversampling was used for DATA2, DATA3, and DATA4.

These datasets, with well-defined pre-treatment input features and post-treatment binary response labels, support single-cell-level supervised learning and enable reliable evaluation using metrics such as AUC, accuracy, and F1 score. The original drug response values in the GDSC and CCLE databases were reported as continuous variables such as IC50 or percent viability. For this study, we adopted binary drug response labels (0 = resistant, 1 = sensitive) derived from existing annotations in the datasets or binarized based on established thresholds as defined in prior studies (e.g., top/bottom quantiles of response distributions). These binary labels serve as ground truth for model training and evaluation.

In our study design, scRNA-seq data is collected before applying any drug treatment. This allows us to capture the baseline transcriptional state of each cell. The same cell population is then subjected to drug treatment, after which cellular response is assessed. To determine which cells responded by dying, we rely on dataset-provided viability labels, which were derived by the original authors using post-treatment assays. Although we do not directly observe or trace individual cells over time, the annotations provided link each cell's pre-treatment transcriptome to its post-treatment viability outcome.

Comprehensive datasets from the GDSC and CCLE databases were used in this study. These datasets included drug response profiles for 1,280 cancer cell lines, 1,557 drugs, and 15,962 genes. Drug-sensitive cell lines were annotated with a binary label of 1, whereas drug-resistant cell lines were labeled as 0. Predictive model performance may be negatively impacted by the notable class imbalance in the training dataset caused by the varying distribution of sensitive and resistant cell lines across drug treatments.

Oversampling and the Synthetic Minority Oversampling Technique (SMOTE) (29) were two sampling strategies used to optimize the training dataset structure to lessen the problem of class imbalance. SMOTE used a more complex technique by creating synthetic samples inside the minority class's feature space, increasing the diversity and representativeness of the training data, whereas oversampling duplicated samples from the minority class to increase their representation.

The bulk RNA-Seq datasets used for pre-training were split into 80% training and 20% test sets. Similarly, the single-cell RNA-Seq datasets used for fine-tuning and testing were also divided with an 80/20 ratio. This consistent splitting ensures that both stages (pre-training and fine-tuning/testing) are clearly separated and reproducible.

To address class imbalance in the datasets, different oversampling techniques were applied. Specifically, SMOTE was used for DATA1, while standard oversampling was used for DATA2, DATA3, and DATA4. The decision to use SMOTE for DATA1 was based on the significant class imbalance in this dataset, where a large proportion of drug-resistant cells were underrepresented. SMOTE was chosen to generate synthetic samples for the minority class, allowing the model to learn more effectively from the drug-resistant cells. For the other datasets (DATA2–4), the class imbalance was less pronounced, so standard oversampling (duplicating minority class samples) was deemed sufficient.

While SMOTE helps to balance the dataset by generating synthetic samples, it can also introduce some risk of overfitting, especially if the synthetic samples do not accurately represent the minority class. To ensure that our model's performance improvements are robust, we plan to evaluate alternative balancing strategies (such as undersampling or different oversampling techniques) in future work, and will report these results for comparison.

To justify the selection of different sampling strategies across datasets, we conducted a comparative experiment to evaluate the impact of SMOTE and simple oversampling on model performance. SMOTE yielded significantly better results for DATA1, which had a severe class imbalance, while simple oversampling consistently outperformed SMOTE for DATA2, DATA3, and DATA4, where the imbalance was less pronounced. These results demonstrate that our tailored preprocessing design is robust and appropriate for the specific class distributions in each dataset. This evidence supports the rationale for adopting SMOTE for severely imbalanced data and standard oversampling for datasets with moderate imbalance. We believe this careful selection improves model generalizability and minimizes the risk of overfitting associated with inappropriate sampling techniques.

Additionally, the following preparation procedures were used to improve the dependability and quality of the single-cell RNA-seq data: (i) To remove noise and non-informative features, cells with <200 recognized genes—a sign of empty droplets—and genes discovered in fewer than three cells were eliminated. These steps are particularly important in single-cell RNA-seq data, as they help to eliminate low-quality cells that could negatively impact downstream analysis. (ii) To prevent confounding effects, cells with mitochondrial gene expression levels higher than 10% of total gene expression were eliminated. This step is commonly used in single-cell data to avoid the inclusion of stressed or dying cells, which may distort the overall analysis. (iii) To make subsequent studies easier, the count matrix was normalized using the total UMI counts per cell and then logarithmically transformed. Normalization is especially crucial in single-cell RNA-seq data to ensure that cell-to-cell comparisons are valid and not influenced by technical variations. All of these preprocessing steps were performed on scRNA-seq data collected prior to drug treatment, ensuring that the model learns from the true pretreatment transcriptional states.

In this study, we use both bulk RNA-seq data and single-cell RNA-seq data for drug response prediction. For the bulk RNA-seq data used in this study, we use data derived from established drug response databases such as GDSC and CCLE. These databases provide drug response labels for cancer cell lines, denoted as Y_g, where each label indicates whether the cell line is drug-sensitive or drug-resistant based on post-treatment assays. For the single-cell RNA-seq data, we use experimental drug response annotations obtained through assays such as Annexin V staining. These provide binary labels for individual cells, denoted as Y_s, representing whether each cell is drug-sensitive (1) or drug-resistant (0) after treatment.

Throughout the manuscript, we distinguish between bulk-level drug response labels (Y_g) and single-cell-level drug response labels (Y_s) to avoid any ambiguity. The main innovation of this study is the creation of the ATSDP-NET model, even though these sampling and preprocessing methods enhanced the dataset's quality and balance. In contrast to conventional techniques, ATSDP-NET incorporates a deep transfer learning framework with a multi-head attention mechanism to precisely extract features specific to single-cell data, allowing for extremely precise drug response predictions. The model's ability to predict drug sensitivity and resistance with more accuracy and interpretability is greatly increased by this focused approach, which successfully tackles the difficulties associated with single-cell feature extraction.

Figure 1 summarizes the main pipeline of ATSDP-NET, which consists of three stages:

The ATSDP-NET framework is structured into three distinct stages: (1) feature extraction from bulk RNA-seq data and drug response prediction, (2) feature extraction from single-cell RNA-seq data and deep transfer learning (DTL) model training, and (3) model transfer and single-cell drug response prediction.

The experimental environment was configured as follows: an 11th Gen Intel^® Core™ i7-11700HX CPU (2.50 GHz), NVIDIA GeForce GTX 1050 Ti GPU, 24 GB RAM, Windows 10 operating system, and PyTorch version 2.2.2.

The ATSDP-NET model was comprehensively evaluated on four publicly available single-cell RNA sequencing (scRNA-seq) datasets, each representing distinct drug treatments. The selected drugs include Cisplatin, I-BET-762, and Docetaxel, which are widely used in cancer therapy. Specifically, DATA1 and DATA2 comprise human oral squamous cell carcinoma (OSCC) cells treated with Cisplatin, DATA3 includes human prostate cancer cells treated with Docetaxel, and DATA4 involves murine acute myeloid leukemia (AML) cells treated with I-BET-762.

To address the class imbalance issue during training, SMOTE sampling was employed for the DATA1 dataset, while oversampling techniques were applied to the DATA2, DATA3, and DATA4 datasets. Each dataset contains binary drug response annotations for individual cells, where the labels were derived from the original data as binary indicators (0 for resistant cells and 1 for sensitive cells).

The hyperparameters of the ATSDP-NET model used in the experiments are summarized in Table 1, which outlines the configurations for each of the four datasets.

To assess the performance of the ATSDP-NET model across different datasets, the model was tested on four datasets: DATA1, DATA2, DATA3, and DATA4. The evaluation aimed to measure the effectiveness of ATSDP-NET in predicting drug responses in various single-cell RNA sequencing (scRNA-seq) datasets under distinct experimental conditions.

Figure 2a consists of two subplots. The radar chart on the left illustrates the performance of the ATSDP-NET model across four datasets (DATA1, DATA2, DATA3, and DATA4) based on five key metrics: Precision, Recall, F1-Score, AUC, and AP. Each dataset is represented by a distinct color, allowing for an intuitive comparison of the model's performance across different datasets. The bar plot on the right presents the specific numerical values for the five performance metrics across each dataset. This visualization provides a clearer understanding of the detailed performance of the model on Precision, Recall, F1-Score, AUC, and AP, facilitating cross-dataset comparisons.

Figure 2b also consists of two subplots and is designed to showcase the classification performance of the model across the four datasets. The ROC curve on the left displays the Receiver Operating Characteristic (ROC) curves for each dataset along with their corresponding AUC values (Area Under the Curve). Higher AUC values indicate better classification performance. Notably, the AUC for DATA3 is the highest at 0.9068, demonstrating the best performance among the datasets. Meanwhile, DATA1, DATA2, and DATA4 also exhibit strong AUC values of 0.8610, 0.8554, and 0.8845, respectively.

The PR curve on the right visualizes the Precision-Recall (PR) curves and their corresponding areas for each dataset. The PR curve reflects the model's precision at varying levels of recall, making it particularly suitable for evaluating imbalanced datasets. It is evident that the PR curve area for DATA2 is the highest, reaching 0.9431, indicating superior performance on this dataset.

To evaluate the robustness of our method, we performed 10-fold stratified cross-validation on each dataset. To account for possible randomness due to data shuffling and model initialization, we repeated the entire training and evaluation process independently 10 times using different random seeds. The mean and standard deviation of performance metrics, including precision, recall, F1-score, AUC, and average precision, are reported in Table 2 to reflect the consistency and reliability of our results.

Table 3 summarizes the five key performance metrics of the ATSDP-NET model across four datasets, including Precision, Recall, F1-score, AUC, and AP, providing a comprehensive evaluation of the model's classification performance across different tasks. Overall, the ATSDP-NET model demonstrates high accuracy and stability across all metrics, highlighting its robustness and adaptability.

The results indicate that the model achieves a well-balanced performance between Precision and Recall, suggesting its ability to maintain predictive accuracy while effectively identifying positive samples. The F1-score further underscores the model's ability to harmonize precision and recall, showcasing superior classification performance on certain datasets. The AUC metric reflects the model's strong discriminatory power in distinguishing between positive and negative samples, while the AP metric validates the high-quality confidence scores of the predictions.

In conclusion, the results presented in Table 3 demonstrate that the ATSDP-NET model exhibits stable and outstanding performance across various datasets. It effectively addresses diverse data distributions and feature patterns, demonstrating strong generalization capabilities and classification performance.

In this section, the impact of varying attention head counts in the multi-head attention mechanism on the performance of the ATSDP-NET model is analyzed. Experiments were conducted on four datasets (DATA1, DATA2, DATA3, and DATA4), with the number of attention heads ranging from 0 to 32 within the ATSDP-NET framework. To comprehensively assess this influence, key performance metrics, including Precision, Recall, F1-score, AUC, and Average Precision (AP), were recorded under different attention head configurations. The trends in these performance metrics as a function of the number of attention heads are visualized in Figure 3, providing insights into the optimal configuration of the multi-head attention mechanism.

Figure 3 illustrates the impact of varying attention head counts on the performance of the ATSDP-NET model across four datasets. For DATA1 and DATA2, the model achieved optimal performance when the number of attention heads was set to 8. Under these conditions, the Precision, Recall, F1-score, AUC, and AP for DATA1 reached 0.8527, 0.7865, 0.7918, 0.8609, and 0.9239, respectively. Similarly, for DATA2, these metrics were 0.8448, 0.7711, 0.7816, 0.8553, and 0.9289, respectively. At this configuration, the ATSDP-NET model exhibited superior performance on both the PR and ROC curves. For DATA3, the optimal performance was observed when the number of attention heads was set to 4, with the highest Precision, Recall, F1-score, AUC, and AP recorded as 0.8332, 0.8086, 0.8051, 0.9067, and 0.8958, respectively. In DATA4, the model performed best with 8 attention heads, achieving Precision, Recall, F1-score, AUC, and AP values of 0.8711, 0.8703, 0.8704, 0.8845, and 0.8475, respectively. At these configurations, the model also achieved its peak performance on both the PR and ROC curves.

These experimental results highlight that 4 attention heads yielded the best performance for DATA3, while 8 attention heads significantly enhanced the accuracy and performance of the ATSDP-NET model for DATA1, DATA2, and DATA4. The findings underscore the critical importance of appropriately configuring the number of attention heads in multi-head attention mechanisms. A well-chosen head count can effectively improve the model's ability to capture and process information, providing valuable insights into the application of multi-head attention mechanisms for addressing complex data challenges.

In this study, we employed several deep learning methods, including Sparse Autoencoders (SA), Graph Attention Networks (GAT), Residual Networks (ResNet18), Transformers, Graph Convolutional Networks (GCN), and Convolutional Neural Networks (CNN), as well as two state-of-the-art drug response prediction models, DeepDR and DrugCell. These models were adapted to single-cell RNA-seq data to serve as baseline comparisons for our proposed ATSDP-NET. The performance of all models was evaluated using key metrics, including Precision, Recall, F1-score, Area Under the Curve (AUC), and Average Precision (AP).

The SA model is traditionally an unsupervised method for dimensionality reduction. To adapt it for drug response prediction, we used the autoencoder to extract sparse representations of the gene expression data, which were then passed through a classification layer. While SA is not inherently designed for classification tasks, its effectiveness for biological feature extraction has been demonstrated, for example, in Lei et al. (30). Including SA as a baseline highlights the advantage of supervised learning in our ATSDP-NET.

The ResNet18 and CNN models were originally designed for image data but were adapted here by reshaping the gene expression matrix into 1D vectors. This allowed us to leverage their feature extraction capabilities for single-cell gene expression profiles. We added fully connected layers for the final drug response prediction. Their potential for this task is supported by previous work, such as Chen and Jiang (31). Although these architectures are not optimized for RNA-seq data, they provide meaningful baselines for assessing whether ATSDP-NET, which is tailored for single-cell data, can outperform general-purpose models.

The Transformer model was employed without positional encoding to simplify the model and test the effectiveness of its self-attention mechanism on unordered gene expression data. Despite the absence of positional encoding, the Transformer could capture global dependencies among genes, as shown in Jiang et al. (32). This setup allowed us to examine whether a streamlined Transformer remains competitive for drug response prediction on single-cell data.

For GCN and GAT, we constructed graphs where each node represents a cell and edges reflect similarity between cells based on their gene expression profiles. An adjacency matrix was generated by applying a similarity threshold (e.g., 0.8) to determine connections. This graph-based representation enables the models to capture local relationships and community structures, enhancing predictive power. The effectiveness of this approach is supported by Yang et al. (33).

DeepDR and DrugCell are state-of-the-art drug response prediction models designed for bulk RNA-seq or multi-omics data. We adapted these models to our single-cell RNA-seq datasets for a comprehensive comparison. All baseline models were fine-tuned and evaluated using the same training, validation, and testing splits as ATSDP-NET to ensure a fair and consistent comparison.

To ensure reproducibility, each model was trained under the same conditions, with consistent hyperparameter tuning and evaluation protocols. The detailed implementation settings and parameter configurations are provided in the Supplementary material.

By adapting and fine-tuning these baseline models, we aimed to demonstrate the advantages of our proposed ATSDP-NET, which is specifically designed for single-cell drug response prediction.

The choice of deep learning methods over traditional baseline approaches, such as PCA+SVM, is motivated by the significant differences in data structure and processing requirements between single-cell RNA-seq data and bulk cell data (e.g., bulk RNA-seq data). Single-cell RNA-seq data typically exhibits high dimensionality and sparsity, making it difficult for traditional methods to effectively handle these characteristics, especially when integrating single-cell and bulk cell data. Deep learning methods, however, can automatically learn feature representations and effectively capture complex non-linear relationships within the data, making them more suitable for such complex datasets. Unlike traditional dimensionality reduction techniques like PCA, deep learning models do not rely on manual feature selection but instead use hierarchical learning mechanisms to automatically extract and integrate important features, thereby significantly improving the accuracy of drug response predictions.

Additionally, deep learning methods possess stronger generalization capabilities, enabling them to handle higher-dimensional and more complex patterns. In single-cell drug response prediction tasks, deep learning effectively integrates various types of data and, through multi-layered neural network learning, captures subtle cellular differences and potential non-linear patterns, further enhancing prediction accuracy. Therefore, the selection of deep learning methods is driven by the need to better handle the integration of single-cell and bulk cell data, while maximizing the accuracy of drug response predictions.

Figure 4a (DATA1) illustrates the differences in five key performance metrics across various models for the DATA1 dataset. The radar chart demonstrates that ATSDP-NET excels in Precision, Recall, AUC, and AP, highlighting its exceptional classification capability for the DATA1 dataset. In contrast, other models, such as SA and ResNet18, exhibit relatively poor performance, particularly in Recall and Precision. Models like DrugCell and DeepDR also show inferior performance compared to ATSDP-NET.

Figure 4b (DATA2) presents the evaluation results for the DATA2 dataset. ATSDP-NET continues to outperform other models in most metrics, with particularly notable advantages in Precision and F1-score. Although models such as Transformer and CNN show competitive performance in some individual metrics, ATSDP-NET consistently outperforms them across all key metrics in the DATA2 dataset, demonstrating its robustness and effectiveness in handling different datasets.

Figure 4c (DATA3) compares the performance of various models on the DATA3 dataset. While the performance differences between models are relatively small, ATSDP-NET still demonstrates a slight advantage in key metrics such as Precision, Recall, and F1-score, further validating its robust classification ability even when the performance differences across models are minimal.

Figure 4d (DATA4) presents the model performance on the DATA4 dataset. ATSDP-NET achieves near-optimal results across all five performance metrics, with particular superiority in Precision and AUC. In contrast, other models, such as GCN and CNN, show significantly lower performance in all metrics, while ResNet18 also lags behind ATSDP-NET.

Table 4 provides a clear comparison of the comprehensive superiority of ATSDP-NET across the five key metrics: Precision, Recall, F1-score, AUC, and AP. Notably, ATSDP-NET shows significant advantages across all metrics in both the DATA1 and DATA4 datasets, emphasizing its adaptability to drug response prediction tasks. This performance superiority is attributed to the multi-head attention mechanism embedded within the model, which effectively captures both global and local features in the data, thereby improving feature extraction and classification accuracy.

In contrast, other models exhibit certain limitations. For instance, while ResNet18 and Transformer show competitive performance in some metrics, they fall short in Recall and F1-score, reflecting their inability to capture the deep features inherent in single-cell data. Traditional deep learning models, such as GCN and CNN, lack specific design optimizations for drug response prediction tasks and therefore fail to deliver satisfactory performance across all datasets.

In conclusion, compared to the eight other models (SA, GAT, ResNet18, Transformer, GCN, CNN, DrugCell, and DeepDR), ATSDP-NET consistently demonstrates superior performance across all four datasets (DATA1, DATA2, DATA3, DATA4). These results underscore the critical role of the multi-head attention mechanism in enhancing the model's robustness and adaptability, enabling ATSDP-NET to more effectively capture significant features within the data, thereby improving classification performance.

In the study of DATA4 (GSE110894), ATSDP-NET was employed to predict the drug response efficacy of a BET inhibitor (I-BET) in 1,419 leukemia cells characterized by AF9 gene fusions (Figure 5a). The experimental design consisted of five distinct conditions: one control group without any specific treatment (101 CELL CONTROL), two drug sensitivity states (DMSO and I-BET 400 NMOL), and two drug resistance states (I-BET resistance and I-BET withdrawal). Notably, the I-BET withdrawal group refers to cells in which drug treatment was discontinued at a specific time point to evaluate cellular responses or state changes post-withdrawal. Comparative analysis with the original study data demonstrated that ATSDP-NET achieved highly consistent results in predicting drug responses.

ATSDP-NET provided two approaches for drug response prediction: a continuous probability score and binary sensitivity/resistance labels. The continuous probability score reflects the likelihood of a cell's sensitivity to the drug, with higher scores indicating greater sensitivity. The binary labels are derived by applying a threshold to the continuous probability scores: cells with scores ranging from 0 to 0.5 are categorized as resistant, whereas those with scores from 0.5 to 1 are classified as sensitive.

In addition, gene scores were utilized to reflect the overall expression levels of differentially expressed genes (DEGs) within clusters of sensitive or resistant cells. The underlying assumption of this gene scoring approach is that accurate predictions should assign the correct drug response labels to individual cells. Consequently, the DEG scores for resistant and sensitive states should exhibit consistent correlations with DEGs derived from the ground truth data.

The analysis revealed that the generated DEG scores outperformed the DEG scores derived from ground truth labels in distinguishing resistant and sensitive cells (Figure 5b). Specifically, in the resistant DEG list, the correlation between the predicted and true DEG scores reached R² = 0.788, while in the sensitive DEG list, the correlation was R² = 0.888 (Figure 5c).

To assess the statistical significance of these correlations, 1,000 random tests were performed by selecting random genes in equal numbers to the predicted DEG list, and the correlations were recalculated. The results demonstrated that, across 1,000 random trials, the p-values for the correlations of resistant and sensitive DEG scores were both < 0.001, indicating that the observed correlations were statistically significant and robust (Figure 5d).

Although ATSDP-NET demonstrates high accuracy in predicting single-cell drug responses, its greater value lies in its ability to uncover the key molecular mechanisms underlying drug resistance (34). In this study, ATSDP-NET was employed to analyze single-cell data from oral squamous cell carcinoma (OSCC) cells treated with cisplatin, aiming to identify critical gene markers associated with cisplatin resistance (35) and assess their potential clinical applications.

Cisplatin, a widely used chemotherapeutic agent, primarily induces cancer cell apoptosis by causing DNA damage (36). However, cancer cells often develop resistance to cisplatin by activating DNA repair pathways or inhibiting apoptosis (37). Through high-resolution single-cell analysis enabled by ATSDP-NET, several key genes (KGs) associated with cisplatin resistance were identified. Many of these genes were significantly enriched in resistant cell populations, suggesting their potential role as mediators of cisplatin resistance.

In this study, ATSDP-NET was applied to analyze the gene expression profiles of cisplatin-sensitive and resistant cells, uncovering critical molecular mechanisms associated with drug resistance and sensitivity. Figure 6a illustrates the classification and prediction results across four distinct cell sources. The ATSDP-NET model effectively distinguished resistant and sensitive cell populations, with predictions highly consistent with ground truth labels, demonstrating robust classification performance and precise identification of drug responses. Furthermore, the predicted continuous response distribution revealed a gradient from resistance to sensitivity, highlighting the heterogeneous nature of drug sensitivity among cancer cells.

Figure 6b illustrates the differentially expressed gene (DEG) scores for sensitive and resistant cells. The high DEG scores observed in sensitive cells suggest that these genes may play critical roles in cisplatin-induced apoptosis. Conversely, the distribution of DEG scores in resistant cells indicates that these genes may be involved in key resistance mechanisms, such as enhanced DNA repair pathways (38) and apoptosis inhibition (39), which enable cells to counteract cisplatin toxicity effectively.

Figure 6c presents a heatmap depicting the expression patterns of key resistance and sensitivity genes, categorized into “resistance-associated key genes (40)” and “sensitivity-associated key genes (41).” In cisplatin-resistant cells (HN120PCR), resistance-associated key genes exhibit elevated expression, while cisplatin-sensitive cells (HN120P) predominantly display higher expression of sensitivity-associated key genes. This contrasting gene expression pattern suggests that resistant cells may upregulate resistance-associated key genes to bolster their ability to withstand cisplatin treatment. In contrast, sensitive cells appear to activate sensitivity-associated key genes to enhance their responsiveness to cisplatin.

Figure 6d provides the names and distribution of specific resistance- and sensitivity-associated key genes. For instance, resistance-related genes such as ZBTB48 (42), NEK6 (43), and FBXO25 (44) show elevated expression in resistant cell populations, potentially contributing to enhanced DNA repair and anti-apoptotic pathways, thereby strengthening cellular resistance to cisplatin. In contrast, sensitivity-related genes such as UCHL1 (45), GNGT1 (46), and EDIL3 (47) exhibit significant expression in sensitive cells, potentially promoting apoptosis and increasing cellular sensitivity to cisplatin.

In summary, the ATSDP-NET model successfully identified key gene groups associated with cisplatin sensitivity and resistance, revealing distinct gene expression patterns between resistant and sensitive cells. Resistant cells appear to upregulate resistance-associated key genes to enhance DNA repair and survival pathways, while sensitive cells activate sensitivity-associated key genes to expedite apoptosis triggered by cisplatin. These findings provide novel insights into the mechanisms underlying cisplatin resistance and sensitivity, offering a foundation for developing personalized therapeutic strategies aimed at targeting resistance-associated genes or activating sensitivity-related genes.

To validate the relationship between the predicted drug responses and the progression of drug treatment, trajectory inference was performed on DATA1 (treated with cisplatin). Pseudotime trajectories were inferred using the Diffusion Pseudotime (DPT) method implemented in Scanpy (48) in Python. Cells were pre-processed with PCA and UMAP for dimensionality reduction, a k-nearest neighbor graph was constructed, and diffusion maps were computed to estimate the pseudotime ordering. Pseudotime analysis revealed a trajectory where cells transitioned progressively from a sensitive state to a resistant state (Figure 7a). When comparing pseudotime results with the predicted drug response probability scores mapped onto the same diffusion UMAP, a clear trend was observed: as pseudotime progressed, the probability of resistance steadily increased, while sensitivity probability decreased correspondingly (Figure 7a).

These findings suggest that cells surviving high-dose drug treatment exhibited a marked drug-resistant phenotype, which aligns closely with the experimental (ground truth) labels of drug response states. This strong correlation highlights the predictive accuracy of ATSDP-NET in capturing the dynamic progression of drug responses at the single-cell level.

To further elucidate the molecular mechanisms underlying the development of drug resistance, the expression dynamics of key genes (KGs) identified by ATSDP-NET were analyzed along the pseudotime trajectory. Figure 7b highlights the expression patterns of two representative resistance-associated genes (ZBTB48 and NEK6) and two sensitivity-associated genes (UCHL1 and GNGT1). The expression levels of these genes align with the pseudotime trajectory trends and are consistent with the predicted drug response probability scores. Resistance-associated genes exhibit elevated expression during the later stages of pseudotime progression, whereas sensitivity-associated genes show higher expression during the early stages. This suggests that these genes play distinct regulatory roles in resistant and sensitive cellular states. It should be noted that pseudotime is an inferred ordering based on transcriptional similarity and does not incorporate actual treatment time or drug exposure information. Therefore, the observed increase in resistance probability may represent an association rather than a causal relationship.

Further analysis revealed that the predicted list of key genes (KGs) demonstrated significantly greater expression differences in distinguishing sensitive and resistant cell states compared to the differentially expressed genes (DEGs) (Figure 7c). This underscores the ability of ATSDP-NET to accurately identify molecular markers that are critical for characterizing drug response states and provides valuable insights into the regulatory mechanisms driving drug sensitivity and resistance.

The experimental results demonstrate that ATSDP-NET not only effectively predicts drug response probabilities but also uncovers the dynamic changes in cellular states through pseudotime analysis, providing a robust molecular basis for understanding the mechanisms underlying drug resistance. Furthermore, within the resistance-associated DEG list, the correlation between predicted DEG scores and actual DEG scores reached an R² of 0.652, while the correlation for the sensitivity-associated DEG list achieved an R² of 0.881 (Figure 8). These findings underscore the model's predictive capabilities, particularly its superior performance in identifying drug-sensitive cells. The significant correlation between sensitivity and resistance scores highlights the potential of ATSDP-NET as a reliable tool for drug response prediction in biomedical research.