INTERNATIONAL CONGRESS OF
PharmacoGenetics
Deep Learning for Single-Cell Genomics: Trends, Challenges, and Opportunities
-
Asal Naghipour-Kordlar,1 Maryam Radmanfard,2,*
1. Faculty of Nursing, Tabriz University of Medical Sciences, Tabriz, Iran
2. Department of Basic Sciences, Ta.C., Islamic Azad University, Tabriz, Iran
- Introduction: Recent progress in single-cell genomics has transformed our ability to interrogate complex biological systems by capturing transcriptomic, epigenomic, and spatial information at single-cell resolution (Ge et al., 2025). These technologies provide unprecedented opportunities to uncover cellular heterogeneity, developmental trajectories, and microenvironmental organization that remain obscured in bulk profiling (Ma & Xu, 2022). Nevertheless, the rapid expansion of single-cell datasets brings substantial analytical challenges. High dimensionality, sparsity, technical variability, and batch effects often limit the performance of conventional computational methods (Ding et al., 2024). To address these issues, deep learning (DL) has emerged as a powerful framework, offering advanced capabilities in feature extraction, representation learning, and multimodal data integration (Ge et al., 2025; Ma & Xu, 2022). Despite these advantages, the application of DL in single-cell research is still in its early stages. Critical concerns regarding model interpretability, reproducibility, and the establishment of robust benchmarking standards remain active areas of discussion (Ding et al., 2024; Ma & Xu, 2022). Tackling these challenges will be essential to fully exploit DL’s potential for advancing our understanding of cellular diversity, disease mechanisms, and the discovery of therapeutic targets (Ding et al., 2024; Ge et al., 2025; Ma & Xu, 2022).
- Methods: In this review, we searched recent studies on deep learning in single-cell genomics using PubMed, Google Scholar, and journals such as Nature. Keywords included deep learning, single-cell genomics. We focused on papers from recent years and included both review articles and original studies that applied DL in this field. The main findings were then summarized into key themes of progress, challenges, and future directions.
- Results: 1. Methodological Advances in Deep Learning for Single-Cell Genomics Recent studies have introduced several deep learning (DL) models tailored for single-cell genomics. Notably, DANCE (Deep generative modeling and clustering of single-cell genomic data) offers a unified framework for simultaneously modeling gene expression and cell clustering, enhancing the understanding of cellular heterogeneity. DANCE supports multiple modules and tasks, facilitating the evaluation of computational methods across various single-cell analysis tasks (Ding et al., 2024). Additionally, the application of multi-view subspace clustering, as demonstrated by scDMSC, addresses the high dimensionality and sparsity inherent in single-cell multi-omics data, providing more accurate clustering results. These advancements contribute to a more nuanced understanding of cellular diversity and function. 2. Integration of Spatial Transcriptomics and Epigenomics The integration of spatial transcriptomics with DL models has enabled the analysis of gene expression in the context of tissue architecture. Studies have highlighted the potential of DL in interpreting spatially resolved transcriptomic data, facilitating insights into tissue organization and function. Furthermore, combining DL with explainable artificial intelligence (XAI) tools has uncovered novel molecular insights into aging processes by analyzing single-cell and epigenetic data. This approach enhances the interpretability of complex models, making biological insights more accessible (Li et al., 2025). 3. Challenges and Limitations Despite the advancements, several challenges persist in the application of DL to single-cell genomics. Issues such as data sparsity, batch effects, and the need for large annotated datasets hinder the development of robust DL models. Moreover, the interpretability of complex DL models remains a significant concern, necessitating the integration of XAI techniques to provide biologically meaningful interpretations. Addressing these challenges is crucial for the broader adoption and application of DL in single-cell genomics (Ma & Xu, 2022). 4. Future Directions Future research should focus on developing DL models that can handle the complexity and heterogeneity of single-cell data. Emphasis on improving model interpretability through XAI, addressing data quality issues, and creating standardized benchmarking datasets will be crucial for advancing the field. Collaborative efforts across disciplines will be essential to overcome existing limitations and fully realize the potential of DL in single-cell genomics (Erfanian et al., 2023). Table: Deep Learning Models in Single-Cell Genomics Model/Tool Focus Area Key Features DANCE Gene expression modeling and clustering Unified framework, supports multiple modules and tasks scDMSC Multi-omics data clustering Multi-view subspace clustering, addresses high dimensionality and sparsity XAI-integrated DL models Aging process analysis Combines DL with explainable AI tools, enhances model interpretability Discussion The rapid development of deep learning (DL) methods has significantly advanced the analysis of single-cell genomics, enabling researchers to extract meaningful patterns from highly complex, sparse, and high-dimensional datasets. Models such as DANCE and scDMSC have demonstrated the ability to uncover cellular heterogeneity and improve clustering accuracy, illustrating the practical benefits of DL in dissecting biological complexity (Ding et al., 2024; Erfanian et al., 2023). Integration with spatial transcriptomics and epigenomics has further extended the utility of DL by allowing researchers to study gene expression in the context of tissue architecture. Combining DL with explainable AI (XAI) techniques enhances model interpretability and biological insight, which is essential for translating computational findings into practical biomedical applications (Li et al., 2025). Despite these advances, significant challenges remain. Data sparsity, batch effects, and limited availability of well-annotated datasets can hinder model performance. Furthermore, the interpretability of complex DL models continues to be a key limitation, highlighting the need for standardized frameworks, reproducibility measures, and benchmarking datasets (Ma & Xu, 2022). Looking forward, future work should focus on developing integrative DL models that can simultaneously handle multimodal single-cell datasets, including transcriptomic, epigenomic, and spatial information. Greater emphasis on explainable, transparent models will be crucial for facilitating adoption in biological and clinical research. Additionally, collaborative efforts across computational and experimental disciplines will accelerate the development of robust, generalizable DL tools capable of addressing current challenges.
- Conclusion: Deep learning (DL) has emerged as a transformative tool in single-cell genomics, offering unprecedented capabilities to analyze high-dimensional, sparse, and heterogeneous data. Methods such as DANCE and scDMSC have demonstrated significant improvements in clustering accuracy and cellular heterogeneity analysis, while integration with spatial transcriptomics and explainable AI (XAI) approaches has enhanced biological interpretability. Despite these advancements, challenges such as data sparsity, batch effects, and model interpretability remain critical barriers. Future research should prioritize the development of integrative DL models capable of handling multimodal datasets, while ensuring transparency, reproducibility, and robust benchmarking. Collaborative efforts between computational and experimental researchers will be essential to fully harness DL’s potential in uncovering cellular diversity, understanding disease mechanisms, and guiding therapeutic discovery.
- Keywords: single-cell genomics, deep learning, spatial transcriptomics, epigenomics, explainable AI
