[Paper Review] Deep Learning in Single-Cell and Spatial Transcriptomics Data Analysis: Advances and Challenges from a Data Science Perspective

Shuang Ge, Shuqing Sun+3 more|arXiv (Cornell University)|Dec 4, 2024

Single-cell and spatial transcriptomicsBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper reviews deep learning (DL) applications in single-cell and spatial transcriptomics, addressing four key data science challenges: data sparsity, diversity, scarcity, and correlation. It evaluates 58 methods across 21 curated datasets from 9 benchmarks, highlighting DL's advantages in handling high-dimensional, noisy, multimodal data, and proposes future directions in AI methodology, benchmarking, and clinical translation.

ABSTRACT

The development of single-cell and spatial transcriptomics has revolutionized our capacity to investigate cellular properties, functions, and interactions in both cellular and spatial contexts. However, the analysis of single-cell and spatial omics data remains challenging. First, single-cell sequencing data are high-dimensional and sparse, often contaminated by noise and uncertainty, obscuring the underlying biological signals. Second, these data often encompass multiple modalities, including gene expression, epigenetic modifications, and spatial locations. Integrating these diverse data modalities is crucial for enhancing prediction accuracy and biological interpretability. Third, while the scale of single-cell sequencing has expanded to millions of cells, high-quality annotated datasets are still limited. Fourth, the complex correlations of biological tissues make it difficult to accurately reconstruct cellular states and spatial contexts. Traditional feature engineering-based analysis methods struggle to deal with the various challenges presented by intricate biological networks. Deep learning has emerged as a powerful tool capable of handling high-dimensional complex data and automatically identifying meaningful patterns, offering significant promise in addressing these challenges. This review systematically analyzes these challenges and discusses related deep learning approaches. Moreover, we have curated 21 datasets from 9 benchmarks, encompassing 58 computational methods, and evaluated their performance on the respective modeling tasks. Finally, we highlight three areas for future development from a technical, dataset, and application perspective. This work will serve as a valuable resource for understanding how deep learning can be effectively utilized in single-cell and spatial transcriptomics analyses, while inspiring novel approaches to address emerging challenges.

Motivation & Objective

• To systematically analyze the four major data science challenges in single-cell and spatial transcriptomics: sparsity, diversity, scarcity, and correlation.
• To evaluate the performance of 58 computational methods across 21 datasets from 9 benchmarks in tasks such as imputation, clustering, and multi-omics integration.
• To compare deep learning approaches with traditional machine learning methods in terms of accuracy, interpretability, and robustness on complex biological data.
• To identify limitations in current benchmarks, including evaluation metrics and dataset representativeness, and advocate for biologically interpretable, fair, and robust benchmarks.
• To outline future research directions in AI methodology, dataset development, and real-world clinical applications of deep learning in single-cell and spatial omics.

Proposed method

• The authors curate 21 datasets from 9 established benchmarks covering tasks such as imputation, cell identity recognition, clustering, gene regulatory network inference, and multi-omics integration.
• They evaluate 58 computational methods, including deep learning models such as autoencoders, variational autoencoders, graph neural networks, generative adversarial networks, and convolutional neural networks.
• Performance is assessed using standard metrics like accuracy, AUC, and RMSE, with emphasis on biological relevance through in vitro validation and simulation-based benchmarks.
• The review integrates mathematical foundations of DL models, comparing them with traditional ML techniques, and discusses their ability to model complex biological dependencies and spatiotemporal correlations.
• The authors propose enhancing interpretability by linking DL models to prior biological knowledge through interpretable parametric rules.
• They advocate for dynamic, continuously updated models and simulation-based data generation to improve generalizability and robustness on unseen or uncertain data.

Experimental results

Research questions

• RQ1 How do deep learning models outperform traditional machine learning methods in handling high-dimensional, sparse, and noisy single-cell and spatial transcriptomics data?
• RQ2 What are the key limitations of current benchmark datasets and evaluation metrics in reflecting the true performance and generalizability of computational methods?
• RQ3 How can multimodal and multi-source data integration be effectively achieved using deep learning to improve biological interpretability and prediction accuracy?
• RQ4 What are the main challenges in modeling spatiotemporal dependencies and incorporating prior biological knowledge into deep learning frameworks for single-cell and spatial omics?
• RQ5 What future directions in AI methodology, dataset curation, and clinical application are most critical for advancing the field of single-cell and spatial transcriptomics?

Key findings

• Deep learning models, particularly autoencoders, graph neural networks, and generative models, demonstrate superior performance in handling high-dimensional, sparse, and noisy single-cell and spatial transcriptomics data compared to traditional methods.
• The review identifies that current benchmarks often lack biological relevance and fail to reflect the heterogeneity introduced by diverse sequencing platforms and continuous spatial-temporal settings.
• Evaluation metrics such as accuracy, AUC, and RMSE are widely used but insufficient; biological validation through in vitro experiments is essential to confirm the significance of model predictions.
• Simulation-based data generation offers a promising approach to creating diverse, labeled benchmark datasets that improve model evaluation and generalizability.
• There is a critical need for fair, robust, and biologically interpretable evaluation metrics that go beyond standard performance scores to reflect real-world applicability.
• Future research should prioritize dynamic, continuously updated deep learning models, integration of prior biological knowledge, and translation of DL methods into clinical and precision medicine applications.