gms | German Medical Science

Introduction: Deep-learning techniques can be powerful, but the complex computational setup and software deployment required [1] have widened the gap between sophisticated data modelers and less tech-savvy users. To change this, offering deep learning methods via the web browser can be a solution that avoids complex setups but still allows interactive data exploration. An increasingly important issue in life science is modeling single-cell RNA sequencing (scRNA-seq). scRNA-seq captures the complex cellular heterogeneity of RNA transcripts by profiling the transcriptomes of individual cells. Deep-learning techniques based on variational autoencoders [2] facilitate biological insight [3], [4] by learning low dimensional latent representations of scRNA-seq data. We implement a novel dimensionality reduction method based on gradient optimization. Together with visualization, it allows users to interactively explore individual genes from scRNA-seq data locally in their web browsers. To improve reproducibility, the data and models are versioned together in a research data management system, and exploration results are eventually added to improve data documentation.

Methods: Gradient-based optimization is at the heart of deep learning and is usually provided by complex frameworks. Earlier approaches for deep learning methods in the browser [5] relied on reimplementing algorithms in Javascript to provide gradients; instead, we utilize the intermediate format WebAssembly [6], which is available in all modern browsers and can target different programming languages. We enhance the gradient optimization of our deep learning method by introducing a novel term that allows optimization with respect to a selected set of genes by domain experts. This approach enables exploration of the impact of these genes on low-dimensional latent representations. The gradient optimization is implemented in Swift combined with the Python-based Shinylive, a serverless framework for providing interactive interfaces. The bundle of all components is delivered via a research data management system and can therefore be used optionally to provide managed datasets. The user is provided with interactive visualizations, and results can be transferred back as data documentation.

Results: As a case study, we have shown that deploying deep learning via WebAssembly can make these methods more accessible in restricted environments or to less tech-savvy scientists. With this technology, domain experts can interactively explore the low-dimensional manifold produced by deep learning techniques in their web browsers. Combining method and data as revision-controlled digital objects further lower barriers to use increases reproducibility, and improves data documentation through optional reintegration of exploration results.

Discussion: WebAssembly is an emerging technology that simplifies the integration of methods developed in various programming languages. While it still has room for improvement, particularly in supporting data science languages, the solution demonstrates the feasibility of making complex models accessible and executable in local environments. Adoption to other deep learning-based models for Multiomics data integration is possible.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.