Wide and deep learning for automatic cell type identification.

Area of Science:

• Computational biology
• Cancer research
• Machine learning

Background:

• Accurate cell type classification is crucial for understanding the tumor microenvironment and patient immune response.
• Single-cell technologies generate high-dimensional data requiring advanced analytical methods.
• Overfitting is a common challenge in machine learning models for biological data.

Purpose of the Study:

• To demonstrate the effectiveness of Wide and Deep Learning (WDL) for cell type classification in cancer research.
• To show that regularization techniques, specifically dropout and L2 regularization, can prevent model overfitting.
• To evaluate WDL's performance against traditional deep learning and state-of-the-art algorithms.

Main Methods:

• Implementation of a Wide and Deep Learning (WDL) model incorporating both deep neural networks and wide linear models.
• Application of regularization techniques, including dropout and L2 regularization, to mitigate overfitting.
• Comparative analysis of WDL against traditional deep learning, CHETAH, and SingleR using cross-platform single-cell data.

Main Results:

• WDL with combined dropout and L2 regularization achieved a stable validation loss, independent of training iterations.
• WDL demonstrated superior cell type prediction accuracy (36.5% to 86.9%) when trained on one platform (10X melanoma) and tested on another (SMART-seq basal cell carcinoma).
• WDL outperformed existing methods, achieving higher accuracy for overall cell types and specific T cell subtypes (CD4, CD8) compared to CHETAH and SingleR.

Conclusions:

• Regularization is essential for building robust predictive models in cancer research.
• WDL offers enhanced predictive performance and classification accuracy, particularly in cross-platform single-cell analyses.
• WDL represents a significant advancement for cell type classification in the complex tumor microenvironment.