Identification of Retinal Cell Types Based on Single-Cell Transcriptomics

Abstract

Purpose : To identify different retinal cell types from patterns of transcriptome data and to evaluate the model based on data from different batches.

Methods : We developed a deep domain adaptation model to identify retinal cells based on single-cell RNA sequencing (scRNA-seq) data (Fig. 1). The dataset included 44,808 single cells from 39 retinal cell types with 24,658 genes collected from mice in seven different batches (B1 to B7). Our unsupervised model included source classification, adversarial, and target virtual adversarial loss functions in the learning process along with domain adaptation strategy based on conditional maximum mean discrepancy (CMMD) loss function to align the source and target distributions to reduce misclassification error and maximize robustness. We evaluated the proposed model using classification accuracy and confusion matrix based on data from different batches.

Results : The number of cells in each batch ranged from 3226 to 8336 (Fig. 2-B). The accuracy of the model based on different pairs of batches is shown in Fig. 2-C to 2-H. The mean accuracy of the model for correctly classifying 39 different retinal cell types was ~92%. Across seven different batches, the identification accuracies of the model ranged from 74% to nearly 100%. Our results outperformed several state-of-the-art models.

Conclusions : We integrated multiple loss functions to a deep learning-based domain adaptation model to identify retinal cell types from scRNA-seq data and achieved a high level of accuracy in detecting correct retinal cell types. The model was relatively resistant to batch effect and could be used in single-cell studies to detect various cell types from different tissues.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Figure 1. Diagram of the proposed deep learning-based domain adaptation model. A total of 44,808 retinal cells with 24,658 unique genes from seven different batches (B1 to B7) were used to develop and evaluate the model. The confusion matrix of 39 retinal cells is presented as the outcome. ReLU: Rectified Linear Unit; FC: Fully Connected.

 

Figure 2. A) Partial structure of the retina. B) Cell counts and cell type membership distribution in different batches of the dataset. C-H) Box plots representing the accuracy of the model based on different pairs of batches (source -> target).