Abstract
Purpose :
To assess the accuracy of a deep learning-based domain adaptation model to learn patterns of gene expression in different subtypes of retinal ganglion cells (RGCs) and utilize the knowledge to determine if similar gene expression patterns can be recognized within unseen subsets of RGCs
Methods :
We utilized single cell RNA-seq profiles from 6,225 RGCs of the right and left eyes of postnatal day five mice, which are publicly available on GEO. We developed a deep unsupervised domain adaptation construct to learn patterns of scRNA-seq of RGC subtypes and then used the learned patterns to identify the subtypes of new single RGCs. The scRNA-seq dataset was divided into the labeled source and unlabeled target subsets. We used four loss functions in the learning process: 1)conditional maximum mean discrepancy (CMMD) loss, to minimize the discrepancy between feature distributions of the data in the source and target domains, 2)min-max loss in the discriminative network, to minimize the class discrimination error between the source domain features and the target domain features, 3)cross-entropy loss for source domain, to minimize the source classification error, and 4)virtual adversarial training (VAT) loss, to prevent overfitting and to boost model generalization to unseen RGC cells(Fig. 1).
Results :
We used 70% of the labeled RGC cells in the source domain and 30% of the unlabeled RGC cells in the target domain. Patterns of gene expression in the source domain dataset(Fig. 2, left) were learnt by the model and the knowledge was transferred to the unlabeled RGCs in the target domain dataset with a classification accuracy of 96%(Fig. 2, right).
Conclusions :
Deep unsupervised domain adaptation can learn cell-specific expression patterns from single RGCs in one dataset and successfully transfer the knowledge to single RGCs in another unseen dataset. Such deep learning models may mitigate systematic bias and allow for transfer of gene expression knowledge among different subclasses of cells or even different tissues.
This is a 2021 Imaging in the Eye Conference abstract.