Application of machine learning to discriminate photoreceptor cell species in xenotransplanted chimeric retinas

Abstract

Photoreceptor transplantation is being studied to improve visual function in retinal diseases causing blindness, such as age-related macular degeneration, hereditary eye diseases, and traumatic retinopathy, among others. Preclinical studies often involve the delivery of exogenous human photoreceptor cells into the retinas of animal models. In such experiments, a key readout is the differential frequency of donor cell somatic integration versus artificial labeling secondary to material transfer of cytosolic or nuclear labels from donor to recipient cells. For this analysis, the ability to recognize photoreceptor nuclei as being of donor (human) versus animal is key, but purely immunohistology discrimination can be challenging because of antigenic species overlap or intercellular antigen transfer. To address this challenge, we sought to develop and validate a computational technique to discriminate between photoreceptor cells of different animal species based on machine learning of nuclear morphology. Here, we aimed to evaluate the feasibility of using computer-assisted detection of separate nuclei and employing random forest classification to automate the species differentiation, among DAPI-stained photoreceptors after xeno-transplantation of human photoreceptors into the retinas of mice and pigs. Our models were trained on single-species samples and validated with mixed-species samples. We then transplanted human embryonic stem cell-derived retinal organoid cells into rodent and pig retinal degeneration models. The random forest model accurately determined cell identity post-xenotransplantation, validated by histological assessment using an anti-human nuclear antibody. Our results support the potential efficacy of employing machine learning image analysis and classification techniques that may promote experimental rigor, minimize observer bias, and enable high throughput semi-automated workflows for transplantation outcomes analysis. The methodological framework reported here may enable a more nuanced and precise analysis of the behavior of transplanted photoreceptors for the purposes of human retinal regeneration.