P28-12 Stronger together: achieving scalability for a novel in vitro human male reproductive toxicity assay by deep learning image analysis pipelines

Declining male fertility, potentially driven by male reproductive toxicity from exposure to harmful environmental factors, chemical compounds, or pharmaceuticals, is a major public health concern. Assessment of nonclinical developmental and reproductive toxicity (DART) for all new pharmaceuticals is required by the European Medicines Agency, however animal testing (typically in mouse, rat, or rabbit) might not always be predictive because of fundamental differences in the reproductive systems from humans. Therefore, there is a pressing need for a fully validated, humanized system to assess male reproductive toxicity. We recently published a protocol for in vitro gametogenesis from human stem cells. Here, we aim to develop a scalable, validated, fully characterized, high-content screening platform for male reproductive toxicity screenings or protective drug screenings, using these human primordial germ cell-like cells (hPGCLCs) as proxy for spermatogonial stem cells (SSCs). hPGCLCs were generated from human pluripotent stem cells (hPSCs). Thereafter, cells were fixed and stained for nuclear-markers POU5F1 (or OCT4), and SOX17. Double positive cells were identified as hPGCLC. Confocal microscopy images were acquired using multiple imaging sites per well. Deep convolutional neural networks were trained to identify individual nuclei and predict germ-cell identity based on marker intensity. In addition, prediction of germ-cell identity by multi-layer perceptron based on generic cell features was explored. Models were compared by accuracy of their prediction relative to the ground truth. We are currently optimizing the image-analysis pipelines and developing a user-friendly interface for comprehensive analysis of the results. Initial results show that the combination of OCT4 and SOX17 staining is able to reliably identify hPGCLC in our cell cultures. The current version of our image analysis pipeline based on OCT4 / SOX17 stainings is able to detect individual hPGCLCs even in highly overgrown areas with great accuracy. We are currently identifying different combinations of cell feature to stratify cell populations by unsuper-vised clustering. Implementing deep-learning based computer vision shows great promise to predict germ cell identity, making our in vitro human male reproductive toxicity into a scalable, reproducible, and cost-effective assay, ultimately yielding a highly effective first-tier DART testing protocol.