Robust phenotyping of
developmental defects using
convolutional neural networks

Early development is governed by a handful of signaling pathways that balance tissue growth, differentiation and morphogenesis.

Loss-of-function of each of these pathways causes characteristic patterning defects that can, however, partially overlap with each other.

The automated phenotyping tool that we developed – EmbryoNet – is based on convolutional neural networks and can detect signaling defects in embryos. EmbryoNet outcompetes human assessors in terms of speed, accuracy and sensitivity for the phenotypic classification of the major signaling pathways in early vertebrate development, and it can also classify incomplete penetrance of phenotypes. EmbryoNet can be adapted to other species separated by hundreds of million years in evolution, enabling the analysis of high-dimensional phenomic data across taxa. In a proof-of-concept screen with two drug libraries, we showed that EmbryoNet faithfully recovered compounds of known function and non-toxic treatments. This approach will likely enable the association of currently used medications with still unknown signaling pathways, opening up the possibility to apply well-characterized compounds in new therapeutic contexts and applications.

We provide EmbryoNet as open-source software, with python packages, a GitHub repository and GUIs for labeling data and phenotype classification. We also provide imaging data from training, testing, and drug screen as resources to the community. Due to the modular open-source nature, EmbryoNet can be easily adapted to a variety of purposes, including structures such as amphibian embryos and organoids, where automated phenotyping will expedite biological and pharmaceutical discovery.