Extended Data Fig. 1: Schematic of the training pipeline for EmbryoNet.
From: EmbryoNet: using deep learning to link embryonic phenotypes to signaling pathways
(a) Overview of a training iteration. Augmented embryos were collected into a training batch with known embryo age. After EmbryoNet processed a batch of input images with ages, network outputs were compared with ground truth values. Based on cross-entropy loss, EmbryoNet weights were updated to minimize the loss. (b) Examples of augmentations used. (c) In our model transition logic, embryos have a limited set of allowed class transitions. All start in the class Unknown and can transition to any other class, from where they can go only to Dead, but not to other classes. Other transitions were assigned a cost. The model with the least cost was selected. (d) Schematic of the classification pipeline. (e) Graphical user interface (GUI) of EmbryoNet. (f,g) Comparison of EmbryoNet’s performance to recognize phenotypes induced by signaling modulation using small-molecule inhibitors, overexpression of signaling antagonists or pathway mutants. Nodal phenotypes (f) induced by small-molecule inhibitor treatment (SB-505124, n=33), injection of a pathway antagonist (lefty1 mRNA, n=27) or in a receptor mutant (MZoep, n=27) were all classified by EmbryoNet as –Nodal with similar accuracy. BMP phenotypes (g) induced by small-molecule inhibitor treatment (LDN-193189, n=45), pathway antagonist injection (chordin mRNA, n=26) or in a pathway ligand mutant (swirl-/-, n=13) were all classified by EmbryoNet as –BMP with similar accuracy. Scale bars: 500 µm.
