Extended Data Fig. 5: Classification of mild Nodal and BMP phenotypes.
From: EmbryoNet: using deep learning to link embryonic phenotypes to signaling pathways

(a) Schematic of the experiment with lower inhibitor doses. Lower doses of the BMP inhibitor LDN-193189 lead to weaker phenotypes, detectable from late gastrulation onwards. While in the severe cases no clear structures are distinguishable, moderate embryos have a head and somites and display the characteristic BMP loss-of-function phenotype with curled-up tails (Kishimoto et al. 1997). Mild embryos have a largely intact body axis only missing the tail. (b-d) Confusion matrices showing the performance of classification of weaker phenotypes. EmbryoNet (b) had a lower performance on milder compared to severe phenotypes, especially for Nodal-inhibited samples. Human annotators were also less consequent as seen from the confusion matrix between the accepted ground truth and a second labeler (c). EmbryoNet-Prime had better success in detecting weak Nodal phenotypes compared to EmbryoNet, but was less performant on –BMP and on average (d).