Fig. 4: Application of Twin Networks to identify drug-induced phenotypes.
From: Uncovering developmental time and tempo using deep learning
a, Strategy of similarity calculation between embryos at the same developmental stage under different drug treatments. An untreated embryo (top) serves as reference to which drug-treated embryos (bottom) are compared. Examples for untreated, BMP-inhibited and PCP-inhibited embryos are shown at 1.25, 10 and 26 hpf. The cosine similarity between a treated embryo and the reference embryo is calculated for every timepoint. Scale bar, 500 μm. b–i, Upper panel, mean similarities and s.d. of similarities for untreated (n = 44) (b) and –BMP (n = 44) (c), –PCP (n = 14) (d), –FGF (n = 44) (e), –Shh (n = 44) (f), –Nodal (n = 44) (g), +RA (n = 44) (h) and –Wnt (n = 18) (i) embryos relative to the reference group of untreated embryos as a function of time. Lower panel, significance levels of the difference from untreated embryos determined using a nonparametric one-sided Mann–Whitney U test over each timepoint of the image series. No adjustments for multiple comparisons were made. j, Dependency of the accuracy of abnormality detection on the number of embryos used to analyze –BMP, –PCP, –FGF, –Shh, –Nodal, +RA and –Wnt embryos. Mean and s.d. are shown for five repetitions with randomly selected samples.
