Figure 1.

Overview of the Method

Unsupervised and supervised learning methods were used for building ML-derived understandings of the developmental process with each day as a potential reference. The number of clusters in each reference day can be decided using S-score and scree plots (not shown in figure) or can be exhaustively searched in a range. Multiple versions of developmental processes were constructed. A method was developed for comparing the multiple candidate processes to choose the one with highest self-consistency as the final ML-derived understanding. Existing knowledge was used in the last step to annotate the ML-derived developmental process and detect possible new findings.