Fig. 1: Overview of the experimental design of LT-scSeq, CLADES’s architecture and its robustness on synthetic datasets.

a General workflow of LT-scSeq experiment using static barcoding techniques with viral integration. DNA barcodes are induced at an early time point, then scRNA-seq data and clonal information are acquired at subsequent time points. b CLADES takes total cell counts and transition directions as input, then uses a neural net to estimate the transition rates between populations, and an ODE module to reconstruct cell counts at each time point. c CLADES is able to infer the dynamic changes of population size on various resolutions and the confidence interval of transition rates. d With the estimated kinetic rates of each clone, CLADES can further: 1) simulate detailed topologies of division and differentiation, and 2) infer the meta-clone level probability of lineage realization. e,f Given cell counts generated by either time-invariant or time-variant rates, the performance of both constant and dynamic mode are validated on test time points. e The constant mode (i) consistently performs better than the dynamic mode (ii) no matter how many time points were used to train the model; the robustness of the constant mode given time-invariant datasets was evaluated as well (iii). f the constant mode (i) performs worse than the dynamic mode (ii) except using only 2 training time points; the robustness of the dynamic mode given time-variant datasets was evaluated as well (iii). Mean_n is the mean value for all the test sets given n training time points. Based on the results of the above analysis, the minimum recommended time points for CLADES to have satisfactory performance is 3 (including the induction time point, which can be inferred from the data itself, Methods). Panel (a) and (d) are created in BioRender. Huang, Y. (2025) https://BioRender.com/ak99vwt.