First, we worked on defining a knowledge graph with attributes added to edges. For example, instead of “gene A activates gene B”, we aim to encode “gene A activates gene B with a certain rate that depends on the state of the local nodes or on the external parameters”. These attributes are suitable for encoding mechanistic constraints of gene expression and regulation, such as “The rate of activation of B by A increases linearly until the concentration of B reaches a certain value. Above this value, the activation rate stays constant.”
Further, in order to gain insight into the kinetics and mechanisms of existing gene interactions (i.e. the network structure), we worked on combining this static information with the temporal information from scRNA-seq data. Using a hierarchical Bayesian model with delays, we represent a gene network as a system of stochastic chemical reactions, where biological constraints and molecular interaction kinetics, such as the cell size and the regulatory mechanisms of gene expression, are incorporated implicitly in the mRNA and protein production reaction rates and delays. To evaluate the accuracy of our modelling framework, we simulate in silico scRNA-seq data of a sparse and modular gene network with a delayed Gillespie algorithm. We then reconstruct pseudotime from this data using the existing software (Cyclum and MONOCLE). The next step is, given the ground truth on which genes interact with each other, to infer the kinetic parameters and the mechanisms of these interactions.
While conducting work described above, we encountered the difficulties in obtaining accurate pseudo-time reconstruction from large networks. To mitigate that, we developed a strategy where pseudo-time is reconstructed from gene clusters, which often correspond to functional gene groups in cellular gene networks, and then the trajectories of each of the gene clusters are connected via the genes that belong to multiple clusters simultaneously. We are now preparing for publication the manuscript that presents this pseudo-time reconstruction strategy.
Finally, we developed a new modelling framework that can capture the fingerprint of temporal heterogeneity of gene expression environment from the snapshot data such as scRNA-seq. In particular, this heterogeneity can affect even the mean expression levels, as a consequence of bimolecular interactions with low-abundant environment species. We also described several examples in which incorporating this heterogeneity allows to explain the mean gene expression levels observed from the data. This work is now a manuscript in preparation.
