GPS

require(GCModeller);

#' GPS: the GNN + Transformer based gene regulatory network modelling toolkit.
imports "GPS" from "biosystem";

GPS: the GNN + Transformer based gene regulatory network modelling toolkit.

This R# package module wraps the GPSModel of the GPS sub-system. The model fuses a local message passing branch (edge type aware aggregation over the prior regulatory network) with a global multi head attention branch (Graphormer style shortest path bias), encodes the obtained gene representations over time with a GRU and finally decodes the expression of every single gene with its own linear readout.

The workflow exposed by this module is:

• gps_config: create the hyper-parameter configuration object

(GPSConfig, registered as the gps_config R# type);

• build_gps: build a model object from a time series gene expression matrix and a

prior regulatory network object;

• train: run the masked reconstruction pre-training followed by the next-step

prediction fine-tuning, and attach the training diagnostics to the returned object.

A trained GPSModel implements the InsilicoPerturbationExperiment interface, so the knockouts, knockdown and overexpress api of the bnlearn package can be applied to it directly for in silico perturbation experiments. The generic writeBin handler of this module persists a model into a zip based binary stream, and readBin.modular_bayesian restores it back to a fully usable model object.