build_gps {GPS} R Documentation

Build a GPS (GNN + Transformer) gene regulatory network model.

Description

Usage

build_gps(x,
    prior = NULL,
    config = NULL);

Arguments

x

the time series gene expression matrix object: rows are genes and columns are samples, and every sample must carry a time point label (for example the matrix loaded by geneExpression::load.expr). At least three distinct time points are required. [as GeneExpressionData]

prior

the prior regulatory network object (for example created by bnlearn::as.prior_net). Gene pairs that are not described by this network are kept as isolated nodes and still participate in the global attention branch. [as PriorNetwork]

config

the hyper-parameter configuration created by gps::gps_config; when Nothing the default GPSConfig values are used. Note that the window length may be clipped when the input has too few time points. [as GPSConfig]

env

the R# runtime environment. [as Environment]

Details

The constructor performs the following preparation steps once and keeps the results, so they are reused by every training epoch and by the virtual perturbation analysis:

  1. preprocessing: CPM normalization, log2(x + 1) transformation, per gene z-score and

the optional high variance gene filter;

  1. graph construction: the prior TF -> target edges are mapped onto the modelled genes,

optionally extended with co-expression edges;

  1. static graph encoding: shortest path distance buckets, Laplacian positional encoding and

the in/out degree buckets;

  1. the strictly time ordered train / validation / test split of the sliding windows.

The returned object is untrained; call gps::train before any prediction or perturbation analysis. It can also be persisted with the writeBin handler and restored with readBin.modular_bayesian.

Authors

biosystem

Value

an untrained GPSModel object

clr value class

Examples

[Package GPS version 1.0.0.0 Index]