Fig. 1: Overview of PDGrapher.
From: Combinatorial prediction of therapeutic perturbations using causally inspired neural networks
a, Given a paired diseased and treated gene expression sample and a proxy causal graph, PDGrapher’s perturbagen discovery module, ƒp, predicts a candidate set of therapeutic targets to shift cell gene expression from a diseased to a treated state. b, Given a disease sample’s gene expression, a proxy causal graph and a set of perturbagens, PDGrapher’s response prediction module, ƒr, predicts the gene expression response of the sample to each perturbagen. ƒr represents perturbagen’s effects in the graph as edge mutilations. c, ƒp is optimized using two losses: a cross-entropy cycle loss to predict a perturbagen \({{\mathcal{U}}}^{{\prime} }\) that aims to shift the diseased cell state closely approximating the treated state, \({\rm{CE}}\left({{\bf{x}}}^{\rm{t}},\,{f}_{\rm{r}}({{\bf{x}}}^{\rm{d}},\,{f}_{\rm{p}}({{\bf{x}}}^{\rm{d}},\,{{\bf{x}}}^{\rm{t}}))\right)\,\,({\rm{with}}\,{f}_{\rm{r}}\,{\rm{frozen}})\), and a cross-entropy supervision loss that directly supervises the prediction of \({{\mathcal{U}}}^{{\prime} }\), \({\rm{CE}}\left({{\mathcal{U}}}^{{\prime} },{f}_{\rm{p}}({{\bf{x}}}^{\rm{d}},\,{{\bf{x}}}^{\rm{t}})\right)\) (see Methods for more details). d, Both ƒr and ƒp follow the standard message-passing framework, where node representations are updated by aggregating the information from neighbours in the graph. GEX, gene expression.
