Loading…
Drug repurposing model that propagates transcriptional signatures through a pretrained graph autoencoder of the human protein interaction network.
No providers recorded yet. Browse all providers
Two drugs that act on the same pathway rarely produce the same transcriptional signature, and a network method that wants to compare them has to represent each one as a different weighted graph. That is the sticking point for graph-embedding approaches to drug repurposing: they learn a fixed low-dimensional embedding of a static protein–protein interaction (PPI) network, so a perturbed version of that network cannot be embedded without recomputing or retraining the model. At screening scale — thousands of compounds, each carrying its own signature — that cost is prohibitive.
GraphPert separates the two problems. A graph autoencoder is pretrained once, by self-supervised link reconstruction, on the unperturbed human interactome. A transcriptional signature is then written into the graph as a symmetric edge-weight modification: each interaction is rescaled by the measured expression changes of both proteins it connects. Running that reweighted graph through the pretrained encoder returns a latent displacement — how far, and in which direction, the perturbation moved the network in embedding space. Compounds are ranked by cosine similarity between their displacement and that of a chosen reference perturbation.
The model was developed at IQUIBA-NEA, a joint institute of Universidad Nacional del Nordeste and CONICET in Corrientes, Argentina, and published in Computers in Biology and Medicine. The published case study is melanoma, with a BRAF V600E knockout in A375 cells as the reference, but that reference is an input to the method rather than a property of the model.
The encoder is a three-layer graph convolutional network narrowing 256 → 128 → 64 channels with LeakyReLU activations and dropout, paired with the inner-product decoder of a standard graph autoencoder; training optimizes link reconstruction with Adam at a learning rate of 1e-3 and weight decay of 5e-4 over 400 epochs. Node features are protein language model embeddings — the released pipeline defaults to the 150M-parameter ESM-2 checkpoint — and the graph is assembled from STRING, restricted to human interactions with experimental support above a confidence threshold. The network reported in the paper spans 12,458 proteins. A single scaling factor, 5.0 by default, sets how strongly a signature perturbs the edge weights.
Evaluation is a retrospective decoy screen. Known MAPK cascade inhibitors are seeded at a 1:9 ratio among Connectivity Map compounds with no known MAPK pathway activity, and the combined library is ranked by similarity to the BRAF V600E knockout signature in A375 melanoma cells. GraphPert recovers the seeded inhibitors with an AUC above 0.80.
The direct use is computational drug repurposing. A group holding a reference perturbation that captures the intervention it wants — a knockout of a driver oncogene, or the signature of a drug already known to work — can rank a compound library by network-level resemblance to that reference and carry the top candidates into cell assays. Because scoring is a forward pass through an already-trained encoder, the library can be large and the reference can be swapped without refitting anything. The same machinery supports mechanism-of-action work: compounds whose displacements cluster together implicate shared downstream modules even when their annotated targets differ.
GraphPert's contribution is architectural rather than a benchmark result. It shows that the expensive part of network-based screening — learning the interactome's structure — can be paid once and then reused across arbitrarily many perturbations, and that routing signatures through that learned structure widens the set of mechanisms a screen can reach, picking up distal-target compounds that direct signature matching ranks low. The supporting evidence is correspondingly narrow: one retrospective benchmark, one reference perturbation, one cell line, and no prospective experimental validation. The code and the Connectivity Map signature files used in the paper are published, but no trained checkpoint is distributed and the repository carries no license, so users reproduce the pretraining themselves and redistribution terms are unstated.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.