PDGrapher
Gonzalez G, Lin X, Herath I, Veselkov K, Bronstein M, Zitnik M
2025 · Nature Biotechnology
Recent Level 3 frontier. Graph neural network that predicts transcriptional responses to candidate interventions and ranks perturbations by predicted reconstruction of a target state.
Abstract
From the original paper, Nature Biotechnology · PubMed
Phenotype-driven approaches identify disease-counteracting compounds by analysing the phenotypic signatures that distinguish diseased from healthy states. Here we introduce PDGrapher, a causally inspired graph neural network model that predicts combinatorial perturbagens (sets of therapeutic targets) capable of reversing disease phenotypes. Unlike methods that learn how perturbations alter phenotypes, PDGrapher solves the inverse problem and predicts the perturbagens needed to achieve a desired response by embedding disease cell states into networks, learning a latent representation of these states, and identifying optimal combinatorial perturbations. In experiments in nine cell lines with chemical perturbations, PDGrapher identifies effective perturbagens in more testing samples than competing methods. It also shows competitive performance on ten genetic perturbation datasets. An advantage of PDGrapher is its direct prediction, in contrast to the indirect and computationally intensive approach common in phenotype-driven models. It trains up to 25× faster than existing methods, providing a fast approach for identifying therapeutic perturbations and advancing phenotype-driven drug discovery.
Summary
PDGrapher is the perturbation-trained frontier of Level 3. It is a graph neural network that predicts downstream transcriptional responses to candidate gene or drug interventions and ranks perturbations by whether their predicted outcome reconstructs a desired treated or healthy state.
Together with classical Level 3 exemplars such as Stable Motifs, PDGrapher spans the class from a curated Boolean control method on one end to a learned neural-network surrogate on the other — both qualify as Level 3 because they encode intervention-dependent forward prediction rather than only static similarity.
Why this level
Level 3 because candidate interventions enter an explicit forward operator () — a graph neural network trained on perturbation data — and the predicted post-intervention transcriptional response is what scores each candidate. Representation family is executable intervention model: the network is queryable with new at no additional experimental cost. PDGrapher is cited in §3.4 as the recent perturbation-trained Level 3 exemplar.