Can AI predict genetic responses without losing its transparency?
In our latest work, we present iPerturb, a new neural network framework that predicts genome-wide transcriptional responses to genetic perturbations. Unlike standard deep learning “black boxes,” iPerturb maps biological data onto real gene regulatory networks, offering both unmatched predictive accuracy and full interpretability. Explore how we’re scaling mechanistic biology to the genome level.
Research Abstract: Building an interpretable model that predicts genome-wide transcriptional responses to perturbations remains a grand challenge in biology. Mechanistic kinetic models are difficult to scale to the genome level, whereas deep neural network–based models often lack interpretability. To bridge this gap, we propose iPerturb, a biologically informed neural network framework for predicting gene expression changes induced by genetic perturbations in human cells. iPerturb first extracts regulatory interactions from public gene-regulation databases to construct a context-specific template gene regulatory network (GRN), then uses perturbation data to learn the kinetic parameters governing each regulatory edge, and finally simulates transcriptional responses to unseen perturbations via multi-hop message passing on the GRN. Evaluated on 4,450 gene-knockdown perturbations across two human cell lines, iPerturb recovers cell-type-specific regulatory programmes and outperforms deep neural network–based methods on every predictive metric tested, while remaining fully interpretable at the level of individual regulatory interactions.
(Upcoming research findings will be posted soon)
