Summary
A bioRxiv preprint introduces SPECTRA, a graph-based model that predicts gene-expression responses to CRISPR perturbations by propagating signals through gene regulatory networks. The authors report improved recovery of sparse perturbation-related changes in a large-scale single-cell CRISPRi benchmark.
Researchers at Human Technopole have introduced SPECTRA, a graph-based machine-learning model for predicting how cells change their gene activity after CRISPR perturbations. The work, posted as a bioRxiv preprint on September 22, 2026, represents a perturbation as a local signal in a gene regulatory network rather than as an opaque shift between an unperturbed and perturbed cellular state.
The authors report that this approach can improve recovery of the sparse gene-expression effects caused by perturbations in a large-scale single-cell CRISPR interference (CRISPRi) benchmark. The study is a computational model evaluation, not a clinical or therapeutic study.
Turning a CRISPR perturbation into a network signal
A transcriptome is the collection of RNA molecules produced by a cell at a given time. Single-cell transcriptomics measures this activity separately in individual cells, making it possible to examine how cells respond to the same intervention. In perturbation experiments, however, the resulting expression changes can be difficult to identify because only a small subset of genes may respond strongly, and the effectiveness of a perturbation can vary between cells.
SPECTRA—short for SPEctral CRISPR Transcriptome Regulatory Autoencoder—uses a gene regulatory network to model these responses. In such a network, genes are represented as nodes and directed relationships describe regulatory influence between them. The model injects the CRISPR perturbation as a localized signal and propagates it through the network with directed graph neural networks. This design is intended to connect the targeted gene with downstream expression changes in a way that remains biologically interpretable.
The system combines a variational encoder for control cells with a directed graph decoder. It also integrates single-cell expression data, pretrained transcriptomic context and prior regulatory-network structure into a node-level representation. In contrast, the authors describe many existing approaches as encoding an intervention mainly as a latent shift from an unperturbed to a perturbed state. That can capture a response without making the affected regulatory route easy to inspect.
How the model was evaluated
The researchers evaluated SPECTRA on a large-scale single-cell CRISPRi benchmark. The main task was recovery of differentially expressed genes (DEGs)—genes whose measured expression changes after a perturbation. Performance was assessed with precision, F1 score and area under the precision-recall curve (AUPRC), metrics that capture how accurately the model identifies the relatively small set of genes expected to respond.
The team also examined whether the model's interpretable predictions agreed with established biology. They tested cases involving gene knockouts whose effects had been demonstrated repeatedly using orthogonal experiments, meaning methods independent of the prediction system itself. The authors report that graph-based signal propagation can improve recovery of sparse perturbation-induced transcriptional effects and describe it as a biologically grounded alternative to latent-shift modelling.
The supplied preprint abstract does not provide the benchmark's sample size or numerical values for the reported precision, F1 and AUPRC results, so the magnitude of the improvement cannot be assessed from the available summary. The evidence currently consists of computational evaluation on single-cell perturbation data and comparison with known biological effects. How performance transfers to other cell types, perturbations or regulatory-network settings is not specified in the source.