Deep Neural Network for perturbation-based Gene Regulatory Network inference

Goal: to develop a new computational algorithm based on a deep neural network for perturbation-based gene regulatory network (GRN) inference.

Background:
In systems biology the idea of networks to describe interactions between molecules is generally very popular and useful. For gene regulation a large number of methods to infer GRNs have been developed with many different approaches. GRNs act as a descriptive model of how a gene, regulator, affects the transcription of another gene, target. It does this via a graph based representation of the interactions where each node corresponds to a gene and each edge to a regulatory interaction. A key feature of GRNs is that unlike many other biological networks they are directed, meaning that while a regulator has an effect on the target, the target usually does not have an effect on the regulator. Furthermore, the GRN usually also describes the nature of the regulation, called sign, such that the GRN can be used to tell if a regulator activates or inhibits the expression of the target. All of these features together make the GRN a powerful tool when it comes to both understanding the basics of a cell as well as more advanced science like drug targeting or describing disease mechanics (Liu et al. 2019).

The Sonnhammer group has developed a number of algorithms and methods to improve the accuracy of GRN inference, and has applied these to e.g. identify new cancer therapy targets (Morgan et al. 2020). GRN inference in our group builds on the idea that with knowledge of the effect and target of a perturbation, the regulatory system can be reverse engineered from gene expression data that represent the system’s response to knock-down perturbations. In a recent breakthrough paper (Seçilmiş, Hillerton, Tjärnberg, et al. 2022), the Sonnhammer group showed that by using the perturbation design in the GRN inference algorithm, far superior accuracy is obtained (Fig. 1). We primarily rely on regression-based solutions for this problem, providing our methods with a number of advantages in that they can infer the relative strength, direction, and sign of a given regulatory interaction. Most of our tools are found in the GeneSPIDER package (Garbulowski et al. 2024).

In recent years the focus of the GRN field has moved towards single cell sequencing data as this data offers a few advantages. Foremost among these are that GRNs exist in different configurations for various cell types, allowing to capture a specific correct network for each cell type. In addition to this the high level of details available on expression in single cell data offers a better possibility at quantifying the expression of a given gene. Of particular interest for this project is the Perturb-Seq method, where a single gene can be knocked down in individual cells via CRISPR interference (CRISPRi), followed by measuring the resulting RNA-seq transcriptomic response per cell. In a recent study (Replogle et al. 2022), more than 8000 genes were perturbed this way in 2 million single cells.

Computational approach:
To obtain the most accurate GRNs, we propose to combine expression data and perturbation design using deep neural network technology from the AI field. Note that previous deep neural network (DNN) approaches to GRN inference have not incorporated the perturbation design, which is crucial to obtain high GRN inference accuracy. A suitable architecture of the DNN can be a variational (graph) autoencoder or a graph neural net. In both cases, it will incorporate the perturbation design as an input, and a special loss function will be designed to optimize the training. The project will explore a range of architectures, training procedures, and optimization techniques. Most of the work will be done in Python using e.g. PyTorch or Tensorflow. Benchmarking will be done with GRNbenchmark.org (Seçilmiş, Hillerton, and Sonnhammer 2022)