Deep learning modeling of spatial biology data for GRN inference of cancer

Background:
Cancer systems biology is based on applying data-driven methods to understand the complex cellular systems that may cause cancer when the healthy regulatory wiring becomes aberrant. Understanding the regulatory mechanisms underlying cancer development is crucial to identify new target genes and cancer therapies. This project proposes to harness the increasingly available spatial biology data using a new AI-based approach to construct gene regulatory networks (GRNs). By comparing GRNs between different stages of cancer development, causative shifts in the regulatory circuitry can be detected, and such data-driven insights can lead to new cancer therapies. A better knowledge of the human regulatory interactions, the “regulome”, is paramount to address the challenge of cancer, for which more than 19.3 million new cases are diagnosed annually, causing 10 million deaths worldwide.

Spatial biology, voted Method of the Year by Nature in 2021, has revolutionized researchers’ ability to relate cells’ transcriptomes to their tissue context and phenotype. Due to this unique capability, the popularity of this data-driven approach has steadily increased during the last 10 years, and it is now a mainstream omics platform. Around 10,000 papers from the last decade are found by searching ‘spatial transcriptomics’ in PubMed. A large body of Spatial Transcriptomics (ST) datasets has already been amassed. Public databases like STOmicsDB, SOAR, SpatialDB, DeepSpaceDB, 10xGenomics, and Broad Institute’s Single Cell Portal are making a vast number of spatial samples available, currently over 20,000, mostly from human and mouse. The upward trend of ST will result in many more datasets in the near future.

Understanding the genetic mechanisms underlying cancer formation is crucial to identify new target genes and cancer therapies. Systems biology approaches have been applied to cancer-derived omics data and comparing GRNs between different subtypes of the cancer or to healthy tissue. To unravel the precise regulatory influences between the components in a dynamic system, it is necessary to build detailed quantitative GRN models using gene expression data, which may be complemented with other omics data such as chromatin accessibility. The best technology for reconstructing such GRNs are data-driven approaches based on gene expression responses to system perturbations1. In a breakthrough paper2, the Sonnhammer group showed that by using the perturbation design in the GRN inference algorithm, far superior accuracy is obtained.

Project description:
To obtain accurate GRNs, the student will combine expression data and perturbation design using deep neural network (DNN) technology. Notably, previous DNN approaches to GRN inference have not incorporated the perturbation design, which is crucial to obtain high GRN inference accuracy. The project will start by employing existing DNN-based GRN inference packages like SFINN, GCNG, and SCING3–5 to predict GRNs. Then adaptations to use the perturbation design will be explored, both by adding a loss term during training and by using different input layers depending on whether the gene is perturbed or not. The perturbation-augmented DNN will be benchmarked against traditional perturbation-based methods.

Gene perturbations using CRISPR are not possible in spatial data, but the perturbation design can be reconstructed based on the principle that if a gene’s expression is not changing significantly between conditions, then it is not perturbed. The single-cell like spatial transcriptomics data together with the inferred perturbation design will be collected per region, where each region represents a cancer stage, and used as input to a deep neural network (DNN) model outlined below in order to infer a reliable GRN for each region.