Biological context

Tumor plasticity is becoming a critical issue for caring patients suffering from cancer. Most of the anti-cancer treatments currently fail due to the re-emergence of resistant tumor cells that adapted to those treatments. Recent studies using single cell approaches provide a better characterization of the different cell states that compose the tumor bulk and reflects this plasticity. In glioblastoma (GB), inter- and intra-individual heterogeneities are now better understood with the observation of 4 GB cell states [Neftel, 2019]. Each GB state (OPC-like, NPC-like, AC-like & MES-like), characterized by a transcriptomic program, could evolve to other cell states. The opening question is to understand how cell states could be associated with GB features. Indeed, GB enriched in MES-like cells were correlated with the abundance of infiltrating immune cells, suggesting that those patients would be good responders for immunotherapies. Our recent study shows that inhibition of IRE1, an endoplasmic reticulum stress sensor, impacts on GB plasticity by inducing a stem/progenitor state.

Bioinformatics objectives

The available data include:
(i) public GB transcriptome datasets based on different technologies: microarray, RNA-Seq or single cell RNA-Seq (scRNA-Seq). These datasets include samples from patients (bulk tumors).

(ii) home-made GB transcriptome datasets based on different technologies: microarrays (Agilent, Illumina) or RNA-Seq. These datasets include samples from cell lines cultured at the laboratory and submitted to drugs developed at the laboratory (IRE1 inhibitors) + control conditions.

(iii) paired proteome datasets for the GB cell lines cultured at the laboratory (treated conditions + control conditions)

(iv) molecular signatures defining GB cell states [OPC-like, NPC-like, AC-like & MES-like] or IRE1 activity [high/low]. The 4 GB cell states were obtained through the analysis of single cell RNA-Seq of bulk tumors and the transcriptome signatures associated with IRE1 activity were previously described by our group (36 marker genes).

As these molecular signatures are based on gene expression, they can be used to characterize any GB sample with transcriptome data available (microarray or RNA-Seq). The first objective of this proposal is therefore to use deconvolution methods (e.g., the sigScores function of the scalop R package [https://rdrr.io/github/jlaffy/scalop/] and the EcoTyper machine learning framework [https://github.com/digitalcytometry/ecotyper]) to determine the cell state composition of the different GB samples. Likewise, these samples will also be classified according to IRE1 activity using the transcriptome signature previously described by our group (36 marker genes).

The second aim will be to correlate cell state composition with IRE1 activity. One major question here is to address whether to analyze the different datasets separately or integrate them in a meta-analysis (at least for the datasets based on a same technology) using a home-made pipeline developed by a former M2 bioinfo student.

Finally, the third objective is to investigate the possibility to identify specific protein markers associated to each cellular state using Machine Learning approaches (caret R package) or discriminant analysis applied to combined transcriptome and proteome datasets.