In 2023, Open Problems hosted a competition with the aim of prediction how small modules affect gene expression profiles. See the Kaggle
competition page. However, the provided dataset contained serious biases, raised in [our previous report] (https://storage.googleapis.com/kaggle-forum-message-attachments/2559525/20057/OP2_Write_Up_JN_AP.pdf). In this study, we aimed at enhancing the quality of the perturbation data by addressing the potential biases. The notebook of the analysis can be found in the notebooks. The results of this analysis was contributed to post competition analysis report submitted to NeurIPS 2023 competition
track.
The analysis primarily focuses on the perturbation data, with an emphasis on identifying potential outliers. Extensive explanatory data analysis (EDA) was conducted on single-cell data, pseudobulked data, and the results of the differential expression (DE) analysis. Furthermore, the validity of cell type annotations was rigorously evaluated by cross-checking across donors, Leiden clusters, and biomarkers. Also, new cell type annotations were proposed using CellTypist, revealing higher consistency. Additionally, cluster-specific and compound-specific evaluations were carried out to pinpoint outliers. Subsequently, the single-cell data was pseudobulked and subjected to DE analysis using both the Voom+Limma and EdgeR pipelines
In addition, in the folder 'src/task/methods`, there is the clean implementation of deep NN model proposed by our team which ranked the 20th place in the competition.
| name | roles |
|---|---|
| Jalil Nourisa | author |