Note: This project was done by Saehui Hwang, Archie Shahidullah, and Liam Silvera. We placed second in the Bioinformatics track.
This project analyzes gene expression in response to treatments thought to mimic a COVID-19 inflammatory response in cells. We use a deep autoencoder to generate embeddings for gene expressions, which are then used to create a network in VIP.
The data were generously provided to us by Matt Thomson. It consisted of transcription-based gene expression values for ~47000 cells along ~37000 genes. This was later reduced by filtering. The structure of the raw data is shown below.
After reducing the gene expression vectors (each row) to 128-dimensional latent encodings with our autoencoder, we performed clustering in Virtualitics' VIP software to identify trends for future investigation.
An example of a graph indicating cell response to the drug Alprostadil is below.
We concluded that the drug Dasatinib consistently clustered close to the cytokine controls, indicating a lack of significant gene expressions to this drug. It was also noted that Myeloid cells produced markedly different responses for all drugs versus T and B cells. It is noted this project only indicates a response to drugs to mark certain treatments for future study, not whether this response is an effective immune reaction to COVID-19.