Grant: Travel Awards - Institute Community Support

Genes encode information to form proteins. Expression of genes as proteins results in observable biological properties. Studying how all the genes in our genome are expressed in different tissues and diseases can help us find better treatments and to understand our own biology. However, we have over 19,000 protein-coding genes in our genome, and understanding the behaviour of all of these genes at the same time is a challenging task. This makes it particularly difficult to study rare cancers and to understand the biology of aggressive cancers that have not been studied rigorously through molecular methods. The encoded information in every gene is translated into protein via multiple readouts called transcripts. These transcripts can be used to quantify the amount that each gene is expressed. We taught a neural network to take the expression values for all the genes (as the transcriptome) in various cancers and healthy tissues, and compress them into a smaller set of values. We then decoded the whole transcriptome back from this compressed set, and tested which genes were de-compressed accurately and which ones were not. We found that genes that are known to be essential for cells to survive, are learnt well, whereas obsolete genes and genes with minimal value in cellular development are learnt poorly. This suggests that our method is able to automatically learn which genes are relevant to the biology of individual cancers and healthy tissues, and discards the genes that do not play any important role in this dataset. Our results show potential in studying rare, advanced cancers and dissecting their biology using machine learning. This will eventually help us find better, targeted cures for the same in a short span of time.