A little diversion from science: I bought myself a kindle recently and thought I'd try to draw something on it (my wife decided to chip in too):
Sumit Mukherjee
- 15 Followers
- 255 Following
- 9 Posts
Staff ML scientist @insitro. Ex-Microsoft AI4Good. UW ECE PhD. Interested in ML, Statistical genetics, PPML, AI ethics. Biriyani expert. Views are my own.
It's not my best work but I thought it was kinda fun to imagine a very illogical scenery.
bioRxivMulti-ContrastiveVAE disentangles perturbation effects in single cell images from optical pooled screens https://www.biorxiv.org/content/10.1101/2023.11.28.569094v1?med=mas
A poster version of this work was recently presented at #ASHG23 in case someone wants a quick summary.
In this work we take a systematic pass at this by evaluating phenotypes along two axes: i) heritability, ii) relevance to specific diseases. We achieve these using common tools used in genetic discovery, making our workflow simple and easy to build on.
Prior to this pre-print, there have been several papers that have used embeddings for genetic discovery. However, there isn't any papers (to our knowledge) on evaluating their utility systematically for the task of disease specific genetic discovery.
Link to github repo: https://github.com/insitro/EmbedGEM
GitHub - insitro/EmbedGEM: EmbedGEM: A framework to evaluate the utility of embeddings for genetic discovery
EmbedGEM: A framework to evaluate the utility of embeddings for genetic discovery - GitHub - insitro/EmbedGEM: EmbedGEM: A framework to evaluate the utility of embeddings for genetic discovery
Are you interested in doing genetic discovery on ML derived embeddings but don't know how to interpret the results or compare them against other phenotypes? Well, look no further. Read my latest paper with my colleagues at insitro presenting the first formal framework to do this: https://www.biorxiv.org/content/10.1101/2023.11.24.568344v1