No. of Recommendations: 6
Hi again, everyone,
This is the last thing I will post here on this (I promise) but last week I published the results of my first ML experiment: Training a graph convolutional network to rank drugs for usefulness to a given disease. And my results are better than random! It's still a little like watching magic, but I am starting to understand more. This is the first time I have ever personally reproduced the results of a scientific paper, neat!
https://whafa.substack.com/p/phase-0-gp-kg-reproMy work for the next few iterations is clear-cut: I am trying to fully repro the results of the more recent paper (published just a few weeks ago), which uses a different kind of neural net architecture to do the same thing with better results. After that, I adapt everything from Alzheimer's Disease (AD) to Colorectal Cancer (CRC), personalize the loss function to Maureen's specific genetic data, and then I will be out of road -- there are not even dirt paths past that point, I need to validate everything myself because nobody else has even tried to produce an answer. I have already compared the CRC drug rankings with her own phenotypic data (from RGCC assay), and they agree enough to give me confidence this process maps to cancer.
So I could really use your brain, if you are into this kind of thing. Please feel free to subscribe to my substack and offer any useful input into this process! Just never tell me the odds. Thanks.
