search

gitter-lab / nn4dms

Neural networks for deep mutational scanning data
MIT License
66 stars 16 forks source link

GCN Question: How is the whole graph information combined to make predictions? #2

Closed jrhorne closed 3 years ago

jrhorne commented 3 years ago

Hello! Awesome work!

I had a quick question from the manuscript where it states

"The final output of the network is computed at the graph-level with a single function score prediction for the entire graph."

— how exactly is the graph's node information combined into the final prediction? I tried looking in this repo, but I'm not familiar with TF v1 to know where to look. Is the node information all combined (added, averaged, concatenated, etc.) then connected to the final dense layer?

Sorry if I just missed this somewhere; thanks!

samgelman commented 3 years ago

Hi -- thanks for your interest in this work!

The node information is concatenated / flattened before being fed into the final dense layer. The GCN architectures are defined in the network_specs directory and the actual GCN layer is implemented in my_pipgcn.py.