Closed agitter closed 1 year ago
It is very hard to train our model even for a single plate. I have tried different learning rates but both loss
and val_loss
are not changing at all. The val accuracy
is always 0.
We might want to reconsider this idea.
What we see in the UMAP images and hierarchical clustering is consistent with this result. Many compounds may have a minimal effect on the cells.
It may instead be more realistic to focus on the few compounds that do affect the cells in a unique way. We could try to predict those outliers based on the compound features, which would be more similar to traditional virtual screening. Here, however, we could also tune how the network represents the images and the clusters as we train a mapping from compound feature to effect on the cells.
Can try predicting the compound from the image. Initially start with a single plate again before deciding whether to scale up.