Testing

[iris-ywang]

So far we have PA-VAE up until 11340 drugs (see #1 ). The models are qsar-tested on 500 test samples.

• Aug 16 10:34 averaged_metrics_for_zinc_paired_encoder2_252_to_11340_test_size_500.npy

For SA-VAE, the models from 12600 - 201600 are all qsar-tested on 4500 test samples.

• Jul 25 12:07 averaged_metrics_for_all_zinc_encoder_iris2.npy
• pasted in excel: (local) SA_VAE model results.xlsx)

(the results not looking good)

To do:

• [x] Write the methods to check the decoded SMILES's similarity using Tanimoto similarity (jaccard_similarity on Scikit-learn)
• one using one-hot version of the smiles and the other using Morgan FP version
• [x] Run SA-VAE QSAR-Eval again on testsizes = 500, like for PA-VAE.

[iris-ywang]

Note on AE-similarity:

1. X_decoded is often non-integer. For each row (letter), the highest value/prob is taken as 1 and the rest of the values are put as 0.
2. X_one_hot_decoded might not be able to be translated back to SMILE because they just don't make sense. So I can't use MFP in RDKIT to compare input SMILES and output SMILES for a AE.