Handle mlm and fingerprint data

[calebchin]

• In preprocess → separately mask out the string and write separate data to GCS component
  • Do the same for the fingerprints
• In training -> for each step, use the appropriate columns

Post preprocess GCS schema
smiles smiles_masked protein_name binds masked_target fingerprint

All modes are defined by the head

1. MLM mode:

• Takes in masked smiles string and tries to predict masked part of the string (masked_target)

2. Fingerprint

• Takes in regular smiles string and tries to predict ECPF vector

3. Classification

• Takes in smiles string and tries to predict bind

Double check in deployment that it uses the classification model.