evaluate RDkit and chemprop models

[kaichop]

Start from the antibiotic prediction paper and see whether we can reproduce a similar approach of building a prediction model using chemoprop, and then add features extracted from RDkit from all molecules.

[wangwpi]

I'm going to experiment the parameters of morgan fingerprint (radius and bit) to see if we could find the best radius and bit for the morgan fingerprint. This is a good tutorial for the fingerprints from the kaggle challenge: https://www.kaggle.com/code/towardsentropy/fingerprint-tips-and-tricks

[Euchiz]

I've submitted the results from a chemprop-base-1v5 model with 0.388 public score. Still running a similar version that adds building blocks into features. Code uploaded

[wangwpi]

I tested the simple NN model using new morgan fingerprint (radius=4, bits=2048), however it only achived 0.343 public score, which is even lower than using simpler morgan fingerprint (radius=3, bits=1024). We are hitting the ceiling of simple NN model and morgan fingerprint. We could just move on for the GNN or transformer model.

[Euchiz]

I've submitted the results from a chemprop-base-1v5 model with 0.388 public score. Still running a similar version that adds building blocks into features. Code uploaded

I've submitted the results from a chemprop-base-1v5 model with 0.388 public score. Still running a similar version that adds building blocks into features. Code uploaded

Just submitted a 1:1 model with building-block features. Got 0.410 score. Still trying higher ratio (a bit difficult as it requires more ram in one run. considering training on small subsets consecutively).