Installation

 mamba env create -f environment.yml
 conda activate chemprop

Alternatively install from sourece

git clone https://github.com/chemprop/chemprop.git
cd chemprop
conda env create -f environment.yml
conda activate chemprop

You can download the data used for the models described here from the GitHub OPERA github page https://github.com/kmansouri/OPERA/

TO DO

• Cleaning up repo for chemprop workflow --> Carl

• Hyperparameter optimization --> Carl

• Comparison with Opera models (use same splits for STL) --> Enrico

• Varying the number of tasks in the multitask learning --> Enrico main (Riccardo Carl supporting)

• SAMPL challenge benchmark prediction comparison (check if no overlap OPERA-SAMPL) --> Riccardo

• Make notebook to make the main plot to use as a template for all the following tasks --> Enrico + Riccardo

• Try to do PCA/clustering to select very few datapoints to decide what models will give best performance --> Riccardo at first then Enrico

• Assess effect of variance when taking small subset --> Carl

• Assess effect of training set size --> Domen

Poster: https://www.overleaf.com/3869273436fsspfnqhngvr#806db5

• Title: "Deep learning with 20 datapoints"?
• Abstract (inlcuding motivation)
• Methods
• Results
  • External dataset plot
• Conclusions with perspective outlook
  • All the ideas that we don't have time to try

Meeting:

• 23.05.24 1pm - 2pm First sharing of results

  • Comparison with SAMPL: both MTL and STL give very good results, molecules in SAMPL6 are very similar to training data
    • (for the report) get molecules from the training set that are closest to SAMPL6 molecules
    • (for the report) See how well Opera represents a more "general" chemical space by comparing the PCAs
    • (for the report) Find a dataset that IS different? (so maybe MTL will work better than STL)

• 28.05.24 1pm - 5pm Making the poster

Deadline: May 30th 2024 1:45pm (poster session)