genesis_wb_chem

Temporary repo for the small molecule models, data and workflows until access is granted to the main repo.

In this repo are 2 key chemical packages: chemprop and molxpt

chemprop

A chemical property prediction package popular among pharma cheminformaticians.
Ref: Chemprop: a Machine Learning Package for Chemical Property Prediction
This is the v2 followup to the original Cell paper discovering a novel and structurally diverse antibiotic halicin

Example NBs/workflows:

1. Load data: load example datasets (CSV) into Delta tables on Unity Catalog

Single-task training and inferencing

2. Chemprop: fit ClinTox classifier (single-task)
3. Chemprop: inference ClinTox: use ClinTox classifier from NB #2 to predict ClinTox properties of DrugBank

Multi-task training and inferencing

4. Chemprop: multitask training: do multi-task regression on 10 continuous endpoints from ADMET-AI
5. Chemprop: multitask inference: use multi-task regressor from NB #4 to predict 10 continuous endpoints on DrugBank

Inferencing using Chemprop example models (.ckpt) and CSV

6. Chemprop: inference reaction: use reaction regressor from chemprop to predict reactivity (Ea)
7. Chemprop: multicomponent reaction: use multicomponent regressor from chemprop to predict solubility of compound (component 1) in solvent (component 2)

---

molxpt

A multimodal model pretrained on text (PubMed) and molecular SMILES (PubChem) so you can prompt chemical generation tasks with natural language.
Ref: molXPT: Wrapping Molecules with Text for Generative Pre-training

Example NB:

Load molXPT for inferencing including text2molecule, molecule2text and molecular embeddings generation