divnori/protac-design

UROP Project @ Coley Group

Protac-Design

Description

This repo contains the code behind our workshop paper at the NeurIPS 2022 AI4Science Workshop, Link to Paper. It is organized into the following notebooks:

• surrogate_model.ipynb: Contains the code for processing the raw PROTAC data and training the DC₅₀ surrogate model. Note that you will need to download the public PROTAC data from PROTAC-DB in order to reproduce the results.
• molecule_metrics.ipynb: Contains code for computing metrics on a set of generated molecules. Metrics include percentage predicted active, percentage of duplicate molecules, percentage of molecules regenerated from training set, average number of atoms, chemical diversity, and drug-likeness.
• binary_label_metrics.py: Contains useful functions for analyzing performance of binary classification models.

Then there are additional files in the repo:

• surrogate_model.pkl: Contains the pre-trained surrogate model for DC₅₀ prediction.
• features.pkl: Contains list of features used in surrogate model training; required to reproduce reinforcement learning jobs using protac scoring function.

Instructions

1. Before running any of the notebooks, you will need to download the PROTAC data from the public PROTAC-DB database.
2. You will then need to create a conda environment containing the following main packages: rdkit, pandas, sklearn, scipy, ipython, and optuna. See the instructions in the next section for setting this up.
3. Open the notebooks on your favorite platform, and make sure to select the right kernel before executing.

Environment

To set up the environment for running the notebooks in this repo, you can follow the following set of instructions:

conda create -n protacs-env -c conda-forge scikit-learn optuna rdkit
conda activate protacs-env
conda install pandas scipy 

Citation

Nori, Divya et al. (2022) "De novo PROTAC design using graph-based deep generative models." NeurIPS 2022 AI4Science Workshop.

Additional data

Additional data, including saved GraphINVENT model states, generated structures, analysis scripts, and training data, are available on Zenodo here.

Authors

• Divya Nori
• Rocío Mercado

protac-design's Issues

GraphInvent -finetune issue

Hi,
Its a outstanding work regarding protac. I just went through this code, i faced some difficulty when execulting GraphInvent.

So during finetune part. I have to specify these(mention below which is in comment) in constants.py file.
"""
scoring_model_activity = pickle.load(open('/home/gridsan/dnori/GraphINVENT/data/protac_scoring_models/Protac_Scoring_Model_1024_100nM.pkl', 'rb'))
scoring_model_structure = pickle.load(open('/home/gridsan/dnori/GraphINVENT/data/protac_scoring_models/Protac_Scoring_Model_Structure.pkl', 'rb'))
with open('/home/gridsan/dnori/GraphINVENT/data/protac_scoring_models/features_1024_100nM.pkl','rb') as fp:
features_activity = pickle.load(fp)
with open('/home/gridsan/dnori/GraphINVENT/data/protac_scoring_models/features_structure.pkl','rb') as fp:
"""

Question 1. How can i get these files or can i use fetaure.pkl and surrogate.pkl files here ? if yes then for which pkl to use
Question 2. How can i get protac _activity result that i s mention in scoring_funtion.py file?
Question 3. after running finetune , in return i am getting files like agent.smi , BASF.smi , .valid and .likelihood , So what are those ?