Request for Guidance: REINVENT4 + MOL2MOL + Transfer Learning + QSAR + Docking

[bishalbudha]

Hello everyone,

I am quite new to programming, and I only have basic experience with Linux, Conda, and command-line tools. Previously, I read some literature about REINVENT4 and even proposed my dissertation project based on it. However, despite spending the last two months studying tutorials and documentation, I have not been able to make significant progress. My submission deadline is approaching while most of my colleagues are already preparing for their defense. I am very worried about possibly losing a year in my graduation timeline. I would be extremely grateful if anyone here could help me with some of my questions and confusions.

My case:

1. My target is EGFR.
2. I have already developed a machine-learning based QSAR model, which is saved as a .pkl file.
3. I also have SMILES files containing ~300 inhibitors for EGFR.
4. I would now like to perform transfer learning using MOL2MOL inside REINVENT.
5. Unfortunately, I have not found many detailed tutorials, so I am reaching out to the community for help.

My questions:

1️⃣ After cloning the GitHub repo and setting up the environment, I’m stuck.
What steps should I follow next? How many TOML files do I need to create? Which TOML files are mandatory, and how should I configure them?

2️⃣ There are multiple pretrained models in MOL2MOL. Which model should I prefer for transfer learning? Is there any pretrained model that allows generation without applying benchmark similarity thresholds?

3️⃣ I want to integrate my ML-based QSAR model into reinforcement learning. How can I use the scores from my .pkl model inside REINVENT4 for RL?

4️⃣ I want to combine multiple scoring functions during RL: QED, SA, QSAR score (from my model), and docking score. How can I integrate all these scoring components together? How should I generate the corresponding RL file? What is the inbuilt docking protocol inside REINVENT4 for RL? Is any additional setup needed for docking?

I have tried my best to study and decode the code in this repository, but I could not proceed further on my own. Unfortunately, I don't have local guidance or supervision for this project, so I am staying optimistic that the community here might help.

Thank you very much in advance!

[halx]

Only briefly,

Example configuration files are in config/toml. TL does not make much sense for Mol2Mol as that is used for molecule optimisation and you should have an idea at this stage which molecules you would want to progress as leads. You don't say what code has been used ib creating you QSAR model. For combing multiple scoring see the example configuration files. Use the prior with ecfp and pubchem in the filename.

[bishalbudha]

Thank you halx!
I will be seeking more help and clarifications.