LMLF - Language Model with Logical Feedback

This folder contains 1 utils folder, 1 jsonl file and 1 Python file for the LMLF procedure.

utils

This folder contains 3 utility files for the LMLF procedure.

• json_utils.py: Contains utility functions for reading and writing JSON files.
• LLM.py: Contains utility functions for the LLM queries and prompts.
• properties.py: Contains utility functions for the molecules' properties. I have included a dummy function with a RNG to fetch docking scores in case users don't have Gnina installed.

inhibitors.jsonl

This file contains the material properties in JSONL format. Each line is a JSON object with the following keys:

• composition: The composition of the material.
• dielectric_constant: The dielectric constant of the material.
• label: The label of the material.

lmlf.py

How to run the code?

• Change the directory to the current folder.
• Add in your OpenAI API Key in the lmlf.py file. (Line 8)
• Add absolute path of the inhibitors.jsonl file in the lmlf.py file. (Line 9)
• Enter whatever constraints you wish to optimize the molecules for. (Line 28)
• Enter initial constraints. (Line 31)
• Run the following command:

python3 lmlf.py

This file contains the main LMLF code. It uses the utils to create a feedback loop between the LLM and the material properties.

NOTE: The results will not be updated in the main inhibitors.jsonl file. A new folder is created with each run of the LMLF procedure. The folder is named inhibitors_<timestamp> and contains the updated inhibitors.jsonl file.

The LMLF procedure is as follows:

1. Initialize constraints and background knowledge. This includes the molecular properties that we want to optimize and some known molecules + properties.
2. Load up molecules' properties from inhibitors.jsonl.
3. Initialize the LLM.
4. Randomly sample a molecule with label=1 from inhibitors.jsonl and query the LLM to generate a similar but novel molecule which satisfies the constraints.
5. Get the properties of the generated molecule using utils.
6. Update the molecules' properties in inhibitors.jsonl with the generated molecules.
7. Update constraints, background knowledge and revalidate the database, i.e. inhibitors.jsonl.
8. Repeat steps 4-7 for a number of iterations.