LLM4Mol

Instruction to execution

1. Message (Prompt) template design.
   1. Message templates (code/prompt_template).
   2. Prompt templates (code/prompt_template).
2. Caption generation. (Optional)
   1. We apply pre-trained generative model to generate captions for molecules (smiles2caption).
   2. Generated captions are saved in input/caption.
3. Description generation. (Optional)
   1. We generate descriptions of molecule graph structure accompanied by atom features (rdkit.Chem).
   2. Generated descriptions are saved in input/description.
4. Predictor Message generation.
   1. Generated Predictor messages are saved in input/message.
   2. Generated prompts are saved in input/prompt.
   3. llm.template = IF(D): ask LLM to provide useful features for the task, (with given description).
   4. llm.template = IP(D): ask LLM to make predictions for the task, (with given description).
   5. llm.template = IE(D): ask LLM to make predictions for the task and provide explanations, (with given description).
   6. llm.template = FS(D)-3: given 3 example knowledge, ask LLM to make predictions for the task and provide explanations, (with given description).
5. Query LLMs for response.
   1. Considering the consistency and popularity, we use ChatGPT for now.
6. Make predictions based on LMs.
   1. Generated embeddings are saved as output/prt_lms/ogbg-molbace/IF/microsoft/deberta-base-seed42.emb
7. Make predictions based on GNNs.
   1. Generated predictions are saved as output/gnns/ogbg-molbace/gin-v-raw-seed42/predictions.pt

Notebook files in code/notebook are for demo test.

Example execution code

Generate Caption

Example:

python -m code.generate_caption dataset ogbg-molbace demo_test True device 0

Generate Description and Compress it using ChatGPT

Example:

python -m code.generate_description dataset ogbg-molbace demo_test True
python -m code.generate_compressor_message dataset ogbg-molbace demo_test True
python -m code.query_chatgpt dataset ogbg-molbace llm.template compress_des demo_test True

Generate Prompt

We no longer use it since generated prompts differ from SOTA messages for ChatGPT. We use Message (below) instead.

Example:

python -m code.generate_prompt dataset ogbg-molbace demo_test True

Generate Predictor Message

Example:

python -m code.generate_predictor_message dataset ogbg-molbace demo_test True

Query ChatGPT

Example:

python -m code.query_chatgpt dataset ogbg-molbace llm.template IF demo_test True

Run LM model to evaluate Text quality

WANDB_DISABLED=True TOKENIZERS_PARALLELISM=False CUDA_VISIBLE_DEVICES=0,1 python -m code.mainLM dataset ogbg-molbace data.text raw seed 42

Run GNN model to evaluate

python -m code.mainGNN dataset ogbg-molbace data.feature IF seed 42