Cross-Modal Text-Molecule Retrieval

This repository contains the code for our paper “Towards Cross-Modal Text-Molecule Retrieval with Better Modality Alignment” (BIBM 2024 regular paper).

Our implementation is built on the source code from text2mol, MoMu-GraphTextRetrieval and ACME. Thanks for their work.

Dataset

We use ChEBI-20 dataset from text2mol to conduct the main experiment and PCdes dataset from KV-PLM to conduct comparison with pretrain-finetune paradigm based models.

You need to download the ChEBI-20 dataset from text2mol and put it in the data_dir.

How to Run?

To train and test our model, you can simply run:

bash scripts/train.sh

The model is tested after 60 epochs have been trained, so you can get the results of the text-to-molecule retrieval.

To finetune a trained model on kv_data with paragraph-level and testing:

bash scripts/finetune_para.sh

To finetune a trained model on kv_data with sentence-level and testing:

bash scripts/finetune_sent.sh