The article presents MolCA, a molecular graph-language modeling method which enhances Language Models’ (LMs) understanding of molecules. Traditional LMs lack a comprehension of 2D graph perception, which is essential for understanding molecular structures. MolCA addresses this by employing a cross-modal projector and a uni-modal adapter to connect a graph encoder’s representation space with an LM’s text space. This not only retains the LM’s ability of open-ended text generation but also enriches it with 2D graph information. The effectiveness of MolCA is demonstrated through tasks of molecule captioning, IUPAC name prediction, and molecule-text retrieval, where it outperforms the baselines.
Publication date: 20 Oct 2023
Project Page: https://github.com/acharkq/MolCA
Paper: https://arxiv.org/pdf/2310.12798
