@inproceedings{nam-etal-2026-fact,
    title = "{FACT}: Functional Group Alignment and Consistency in Token Space for Structure-aware Molecular Representation Learning",
    author = "Nam, Hyeonyeong  and
      Choi, Woojae  and
      Lee, Deok-Joong  and
      Son, Young-Han  and
      Lee, Sangwoon  and
      Kang, Bogyeong  and
      Jo, Eunjung  and
      Kam, Tae-Eui",
    editor = "Demner-Fushman, Dina  and
      Ananiadou, Sophia  and
      Roberts, Kirk  and
      Tsujii, Junichi",
    booktitle = "{B}io{NLP} 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.bionlp-1.56/",
    pages = "695--703",
    ISBN = "979-8-89176-434-7",
    abstract = "Molecular representation learning aims to capture chemically meaningful structures for various downstream tasks such as accurate molecular property prediction. However, incorporating functional group (FG) information into SMILES-based models remains challenging. The absence of explicit alignment between graph-defined FG atom sets and tokens in sequence prevents complete substructure masking, while multiple valid SMILES forms of the same molecule lead to inconsistent FG representations in token space. To address these challenges, we propose FACT (Functional Group Alignment and Consistency in Token Space), an end-to-end framework for structure-aware SMILES-based representation learning. FACT introduces an atom?token alignment module for complete FG span masking during pre-training and enforces FG consistency across different SMILES forms during fine-tuning. Experiments on MoleculeNet benchmarks show that FACT achieves state-of-the-art or competitive performance on eight tasks, demonstrating the effectiveness of alignment and consistency learning for molecular representation."
}