CrossMol: Cross-Modal Mask-Predict Pre-training For 3D Molecular Data.

CrossMol achieves state-of-the-art performance on molecular tasks by effectively integrating 3D structures and SMILES modalities through cross-modal mask-predict pre-training.

• Why it matters: Understanding the complex relationships between different molecular data modalities is crucial for accurate molecular modeling, yet current methods treat these modalities as independent, limiting their effectiveness.
• What they did: The authors developed CrossMol, a pre-training model that uses a mask-predict approach to fill in missing 3D structural information from semantic data like SMILES, incorporating a reweighted distance prediction loss for enhanced structural modeling.
• The result: This approach significantly improves downstream task performance, enabling more precise molecular understanding and advancing the development of molecular analysis tools.