MOLAR: Learning Multimodal Molecular Representations from Noisy Labels
Researchers have introduced MOLAR, a novel framework designed to tackle the challenge of noisy labels in multimodal molecular representation learning. This approach disentangles the inference of clean properties from the observation of recorded labels, allowing graph and text modalities to contribute residual evidence to a clean-property distribution. MOLAR's formulation enables the derivation of posterior label reliability and modality-specific molecular evidence during training. Experiments on various benchmarks demonstrate that MOLAR consistently surpasses existing methods and provides interpretable diagnostics for reliability and modality evidence. AI
IMPACT This framework could improve the accuracy of molecular property prediction by addressing the common issue of noisy labels in training data.