MOLAR Learns Multimodal Molecular Representations Despite Noisy Labels.

Predicting molecular properties often relies on labels derived from assays, databases, or weak annotation pipelines, which are frequently noisy. If models treat these recorded labels as perfectly reliable, they risk memorizing corrupted data and learning misleading molecular characteristics. This problem is exacerbated in multimodal molecular representation learning, where errors from noisy labels can spread across different data types, such as graph and text information, during fusion or alignment processes. To address this, researchers propose MOLAR, a framework specifically designed to learn multimodal molecular representations while accounting for noisy labels. MOLAR distinguishes between inferring the true, latent clean property and observing the recorded, potentially noisy label. It allows graph and text data to contribute independent evidence towards a clean-property distribution. A separate categorical label-observation channel then maps this inferred distribution to the recorded labels for training purposes. This innovative formulation enables the model to derive both the posterior reliability of labels and modality-specific molecular evidence. Empirical evaluations on benchmarks with naturally noisy molecular data, as well as controlled label-flipping scenarios, demonstrate that MOLAR consistently surpasses other representative baseline methods. Furthermore, visualization analyses show that MOLAR provides interpretable diagnostics regarding label reliability and the evidence contributed by each modality.