molecular-representation

MOLAR proposes a noise-aware framework for learning multimodal molecular representations from noisy labels by separating clean-property inference from observed label noise, outperforming baselines on molecular benchmarks.