While molecular pre-training has shown great potential in enhancing drug
discovery, the lack of a solid physical interpretation in current methods
raises concerns about whether the learned representation truly captures the
underlying explanatory factors in observed data, ultimately res