S²VM self-supervised learning from large unlabeled drug pairs, achieving SOTA DDI prediction and superior generalization/interpretability on novel/few-shot drugs.