Text this: PreMode predicts mode-of-action of missense variants by deep graph representation learning of protein sequence and structural context