Schizophrenia (SCZ) patients typically vary significantly in symptom severity. Despite numerous studies demonstrate SCZ is linked to brain structure abnormalities, relationships are obscure. In this paper, we establish relationships between structural abnormalities and symptom severity. All analyses are performed in two datasets (discovery: 326 SCZ and 298 normal control (NC); replication: 216 SCZ and 173 NC). We first build normative models in NC group, based on which we calculate atrophy values of cortical thickness, surface area, and gray matter volume in SCZ. Finally, we use atrophy values to predict symptom severity via generalized additive models and further evaluate the marginal effect of each structural feature. We found atrophy values could reliably predict symptom severity across two datasets (discovery: Pearson r = 0.29, P < 1 × 10^-5; replication: r = 0.26, P = 3 × 10^-5). Our findings could aid in understanding the pathogenesis of symptoms in SCZ.