Infusing theory into deep learning for interpretable reactivity prediction