PARM is a deep-learning model trained on data from massively parallel reporter assays to help predict promoter activity in different human cell types, design synthetic promoters and identify key features of regulatory promoter grammar.