Expanding functional protein sequence spaces using generative adversarial networks

De novo protein design for catalysis of any desired chemical reaction is a long-standing goal in engineering because the broad spectrum technological, scientific and medical applications. However, mapping sequence to function currently neither computationally nor experimentally tangible. Here, we develop ProteinGAN, self-attention-based variant generative adversarial network that able ‘learn’ natural diversity enables generation functional sequences. ProteinGAN learns evolutionary relationships sequences directly from complex multidimensional amino-acid space creates new, highly diverse variants with natural-like physical properties. Using malate dehydrogenase (MDH) as template enzyme, show 24% (13 out 55 tested) ProteinGAN-generated tested are soluble display MDH catalytic activity conditions vitro, including mutated 106 substitutions. therefore demonstrates potential artificial intelligence rapidly generate proteins within allowed biological constraints space. A protein’s three-dimensional structure properties defined by its sequence, but intensive task. new approach generates variations, which tested.