Аннотация:Computational enzyme design is an ambitious but still not overcomed challenge. Recently massive combinatorial virtual screening of single point mutations successfully yielded a variant of Ig scavenger with performance increased by two orders of magnitude. However, scaling this ap-proach to even double mutations is unfeasible because of exponentially growing combinatorial space. One of the possible approaches to address this issue is by dissecting the internal logic in the data from single mutagenesis study and then applying it to predict the effect of multiple simultane-ous mutations on enzyme activity without explicitly calculating it. Methods of machine learning are perfectly suited for this task. Though some efforts were made to predict enzyme activity based on its sequence, enzymatic activity is directly manifested on the level of structure with crucial role of electrostatics. That is why we believe that structure-based descriptor is more suited for this task. What is more, it acts on the level of chemistry and is agnostic of sequence and alphabet, thus allow-ing to capture effects of various cofactors such as metal ions and to study non-homologous enzymes with similar function. In this work we intended to test whether modern architectures of deep convolutional neural networks (CNNs) can be used to correctly predict mutation effect on enzyme’s catalytic activity uti-lizing only 3D structures as input data. Unfortunately, to date there is no large enough dataset link-ing specific enzyme variants’ structure and activity; thus we generated 192 variants of diiso-propylfluorophosphatase. This enzyme was selected as a model system because corresponding reac-tion comprises only one chemical step and is well-studied. Single amino acid substitutions were in-troduced to all positions in 12Å radius from the active center. For each variant we computed reac-tion barrier with the help of hybrid QM/MM metadynamics. Machine learning was performed with recently developed unusual spherical CNN architec-ture, and proteins’ structures were represented as a grid of voxels comprising information about all atoms’ radius, charge and mass. We achieved accuracy over 60% in the three-class classification test (worse, unchanged, bet-ter). Though it is only the initial step, we showed applicability of deep spherical CNNs to computa-tional enzyme design. This study was supported by RFBR Grant#17-54-30025. Calculations were carried out using the equipment of the shared research facilities of HPC computing resources at Lomonosov Moscow State University supported by the project RFMEFI62117X0011.
