For medical and biotechnological uses, protein design has grown in significance.The development of a novel protein necessitates laborious and time-consuming computational or experimental techniques since proteins are formed by complex systems.At the same time, machine learning has made significant advancements in the field of generative modelling in recent years, enabling the solution of complicated problems by utilising vast amounts of available data.
In this research, the authors have taken a glance on the issue of designing all-purpose proteins based on the hierarchical Gene Ontology’s functional labels.They have developed an evaluation strategy using a number of metrics with biological and statistical inspirations since there isn’t a conventional technique to assess generative models in this field.
The conditional generative adversarial network ProteoGAN is then created, and it is demonstrated that it outperforms various established and more current deep-learning baselines for generating protein sequences.Through the examination of hyperparameters and ablation baselines, we provide further insights into the model.Finally, the suthors have proposed a functionally conditional model that, by combining labels, could produce proteins with novel functionalities and offer initial steps in this study field.
The Code and the data are available at the following URL: https://github.com/timkucera/proteogan
Reference
Kucera Tim et.al.(2022)Conditional generative modeling for de novo protein design with hierarchical functions. Bioinformatics 38(13):3454-3461.