Data augmentation for improving multi-step enzymatic reactions analysis

Abstract

In the field of synthetic biology, the application of computer-based approaches can accelerate the design-build-test-learn cycle. For example, utilizing models that predict the yield of reaction products proves highly practical, guiding subsequent optimal experiments. However, obtaining biological experimental data is time-consuming and costly, often insufficient for training machine learning models. This study addresses the challenge of data scarcity by combining masking pretext tasks of self-supervised learning with prior knowledge of multi-step enzyme reactions to augment experimental data. The synthetic data exhibits statistically similar characteristics to the original data, enhancing the performance of various enzymatic reactions analysis tasks. Consequently, this data augmentation technique is expected to be valuable in overcoming data scarcity issues in the field of synthetic biology and life sciences.