Improving genome-scale metabolic models of incomplete genomes with deep learning
02:04 Share

Link to bioRxiv paper: http://biorxiv.org/cgi/content/short/2023.07.10.548314v1?rss=1

Authors: Boer, M. D., Melkonian, C., Haas, A. F., Zeiforopoulos, H., Garza, D., Dutilh, B. E.

Abstract: Deciphering the metabolism of microbial species is crucial for understanding their function within complex ecosystems. Genome-scale metabolic models (GSMMs), which predict metabolic traits based on the enzymes encoded in a genome, are promising tools to study microbial ecosystems when genome sequences can be obtained. However, constructing GSMMs for uncultured bacteria is challenging, as metagenome-assembled genomes (MAGs) are typically incomplete, leading to metabolic reconstructions with numerous gaps. Existing methodologies often fill these gaps with the minimum set of reactions necessary to simulate an objective function such as growth. Here we introduce an artificial intelligence-based alternative: the Deep Neural Network Guided Imputation Of Reactomes (DNNGIOR). The DNNGIOR neural network learns weights for missing reactions in incomplete GSMMs from genomes spanning the bacterial domain. We demonstrate that reactions that occur in greater than 30% of the training genomes can be accurately predicted (F1 score = 0.85). The weights generated by the DNNGIOR neural network improved the gap-filling of incomplete GSMMs, when assessed on a large and phylogenetically diverse testing dataset and on a small set of high-quality manually curated models. The accuracy of DNNGIOR was on average 14 times greater than the standard unweighted gap-filling for draft reconstructions and 2-9 times greater for manually curated models. DNNGIOR is available at https://github.com/MGXlab/DNNGIOR or as a pip package (https://pypi.org/project/dnngior/).

Copy rights belong to original authors. Visit the link for more info

Podcast created by Paper Player, LLC