Transformation and Innovation in Machine Learning Models: How Active Learning Workflow can Change the Way of Optimization in Genetic and Metabolic Networks

The development of digitization has turned the twenty-first century into a data-centric world, with AI and ML helping improve lives in every sector. Healthcare and biotech sectors are also adapting quickly to solve problems by seeking solutions in design and optimization, ranging from genomic studies, protein, enzyme, and metabolic engineering to solving complex genetic circuit design and optimization. The task is not simple as it first appears because there are few empirically labeled datasets and few programmers in the biotech sector.. So, it is necessary to approach this issue efficiently, which is less labor and cost-intensive, to bridge the gap between ML algorithms and their applications.

A recent publication in Nature Communication has succeeded in improving the current bottlenecks in the system. With the collaboration of INRAe Institute in Paris, they built a METIS application to democratize and standardize machine learning. METIS is the short form of Machine-learning guided Experimental Trials for Improvement of Systems, It is an active machine learning workflow with an optimized metabolic network that interactively suggests the next set of experiments after training on the previous set of experiments. It uses machine learning algorithms with customized adjustments according to user experience and experimental facilities with minimal datasets.

For the validation part, the researchers used the existing dataset from a recent optimization of an E. coli extract-based in vitro TXTL(Transcription-Translation) system. Using METIS optimization, the training datasets were reduced from 1000 data to a mere 10 data points, which showed similar yield optimization within 10 learning cycles. “Research found a model that is even less dependent on data,” says one of the study’s authors. To show the versatility of METIS on various biological systems, they improved a complex metabolic network called the CETCH (hydroxybutyric-COA) cycle. A CO2 fixation cycle comprising 17 different enzymes plus 10 cofactors is more efficient than natural photosynthesis like CBB(Calvin- Benson-Bassham) cycle. METIS let to improve the productivity of the CETCH cycle tenfold with only 1000 experiments by exploring combinatorial space of 1025 conditions making it the most efficient Co2 fixing in vitro system to date.

METIS is a straightforward, versatile active learning workflow that experimentalists can use with no experience in programming and knowledge of advanced computational skills. METIS can run on Google collab, a free online open platform to run python codes without the need to install or have local computational power. The notebooks can further be used by researchers for data-driven predictions utilizing the workflow. Apart from network optimization with minimal experimental datasets, it can also help discover unknown interactions and bottlenecks in these networks, paving the way for hypothesis-driven improvement. Also, this paves the way for the study, prototyping, and optimizing the system efficiently and systematically.

This Article is written as a summary article by Marktechpost Staff based on the research paper 'A versatile active learning workflow for
optimization of genetic and metabolic networks'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper and reference article.

Please Don't Forget To Join Our ML Subreddit