how close we already are to this vision via AI-led full sequence design of a synthetic yeast chromosome

点击翻译按钮获取中文摘要

In designing, building, and testing AI-written regions of yeast chromosomes, I will learn what genomic knowledge and AI design capabilities are currently missing that would enable bespoke chromosome design in the next decade. I will develop solutions to these current gaps, so that by the end of the project we can use AI to generate the full sequence of a synthetic chromosome arm, providing the AI with only a list of genes to be included, the peptide sequences of these, their target expression and any regulation required. The project will build on work at Imperial for building and testing synthetic chromosome regions in yeast (Ellis), and in designing AI-driven tools for engineering biology (Stan). Importantly, it will build on the existing success of key published AI tools that use GANs and CNNs to write synthetic yeast promoters [1], 5'UTRs [2] and 3'UTRs [3] to specification. These 3 parts account for >85% of non-coding DNA in a yeast chromosome.As a pilot study, I aim to integrate published AI tools for genetic part design in S. cerevisiae yeast to create a program that suggests the full sequence encoding a cluster of metabolic genes so that they are expressed at levels that match what is typical in wildtype yeast. This will be done using a yeast strain which has been modified by relocating all the genes encoding the synthesis of tryptophan and histidine to 'testbed' clusters - where the pathways are expressed from constitutive promoters (Ellis lab). To identify design failures, the synthetic DNA will be purchased and swapped-in to replace the current cluster DNA, and cells will be tested for viability, growth rate and functionality of the encoding pathways when all or some of the DNA is AI-designed.To improve the design of regulatory DNA, I will extend the work on a state-of-the-art promoter-writing Generative Adversarial Network (GAN) developed by Zrimec and collaborators [4], which, in 57% of the cases, generated highly-expressed synthetic sequences surpassing the expression levels of highly-expressed natural controls. This will allow me to use natural regulatory sequence data to train a deep generative model, which will learn feasible and biologically consistent candidatesequences. The generator's ability will be guided by a predictor model that fine-tunes generated sequences toward target gene expression. Multiple designs for each regulated promoter will be synthesised and tested in the yeast in their cluster for functionality, to iteratively improve the AI-based design for regulated promoters and understand how to incorporate this into AI chromosome design.