An AI model designed 15 DNA sequences that activated a reporter gene in mouse embryonic tissues

An AI model designed 15 DNA sequences that activated a reporter gene in mouse embryonic tissues

On August 25, Nature Genetics published a paper on 15 synthetic enhancers. These model-designed DNA sequences regulate gene activation. In mouse embryos, each sequence activated a reporter gene in the heart, limb buds, or central nervous system.

Cells in the heart, limbs, and nervous system contain the same DNA, but different genes are active in each tissue. Enhancers are DNA sequences that bind regulatory proteins and thereby help activate a gene. Previous models had designed similar sequences for use in cell cultures and Drosophila embryos. The authors tested whether designed sequences could also function in mammalian tissue.

Shengzhi Chen's team trained the neural networks in two stages. First, the models learned from DNA accessibility maps to predict where gene-regulating proteins could bind in three tissues of a mouse embryo. The researchers then fine-tuned them using 311–432 enhancers for each tissue whose activity had already been tested experimentally. The large genomic dataset helped the models identify sequence properties, while the smaller set of functional experiments helped them distinguish sequences that actually activate a gene in the intended tissue. Among the enhancers with the highest-confidence predictions, at least 70,6% were active in the intended tissue.

The Ledidi program, which modifies DNA sequences according to model scores, then started with 1 200 random templates for each tissue. It substituted DNA bases to increase the score for one tissue while reducing the scores for the other two. The authors selected five variants for the heart, five for the limbs, and five for the central nervous system. All 15 sequences were derived from random templates.

Each sequence was placed upstream of the same reporter gene, a test gene that stains a cell when activated. The resulting construct was inserted at the same location in the mouse genome. On developmental day 11,5, all 15 sequences produced a reproducible signal in the assigned tissue in at least three embryos. Four of the five heart enhancers and four of the five central nervous system enhancers were active only in the intended tissue. The limb variants remained most active in the limbs, although they also produced a weak signal in other embryonic tissues.

Targeting an enhancer to a specific part of the nervous system proved more difficult. One sequence designed for the forebrain also strongly activated the reporter in the limbs. Of the two midbrain variants, one did not activate the reporter, while the other construct contained a duplication of 419 base pairs.