Machine learning guided design of strong transcriptional activation domains

Researchers have employed artificial intelligence to streamline the creation of structured proteins, but designing functional intrinsically disordered proteins remains challenging due to their weak structural constraints and unclear sequence-to-function relationships. Activation domains (ADs), transcription factor regions that recruit coactivators, serve as an ideal test subject for the creation of new intrinsically disordered sequences.

ADs have been a focus of research for many years and have practical applications in fields such as synthetic biology, biotechnology, and agriculture. To generate 500,000 candidate sequences resembling known ADs, the researchers utilized computational methods and evaluated these sequences using a range of existing machine learning models, including three neural networks.

Following this initial screening, 18,750 sequences were chosen for experimental validation in yeast. Surprisingly, 90% of these 16,770 tested sequences proved to function as ADs. Sequences that were predicted to be active by multiple models exhibited stronger functionality, leading to the development of a fast and accurate consensus AD predictor named MADRIGAL.

The researchers observed that several of their designed ADs maintained functionality across various organisms, including fungi, plants, and human cells. This study demonstrates a broadly applicable approach to integrating sequence-to-function models into the design process, enabling the efficient generation of functional intrinsically disordered sequences.