Prediction and design of transcriptional repressor domains with large-scale mutational scans and deep learning

Accurate models of transcriptional repressor activity could facilitate the development of impactful diagnostics and tools for controlling gene expression. Here, we describe our work to train a deep learning model from high-throughput mutational scanning data to predict repressor function at amino acid (AA) resolution. We measured the repressor activity of 115,000 variant repressor domains (RDs) from more than 50 RDs in human cells. These data identify thousands of clinical variants with loss of function (LoF) or gain of function (GoF), including TWIST1 HLH variants associated with Saethre-Chotzen syndrome and MECP2 domain variants associated with Rett syndrome. We also leverage these data to annotate short linear interacting motifs (SLiMs) in disordered domains that are critical for repression. We then train a deep learning model called TENet (Transcriptional Effector Network) to predict repressor function from AA sequence: the neural network recognizes sequence, structure, and biochemical properties contributing to repressor activity. We employ our model to guide the strategic mutation of RDs with desired strengths through directed evolution, which we illustrate by tuning the activity of both structured and disordered domains. Our work highlights critical considerations for improving functional predictions of variant effects and advancing synthetic design of precise regulatory machinery.