C2 - Epigenetic Editing and RNA Editing
1669: Accelerating the Discovery of Novel Hypercompact Transcriptional Activators with Machine Learning
Type: Poster Session
Poster Board Number: 1669
Presentation Details
Session Title: Friday Posters: Epigenetic Editing and RNA Editing
Due to the programmability and versatility of CRISPR-Cas systems, gene therapy research is emerging as an increasingly promising strategy to treat genetic disease. However, many genetic diseases, such as those caused by haploinsufficiency, cannot be treated with traditional knock-out based CRISPR gene therapies, since they are caused by allelic loss-of-function. Due to its ability to both activate as well as suppress gene expression, epigenetic editing using CRISPR-dCas systems has the potential to address many genetic diseases that are unsuitable for traditional gene therapies. However, the versatility of epigenetic editing is constrained, especially in the context of gene activation, by the limited number and large size of most modulator peptides, severely limiting their therapeutic utility.
To address these limitations, we have built a generative AI platform capable of designing de novo hypercompact modulator peptides with the ability to transcriptionally upregulate a genetic locus. We first collected a large corpus of training data by performing high-throughput screens to discover novel transcriptional activators among peptides derived from human, viral, and archaeal proteomes. We then trained a machine learning ensemble model, composed of a decision tree model and a convolutional neural network, which leverages transfer learning (via large protein language model embeddings) to predict transcriptional activators based on peptide sequence alone. By exploiting a novel sampling algorithm, which we call evolutionary Monte Carlo search, to more efficiently traverse the predicted activator fitness landscape, we used this machine learning platform to generate a library of several thousand hypercompact peptides predicted to be transcriptional activators.
We experimentally screened these peptides and validated that our generative AI approach dramatically increased discovery rate (up to 45-fold) resulting in the discovery of hundreds of novel transcriptional activators sharing little sequence similarity with known naturally occurring peptides. We next investigated the evolutionary, biochemical, and biophysical properties of the synthetic activator library, revealing that validated activators consistently lack conserved functional domains but do share certain biochemical features, such as strong negative electrostatic potential. We subsequently selected 10 of our top synthetic activators for further characterization, and assessed their activation strength by screening them at an artificial GFP locus as well as at an endogenous human locus (CD45), comparing their potency to gold standard activators (e.g. VPR, VP64, ⋯). Additionally, using RNA-seq, we assessed off-target gene expression changes to evaluate the specificity (and thus safety) of on-target transcriptional upregulation using a synthetic activator. These results demonstrate the capability of machine learning to accelerate the discovery of novel functional peptides to expand our toolbox of epigenetic modulators for future therapeutic applications.
M. Zaki Jawaid, T. Blair Gainous, Kavita Jadhav, Aayushma Gautam, Chris Still, Timothy P. Daley, Alexandra Collin de l'Hortet, Dan O. Hart, Robin W. Yeo, Tyler Borrman
Epic Bio, South San Francisco, CA"
Find This Session