Abstract
The proposed work in this MIRA application focuses on my long-term goal to create programmable toolkits for
proteome editing. Inspired by research into developing broad-targeting CRISPR enzymes for programmable
genome editing, my laboratory designs analogous peptide-guided enzymes for post-translationally modifying
target proteins from sequence alone, enabling targeting of both structurally-stable and disordered proteins. To
this point, we leverage generative artificial intelligence approaches to design these peptides, including protein
language model (pLM)-based interface predictors and diffusion models for de novo peptide binder generation.
Integrated with these efforts, we experimentally engineer peptide-E3 ubiquitin ligase fusions (termed
ubiquibodies or uAbs) that can degrade diverse pathogenic proteins. Building upon this work, we will combine
our current technologies and expand our research into three new areas to enable broad-scale proteome
editing applications via generative protein language modeling. First, we will build upon our current
experimental efforts to enable peptide-guided deubiquitination, thus enabling both protein degradation with
our uAbs and now protein stabilization of target proteins. To accomplish this goal, we will explore fusion of
generated peptide binders to deubiquitinase domains, as well as to designed recruiter domains of
endogenous deubiquitinases. The results of this research will establish easy-to-design “on” and “off” switches
for controlling protein states in cells. Second, we will develop a new line of investigation, specifically to
enable mutant-specific targeting of protein variants. Here, we will leverage the variant prediction capabilities
of state-of-the-art pLMs, alongside Gaussian Process-based predictors, to prioritize binders that discriminate
between wild-type proteins and highly-similar mutant variants that confer pathogenic properties. These
binders can then be integrated into our degradation and stabilization architectures for experimental validation.
We envision that this new line of investigation will enable selective degradation of mutant, disease-causing
proteins, such as mutant Huntingtin protein or mutant KRAS. Third, we will establish a specific focus on
selectively targeting dysregulated, post-translationally-modified proteins, which have numerous implications in
various diseases, such as aging and cancer. To enable discrimination between wild-type proteins and their
modified counterparts, we will train new PTM-specific protein language models via both masked language
modeling and autoregressive training tasks. By integrating SwissProt annotations with unique per amino
acid-based tokenization strategies, the results of this research will enable the design of degraders that are
selective to modified proteins, such as phosphorylated oncogenic drivers, leaving wild-type isoforms intact.
Together, the work we propose in this MIRA application promises to yield robust, scalable, and most
importantly, programmable tools for protein targeting and editing applications.