Overview
How do proteins recognize one another?
Protein-protein interactions control critical cellular processes — from signal transduction to gene regulation to cell survival. Understanding how sequence and structure encode interaction specificity is essential for comprehending protein evolution and for designing therapeutics that target disease-relevant interactions.
Research in the Keating laboratory focuses on elucidating the mechanisms of interaction specificity that allow proteins to select correct binding partners among many possibilities. We integrate computation and experiment to understand, predict, and re-design protein complexes, with particular focus on biomedically important interaction domains.
Our approach combines molecular biology, protein chemistry, structural biology, machine learning, and high-throughput methods to build predictive models of protein recognition and deploy these insights for applications in research and therapy.
bZIP Transcription Factors
The alpha-helical coiled coil is a prevalent, experimentally tractable interaction motif. We’ve mapped the dimerization specificity of ~53 human bZIP proteins, revealing how sequence encodes selective recognition and how these interaction networks have diverged across species. Using computational design and high-throughput validation, we’ve developed peptide inhibitors selective for individual bZIP targets (JUN, FOS, XBP1, MAF) relevant to disease.
Bcl-2 Family Proteins & Apoptosis Control
We’ve comprehensively mapped BH3-binding specificity across human and viral Bcl-2 proteins and built predictive models to guide inhibitor design. Our designed peptide inhibitors of anti-apoptotic Bcl-2 proteins — particularly Mcl-1 — induce cell death selectively in cancer cells dependent on these proteins for survival. Collaborations with the Walensky lab have led to chemically modified, cell-permeable versions of these inhibitors now advancing toward clinical development.
Short Linear Motif Recognition
We’re systematically defining how proteins recognize short linear sequences across diverse biological contexts:
LC3B and autophagy — We screened ~500,000 proteome-derived peptides to define the binding determinants of LC3B, the core autophagy protein. Our work refined the LC3-interacting region (LIR) motif definition, revealed unexpected specificity rules, and enabled design of synthetic LC3B binders — expanding our understanding of autophagy partner networks.
EVH1 and MATH domains — We’re mapping the specificity of domains that bind short linear motifs in signaling and scaffolding complexes, combining proteome screening, structural modeling, and biophysics to understand how flanking sequences and core motifs together encode binding specificity.
Disease Mutations & Structural Mechanisms
We apply computational design and structural modeling to understand how disease-associated mutations alter protein function. For example, we used in silico modeling coupled with in vivo studies to show that two different STAT5B mutations at the same position have opposing functional consequences — one a gain-of-function driver of T-cell leukemia, the other a loss-of-function variant. This work illustrates how structure-based approaches reveal mechanistic roles of disease variants.
Machine Learning for Protein Design & Prediction
The emergence of AlphaFold and related deep-learning tools has transformed structural prediction, but these models have important limitations. We’re advancing the field in several ways:
Improved sequence design — We developed PottsMPNN, which learns Potts energy functions from protein backbones to generate sequences with better fold compatibility and more accurate energy predictions than models optimized solely for native sequence recovery.
Understanding model performance — We benchmarked AlphaFold2-Multimer, AlphaFold3, and other methods on protein-peptide interactions, revealing training biases and showing how multiple sequence alignment quality impacts prediction accuracy — insights critical for interpreting model outputs.
Peptide Inhibitors of Bacterial Defense Systems
We’ve extended our fragment-based peptide design approach to target toxin-antitoxin (TA) defense systems — widespread bacterial mechanisms that impede phage infection. Using deep-learning scoring functions, we designed peptide inhibitors of the RelE toxin that effectively block antiphage defense. This work showcases how computational design can generate counter-defense elements to enhance phage therapy efficacy.
Across these projects, our premise is simple: high-quality, high-throughput data enables predictive model building, and predictive models are essential for design and discovery.