How do proteins recognize one another?

Protein-protein interactions play critical roles inside and outside the cell, maintaining structure, transducing signals, regulating gene expression, executing metabolic programs, and enabling cell growth and motility. An outstanding goal in biology is to understand how protein interactions are determined by sequence and structure, and how protein-protein recognition contributes to biological function in healthy and diseased organisms.

Research in the Keating laboratory is focused on understanding the mechanisms of interaction specificity that allow proteins to engage the correct binding partners among myriad possibilities. Interestingly, many proteins interact selectively with some - but not all - members of structurally conserved families. Elucidating the origins and logic of selective recognition is important for understanding the evolution of protein families in which gene duplication and divergence have provided access to new activities. A thorough understanding of how sequence and structure encode specificity is also important for designing synthetic molecules that engage specific targets, for applications in research, diagnosis and therapy.

Our approach to studying protein-protein interactions focuses on selected families of biomedically important interaction domains and integrates computation and experiment to understand, predict and re-design protein complexes.

To achieve a deeper and useful understanding of how protein sequence and structure determine interaction specificity our projects focus on:

  • Obtaining high-quality interaction data in high throughput

  • Building predictive models that capture how sequence encodes interaction specificity

  • Designing novel proteins or peptides with desired interaction profiles

  • Deploying protein-interaction inhibitors for biomedical research

  • Modeling how protein interaction networks have evolved


Our premise is that technology can provide high-quality data in increasingly high throughput, data can inform model building, and models are essential for interaction prediction and design. We use the tools of molecular biology, protein chemistry, biophysics, structural biology, molecular modeling, deep sequencing, data mining, bioinformatics and machine learning to improve our understanding and our abilities to manipulate protein function.

Sections below give examples of our approaches and findings. See our Publications page for our latest contributions.

bZIP Transcription Factors

The alpha-helical coiled coil is a common interaction motif found in proteins with many different functions. Coiled-coil structures consist of two or more helices that wrap around one another with a superhelical twist. Their simple sequence patterns and symmetrical structures make coiled coils amenable to computational modeling. Short coiled coils fold autonomously and reversibly, making them experimentally tractable as well. A rich body of literature dating to the 1950s reports many sequence/structure/function relationships for coiled coils, as well as a plethora of important roles in biology.

Mapping the dimerization specificity of bZIP transcription factors in humans and model organisms.

Basic region leucine zipper (bZIP) transcription factors dimerize using a coiled-coil motif and are critical in all eukaryotes for regulating essential processes. bZIP proteins dimerize to bind to DNA, and 53 human bZIP paralogs can give rise to over 1400 different possible homo and heterodimers.


Structure of the bZIP transcription factor AP1 (a Fos/Jun heterodimer) bound to DNA. Approximately 53 bZIPs are encoded in the human genome, and these can form a variety of homo- and hetero-dimers, with dimerization mediated by a parallel coiled coil. (Structure by Glover & Harrison, 1995)

We experimentally mapped the protein-protein interactions of the human bZIPs and uncovered a high degree of interaction specificity encoded directly in bZIP sequences. We also quantified bZIP homo- and heterodimerization for all pairs of bZIP proteins in four metazoan species and two single-cell organisms, providing a uniquely detailed view of the evolution of a protein interaction network. Importantly, we found that interactions are not highly conserved between orthologous protein pairs, which has significant implications when considering the transfer of protein interaction data from one species to another. In collaborative work with the Andari lab (U. Wisconsin) we used library selection methods to determine the DNA binding specificity of 22 homodimers and 80 heterodimers for human proteins. Our bZIP protein-protein interaction measurements for ~5,000 metazoan bZIP pairs provided data that we have used to improve algorithms for predicting and designing bZIP coiled-coil interactions.

bZIP interactome networks for human proteins: protein-protein (left) and protein-DNA (right) interactions. Reinke AW, Baek J, Ashenberg O, Keating AE. (2013) Networks of bZIP protein-protein interactions diversified over a billion years of evolution. Science 340, 730. Rodríguez-Martínez JA, Reinke AW, Bhimsaria D, Keating AE, (2017) Ansari AZ. Combinatorial bZIP dimers display complex DNA-binding specificity landscapes. Elife pii: e19272.

Designing selective peptide inhibitors of bZIP transcription factor dimerization.

Designing peptides that can interact selectively with just one protein member of a large superfamily is challenging. Our group developed novel computational methods for the specificity design problem. Some of the methods are rooted in structure-based scoring of different interactions whereas others are based on models learned from high-throughput data. We perform design using integer linear programming to optimize interaction with a desired target while disfavoring interaction with other related bZIPs. We have applied our method to design peptide inhibitors for human and viral transcription factors important for disease including JUN, FOS, XBP1 and MAF. We have tested our designs using high-throughput microarrays and solution biochemistry, confirming tight and selective binding to many targets. The figure below shows three design successes, with the curves illustrating binding of designed peptides to the indicated bZIP targets and the heat map illustrating specificity of the design for the target over other human bZIPs.

Computational optimization gives designed peptides that bind selectively to human bZIP transcription factor targets. Grigoryan G, Reinke AW, Keating AE. (2009) Design of protein-interaction specificity gives selective bZIP-binding peptides. Nature 458, 859. Potapov V, Kaplan JB, Keating AE. (2015) Data-driven prediction and design of bZIP coiled-coil interactions. PLoS Comput Biol. 11, e1004046.

Developing modular coiled-coil interaction parts for synthetic biology and nanotechnology

The coiled-coil protein-interaction motif is found in thousands of natural proteins and is also a useful, modular protein component for molecular engineering. The short length, simple structure, high affinity and adjustable interaction specificity of the coiled coil have led to many applications of this motif in structural, cell and synthetic biology. Using tools and reagents from our work on bZIP transcription factors, we developed a molecular toolkit of 30 biochemically characterized coiled-coil dimers made from synthetic zipper proteins, or “SYNZIPs.” The figure below shows the crystal structure of one such complex. We developed sets of non-cross-reacting parallel heterodimers and anti-parallel homodimers and, in collaboration with the Lim lab (UCSF), demonstrated that our molecular parts can be used to rationally construct signal transduction pathways in yeast. We also used SYNZIPs to construct a nanotriangle, a model of which is shown below. We described our reagents in “specification sheets,” available at this site, which include information about interaction affinity, structure, and performance in various interaction assays.

SYNZIP peptides form selective heterodimers and can be used for application in synthetic biology and materials science. Reinke AW, Grant RA, Keating AE. (2010) A synthetic coiled-coil interactome provides heterospecific modules for molecular engineering. J Am Chem Soc. 132, 6025. Park WM, Bedewy M, Berggren KK, Keating AE. (2017) Modular assembly of a protein nanotriangle using orthogonally interacting coiled coils. Sci Rep. 7, 10577.

Bcl-2 family proteins


Bcl-2 family anti-apoptotic family members have a conserved globular structure and can bind and sequester pro-death proteins via their Bcl-2 homology 3 motifs (BH3 motifs). This protein-protein interaction controls critical cellular life-vs.-death decisions, making inhibition of Bcl-2 family interactions a promising therapeutic strategy for treating cancers and other diseases.

From a molecular recognition perspective, the docking of a single BH3 helix into a globular receptor is simple enough to allow comprehensive experimental analysis and extensive computational sampling, yet complex enough to extend our understanding of specificity in molecular recognition. We use a variety of computational and experimental methods to understand and manipulate Bcl-2 family interaction specificity.

Computational modeling of Bcl-2 family protein binding specificity. DeBartolo J, Dutta S, Reich L, Keating AE. (2012) Predictive Bcl-2 family binding models rooted in experiment or structure. J Mol Biol. 422, 124. Xue, V PhD thesis 2018.

Data and models describe the specificity landscape of Bcl-2 family protein interactions

Our laboratory has experimentally mapped BH3 binding preferences for five human and three viral Bcl-2 proteins using fluorescence assays, peptide SPOT arrays and peptide library screening, and we have built several kinds of computational models to predict Bcl-2 family binding specificity from sequence. We have applied our models to discover new BH3-like peptides in the human proteome and measured their interaction specificities. Our models guide our work on Bcl-2 protein inhibitor design, described below.

Computational modeling of Bcl-2 family protein binding specificity. DeBartolo J, Dutta S, Reich L, Keating AE. (2012) Predictive Bcl-2 family binding models rooted in experiment or structure. J Mol Biol. 422, 124. Xue, V PhD thesis 2018.

Designing potent and selective peptide inhibitors of Bcl-2 family proteins.

Over-expression of anti-apoptotic Bcl-2 family proteins confers resistance to pro-death signaling and chemotherapy in many cancers, making these proteins promising targets for pharmaceutical intervention. Using integrative computational and experimental methods and rational structure-based design, we have developed potent and selective peptide inhibitors of six anti-apoptotic Bcl-2 family proteins (four human targets and two proteins from oncogenic viruses). Our peptides induce mitochondrial outer membrane permeabilization in a range of cells with the specificity pattern expected based on their biochemical binding profiles. Bcl-2 family protein Mcl-1 is a particularly important therapeutic target. Working with the Walensky and Leati labs, we started with our successful design of potent and selective Mcl-1 inhibitors and transformed these into chemically modified peptides that are selectively cytotoxic to cancer cells that rely on Mcl-1 for survival.  

Computational modeling and library screening led to the design of peptides that bind tightly and selectively to Mcl-1, a protein that contributes to the chemoresistance of many cancers (left). Subsequent studies generated cell permeable, cytotoxic peptides (right). Foight GW, Ryan JA, Gullá SV, Letai A, Keating AE. (2014) Designed BH3 peptides with high affinity and specificity for targeting Mcl-1 in cells. ACS Chem Biol. 9, 1962. Rezaei Araghi R, Bird GH, Ryan JA, Jenson JM, Godes M, Pritz JR, Grant RA, Letai A, Walensky LD, Keating AE. (2018) Iterative optimization yields Mcl-1-targeting stapled peptides with selective cytotoxicity to Mcl-1-dependent cancer cells. Proc Natl Acad Sci U S A. 115, E886.

Short Linear Binding Domains

​We have recently become interested in the interaction specificity of protein domains that bind to short linear motifs. This type of protein-peptide interaction is common in eukaryotic signaling and scaffolding complexes. We are focusing on EVH1 domains in the Ena-VASP family of proteins and MATH domains in the TRAF family. These domains play key recruiting functions and, despite the report of consensus core binding motifs for both families, the full determinants of interaction specificity are not known. We aim to more comprehensively define what is required for binding, provide a biophysical model of how specificity is achieved, design synthetic peptides or mini-proteins that can inhibit domain interactions, and map the evolution of domain-peptide interaction networks using proteome screening, bioinformatics, and structure-based modeling. We are particularly interested in understanding the role of flanking sequences that surround the core binding motif.