How to use computational tools to maximize the coverage of protein sequence/structure/function space Murray Lab: Nebojsa Mirkovic, Tonya Silkov, Hunjoong Lee, Frank Indiviglio, Janey Li Honig Lab: Markus Fischer and Donald Petrey PSI Bottlenecks 1) Not enough connection between modeling and biology/experiment 2) “Modelability” not used in defining families or a dynamic target selection strategy 3) Incomplete use of functional information in model building

denotes a phosphoinositide headgroup Phosphoinositide signaling processes

Intracellular membranes contain distinct lipid compositions and carry different charge densities Binding behavior of a +8e peptide to membranes carrying different negative charge densities Biophysical properties of cellular protein/membrane interactions

Motif 1Motif 2 C1/DAGC2/Ca 2+ Protein kinase C– , ,  PH/PIP 2 C2/Ca 2+ Phospholipase C–  PH/PIP 2 PX/PI3PPhospholipase D FYVE/PI3PPH/PIFGD1(a Rho/Rac GEF) Basic/PSPH/PIP 2 GPCR kinase C2/Ca 2+ NonpolarCytosolic phospholipase A2 ENTH/PIP 2 Prot/prot Epsin1, AP180 MyristateBasic/PSSrc, MARCKS, (HIV-1 Gag) Proteins that function in phosphoinositide pathways contain multiple membrane binding motifs Multiple inputs: Temporal and spatial control of subcellular targeting through coincidence counting

+25 mV -25 mV Many peripheral proteins, especially those involved in subcellular targeting, are either highly basic or charge polarized.

Quantitative physical theory for the interaction of proteins with membrane surfaces

Connection among biophysical properties, membrane binding behavior, and subcellular localization No calcium Calcium Phospholipase C  C2 domains Homology models of all isoforms 5-lipoxygenase C2 domain Homology model

Structural genomics and proteomics-level studies of lipid-interacting domains: Northeast Structural Genomics and Arabidopsis 2010 Apply what we have learned to whole families BAR domains C1 domains C2 domains ENTH domains FERM domains FYVE domains GRAM domains High-throughout comparative modeling: Leverage structure information PDZ domains PH domains PHD domains PX domains Sec14 domains START domains VHS domains

All lipid-binding domains in all model genomes Use what we have learned computationally and experimentally to develop: 1. More complete lists of peripheral proteins of known structure from the PDB; 2. Detect and model all instances of peripheral proteins in sequence databases; 3. Discover new instances, novel functionalities, new families; 4. Create databases to house this information; 5. Use this information to annotate protein sequences of unknown function.

PDB Structure Sequence Homologues Non-redundant & unsolved Models Model quality Secondary structure Multiple alignments Modeling alignments Homologous structures Data on homologues (species, IDs, coverage, length, e-value, seq. is.) Leverage: unique models MarkUs: Function annotation Family analysis Specialized databases Web-accessible models database DSSP PSI-BLAST Modeller or Nest PROSA, pG score ClustalW pG > 0.7 Target reprioritization Nebojsa Mirkovic Proteins 66:766 SkyLine: High-throughput comparative modeling “Modelability”: Create “reliable” models using known structures as templates

NESG Models Database Frank Indiviglio

Models Database: “Leverage”: Number and quality of 3D models produced from a set of structures as templates PSI1 and PSI2: NESG leverage ~220 sequence unique models Hunjoong Lee

Alternative models based on different PDB templates, reliability measures and sequence coverage

Additional search mechanisms: Expand methodology to the entire PDB, create specialized family and genome databases

2.3x10 -9 M 2.6x10 -9 M C2 domains from phospholipase C isoforms: Comparative functionality KdKd KdKd

8.9x10 -8 M → 6.2x10 -9 M 4.0x10 -8 M C2 domains from phospholipase C isoforms: Comparative functionality KdKd

2.3x10 -9 M Differences between d1 and d4 : Detection of specificity determinants leads to hypotheses for differential regulation 8.9x10 -8 M → 6.2x10 -9 M KdKd KdKd

FYVE domain family: Electrostatic properties of models correlate with in vitro binding measurements and subcellular localization: Comparison of different members Whole family modeling: FYVE domains

FYVE domain family: Electrostatic properties of models correlate with in vitro binding measurements and subcellular localization: Residue substitution of a single family member

Model/ComputationExperiment Structure There is no straightforward prescription: Each family has to be dealt with individually “Modelability”: Create “reliable” models using known structures as templates Dynamic target re-prioritization is an important strategy

START domain leverage Modelability (7378) versus 30% sequence identity (2767)

Characterize different START domains based on structural information Discriminate whether START domains bind cholesterol or PC (PI) or other ligands Provide leads for chemical library studies for function-interfering compounds Detailed computational analysis and function annotation Fine-grain structure analysis in the absence and presence of potential ligand Experimental characterization: Protein production, SPR analysis, cellular studies Collaborations with Experimental Groups Cho Lab: High-throughput analysis of Human and Arabidopsis START domains Clark Lab: Docking studies of ubiquinone into nematode START domain, electron transport

START domains in the Arabidopsis thaliana genome SkyLine produces quality models for 58 non-redundant sequences versus 35 Arabidopsis START domains detected by sequence searches (Genome Biology 5:R41) Key Findings (Tonya Silkov) 1.45 sequences are of the Birch antigen class 2. Two sequences correspond to AHA1 domains (Activator of Hsp90 ATPase) SCOP classifies AHA domains as belonging to the Birch antigen superfamily 3.Two sequences predicted in databases as integral membrane proteins of unknown function 4.Five sequences for related models apparently represent a group of uncharacterized plant START domains

Fig. 1 ENTH domainANTH domainVHS domain Cross-genomic studies Structure similarity among lipid-binding domains Tonya Silkov PIP 2

J Biol Chem. 278:28993 with Cho Lab Helix 0 ANTH ENTH ENTH and ANTH: similar topology, different membrane binding mechanism

Helix 0 From above Tonya Silkov ENTH ANTH ENTHANTH Cho Lab: First 25 amino acids are required for both PIP2 binding and membrane penetration. Produce enough protein to obtain crystals. Arabidopsis domain with novel dual ENTH and ANTH functionality

Fig. 1 ENTH domainANTH domainVHS domain A novel functional subclass of VHS domains Tonya Silkov

KIAA1530 (Homo sapiens) XP_ (Strongylocentrotus purpuratus) CAB71110 (Arabidopsis thaliana) XP_ (Gallus gallus) Tonya Silkov A new VHS-related family, “VR domains”, found in other genomes

Among this subset of VHS domains, the basic surface patch is conserved Hypothesis: It constitutes a phosphoinositide-specific binding site VR domain family of membrane-binding VHS domains Tonya Silkov Human and Arabidopsis constructs are being examined in the Cho lab

The ability to construct a quality model of a sequence is a more strategic definition of a protein family member Allows for the discovery of distantly related members With function annotation, allows for the discovery of new sub-groups Structures + Sequences -> Models + Function annotation (Markus) More comprehensive coverage of protein sequence/structure/function space By constantly updating resources as new information becomes available, we produce a more relevant (dynamic) target selection strategy