1.
~400,000 peptide mass spectra

2. A few diverse examples of proteins: A muscle protein:
aspirin
A virus protein shell (“capsid”):
Watercolors by David Goodsell, Scripps

3. Outline Part I What dictates the 3D shape (“fold”) of proteins?
1. Primary structure of proteins
- amino acids & peptide bonds
2. Secondary structure of proteins
- “local” folding topology & predicting 2° structure
3. Tertiary structure of proteins
- “global” folding topology
- X-ray crystallography & NMR
- aligning structure computationally
- protein folding
- designing new structures
Part II
How do proteins interact with each other in the cell?

4. The levels of protein structure:

5. solvent accessible
surface
“ribbon” = Ca backbone
Different
representations
of a typical
globular protein
(myoglobin)
ribbon + stick-figure
side chains
all atoms drawn
at van der waals radii

6. Due to resonance forms of the peptide bond:
Peptide bonds (N-CO) are planar, so only allowed rotation
along amino acid backbone is around Ca-N and Ca-CO bonds
==> by convention angles called F & Y
Protein folding = the selection of F/Y angles & side chain angles
leading to low energy packing of the atoms

7. A Ramachandran plot shows only certain F/Y combinations are sampled,
dictated by steric hindrance of atoms neighboring peptide bond
Favored regions correspond to secondary structures
==> allowable “local” structural conformations

8. 3 of the most common secondary structures
a helix
3.6 aa’s/turn

9. Amino acids vary in their intrinsic propensities to adopt
the different secondary structures

10. Given aa sequence, how to predict 2° structure? ==> PhD
input = 13 aa sliding window
- neural network, predicts 3 states: a helix, b strand, coil
& relative level of solvent accessibility
==> 3 state prediction accuracy ~72%

11. Some proteins have unusual secondary structures that
span membrane => membrane proteins
How to identify transmembrane segments in a protein?
A 0.1
B 0.3
C 0.4
A 0.4
B 0.1
C 0.3
Current best approach, TMHMM
is based on Hidden Markov models.
transition
probabilities
Hidden states
emission Y
A generic HMM: X
Hidden state seq:
Observable seq:
XXXXYYYYXXXY
CCBCCAAABCAC
Goal = recover hidden state sequence by
analyzing emissions

12. TMHMM hidden Markov model inside & outside loop models, helix cap
HMM for
5-25 aa
helix core
Correctly predicts >90 % of the transmembrane helices
Discriminates between soluble and membrane proteins with
false positive rate ~1%
Krogh et al, J Mol Biol. 305: (2001)

13. Packing of secondary structures leads to more complex
3D assemblies (“motifs”):

14. = 3D packing of secondary structural elements
Tertiary structure
= 3D packing of secondary structural elements
- Hydrophobic residues (Phe, Ile, Leu, Trp) buried in the core
- Core densely packed; not room even for H2O, comparable to a typical crystal
- Core atoms so close that van der Waals bonds contribute significantly
- Charged and polar R groups (e.g., Arg, Lys, Glu, Asp, His) on outside and hydrated

15. atomic layers in crystal
Experimental approaches to protein structure I
X-ray crystallography
crystal of pure
protein
Rotate crystal, collect
amplitudes of diffracted
X-rays as function of
incident angle of X-rays
Find phases of
diffracted X-rays
(by experiment or
computation)
With phases &
amplitudes,
Fourier transform to
find distribution of
electrons
(“electron density”)
in protein
Electrons in crystal diffract X-rays
according to Bragg’s Law: nl = 2d sinq
wavelength
angle of
X-rays
to plane of
atoms
distance between
atomic layers in crystal
Build atomic model into
electron density, refine
From B. Rupp’s X-ray crystallography intro:

16. Experimental approaches to protein structure II
Nuclear magnetic resonance
protein in
solution
in center
Vary radio wave pulses,
Measure field generated
in response over time
=> function of chemical
environment of each nucleus
Assign identities
to nuclei, measure
distances between
amino acid atoms
Use distance
geometry to solve
for ensemble of
3D structures
consistent with
distance constraints
very strong
magnet
coils to send/detect
radio waves
Basic principle: Atomic nuclei w/ odd mass #’s have spin
==> charged, spinning particles & produce magnetic field
In an external magnetic field, this nuclear magnetic field precesses around an axis
Can observe this process by applying radio wave pulses at
frequencies related to precession frequencies & measuring
the resulting induced electric current
Flemming Poulson, A Brief Introduction to NMR spectroscopy of proteins.

17. 3 broadest classes of protein 3D structures
Fibrous
e.g., collagen
Membrane
e.g, K+ channel
& Globular ...

18. Examples of globular protein “folds”
all a
a/b
all b
a+b

19. >24,000 experimentally determined protein structures
stored in PDB database:

20. Atomic coordinates of a protein structure (PDB format)
- first 3 aa’s = Met-Glu-Ala...
atomic coordinates
aa type & #
occupancy
atom type
atom # & name
x y z
B-factor
ATOM N MET A N
ATOM CA MET A C
ATOM C MET A C
ATOM O MET A O
ATOM CB MET A C
ATOM CG MET A C
ATOM SD MET A S
ATOM CE MET A C
ATOM N GLU A N
ATOM CA GLU A C
ATOM C GLU A C
ATOM O GLU A O
ATOM CB GLU A C
ATOM CG GLU A C
ATOM CD GLU A C
ATOM OE1 GLU A O
ATOM OE2 GLU A O
ATOM N ALA A N
ATOM CA ALA A C
ATOM C ALA A C
ATOM O ALA A O
ATOM CB ALA A C

21. Some of the major computational questions in structural biology
1. How to distinguish membrane proteins from soluble proteins ?
2. How to align protein structures & start organizing them into families, etc. ?
3. How to predict folded protein structure from the linear amino acid sequence?
4. How to identify the active/functional region of the protein from the structure?
5. How to predict the interactions of drugs or other proteins from the structure?
6. How to computationally predict the structural consequences of mutations?
7. How to predict protein function from structure?
8. How to design new or unnatural protein structures?

22. How to find the best superposition of 2 protein structures?
Note: superimposing 2 structures is easy if you know the equivalent amino acids
-> the hard part is to find this mapping of atoms from 1 structure to the other
One now-classic approach: DALI
Align sequence #1
to sequence #2
so as to maximize
similarity in
contact patterns
Amino acid #
Protein #1 structure
Ca coordinates only
Amino acid #
Calculate matrix of all
pairwise Ca-Ca distances
Repeat for protein # 2
Holm & Sander, J Mol Biol. 233: (1993)

23. Best structural alignment corresponds to maximizing
i, j = aligned pairs of matched residues
i = iA, iB j=jA,jB
f = similarity of 2 Ca-Ca distance matrices, dAij and dBij
In the simplest case,
where dAij and dBij are equivalenced residues in proteins A and B.
and q R = minimum level of similarity
Choose mapping
of residues
(e.g. iA to iB)
to minimize
dAij- dBij iA iB
dAij
dBij jA jB
Protein A
Protein B

24. The ability to compare structures has led to recognition
of a hierarchy of 3° structures (“folds”)
Class
As organized in the
CATH or SCOP or FSSP
databases:
Architecture
Manual classification at
architecture level,
automated at topology level
Topology
Homologous
Superfamily H
flavodoxin homologues

25. Protein Folding Classic experiment from 1960’s (Chris Anfinsen):
Purified small protein RNaseA,
Refolded in a few minutes in solution
==> all information necessary for correct folding
was captured in the linear amino acids sequence
Corollary:
Proteins do not fold by randomly testing conformations.
Given a 100 amino acid protein,
& 10 possible conformations / amino acids
= possible conformations for the protein
==> not possible to randomly sample, clearly constrained search

26. An energetic view of the folding process
Fast
Slow
Large # of
conformationally
different
molecules
Collection of
similar
conformations
interconverting
Unique or
small # of final
conformations
optimize packing T
“hydrophobic
collapse”
free energy U M F
Molten
globule
Transition
state
Unfolded
Folded
folding trajectory
Local secondary
structures form first
Adapted from Branden & Tooze

27. One long-time goal of biologists/biophysicists:
Solve the Protein Folding Problem =
computationally predict protein 3D fold
from 1D amino acid sequence
Two general approaches:
1st principles/ab initio:
e.g., atomistic molecular dynamics simulations
of proteins, modeling force fields w/ electrostatic,
van der waals forces, solvent, etc. over long time
Empirical:
- fold recognition/threading
- reverses the process: given set of structures,
learn empirical rules that predict folds
Empirical currently more successful at predicting final
structure, but no information about folding trajectory

28. An example of a successful design of a new protein fold
by a combination of empirical & ab initio structural modeling
designed 93 amino acid
protein with topology
not in PDB dbase
designed model
solved structure
Kuhlman et al, Science, 302: (2003)

29. The Kuhlman et al. design strategy
Starting model = Choose predefined 3D topology
Assemble 3D model from 3 and 9 amino acid fragments of known structure
==> Generated 172 backbone-only starting models
Initialization
Choose optimal sequence for each starting model using energy function that captures:
12-6 Lennard-Jones potential
orientation-dependent hydrogen bonding term
implicit solvation model
Choose amino acid side chain orientations (“rotamers”) by sampling from known structures
Iterate between:
Optimize choice of amino acid sequence for a fixed backbone conformation
Optimize amino acid backbone coordinates for a fixed sequence
Same energy function used at all stages
Only previous lowest energy sequence/structure optimized at each stage
Final designed sequence not similar to any known protein sequence
Kuhlman et al, Science, 302: (2003)

30. References
A good introduction to structural biology =
Introduction to Protein Structure
- Carl Branden & John Tooze
Web resources:
Protein Data Bank = > 24,000 protein structures, atomic
coordinates, & the “protein of the month”
CATH/SCOP protein structure hierarchies:
Several of the illustrations in this tutorial were taken from
Lehninger Principles of Biochemistry, by Nelson & Cox

31. Part II

32. Macrophage (“white blood cell”) Blood serum
Bacterium
“Macrophage and Bacterium 2,000,000X”
Watercolor by David S. Goodsell, 2002

33. Typical size ranges of known protein structures & assemblies
single
protein
domain
dimeric
protein
aquaporin
(membrane channel)
Ribosome
From a (recommended) review article==>Sali et al. Nature 422: (2003)

34. Outline Part I What dictates the 3D shape (“fold”) of proteins?
Part II
How do proteins interact with each other in the cell?
4. “Quaternary” structure of proteins
& protein interactions
5. Experimental approaches to determine interactions
- yeast 2 hybrid, mass spectrometry
6. Testing the accuracy of the interactions
7. Moving back to the atomic resolution world
- electron microscopy & tomography
- modeling structures of complexes

35. Why study interactions?
Proteins interact all the time (e.g., bump into each other non-specifically)
We’re interested in specific interactions
==> e.g., those w/ downstream consequences
For example, consequences might include:
Inducing a change in the structure of an interaction partner
Stabilizing or destabilizing an interaction partner
Modifying the activity of a protein (activate, inhibit, or otherwise regulate)
Cause interaction partner to move to another location
Cut interaction partner
Chemically modify interaction partner (phosphorylate, dephosphorylate,
glycosylate, deglycosylate, ubiquitinate, sumoylate, etc...
==> more than 200 modifications to proteins known, many catalyzed
by other proteins
So, defining interactions helps to define these processes &
their functional consequences

36. Experimental/Computational methods for observing/inferring protein interactions
Sali et al. Nature 422: (2003)

37. X-ray structure
of ATP synthase
Schematic
version
Network
representation a b g d b2 e a
c12
Total set = protein complex
Sum of direct + indirect
interactions

38. Some methods measure direct interactions, some indirect
Xenarios & Eisenberg, Curr. Op. Biotech. 12:334-9 (2001)

39. Interactions between
yeast proteins

40. Experimental approaches to protein interactions I
Yeast two-hybrid +
DBD
“Bait”
“Prey”
Act
DNA binding
domain
Transcription
activation
domain
Prey
Act
Core transcription
machinery
Bait
DBD
transcription
operator or
upstream activating
sequence
Reporter gene
Basic idea = screen library of “prey” proteins to test which ones interact with a given “bait” protein
Fields & Song, Nature 340:245-6 (1989)

41. Experimental approaches to protein interactions I
High-throughput yeast two-hybrid I
Haploid yeast
cells expressing
activation domain-
prey fusion proteins
Diploid yeast
probed with
DNA-binding domain-
Pcf11 bait
fusion protein
Uetz et al. Nature 403 (2000)

42. Uetz et al. Nature 403 (2000)

43. was the apparent inconsistency among the interaction sets
A second group (Ito et al.), with a related yeast two-hybrid approach, also mapped
a large number of interactions, then compared the interactions w/ the Uetz data:
A surprise at the time
was the apparent inconsistency
among the interaction sets
==> either # of
potential interactions is large or
false positive rate high
(or both)
Ito et al. PNAS 98: (2001)

44. Experimental approaches to protein interactions II
Mapping complexes by mass spectrometry I
“Bait”
protein
Tag
Interaction
partners
co-purified
with “bait”
Affinity
column
493 bait proteins
3617 “interactions”
protein 1
protein 2
Ho et al. Nature 415 (2002)
SDS-
page
protein 3
Trypsin digest,
identify peptides by
mass spectrometry
protein 4
protein 5
protein 6

45. Experimental approaches to protein interactions I
A variant: Tandem affinity purification (TAP) + Mass spectrometry
Tag1
Tag2
Bait
Affinity
column2
protein 1
Affinity
column1
protein 2
SDS-
page
protein 3
protein 4
+ protease
protein 5
protein 6
Trypsin digest,
identify peptides by
mass spectrometry
Affinity
column1
Rigout et al., Nature Biotech. 17: (1999)

46. Gavin et al. Nature 415 (2002)

47. How accurate are these high-throughput screens?
Can compare to known interactions, but these are incomplete
A different strategy is to identify properties that correlate with interactions
& test versus those properties
Three tests:
1. Comparison of interactions to a reference interaction set
2. Comparison of mRNA co-expression of interacting partners
3. Comparison of functions of predicted interaction partners

48. (tends to underestimate accuracy)
Test #1
Estimate accuracy by comparing to a well-determined
reference set of interactions
(tends to underestimate accuracy)
von Mering, Krause et al. Nature May 8, 2002

49. Estimating interaction assay accuracy by
Test #2
Estimating interaction assay accuracy by
assessing mRNA co-expression of putative interaction partners
Correlation coefficient between expression vectors
derived from many DNA microarray experiments
True interactions
Random
Protein
Pairs
Estimate % false positives
from observed vs. expected
genes w/ correlated
expression
Estimated false positive rates based on this test:
Mrowka et al. Genome Research 11: (2001)

50. of those from random & well-characterized interactions
A related strategy: fit distribution of co-expression relationships as mix
of those from random & well-characterized interactions
==> Mixture % indicates accuracy.
Deane, Salwinski et al. Mol. Cell. Proteomics (2002)

51. Estimated true positive rates based on this test
>1 independent expmt
>2 independent expmt
Genome-wide
yeast two-hybrid
At least 1 small-scale expmt
>1 independent experiment
Paralogs also interact
Increasing # of
Interaction Sequence
Tags
Deane, Salwinski et al. Mol. Cell. Proteomics (2002)

52. S U pw1 pw2 Test #3 Swi4 Cdc27 Cell cycle MAPK signaling pathway
Estimate accuracy by measuring functional similarity of putative partners
==> in particular, measure tendency to be in same cellular system or process
From literature & pathway databases (KEGG/GO), we know ~ yeast protein functions:
Swi4
Cdc27
Cell cycle
MAPK signaling pathway
Cell cycle
Ubiquitin-mediated proteolysis
Pathways
of A
Pathways
of B
pw1
pw2
Jaccard coefficient = # pathways in common / # total pathways 1 n S
pw1
pw2 U
<pathway similarity> = n
pairs
Systematically test every pair of characterized proteins

53. Quality of the observed protein-protein interactions
as measured by the pathway overlap test
max agreement
of interacting
proteins’ pathways
Small-scale experiments
Large-scale
yeast two-hybrid
interaction experiments
Date & Marcotte, Nature Biotech. 2003

54. The various accuracy tests agree to a first approximation
(at least as regards the ranking of accuracies)
Estimated True Positive Rate
via
Co-expression Test set Pathways
Authors Method # interactions Mrowka Deane vonMering Date
Ito et al. Y2H % 22% % ~18%
Ho et al. MS ~3617 ~10% 1-3%
Gavin et al. MS ~1440 ~85% ~10%
Uetz et al. Y2H % 50% ~57%
Tong et al. synthetic lethal ~20%
>1 independent experiment ~ % ~30-40% ~87%
>2 independent experiments % ~60-70% ~95%

55. The current highest throughput protein interaction screens:
Authors Method # interactions
Ito et al. Y2H
Yeast Tong et al. SL ~4000
Ho et al. MS ~3617
Gavin et al. MS ~1440
Uetz et al. Y2H
Fromont-Racine et al. Y2H
Tong et al. SL
Newman et al. Y2H
C. elegans Li et al. Y2H ~4000
Walhout et al. Y2H
Davy et al. Y2H
Fly Giot et al. Y2H 20,405
Human Bouwmeester et al. MS
& several others, including Hepatitis C & H. pylori
Y2H = yeast two hybrid
MS = mass spectrometry
SL = synthetic lethal

56. ==> ~1/3 of the way to a complete map!
How many meaningful physical protein-protein
interactions are there?
At a rough estimate:
Human
Yeast
~5,800 genes
~40,000 genes
~5,800 proteins
x 2-10 interactions/protein
>>40,000 proteins
x 2-10 interactions/protein
~12, ,000 interactions
>>80, ,000 interactions
>10-20,000 known,
perhaps ~1/2 correct
==> ~1/3 of the way to a complete map!
<5,000 known
==> approx. 1% of
the complete map!
==> We’re a long ways from the complete map of the human “interactome”

57. electron microscopy or
Can we relate these interactions back to the protein structure?
==> A growing area of research is combination of low resolution structure
with atomic models to build structures of protein complexes:
For example:
Low resolution
electron density
map
from
electron microscopy or
electron tomography
Experimental or
computational
protein models
Rough estimate
of atomic model
of protein complex

58. Reconstructed electron density map
Example 1 – Electron microscopy of a protein complex
Experimental electron microscopy data
Reconstructed electron density map
of protein complex
Dock atomic
models into
electron
density maps
Sali et al. Nature 422: (2003)

59. Example 2- Electron tomography of a protein complex/assembly
Measure projections
of molecules after
illuminating with
electron beam
from different angles
Reconstruct density
distribution (“tomogram”)
as sum of back-projected
densities
Sali et al. Nature 422: (2003)

60. Reconstructing cellular organization of molecular complexes by
fitting structures into electron tomograms
“noisy” tomogram
(3D density map)
of single cell
Fit known structures
(“templates”) into density
Sali et al. Nature 422: (2003)

61. Some Protein Interaction Resources on the Internet
Protein interaction databases
Biomolecular Interaction Network Database (BIND)
Currently 73,000 interactions
Database of Interacting Proteins (DIP)
Currently 44,000 interactions
Protein Quaternary structure database (PSQ)
Atomic structures of interacting proteins
Interactive visualization of networks
Cytoscape:
Interactive display of protein networks
LGL (Large Graph Layout):
Visualization of networks with up millions of edges, 100,000’s of vertices