1. Protein Structure Prediction
Matthew Betts
Russell Group, University of Heidelberg, Germany

2. Structure Function Sequence Active/inactive? Binds/does not bind?
Substrate specificity?
Function
Sequence

3. What is this about?
What we do to find out what a protein might be doing
Looking at sequences, with a particular emphasis on finding out something about the protein structure
Some background for practical work

4. Given a sequence, what should you look for?
Functional domains (Pfam, SMART, COGS, CDD, etc.)
Intrinsic features
Signal peptide, transit peptides (signalP)
Transmembrane segments (TMpred, etc)
Coiled-coils (coils server)
Low complexity regions, disorder (e.g. SEG, disembl)
Hints about structure?

5. Given a sequence, what should you look for?
“Low sequence complexity”
(Linker regions? Flexible? Junk?
Transmembrane segment
(crosses the membrane)
Signal peptide
(secreted or membrane
attached)
Tyrosine kinase
(phosphorylates Tyr)
Immunoglobulin domains
(bind ligands?)
SMART domain ‘bubblegram’ for human fibroblast growth factor (FGF) receptor 1
(type P11362 into web site: smart.embl.de)

6. What about structure? 3D 3D 3D
Intrinsic features general mean trouble for structure determination, so they are usually skipped
Knock on effect is that structures for large, flexible multi-domain proteins are rare
Structure determination/prediction therefore typically restricted to parts (with exceptions obviously)

7. Structure prediction
algorithm
Sequence
Structure

8. Best predictions are by homology
Is your sequence homologous to a known structure?
If yes, then often very good models of structure can be constructed.
This is what we will do in the practical

9. Homology Modelling
algorithm +

10. Homology Modelling Steps
Identify a homologue of known structure
Get the best alignment of your sequence to the structure
Model building
Side-chain replacement
Loop building
Optimisation/relaxation/minimisation

12. Problems with loops
Two subtilisin-like serine proteases

13. Sanchez et al, Nature Struct. Biol. (Suppl), 7, 986-990, 2001

14. The Twilight Zone Sander & Schneider (EMBL, ca. 1990)
Compared all known structures to each other using sequence comparison.
For each fragment of a particular length & sequence identity, simply asked the question: is the structure similar or different.
The line to the right is where one can be 90% confident that an alignment of a particular length & sequence identity
Below the line, structures can be either similar or different: the twilight zone.
(Basis for much of the sequence alignment statistics that are now in use today)
Based on Sander & Schneider, Proteins, 9, 56, 1991

15. Similar structures within the twilight zone
sequence identity: % % %
…can we find these similarities without known structures if sequence searches fail?
Russell et al, J.Mol. Biol., 1997

16. Fold Recognition (‘Threading’)? ? ? ? ?
>C562_RHOSH
TQEPGYTRLQITLHWAIAGL…
Does the sequence
“fit” on any of a
library of known
3D structures?

17. Fold Recognition (‘Threading’)
Jones, Taylor, Thornton,
Nature, 358, 86-89, 1992.

18. Residue pair potentials
Phe
GOOD
Asp
Asp
Phe
BAD
Arg

19. Fold Recognition Executive Summary
Works some of the time
Probably best at identifying distant homologues, where sequence identity is in the twilight zone
Useful sites:
3D-PSSM, FUGUE, (Gen)-Threader
Meta predictions are the best - combine all and get a consensus
E.g. bioinfo.pl/meta

20. If no homology…
Is your sequence homologous to a known structure?
If no then actual models are less accurate, but structural insights still possible
First, secondary structure prediction

21. Secondary-structure prediction
algorithm
Neural networks
Inductive logic programming
Spin-glass theory
Human intuition

22. Secondary-structure prediction
E.g. Chou & Fasman, 1974
Helix forming: Glu, Ala, Leu
Helix breaking: Pro, Gly
Strand forming: Met, Val, Ile
Strand breaking: Glu, Lys, Ser, His, Asn
Etc.
Numerical approach + simple protocol =
prediction of secondary structure
Said “80%” accuracy. Reality: 50-60%
Tested the method on the same proteins used to
derive the parameters… big no-no.

23. Homologous proteins add
a lot of information
70% accuracy!
SS pred

24. What about de novo or ab initio prediction?
Can you simulate folding using physics to predict the structure of a protein
No, not usually.
However, advances have been made…
David Baker, co-workers and subsequent followers: fragment based structure prediction. De novo not ab initio

25. Predicting Fragments
Preferences learned from all stretches with a similar structure

26. Assembling Fragments Database of structures
Fragments matching the target sequence
Assembly of fragments
Selection of best model

27. The Prediction Irony
General trend: increasing accuracy is more a function of data than algorithms
In other words: as we know more structure, and indeed even sequence data, we get better at predicting
Probably we will have a perfect algorithm for protein structure prediction when we know all of the answers
Structural genomics & the generally increased pace of structure predictions means there aren’t many really “new” structures anymore

28. Things to Remember
Methods have mostly been developed for soluble, globular proteins or domains
Problems with membrane proteins, low-complexity, etc.
Many segments in proteins should be studied with other methods:
Signal peptides
TM regions
Coiled-coils
Intrinsic Disorder (e.g.

29. What we use this for…

30. Understand molecular interactions Predict molecular interactions
We aim to:
Understand molecular interactions
Predict molecular interactions
Focus on those interactions of biomedical importance
Apply tools to large datasets
Use interaction networks predictively
To predict new interactions
To predict other details like pathologies, toxicities

31. Modelling or predicting interactions by homology
Your favourite protein N C
Your second favourite protein N C
Match to
known structure
Match to
known structure
Templates
in contact?
Histidyl adenylate
tRNA Synthetase
Modelled Interaction

32. Prediction of Structures of Complexes
Five component complex
X-ray
Two-hybrid network
homology
(e.g. blast) +
Electron microscopy & Mass Spectometry
Russell et al, Curr. Opin. Struct Biol. 2004
Aloy & Russell, Nature Rev. Mol. Cell. Biol. 2006
Taverner et al, Adv Chem. Res. 2008

33. Adding Mechanisms to Interaction Networks
Who interacts with whom?
What does the interaction look like?
Ga/q
RGS-4 P
Ga/i
How strong? How fast?
RGS-3
Which piece from which protein?

34. Bridging the information gap
Modelled complexes
Aloy & Russell, Nature Rev. Mol. Cell. Biol., 2006.

35. From Proteomics to Cellular Anatomy?
Kuehner et al, Science, 2010

36. From Proteomics to Cellular Anatomy?
Kuehner et al, Science, 2010

37. Some Links
Guide to the amino acids
Guide to Structure Prediction
meta.bioinfo.pl
Meta server (runs virtually all reliable prediction methods)

38. Structure Prediction Practical
Structure
Active/inactive?
Binds/does not bind?
Substrate specificity?
Function
Sequence
In groups of two or more you will attempt to answer functional questions about a particular protein target

39. Acknowledgements www.russelllab.org Current group members
Rob Russell (the boss), Matthew Betts, Leonardo Trabuco, Oliver Wichmann, Mathias Utz, Yvonne Lara
Alumni
Chad Davis, Olga Kalinina, Ricardo de la Vega, Victor Neduva, Evangelia Petsalaki Damien Devos
Complex modeling & interactions collaborators
Patrick Aloy (IRB Barcelona)
Anne-Claude Gavin (EMBL Heidelberg)
Peer Bork (EMBL Heidelberg)
Luis Serrano (CRG Barcelona)
Achilleas Frangakis (Uni Frankfurt)
Bettina Boettcher (Edinburgh)