1. Remote Homology Detection of Beta-Structural Motifs Using Random Fields
Matt Menke, Tufts
Bonnie Berger, MIT
Lenore Cowen, Tufts
ISMB 3Dsig 2010
July 10, 2010

2. Inferring structural similarity from homology is hard at the SCOP superfamily/fold level

3. Profile HMMs 3

4. HMM is trained from Sequence Alignment of Known Structures
But: cannot capture pariwise
long-range beta-sheet interactions!

5. Pectate Lyase C (Yoder et al. 1993)
HMMs cannot capture statistical preferences from residues close in space but far, and a variable distance apart in seq.
Pectate Lyase C (Yoder et al. 1993)

6. Look at Just Pairs or Generalize to Markov Random Fields
Only look at Pairs:
Generalize to Markov Random Fields
Liu et al. 2009
Zhao et al. 2010
Menke et al. 2010
(This work) B3 T2 B2 B1
[Bradley, Cowen, Menke, King, Berger, PNAS, 2001, 98:26, 14,819-14,824 ;
Cowen, Bradley, Menke, King, Berger (2002), J Comp Biol, 9, ]

7. Let’s look at what this would mean for propeller folds

8. SCOP (http://scop.mrc-lmb.cam.ac.uk/scop
Goal: capture HMM sequence information and pairwise information in beta-structural motifs at the same time!
SCOP (

9. Structural Motifs Using Random Fields
SMURF

10. Structural Motifs Using Random Fields
Can we get
the benefit
of pairwise
correlations
without having
to throw away
all sequence info?

11. The template is learned from solved structures in the PDB

12. The template is learned from solved structures in the PDB: Aligned with Matt

13. Digression: Matt structural alignment program
Menke, Berger, Cowen, (PLOS Combio 2008)
Specifically designed to align more distant homologs
AFP chaining using dynamic programming with “translations and twists”
(flexibility)

18. The template is learned from solved structures in the PDB: Aligned with Matt

19. Two beta tables are learned from amphapathic beta sheets that are not propellers from solved structures in the PDB. A C D E F G H I K L M N P Q R S T V W Y
0.78
0.18
0.14
0.15
0.59
0.70
0.06
1.06
0.07
1.19
0.17
0.12
0.05
0.11
0.08
0.22
0.25
1.53
0.27
0.24
0.03
0.28
0.34
0.02
0.01
0.39
0.10
0.16
0.26
0.40
0.57
0.19
0.66
0.61
0.13
1.35
0.43
0.58
0.77
1.13
0.23
0.09
0.31
1.27
0.48
0.04
2.27
2.21
0.38
0.29
0.45
2.56
0.42
0.00
2.96
0.33
0.36
2.64
0.50
0.49
0.44
3.74
0.64
Two pairwise
Exposed Residue A C D E F G H I K L M N P Q R S T V W Y
0.27
0.04
0.13
0.28
0.22
0.18
0.11
0.31
0.23
0.38
0.06
0.37
0.49
0.25
0.08
0.05
0.07
0.03
0.02
0.01
0.10
0.09
0.71
0.12
0.15
0.50
0.36
0.41
0.24
0.43
0.21
1.92
0.14
1.49
0.60
1.01
0.63
0.32
0.16
0.34
0.19
0.29
0.33
0.17
0.20
0.48
0.57
0.30
0.59
0.40
0.46
0.42
0.70
1.17
0.52
0.26
0.62
0.39
0.47
0.68
0.72
0.91
0.88
1.60
0.82
0.87
0.64
Buried Residue

20. Computing a Score
Sequences are scored by computing their best “threading” or “parse” against the template as a sum of HMM(score) + pairwise(score)
No longer polynomial time (multi-dimensional dynamic programming)
Tractable on propellers because paired beta-strands don’t interleave too much

21. Let’s look at what this would mean for propeller folds

22. Let’s look at what this would mean for propeller folds
Training set for HMM score: leave-superfamily-out cross validation
Training set for pairwise score: amphapathic beta-sheets from NON-propellers

23. Results on Propellers 6-bladed 7-bladed TNeg Hmmer Smurf 97% 52 80 87
96% 56
95% 64 93
94% 68 84 90
93%
92% 88 97
91% 92
90%
100

24. Results on Propellers
Note that this is “6 (or 7)” bladed propeller versus non-propeller– distinguishing the number of blades in the propeller seems to be a much harder problem….

25. Different propeller closures
1jof trc

26. So: what new sequences fold into propellers?
We predict a double propeller motif in the N-terminal region of a hybrid 2-component sensor protein.

27. What are these proteins?
First found in a benign bacteria in human gut.
May be involved in adapting to changes in diet/efficiently processing different sugars
Found in other bacterial species: help sense and adapt to environmental changes.
Big stretch (I am not a biologist): help to study human obesity epidemic??

28. Popular Domains HisKA histidine kinase domain
GGDEF adenylyl cyclase signalling domain
SpoIIE sporulation domain
Gaf domain
PAS domain
HATPase domain

29. Species distribution

30. Distinguishing Number of Blades
The automatic SMURF consensus 7-bladed template only learns 6 blades.
Sequence motifs are similar– the same Pfam motif occurs in propellers with different numbers of blades
The fix: throw out propellers with a “funky” 7th blade by hand and build a new template. Now 6-bladed propellers don’t like the 7-bladed template
Double propellers we found are probably 7-7 (but 7-6 is also plausible).

31. Predict propellers with Smurf!
Accepts sequences in FASTA format
6,7,8-bladed templates, as well as all 9 double-propeller template
pairwise tables
long list of predicted propeller sequences

32. What’s Next for SMURF? Long-range dependencies
Deeply interleaved β-strand pairs

33. Conclusions
Combining an HMM score with a pairwise score can help recognize beta-structures
Computing this score exactly with a random field is highly computationally intensive
We will begin to look at when it is feasible and when we should use heuristics.
Also: add side-chain packing, other model refinements.

34. More Questions
When should we over-weight the HMM versus the pair portion of the score?
-- the case of 8-bladed propellers
Are there other ways to incorporate pairwise dependencies into HMMs?

35. An Hmm is only as good as its training data
An Hmm is only as good as its training data– or is it?
Idea: we augment the training set, using the simplest model of evolution!
See Kumar and Cowen’s ISMB proceedings paper!

36. Acknowledgements
National Institutes of Health
Thank you!