Bioinformatics CS 40400

1
BioinformaticsCS 40400

• Gianluca Pollastri
• office CS A1.07
• email gianluca.pollastri_at_ucd.ie

2
Lecture notes

• http//gruyere.ucd.ie/2007_courses/40400/
• confidential..

3
Recommended/useful readings

• No book is actually required
• Introduction to Computational Molecular Biology
• Setubal, Meidanis
• Introduction to Bioinformatics
• Lesk
• Bioinformatics the Machine Learning approach
• Baldi, Brunak
• Biological sequence analysis (but this is a tough
  one)
• Eddy, Durbin, Krogh, Mitchison

4
(No Transcript)
5
The course so far..

• Introduction proteins, RNA, DNA. GenBank,
  SWISS-PROT, PDB, ExPaSy.
• Sequence comparison. Needleman-Wunsch and
  Smith-Waterman algorithms. Semiglobal comparison.
  Variations to basic algorithms. Approximate
  algorithms BLAST. Multiple sequence alignments.
  Approximate algorithms ClustalW
• Molecular phylogenetics. Distance based
  algorithms UPGMA, Neighbour Joining. Maximum
  parsimony. Maximum Likelihood. Rooting trees.
  Estimating times. Bootstrapping.

6
Whats next?

• Protein structure prediction Comparative
  Modelling Threading De novo. Introduction to
  artificial neural networks. Prediction of protein
  structural features by machine learning
  algorithms secondary structure, solvent
  accessibility, contact maps.

7
Protein Structure Prediction and Structural
Genomics
Loosely based on

• David Baker and Andrej Sali

8

• MKTLVHVASV EKGRSYEDFQ KVYNAIALKL REDDEYENYI
  GYGDDLVRLA
• WHISGTWDKH DNTGGSYGGT YRFKKEFNDP SNAGLQNGFK
  FLEPIHKEFP
• WISSGDLFSL GGVTAVQEMQ GPKIPWRCGR VDTPEDTTPD
  NGRLPDADKD
• AGYVRTFFQR LNMNDREVVA LMGAHALGKT HLKNSGYEGP
  WGAANNVFTN
• EFYLNLLNED WKLEKNDANN EQWDSKSGYM MLPTDYSLIQ
  DPKYLSIVKE
• YANDQDKFFK DFSKAFEKLL ENGITFPKDA PSPFIFKTLE EQGL

9
(No Transcript)
10
Why?

• 1.7M protein sequences, cheap product of genome
  sequencing projects.
• 29k high resolution protein structures.
  Determined by X-ray christallography, NMR
  painful, costly and time consuming.
• Sequence determines Structure, Structure
  determines Function. This is why we want to know
  the structure..

11
(No Transcript)
12
ATOM 1 N LEU A 4 12.803 88.583
75.298 1.00 25.27 N ATOM 2 CA
LEU A 4 12.284 89.166 74.064 1.00 25.73
C ATOM 3 C LEU A 4
11.896 88.062 73.094 1.00 24.57 C
ATOM 4 O LEU A 4 12.740 87.441
72.459 1.00 21.83 O ATOM 5 CB
LEU A 4 13.283 90.098 73.387 1.00 25.28
C ATOM 6 CG LEU A 4
12.714 91.348 72.710 1.00 30.06 C
ATOM 7 CD1 LEU A 4 13.446 91.644
71.405 1.00 23.16 C ATOM 8 CD2
LEU A 4 11.221 91.221 72.456 1.00 27.13
C ATOM 9 N VAL A 5
10.588 87.839 72.988 1.00 19.24 N
ATOM 10 CA VAL A 5 10.180 86.742
72.108 1.00 22.98 C ATOM 11 C
VAL A 5 9.286 87.293 71.005 1.00 25.18
C ATOM 12 O VAL A 5
8.388 88.103 71.215 1.00 19.54 O
ATOM 13 CB VAL A 5 9.527 85.607
72.915 1.00 36.54 C ATOM 14 CG1
VAL A 5 8.876 86.145 74.185 1.00 68.19
C ATOM 15 CG2 VAL A 5
8.518 84.844 72.075 1.00 42.84 C
ATOM 16 N HIS A 6 9.594 86.832
69.801 1.00 19.98 N ATOM 17 CA
HIS A 6 8.898 87.164 68.570 1.00 14.76
C ATOM 18 C HIS A 6
8.153 85.933 68.072 1.00 13.19 C
ATOM 19 O HIS A 6 8.794 85.029
67.536 1.00 12.12 O ATOM 20 CB
HIS A 6 9.900 87.636 67.521 1.00 15.61
C ATOM 21 CG HIS A 6
10.488 88.969 67.851 1.00 16.42 C
ATOM 22 ND1 HIS A 6 11.808 89.287
67.631 1.00 17.91 N ATOM 23 CD2
HIS A 6 9.916 90.073 68.382 1.00 12.99
C ATOM 24 CE1 HIS A 6
12.036 90.531 68.009 1.00 10.96 C
ATOM 25 NE2 HIS A 6 10.904 91.032
68.472 1.00 17.40 N ATOM 26 N
VAL A 7 6.839 85.922 68.277 1.00 10.72
N ATOM 27 CA VAL A 7
6.048 84.781 67.851 1.00 11.90 C
ATOM 28 C VAL A 7 5.539 85.014
66.423 1.00 17.11 C ATOM 29 O
VAL A 7 4.938 86.053 66.131 1.00 8.14
O ATOM 30 CB VAL A 7
4.833 84.488 68.746 1.00 12.98 C
ATOM 31 CG1 VAL A 7 4.223 83.146
68.336 1.00 11.94 C ATOM 32 CG2
VAL A 7 5.188 84.475 70.218 1.00 14.19
C
13
(No Transcript)
14
Simulating nature?

• We probably dont know the physics well enough
  (or rather, we know it well on an intractably
  small scale)
• Ugly landscapes to search.
• An enormous amount of time steps needed.
• Computationally intractable.
• We need to
• start close to the solution
• approximate/simplify

15
Methods for 3D prediction

• If there are proteins of known structure that
  look like the one I want to model, comparative
  modelling (CM) or threading/fold recognition
  methods available (starting close to the solution
  and possibly simplify).
• If there arent, we use de novo (or ab initio)
  methods (cant start close to the solution we
  need to simplify/approximate).

16
Comparative Modelling (CM)

• Find proteins of known structure whose sequence
  looks like my sequence (templates).
• Align sequence and template(s)
• Build a model
• Figure out if the model makes sense

17
Structure more conserved than sequence

• If two sequences are more than 30 similar,
  strong structural similarity almost guaranteed.
• (Average similarity of unrelated sequences around
  7)

18
CM

• Find proteins of known structure whose sequence
  looks at least 30 like my sequence (templates).
• Align sequence and template(s)
• ..

19
Sequence similarity

• Query 3 FEFHGYARSGVIMNDSGASTKSGAYITPAGETGGAIGRL
  GNQADTYVEMNLEHKQTLDNG 62
• FEFH YAR V MND A K AY PA E A
  RL NQAD YVEMNLEHKQ LDN
• Sbjct 3 FEFHHYARCHVHMNDCHACCKCHAYHCPAHECHHAHHRL
  HNQADCYVEMNLEHKQCLDNH 62
• Query 63 ATTRFKVMVADGQTSYNDWTASTSDLNVRQAFVELGNLP
  TFAGPFKGSTLWAGKRFDRDN 122
• A RFKVMVAD Q YNDW A DLNVRQAFVEL
  NLP FA PFK LWA KRFDRDN
• Sbjct 63 ACCRFKVMVADHQCCYNDWCACCCDLNVRQAFVELHNLP
  CFAHPFKH--LWAHKRFDRDN 120
• Query 123 FDIHWIDSDVVFLAGTGGGIYDVKWNDGLRSNFSLYGRN
  FGDIDDSSNSVQNYILTMNHF 182
• FD HW D DVVFLA YDVKWND LR NF LY
  RNF D DD N VQNY L MNHF
• Sbjct 121 FDHHWHDCDVVFLAHCHHHHYDVKWNDHLRCNFCLYHRN
  FHDHDDCCNCVQNYHLCMNHF 180

20
CM finding templates, aligning sequence and
templates

• Sequence comparison methods
• Exact version (complexity o(nm) where n and m are
  sequence lengths) a bit demanding.
• Linear approximations (blast, psi-blast)
  aligning a sequence vs 1M sequences takes tens of
  seconds.

21
CM building/assessing model

• Copy template or parts thereof (to start close to
  the solution)..
• Fondle it a bit and assess fondling by physical
  energy or pseudo-energy.

22
Threading/Fold Recognition

• If no sequence similarity is detected
• Find proteins of known structure (templates) by
  some other method (not sequence comparison)
• Align sequence and template(s)
• Build a model
• Figure out if the model makes sense

23
Threading finding templates

• Libraries of folds.
• Thread the sequence into each of the folds and
  check if it has low energy in one or more of them.

24
Threading

• Energy computations constrained by folds.
• This is a lot simpler (quicker) than
  unconstrained search.
• Still 1-2k folds

25
Threading building/assessing model

• Copy template or parts thereof (to start close to
  the solution)..
• Fondle it a bit more than in CM and assess
  fondling by physical energy or pseudo-energy.

26
De novo prediction

• No sequence similarity with proteins of known
  structure detected.
• No fold where threading is possible at acceptable
  energy levels.

27
De novo prediction note

• Sequence similarity methods are not perfect.
• Threading methods are far from perfect
  (especially energy functions).
• Often de novo methods are used for proteins whose
  structure does resemble a known one.

28
De novo how does it work?

• It usually does not.
• (neither does threading)
• Search for a minimum of some energy function. Key
  actors
• How we search the space of 3D configurations.
• The energy function we use.

29
De novo simplify

• An all-atom model is computationally heavy
• Only some atoms are modelled (e.g. backbone
  atoms).

30
De novo simplify more

• An all-atom model is computationally heavy
• Whole stretches of atoms are modelled together.

31
Choosing the stretches

• Regular local structures (helices, strands) are a
  natural modelling unit.
• We dont know where they are.
• Machine learning.
• Huge field.

32
(No Transcript)
33
(No Transcript)
34
The energy function

• Purely physical functions are not accurate enough
  for all-atom models, are very inaccurate for
  coarser models, and generally dont provide
  decent landscapes to search.
• Pseudo-energies huge room for machine learning,
  e.g. contact map prediction.

35
Contact maps

• Amino acid adjacency map.
• Invariant to rotations and translations, unlike
  xyz coordinates.
• Maps with 50-60 uniform random noise compatible
  with correct 3D structure.

36
Do CM, threading, de novo work?

• CM works, so long as the template is correct (it
  often is)
• Threading works, so long as the template is
  correct (it never is)
• De novo in some cases produces one model out of 5
  that is correct over a short stretch of amino
  acids (80?).

37
(No Transcript)