Fold Prediction

1
Fold Prediction

• -Huzefa Rangwala
• (rangwala_at_cs.umn.edu

2
Problem Definition

• Hierarchy
• According to the SCOP database proteins are
  classified to reflect both the structural
  evolutionary relationships
• Same family gt clear evolutionary relationship
• Same super-family gt probably common evolutionary
  origin
• Same fold gt Major structural similarity
• Fold Prediction
• To recognize similarity at all the above three
  levels by defining independent problems.
• Sequence-Structure Homology Recognition
• To recognize similarity at only family and
  super-family levels
• Does this imply that at the fold prediction level
  structural information will play a significant
  part compared to just sequence information?

3
Techniques (Groups)

• Pairwise sequence comparisons recognize close
  homologs
• Use of Profiles/ HMM relying only on sequence
  information are able to capture distant homolog
  relationships
• Second scheme some structural information
  improve the accuracy
• Eg- Fugue, HMAP, 3D-PSSM (Todays focus)
• Threading methods (sequence comparisons with a
  structural template) Show a significant
  performance for the fold level

4
Schemes to be discussed

• 3D-PSSM
• FUGUE
• HMAP
• AGAPE

5
Evaluation Ranking

• Each of the schemes will give us a similarity
  score between two protein domains whose
  relationship we are trying to figure out i.e
  whether they are part of the same family,
  superfamily or fold?
• We sort the pairs of relationships based on the
  similarity scores and get an ordered rank to
  compute true positives and false positives. Based
  on these two statistics we compute the ROC,
  Coverage, Specificity or Sensitivity.
• Defining True Positive False Positive
• TP At the superfamily level domain pairs that
  share the same superfamily but different family
  level. (Same family relationships are to trivial
  to predict)
• FP Having different fold levels are considered
  as FPs. If the pairs have the same folds but
  different superfamily they are neglected.
• Similar definitions hold for the other benchmark
  levels
• Different schemes use this same strategy but
  restrict the pairs they compare. HMAP, 3D-PSSM
  does the overall ranking to compute the
  statistics whereas FUGUE defined these statistics
  on a per test domain level

6
Evaluation Alignment Accuracy

• HMAP and AGAPE talk about evaluation based on
  comparison to structural alignments.
• HMAP uses the structural alignment between PrISM
  between pairs as the gold standard.
• Two measures Full accuracy and core accuracy
  (positions within the secondary structure
  regions) are evaluated.

7
SS 3D-PSSM (Kelley et al.)

• Goal enhance fold detection by using profiles
  that leverage structural similarity.
• Overview
• Other fold-detection methods compute distance
  between a query sequence (unknown fold) and a
  library sequence (known fold) using a standard
  PSSM.
• 3D-PSSM uses extra information
• Solvation potentials per residue
• Secondary structure
• 3D-PSSM obtained from a super PSSM calculated
  from a structural alignment between sequences in
  the super-family

8
SS 3D-PSSM (residue features)

• Residue features and scoring functions between
  query and library sequences
• Degree of burial - Score is solvation
  potentials per residue (frequency of occurrence
  of amino acid with a specific degree of burial
  relative to all amino acid types with this degree
  of burial).
• Secondary structures - as calculated by STRIDE
  for library, PSI-Pred for query. Score is 1 for
  matching elements, -1 for mismatches.
• 1D / 3D-PSSM - (next slide). Score is profile
  to profile score, as in PSI-Blast.

9
3D Profile Whats going on ?
10
SS 3D-PSSM (building 3D-PSSM)

• 3D-PSSM for a master sequence (built offline)
• Perform a 3D alignment using SAP (Orengo, 1992)
  between master sequence and all sequences in
  superfamily.
• Select similar structures (RMS lt 6A)
• Create a structural MSA
• Start with the closest sequence to the master
• Iteratively align the sequence the sequence
  structurally closest to any of the sequences
  already in the MSA
• The 3D-PSSM is the cumulative PSSM made by
  combining all aligned PSSM's.

11
SS 3D-PSSM (searching)

• For a given query, compute the score to every
  library sequence
• Use a global DP alignment with pairwise scores
• S(Xi, Yj) Spssm(Xi,Yj)
  Ssecondary(Xi,Yj) Ssolvation(Xi,Yj)
• Perform three different DP passes
• Library sequence is matched to query PSSM
• Query sequence is matched to library 1D-PSSM
• Query sequence is matched to library 3D-PSSM
• Take the maximum score from the three passes

12
SS3D-PSSM (evaluation)

• Now that we have a score for a sequence pair, we
  need to determine it's significance.
• We do this by fitting a linear function to
• log(number of hits up to a score) vs.
• log(sum of all scores)
• Effectiveness
• 136 sequences in the test set that met critera
• PSI-Blast assignment failed
• There was a homology to a protein with a 3D
  profile
• Only kept one test sequence per 3D profile.
• 18 of the 136 sequences were correctly classified

13
CKFUGUE Remote
Homology Detection using Known Structure
Information

• Big Picture FUGUE constructs structural profiles
  for families of known structures. Each profile is
  converted into a scoring matrix (like PSI-BLAST)
  which is used to align the sequence. (z-Scores)
  Random alignments are conducted to obtain mean
  scores for a structural profile and a query
  sequence is considered a aligned well if its
  score is significantly higher than the mean.

14
CK Basics of FUGUE
Alignment

• Start with a collection of structure alignments -
  source is pruned version of HOMSTRAD 177
  families (aligned), 706 total structures
• Want to build a scoring matrix based on the
  structures in a family
• Count substitutions between amino acids in
  (structurally) aligned positions in the families
• Catch When residue A and B appear in aligned
  positions, don't just increment the counts - also
  consider the environment in which the two appear.

15
CK Environment

• Three categories of environments were considered,
  each with several classes.
• Secondary structure alpha-helix, beta-strand,
  irregular structure (coil), or positive phi
  main-chain angle (does this mean main chain, i.e.
  no 2ndary structure??) - 4 classes
• Solvent accessibility 7 or greater is
  accessible, o/w inaccessible - 2 classes
• Hydrogen bonds Combination of 3 T/F conditions,
  side-chain to side-chain or not, side-chain to
  main-chain NH or not, side-chain to main-chain CO
  or not - 8 classes
• An environment is a specification of class for
  each of these 3 categories - total of 64
  environments. Each residue in the structural
  families fits into one of these classes

16
CK Substitution Matrix Formulation

• Construct a substitution matrix of log odds
  scores, each entry is log(P(BA,E))/q_B q_B is
  background prob of B
• Background probability is not taken from a Blosum
  matrix calculated from the occurrence of a
  residue in the structure family (eqn 3??)
• Entries are actually a weighted combination of
  the above log-odds score and the score from an 'a
  priori' distribution calculated based on another
  study
• Some residues excluded (masked) from substitution
  calculation domain-domain interaction residues
  and interaction with heteroatoms(??)

17
CK Example of Substitution Matrix Calc
18
CK Profile Calculation

• Calculate a weight for each structure based on
  the sum of its dis-similarity to all other
  structures divided by total dis-similarity of all
  structures in the family (dis-similarity is
  fraction of identical residues (??))
• The scoring matrix has as many columns as the
  longest structure and as many rows as there are
  structures in the family alignment. Entries in
  the scoring matrix are calculated as follows
• To calculate entry at column P (position) and row
  B (amino acid), do the following for every
  structure. Its amino acid at position P is A in
  environment E. Add to the scoring matrix entry
  the weight of the structure multiplied by
  S(A,E?B) a score from the substitution matrix.

19
CK Gap Penalty Calculations

• Gap penalties modified based on secondary
  structure of position
• Insertion/Deletion highly penalized in the middle
  of secondary structure region highly penalized,
  in unordered (coil) regions lightly penalized
• Overall penalty at a position determined by
  weighted sum of penalties for each structure at
  that position
• Numerical penalty scores (high, low, very low,
  etc) determined numerically from random
  simulation
• Position specific gap scores calculated for each
  family, modified by solvent accessibility (??)

20
CK Evaluation

• Used a benchmark defined in Shi et. al 976
  domains from SCOP.
• For domain D (out of the test set), run FUGUE
  with D as the query sequence to generate a list
  of scores for each other domain. Sort these
  scores. Calculate specificity and sensitivity
  curves. Each domain D has a psi-blast profile.
• Compared performance at finding relatives in SCOP
  family, super-family, and fold levels
• For family level, count TPs as family members.
  For super-family level, count TPs as super-family
  members - family members. Likewise for fold level
  (fold members - SF - F)
• Outperformed other methods on family and
  super-family, beaten by THREADER at fold level.
• Several other methods used to evaluate various
  aspects of FUGUE

21
HMAP Hybrid multidimensional alignment profiles

• Big Picture A series of hybrid multidimensional
  alignment profiles that combine sequence,
  secondary and tertiary structure information were
  developed. They conclude with their evaluation
  that secondary structure plays a significant part
  in remote homology detection over 3D structures.

22
HMAP The setting
23
KD,NWHMAP Details

• Selecting the correct subset of sequences
  structures
• No sequence should have greater than 40 sequence
  identity
• Structures of at least 2.5 Angstrom
• More than 50 SSE are excluded
• Sequence-based profiles are generated using
  PSI-BLAST
• Database is restricted by CD-HIT to 65 identity

24
KD,NWHMAP Details

• Multiple structure based sequence alignments and
  structure based profiles
• For each template structure based on the PSD
  score (lt1) get the closest neighbors.
• Sequence for each of these neighbors is used as a
  seed into PSI-BLAST.
• Keep only those sequences above a certain
  threshold
• Purged sequence alignments are combined with a
  sequence using multiple structure alignment as a
  guide. (SEMSA)
• Get a similar PSSM -gt 3D Profile

25
KD,NWHMAP Details

• Secondary Structure Profiles
• Just like in the previous step from the multiple
  structure alignment, create a secondary profile
  where instead of the 20 amino acids we use the
  three secondary structure elements
• Motifs
• In regions where they are confident (core/motif)
  about the structure from the alignments, they
  incorporate structural information. In regions
  where they are not confident (loops), they revert
  to 1D info.

26
KD,NWHMAP(1,2) HMAP(1,2,3)

• HMAP(1,2) 1d Profile secondary structure
  profiles
• HMAP(1,2,3) interleaving of 1d profile with 3d
  profile based on motif information secondary
  structure information.

27
KD,NWQuery Profiles Alignments

• Query profiles are generated using PSI Blast
  profile and predicted secondary structure
  information (predicted using JNET, PSIPRED and
  PHD - averaging).
• Simple DP but the scoring is a dot product of the
  profiles weighted by the secondary structure as
  well as the conservation factors. (NEXT SLIDE)
• Specific-Structure gap penalties like the FUGUE
  but the implementation is that the penalties are
  based on the secondary structure profiles.

28
Similarity score for HMAP
29
KD,NWPSD Details
30
KD,NWPSD Details

• a of SSEs for protein A
• S(A,B) the score from the double dynamic
  programming algorithm aligning SSEs of the pair
  of proteins A and B.
• x and y are experimentally determined parameters

31
HMAP Evaluation

• HMAP evaluation was done on the ranking to
  compute the coverage vs accuracy defined earlier
  and it was seen that 1d-profile was better than
  PSI-Blast ranking
• The performance of HMAP(1,2) was better than
  HMAP(1,2,3) for the super-family level and
  comparable at the fold level. Can we conclude if
  3D information does add anything significant for
  the remote homology detection problem?

32
DRAGAPE Predicting Without
Folds

• Big Picture Appending profile information with
  secondary structure and solvent accessibility.
  They claim that use of explicit 3D structural
  information as used in other schemes may not be
  very helpful
• They build generalized profiles based on 1D
  structure predictions (i.e prediction of
  secondary structure and solvent accessibility
  using PROFphd). Some preliminary testing is done
  on known structures using DSSP
• Six dimensions are added to the profile based on
  the predictions i.e either buried or exposed and
  helix, strand or other

33
DR Generalized PSSM

• GPSSM(i, j, 1D(i), 1D(j))
• (r) PSSM(i, j) (1-r) SM(1D(i), 1D(j))
• GPPS Generalized Position-Spec. Scoring Matrix
• PSSM from PSI-BLAST
• SM 1D substitution matrix (learnt by aligning
  predicted predicted or known-known or
  combinations of 1D structures)
• i query position j database position

34
DR Methods Fold recognition
(SW-PSSM)

• Obtaining and comparing 1D structures
• O2O Observed against Observed (DSSP)
• P2O Predicted (PROFphd) against Observed
• P2P
• P2P-Bis
• Predicted profile against predicted DB sequence
  AND
• Predicted profile sequence against DB profile

35
DR A simple DP solution
Image from Protein fold recognition by
prediction-based threading. Journal of molecular
biology 0022-2836 Rost yr 1997 vol 270 iss 3
pg 471
36
DRStatistical Significance

• Fitting the scores to an extreme value
  distribution and modifying the scores to get a
  significance estimate thereby gaining an
  improvement in the ranking.
• They introduced bidirectional scoring which shows
  improved performance

37
DR(Good?) Prediction Errors

• P2P does better without using observed because
  the errors correlate
• i.e., 1D prediction mistakes correlate between
  proteins with dissimilar sequences but similar
  structures
• Meanings/implications (?)

38
DRResults

• Using 1D data is
• Always better that sequence only methods (BLAST)
• Currently better than some that use 3D
  information
• However, in the long-run, 3D-including methods
  will probably be better
• Slow uses DP
• Why? Though all the other methods also use some
  sort of DP

39
Statistical Significance

• Fugue z-Score
• HMAP Mixed Distribution Fit
• The scores are made to fit a mixed distribution
  which was estimated empirically based on
  shuffling amino acid profiles and aligning
  shuffled sequences to the original with positions
  of secondary structure elements fixed.
• 3D-PSSM
• The significance of a match was evaluated by
  fitting a linear relationship between log (number
  of hits up to a score) against log (total score)
• AGAPE Extreme Value Distribution
• The significance of a match was evaluating by
  fitting it to an extreme value distribution.