BNFO 602, Lecture 3 Usman Roshan Some of the slides are based upon material by David Wishart of University of Alberta and Ron Shamir of Tel Aviv University
Previously… Pairwise sequence alignment problem –Applications of alignments –Dynamic programming matrix –Traceback –Local alignment --- finding maximal local matches
Previously… Database searching –FASTA –BLAST Scoring matrices –PAM
Multiple sequence alignment “Two sequences whisper, multiple sequences shout out loud”---Arthur Lesk Computationally very hard---NP-hard
Formally…
Multiple sequence alignment Unaligned sequences GGCTT TAGGCCTT TAGCCCTTA ACACTTC ACTT Aligned sequences _G_ _ GCTT_ TAGGCCTT_ TAGCCCTTA A_ _CACTTC A_ _C_ CTT_ Conserved regions help us to identify functionality
Sum of pairs score
What is the sum of pairs score of this alignment?
Tree alignment score
Tree Alignment TAGGCCTT (Human) TAGCCCTTA (Monkey) ACCTT (Cat) ACACTTC (Lion) GGCTT (Mouse)
Tree Alignment TAGGCCTT_ (Human) TAGCCCTTA (Monkey) A__C_CTT_ (Cat) A__CACTTC (Lion) _G__GCTT_ (Mouse) TAGGCCTT_A__CACTT_ TGGGGCTT_ AGGGACTT_ Tree alignment score = 14
Tree Alignment---depends on tree TAGGCCTT_ (Human) TAGCCCTTA (Monkey) A__C_CTT_ (Cat) A__CACTTC (Lion) _G__GCTT_ (Mouse) TA_CCCTT_ TA_CCCTTA TA_CCCTT_ TA_CCCTTA Tree alignment score = 15 Switch monkey and cat
Profiles Before we see how to construct multiple alignments, how do we align two alignments? Idea: summarize an alignment using its profile and align the two profiles
Profile alignment
Iterative alignment (heuristic for sum-of-pairs) Pick a random sequence from input set S Do (n-1) pairwise alignments and align to closest one t in S Remove t from S and compute profile of alignment While sequences remaining in S –Do |S| pairwise alignments and align to closest one t –Remove t from S
Iterative alignment Once alignment is computed randomly divide it into two parts Compute profile of each sub-alignment and realign the profiles If sum-of-pairs of the new alignment is better than the previous then keep, otherwise continue with a different division until specified iteration limit
Progressive alignment Idea: perform profile alignments in the order dictated by a tree Given a guide-tree do a post-order search and align sequences in that order Widely used heuristic Can be used for solving tree alignment
Simultaneous alignment and phylogeny reconstruction Given unaligned sequences produce both alignment and phylogeny Known as the generalized tree alignment problem---MAX-SNP hard Iterative improvement heuristic: –Take starting tree –Modify it using say NNI, SPR, or TBR –Compute tree alignment score –If better then select tree otherwise continue until reached a local minimum
Median alignment Idea: iterate over the phylogeny and align every triplet of sequences---takes o(m 3 ) (in general for n sequences it takes O(2 n m n ) time Same profiles can be used as in progressive alignment Produces better tree alignment scores (as observed in experiments) Iteration continues for a specified limit
Popular alignment programs ClustalW: most popular, progressive alignment MUSCLE: fast and accurate, progressive and iterative combination T-COFFEE: slow but accurate, consistency based alignment (align sequences in multiple alignment to be close to the optimal pairwise alignment) PROBCONS: slow but highly accurate, probabilistic consistency progressive based scheme DIALIGN: very good for local alignments
MUSCLE
Profile sum-of-pairs score Log expectation score used by MUSCLE
Evaluation of multiple sequence alignments Compare to benchmark “true” alignments Use simulation Measure conservation of an alignment Measure accuracy of phylogenetic trees How well does it align motifs? More…
BAliBASE Most popular benchmark of alignments Alignments are based upon structure BAliBASE currently consists of 142 reference alignments, containing over 1000 sequences. Of the 200,000 residues in the database, 58% are defined within the core blocks. The remaining 42% are in ambiguous regions that cannot be reliably aligned. The alignments are divided into four hierarchical reference sets, reference 1 providing the basis for construction of the following sets. Each of the main sets may be further sub-divided into smaller groups, according to sequence length and percent similarity.
BAliBASE The sequences included in the database are selected from alignments in either the FSSP or HOMSTRAD structural databases, or from manually constructed structural alignments taken from the literature. When sufficient structures are not available, additional sequences are included from the HSSP database (Schneider et al., 1997). The VAST Web server (Madej, 1995) is used to confirm that the sequences in each alignment are structural neighbours and can be structurally superimposed. Functional sites are identified using the PDBsum database (Laskowski et al., 1997) and the alignments are manually verified and adjusted, in order to ensure that conserved residues are aligned as well as the secondary structure elements.FSSP HOMSTRADHSSP VAST PDBsum
BAliBASE Reference 1 contains alignments of (less than 6) equi- distant sequences, ie. the percent identity between two sequences is within a specified range. All the sequences are of similar length, with no large insertions or extensions. Reference 2 aligns up to three "orphan" sequences (less than 25% identical) from reference 1 with a family of at least 15 closely related sequences. Reference 3 consists of up to 4 sub-groups, with less than 25% residue identity between sequences from different groups. The alignments are constructed by adding homologous family members to the more distantly related sequences in reference 1. Reference 4 is divided into two sub-categories containing alignments of up to 20 sequences including N/C-terminal extensions (up to 400 residues), and insertions (up to 100 residues).
Comparison of alignments on BAliBASE
Parsimonious aligner (PAl) 1.Construct progressive alignment A 2.Construct MP tree T on A 3.Construct progressive alignment A’ on guide-tree T 4.Set A=A’ and go to 3 5.Output alignment and tree with best MP score
PAl Faster than iterative improvement Speed and accuracy both depend upon progressive alignment and MP heuristic In practice MUSCLE and TNT are used for constructing alignments and MP trees How does PAl compare against traditional methods? PAl not designed for aligning structural regions but focuses on evolutionary conserved regions Let’s look at performance under simulation
Evaluating alignments under simulation We first need a way to evolve sequences with insertions and deletions NOTE: evolutionary models we have encountered so far do not account for insertions and deletions Not known exactly how to model insertions and deletions
ROSE Evolve sequences under an i.i.d. Markov Model Root sequence: probabilities given by a probability vector (for proteins default is Dayhoff et. al. values) Substitutions –Edge length are integers –Probability matrix M is given as input (default is PAM1*) –For edge of length b probabilty of x y is given by M b xy Insertion and deletions: –Insertions and deletions follow the same probabilistic model –For each edge probability to insert is i ins. –Length of insertion is given by discrete probability distribution (normally exponential) –For edge of length b this is repeated b times. Model tree can be specified as input
Evaluation of alignments Let’s simulate alignments and phylogenies and compare them under simulation!!
Parameters for simulation study Model trees: uniform random distribution and uniformly selected random edge lengths Model of evolution: PAM with insertions and deletions probabilities selected from a gamma distribution (see ROSE software package) Replicate settings: Settings of 50, 100, and 400 taxa, mean sequence lengths of 200 and 500 and avg branch lengths of 10, 25, and 50 were selected. For each setting 10 datasets were produced
Phylogeny accuracy
Alignment accuracy
Running time
Conclusions DIALIGN seems to perform best followed by PAl, MUSCLE, and PROBCONS DIALIGN, however, is slower than PAl Does this mean DIALIGN is the best alignment program?
Conclusions DIALIGN seems to perform best followed by PAl, MUSCLE, and PROBCONS DIALIGN, however, is slower than PAl Does this mean DIALIGN is the best alignment program? Not necessarily: experiments were performed under uniform random trees with uniform random edge lengths. Not clear if this emulates the real deal.
Conclusions
Sum-of-pairs vs MP score
Conclusions Optimizing MP scores under this simulation model leads to better phylogenies and alignments
Conclusions Optimizing MP scores under this simulation model leads to better phylogenies and alignments What other models can we try?
Conclusions Optimizing MP scores under this simulation model leads to better phylogenies and alignments What other models can we try? Real data phylogenies as model trees Birth-death model trees Other distributions for model trees… Branch lengths: similar issues… Evolutionary model parameters estimated from real data