Presentation is loading. Please wait.

Presentation is loading. Please wait.

Guiding Motif Discovery by Iterative Pattern Refinement Zhiping Wang, Mehmet Dalkilic, Sun Kim School of Informatics, Indiana University.

Similar presentations


Presentation on theme: "Guiding Motif Discovery by Iterative Pattern Refinement Zhiping Wang, Mehmet Dalkilic, Sun Kim School of Informatics, Indiana University."— Presentation transcript:

1 Guiding Motif Discovery by Iterative Pattern Refinement Zhiping Wang, Mehmet Dalkilic, Sun Kim School of Informatics, Indiana University

2 Outline Introduction and motivation Our framework for motif discovery 1. Initial pattern discovery 2. Build seed motif 3. Extract subsequences 4. Motifs discovery 5. Iterative refinement Experiment and Result Discussion and Future work

3 Introduction – motifs & their applications Protein motifs are short patterns conserved in proteins. They are generally important for the function of a protein or the maintenance of protein structures. 1. Enzyme catalytic sites 2. Regions involved in binding a molecule (ADP/ATP, DNA…) or another protein. 3. A fold important for general 3D structure. Distinguish protein groups based on such patterns. Classify a sequenced protein to a specific family of proteins.

4 Introduction - motif discovery PROSITE: find patterns manually Deterministic algorithm, expectation maximization based: MEME (time consuming) Stochastic algorithm (Gibbs sampling algorithm), random jumps in the search space: Gibbs Sampler; AlignACE The performance varies with the input sequences’ characteristics. ( For example, all known motifs in disease resistance genes in Arabidopsis thaliana were successfully found using MEME after splitting the sequences into two distinct categories of resistance genes, but no motifs were found by inputting all disease resistance genes as a single input file to MEME.)

5 Motivation Motif discover is, in a sense, to compare two models: a model for the pattern (signal model) and a model for negative examples (noise model). Input sequences determine the background noise model. The performance of motif discovery algorithms can be significantly improved by clustering input sequences into smaller groups. Thus Motivation for our research is to use subsequences, instead of using whole sequences, for motif discovery. However, it is quite difficult to select correct subsequence regions without prior knowledge, e.g., genes of the same type. We use an iterative algorithm to solve this problem.

6 Motivation – an example Figure 1, Motif logo for the multiple sequence alignment of a family Figure 2, Motif logo for conserved subsequences of the protein family. PS00343 (L-P-x-T-G-[STGAVDE])

7 Outline Introduction and motivation Our framework for motif discovery 1. Initial pattern discovery 2. Build seed motif 3. Extract subsequences 4. Motifs discovery 5. Iterative refinement Experiment and Result Discussion and Future work

8 Test Data Preparation 1. Download PROSITE pattern and sequence databases. 2. Parse all true positive sequences for each PROSITE ID and store them as a PROSITE family. 3. All sequences of one family contain the same PROSITE pattern. 4. We used PROSITE families to discovery motifs and test the performance of our framework.

9 Framework Overview 1 STEP1. Extract a set S of subsequences around a set of motifs M. STEP2. Input S to a motif discovery algorithm, producing a new set of motifs M’. STEP3. Search entire sequences for more occurrences of M’, producing M’. Set M’ to M and go to step 1.

10 Framework Overview 2

11 Outline Introduction and motivation Our framework for motif discovery Initial pattern discovery 2. Build seed motif 3. Extract subsequences 4. Motifs discovery 5. Iterative refinement Experiment and Result Discussion and Future work

12 Initial Pattern Discovery - thresholds Three thresholds for pattern discovery: 1. length of patterns (L=3, exact patterns longer than 3 do not occur frequently even in the conserved motif regions). 2. log-odd value of 1 st Markov model to random model (statistically significant patterns occur more frequently than random patterns ). 3. support value (patterns should be present in a certain number of sequences ).

13 Initial Pattern Discovery - algorithm 1. Use thresholds to scan the sequences in one set of sequences, find out qualified patterns in each sequence. 2. Rank the sequences according to how many qualified patterns each sequence has. 3. Save the qualified patterns in the top half sequences and eliminate these sequences. 4. Repeat this algorithm on the rest half set of sequences (go to step 1) until no more patterns can be found. The saved patterns will be used later.

14 Initial Pattern Discovery - example Qualified Patterns (p1, p2, p3)

15 Outline Introduction and motivation Our framework for motif discovery 1. Initial pattern discovery Build seed motif 3. Extract subsequences 4. Motifs discovery 5. Iterative refinement Experiment and Result Discussion and Future work

16 Build Seed Motif 1. Start from the pattern with maximal support, use it as the seed motif. 2. Calculate the scores of the candidate patterns (in sequences not covered by the seed motif) to the seed motif. S i = ΣS i-j Wj (j = 1… n) Si: score of candidate pattern i to seed motif Si-j: score of candidate pattern to j th pattern in the seed motif Wj: the weight (support ratio) of j th pattern in the seed motif 3. Add the pattern which has the highest score (also larger than a score threshold) to the seed motif. 4. Go to step 2, until no more patterns can be added to the seed motif.

17 Build Seed Motif - example Calculate pattern scores (threshold = 5) PatternSequenceSupport (suppose no shared sequences) WeightScore to motif P1CLG4W 1 = 1 P2CLN213 P3ALG210 P4ALN24 S 2-1 = 9+4+0 = 13; S 2 = S 2-1 W 1 = 13 P1 C L G 9 4 0 P2 C L N

18 Build Seed Motif - example Calculate pattern scores (threshold = 5) PatternSequenceSupport (suppose no shared sequences) WeightScore to motif P1CLG4W 1 = 4 / (4+2) P2CLN2W 2 = 2 / (4+2) P3ALG28 P4ALN26 S 3-1 = 10, S 3-2 = 4 S 3 = S 3-1 W 1 + S 3-2 W 2 = 8 > 5 S 4-1 = 4, S 4-2 = 10 S 4 = S 4-1 W 1 + S 4-2 W 2 = 6 > 5

19 Build Seed Motif - example Calculate pattern scores (threshold = 5) PatternSequenceSupport (suppose no shared sequences) WeightScore to motif P1CLG4W 1 = 4 / 8 P2CLN2W 2 = 2 / 8 P3ALG2W 3 = 2 / 8 P4ALN29 S 4-1 = 4, S 4-2 = 10, S 4-3 = 8 S 4 = S 4-1 W 1 + S 4-2 W 2 + S 4-3 W 3 = 9 > 5

20 Build Seed Motif

21 Outline Introduction and motivation Our framework for motif discovery 1. Initial pattern discovery 2. Build seed motif Extract subsequences Motifs discovery Iterative refinement Experiment and Result Discussion and Future work

22 Extract Subsequences

23 Motifs Discovery MEME

24 Iterative refinement sub-sequences MAST motif sub-sequences MEME entire protein family Stable? motif discovery no yes

25 Iterative refinement

26 Outline Introduction and motivation Our framework for motif discovery 1. Initial pattern discovery 2. Build seed motif 3. Extract subsequences 4. Motifs discovery 5. Iterative refinement Experiment and Result Discussion and Future work

27 Experiment 1. We used 108 PROSITE families as test data. 2. Ran MEME directly on these families and got the best motif for each of them. 3. Ran our framework and got the best motif (Because of time constraints, our motif framework performed only single iteration. ) 4. Compared the results.

28 Performance The result of the comparison. http://biokdd.informatics.indiana.edu/zhipwang/paper/appendix http://biokdd.informatics.indiana.edu/zhipwang/paper/appendix 63 patterns ( PS00010 PS00011…) 23 patterns (PS00014 PS00033…) 15 patterns (PS00019 PS00035…) 7 patterns (PS01345 PS01286…) Framework ×× MEME ××

29 Performance The result of the comparison.

30 Discussion To make our experiment more rigorous, we choose only the top motif reported by both MEME and our framework. Among the 22 failed cases, our framework did discover 21 of them, though their rank was not top. One flaw: Local optima This framework is general enough to include any motif discovery and search algorithms that report multiple motifs with a statistical score.

31 Future Work On the theoretical side, we are interested in formalizing and understanding the role of noise. How likely subsequences induced by our initial pattern discovery algorithm can include true motifs? Is convergence to true motif regions guaranteed once the initial set of subsequences contain true motifs? For empirical study, we plan to perform multiple iterations using the whole PROSITE pattern set; embed different motif discovery and search programs into our framework.


Download ppt "Guiding Motif Discovery by Iterative Pattern Refinement Zhiping Wang, Mehmet Dalkilic, Sun Kim School of Informatics, Indiana University."

Similar presentations


Ads by Google