1. Accurate Method for Fast Design of Diagnostic Oligonucleotide Probe Sets for DNA Microarrays Nazif Cihan Tas ctas@cs CMSC 838 Presentation

2. CMSC 838T – Presentation Motivation u DNA microarrays techniques are used intensely for identification of biological agents  Gene Expression Studies  Diagnostic Purposes l Identification of Microorganisms in samples  Item Extraction u Complex Problem  Find the necessary probes and the temperature  Probe sets should be reliably detect and differentiate target sequences  Large Databases  NEW!! Homologous Genes (how to find specific probes)

3. CMSC 838T – Presentation Talk Overview u Overview of talk  Motivation  Problem Statement  Algorithm  Mathematical Aspects  Experimentation  Discussion

4. CMSC 838T – Presentation Problem Statement u Specificity vs. Sensitivity  Specificity: # of non-target match is minimized  Sensitivity: # of selected target sequences is maximized. u Original Problem:

5. CMSC 838T – Presentation Problem Statement u Positive Probes  Database set S 0  Target S 1  For each sequence in S 1, find at least one probe  For S 0 - S 1 try to avoid it (but do not care if happens)  High Specificity: # of non-target matches are minimized  High Sensitivity: # of covered target seq. is maximized S0S0 S1S1

6. CMSC 838T – Presentation Problem Statement u Negative Probes  Determine as few as possible probes which together hybridizes with all sequences in S 0 - S 1 but with NONE in S 1.  High Specificity: No seq. in S 1 may hybridize  High Sensitivity: Max # of seq. in S 0 - S 1 be covered S1S1 S0S0

7. CMSC 838T – Presentation Problem cont. u Extend Problem u Specificity vs. Sensitivity  Specificity: No seq. in S 1 may cross-hybridize with any negative probe  Sensitivity: # of seq. covered in B must be maximized.

8. CMSC 838T – Presentation Probe Design Constraints u Sequence Related  Length of probes  Deviation of melting temperature of probe-target hybrids must be low (for physical reasons)  No self complementary regions longer than four nucleotides (not descriptive enough)  Melting temperatures of target and non-target seq. must be larger than a predefined (too close, too hard to identify) l Ensuring a minimum number of mismatches is enough (homologous sequences) u System Related  Execution Time  Usability

9. CMSC 838T – Presentation Algorithm u Overview  Probe Generation  Hybridization Prediction  Probe Selection

10. CMSC 838T – Presentation Algorithm Probe Generation u Subproblem:  Generate probe candidates for the sequences  Keep the set as small as possible without losing any optimal candidate (exclude infeasible ones) u Suffix Tree  Why? l Allows fast recognition of repetitive subsequences l Identifies non-unique probes (i.e. with more than one target) l Efficient for memory and for T computation (reduce time)  How? l Tree is constructed from the sequences l Traversed (Watson-Crick complement)

11. CMSC 838T – Presentation Suffix Tree u Input: TACTACA  TACTACA  ACTACA  CTACA  TACA  ACA  CA  A u $ denotes end of string u Constructed in linear time

12. CMSC 838T – Presentation Probe Generation u Further Improvements  Filters applied for cut off l Probe length (predefined) l G-C content (for temperature) l Self-complementarity u Probes should not contain complements as subsequences  Finally, remove highly conserved (non-specific) regions  Insert into hashtables according to their lengths

13. CMSC 838T – Presentation Algorithm

14. CMSC 838T – Presentation Algorithm Hybridization Prediction u Subproblem:  Search for the right probe  Search is expensive, Intelligent Hashing used u Design  A frame is moved over target and nontarget seqs. with several lengths l Previous algorithm (Kaderali 2002): Use the suffix tree  At each step, hash values are calculated. If hit, predict melting temperature, store in hybridization matrix.  If there are too many hits for a probe, then it is not unique, remove it  Why intelligent? l Hash time is linear l Allows inexact matching because of hashing (No analysis) u Parallelization  Several threads are searching for probe targets. l Tree and hashtables are fixed.  One thread writes to the final matrix

15. CMSC 838T – Presentation Hybridization Prediction u Empirical Simulation:  One million random probe-target pairings generated  Four mismatches or one insertion or deletion plus one strong central mismatch chosen  T<20 C for 93%  Complexity is O ( |S 0 | |S 1 | ) l Possible probe candidates is |S 1 | (linear) l Each position of database S 0 must be checked

16. CMSC 838T – Presentation Algorithm

17. CMSC 838T – Presentation Algorithm Hybridization Prediction u Complexity u In-exact equality  Only the inner three bands of DP matrix are computed  O(l) where l is length

18. CMSC 838T – Presentation Algorithm Probe Selection u Subproblem:  Use the hybridization matrix to finalize the probe selection l We have positive probes and negative probes to proceed u Algorithm Analysis:  For each probe candidate l g: #of matches in S 1 l b: #of matches in S 0 - S 1 l t: highest melting point in S 1  Probes for which g or b values is too large, are removed  Sort according to g,b and t.  Apply Depth First Search u Advantages  Performs well (No comparison though)  Guarantees to choose all specific probes if any were found. u Disadvantages  can NOT guarantee optimal selection in terms of coverage

19. CMSC 838T – Presentation Negative Probe Selection u Let S 2 =S 0 - S 1 and B subset of S 2. The probes that detect S 1 also detects some of B elements. u Algorithm for Negative Probes  Apply probe generating and preselection for B.  Conduct hybridization on B U S 1.  Remove the probes which hybridizes with S 1.  Sort the remaining probes according to their hit number.  Successively select the probes which covers most target seq. u Guarantees optimal solution for coverage and probe number usage

20. CMSC 838T – Presentation Algorithm Probe Selection

21. CMSC 838T – Presentation Mathematical Aspects

22. CMSC 838T – Presentation Mathematical Aspects

23. CMSC 838T – Presentation Mathematical Aspects

24. CMSC 838T – Presentation Experimentation u Parallelized on SMP platform  Classic worker-producer  Intel Dual Pentium III 933 MHz, 1 GB memory u Test data  ssu rRNA of ARB project  20.282 ssu rRNA sequences  1.401 < lengths < 4.179  %97 of them are shorter than 2.000 bases

25. CMSC 838T – Presentation Discussion u High Performance  Execution is linear with size of database, decreases if longer probes are used u Low Memory Consumption  Depends on the size of the sequence selection, NOT database size u Automatic Design of Group Probes and negative probes u High Quality Probe Design

26. CMSC 838T – Presentation Discussion u Comparison with previous work  vs. ARB l Not suited for large scale probe design  vs. LCF l Does not consider highly conserved data l Memory consumption is high l Works well with short probes only  vs. others l Mostly can not deal with insertion and deletions l Execution is slow