Machine Learning Framework for DNA Computing

1. Machine Learning Framework for DNA Computing 3rd MEC Workshop 2001.11.30 신수용

2. DNAC vs. Lego • Consider each DNA base as “LEGO block” • A, C, G, T DNA Lego (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

3. DNAC vs. Lego • Solution • Sequence of blocks • AACTG  A + A + C + T + G • Fixed length • Computing • Decompose the solution to each block • Rebuild the solution from those blocks (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

4. Lego operation • Hybridization • Combining the lego blocks each other. • Ligation • Given a lego blocks of length n, append a block to it and make a lego blocks of length n+1. • Electrophoresis • Check a lego length. • PCR • Amply (put into) lego blocks. • Selection • Extract a lego which we want to do exactly (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

5. DNAC vs. Lego Computing Process (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

6. Solution Initial Search Space • Most machine learning algorithms can be considered as the process of seeking for the optimal solution in the huge search space. (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

7. Search Space in DNAC • Do parallel search Solution Initial (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

8. Conventional ML algorithmsvs. DNAC algorithms • For n nodes Graph problems • Conventional ML algorithms • DNAC algorithms (ideally) • We can provide blocks unlimitedly Search for solution in the ONLYn! space. Search for solution in the space. (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

9. Conventional ML algorithmsvs. DNAC algorithms n! space space (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/ Exhaustive Search

10. DNA Computing • It is important to limit search space. • Do not exhaustive search. • We need more intelligent search process. • Evolutionary process! (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

11. Molecular Evolutionary Computing • Combining DNA computing with evolutionary computation. • Use huge parallelism of DNA computing and smart techniques of evolutionary computation. • We can think that molecular evolutionary computing as evolutionary computation with unlimited population size (ideally). (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

12. MEC • Search  Selection  Search  Selection … Select proper search process (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

13. MEC • Abstract flow Search Selection Solution (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

14. Theoretical problem in DNAC • Hybridization is not wholly controlled, can we make a solution exactly? • Case by case approach (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

15. Theoretical problem in DNAC • Case (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

16. Source (Limited or Unlimited) Case (Tree view) Reaction stage A Stage 1 Step 4 Stage 2 AA AT AG AC Step 1, 2, 3 Stage 3 AAT AAC Stage 4 . . . . . . Hybridization probability P (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

17. Case1 • Constant hybridization success ratio, unlimited blocks, experiments and step by step search • Select best string at each step (each depth in tree) • hybridize only selected string for next step • Then the number of experiment X follows the negative binomial distribution • Thus, in this case, DNAC guarantees the success of experiments (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

18. Case 2 CASE 2, 3, 4 CASE 1 Add existing pool select expand select (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/

19. Case 2, 3, 4 • Case2 • Uniform (?) • Formulation • Case3, 4 • Verify solutions • Probability • Verify assumptions in real life experiments (C) 2001, SNU Biointelligence Lab, http://bi.snu.ac.kr/