Towards Declarative and Efficient Querying on Protein Structures

Transcription

1 Towards Declarative and Efficient Querying on Protein Structures Jignesh M. Patel University of Michigan Biology Data Types Sequences: AGCGGTA. Structure: Interaction Maps: Micro-arrays: Gene A Gene B Expr Expr When there is something to do in the cell, it is a protein that does it.

2 Role of DBMS in Bioinformatics Large Data Sets Growing exponentially! Data Types Sequences/arrays/-D/text Complex Queries Ad-hoc tools today Integrated querying done using procedural methods Scalability/Parallelism Home-grown techniques often used today Need: Declarative and efficient querying # DNA Base Pairs (Billions) Base Pairs Sequences Source: GenBank # Sequences (Millions) /7/4 Jignesh M. Patel, 4 Protein Querying Protein functionality determined by Sequence composition Geometric structure Need to query across all structures Current methods Non-declarative tools, often not very efficient on large data sets Support querying on only one structure /7/4 Jignesh M. Patel, 4 4

3 Periscope Goal: Design, implement, and evaluate a database management system for declarative and efficient querying on all protein structures Periscope IAKDPQ Declarative and Integrated Query MRIVL Data Parse Optimize Evaluate Answer (quickly) /7/4 Jignesh M. Patel, 4 5 Roadmap Background and Introduction Primary Structure Sequence Matching Querying on Secondary Structure PiQA: : Integrated Query Algebra Summary /7/4 Jignesh M. Patel, 4 6

4 Sequence Matching Find similar sequences Given a protein sequence, find homologous matches in the database Similarity based on local similarity A local-alignment alignment algorithm Operations: Replace, Delete, Insert Score using a substitution matrix Database: THE TRAIN DRIVER S CABIN Query: DRAIN /7/4 Jignesh M. Patel, 4 7 Sequence Matching Find similar sequences Given a protein sequence, find homologous matches in the database Similarity based on local similarity A local-alignment alignment algorithm Operations: Replace, Delete, Insert Score using a substitution matrix Query Target - A B A - B - C - C - D - D - Target: C A B I N R I R D R R Query: D R A I N Score: /7/4 Jignesh M. Patel, 4 8

5 D R A I N C A 4 Smith-Waterman B 4 I 6 5 S-W Matrix: G N 5 9 Query - A B C D Target A - B - C - Substitution Matrix D - G i,j = max G i-, j- + S(q i t j ) G i-, j + S(q i ) G i, j- + S( t j ) C A B I N R I R D R R D R A I N /7/4 Jignesh M. Patel, 4 9 Suffix Trees Compact Patricia trie Every suffix has a path (to leaf) Every subsequence is a prefix of a path Can find subsequences very fast E.g. GTACG Size: ~x Position Database G$ 8L 8L 4N L L L N 5L 5L L L N A TA G C CGCCTAG$ $ CCTAG$ CCTAG$ GCCTAG$ N 4L 4L CTAG$ CTAG$ 6L 6L TAG$ TAG$ A G T A C G C C T A G $ N 7L 7L $ 9L 9L CGCCTAG$ G$ G$ 5N L L L L /7/4 Jignesh M. Patel, 4

6 OASIS Search driven by the suffix tree Fill up the S-W S W columns by traversing down the suffix tree Best-first: expand node in the tree with highest expected score Expected Score = current score + best possible score for unconsumed portion of query Guarantees online behavior! Exploits redundancy in the database /7/4 Jignesh M. Patel, 4 OASIS Query: TACG Unit Edit Distance Matrix: Same Symbol Substitution =, else - Q:.TA... T: TACG Score: ++ = 4 Position Database G$ 8L 8L 4N L L L N 5L 5L L L N A TA G C CGCCTAG$ $ CCTAG$ CCTAG$ GCCTAG$ N 4L 4L CTAG$ CTAG$ 6L 6L TAG$ TAG$ A G T A C G C C T A G $ N 7L 7L Q:.A... T: TACG Score: ++ = $ 9L 9L CGCCTAG$ G$ G$ 5N L L L L

7 Experimental Setup: Comparison Data set: swissprot K proteins, # symbols: 4M, data size: 4MB Index Size: 5MB (~.5 bytes per symbol) Workload: queries from ProClass motif database Short queries lengths 6-56, avg len = 6 PAM scoring matrix BLAST parameters (short query settings, 5-55 residues) E=,, word size = OASIS: minscore = ln( Kmn) ln( E) λ Platform:.7GHz Xeon, Linux, 56MB buffer pool, K page size, clock replacement policy /7/4 Jignesh M. Patel, 4 Mean Time in secs (log scale) Experimental Results Performance Comparison S-W BLAST OASIS. 4 6 Query Length OASIS - orders of magnitude faster than S-W, S and usually also faster than BLAST Sensitivity Comparison OASIS retrieved ~ 6% more matches than BLAST Lots more in our VLDB paper

8 Roadmap Background and Introduction Primary Structure Sequence Matching Querying on Secondary Structure PiQA: : Integrated Query Algebra Summary /7/4 Jignesh M. Patel, 4 5 Querying Protein Secondary Structure Secondary Structure describes protein folds Beta sheets (e); Alpha helices (h); Loops (l) Predicted structure; helps determine protein function Data Model Sequence of Segments h h h h l l e e e e e <h>, < l>, <4 e> Query Language Sequence of regular expression terms Example: Beta sheet of length 4-4 followed at some point by an helix of length -6 Query: <e 4 > <? Inf> > <h 6> /7/4 Jignesh M. Patel, 4 6

9 Schema and Query Evaluation id name A B len 5 6 sec-seq seq lleee hhhee Protein Table 4 e 4 Evaluation Methods CSP: complex scan SSS: scan segment table ISS: use segment index MISS (k): multiple ISS seg-id id type l e h len start-pos (Derived) Segment Table Scan each protein tuple and evaluate query using a finite state automata N-way merge join of k multiple segment index scans + FK join /7/4 Jignesh M. Patel, 4 7 Multiple Index Scan Method: MISS(k) Query: <P >< P > <P n > To satisfy other query predicates CSP Merge based on protein id and ordering constraints Merge (id) INLJ (id) Idx Scan (id) Sort (id, start) Sort (id, start) Sort (id, start) (id, start) Idx Scan (type, len) (id, start) Idx Scan (type, len) (id, start) Idx Scan (type, len) Protein Table /7/4 Segment Table Segment Table Segment Table k k most highly selective query predicates Jignesh M. Patel, 4 8

10 Query Optimizer Chooses best method for given query Optimization: Requires estimates of query predicate selectivities and result cardinality Cost model: CPU and IO Estimation: Two types of histograms Basic: segment predicate selectivity Complex: query selectivity /7/4 Jignesh M. Patel, 4 9 Basic Histogram Estimates selectivity of individual query predicates -D D array Dimensions : Length, Fold type Value: number of each <type,len> > segment Example Pred <e 4 4>, Est: : 5 Pred <e 4>, Est: : Len H E L /7/4 Jignesh M. Patel, 4

11 Complex Histogram Estimates result cardinality Four-dimensional structure Protein id (equi( equi-width buckets) Start position (equi( equi-width buckets) Length ( 5) Same as in the Type ( e,( h, l ) basic histogram Example: Position [x] [y] [z][ ] [ e ][ holds the number of <e< z z> segments whose starting position is in the range of the y bucket and whose protein id lies within the x bucket range /7/4 Jignesh M. Patel, 4 Complex Histogram Query: {<P> <P>} Compute selectivity of query result 4 Dimensions - Protein id -Start pos -Length -Type 5 5 Start Positions Protein Sequence P P Same Bucket Distinct Bucket pos pos+ ( n np ) P start positions laststartbucketnum pos= / IDbucketWidth startbucketwidth sp sp+ ( ) P P start sp= positions n n IDbucketWidth /7/4 Jignesh M. Patel, 4

12 Experimental Evaluation Techniques implemented in Periscope Configuration: Using the SHORE storage manager, 64MB buffer pool size, 6K page size Also implemented in a commercial ORDBMS Used type extensibility to create array-like data types for sequences & user-defined functions for FSM Data Set: PIR 5K protein tuples, 5MB Segment Table: M tuples, 55MB Platform:.7GHz Xeon, Linux, 4GB SCSI disk /7/4 Jignesh M. Patel, 4 Complex Histogram Accuracy Query: {<l 5 5> > <? X> <h 4 4>} Experiment: Vary Gap Predicate Result Cardinality 5,, 5,, 5, Estimate Actual Length of Gap Predicate (X) Histogram Details Data Set: PIR xx5x Size: 5.8MB Build time: secs Estimation time: ms Histogram accurate within 8% of the actual result /7/4 Jignesh M. Patel, 4 4

13 Query Performance Query: 9 Predicates P G P G P G P 4 G 4 P 5 P=predicate, G=Gap pred. Expt: : Vary Selectivity of P sp Alternatives: S (.%), L (7%) Choice of algorithm is critical CSP sensitive to position of L pred. Choice of k in MISS critical Index Probe + reduce # protein tuples fetched - probe cost, sorting and merging Influenced by query and predicate selectivity Time (secs) CSP MISS() MISS() MISS(5) SSSSL SSLSS LSSSS Query More details in our VLDB paper Roadmap Background and Introduction Primary Structure Sequence Matching Querying on Secondary Structure PiQA: : Integrated Query Algebra Summary /7/4 Jignesh M. Patel, 4 6

14 PiQL? Details in our PiQA SSDBM paper An algebra in PNF for queries on primary & secondary structures Examples: Match the primary sequence AAANBPPPPSDF,, but ignore mismatch in the segment NBPPP if it is on a loop (P.p * AAA ) ((P.p * NBPPP )) U (P.s( * <L 5 5>)) (P.p( * PSDF ) BLAST: Match n-grams n + match extension + transitive closure Find all occurrences of TGCTGACTCAGCA within 4 bps upstream of a CA with TATA 5- bps upstream of the CA M * TGCTGACTCAGCA 957 M * TATA 6 M * CA 957 TGC..A 5 6 4,,,, TATA CA /7/4 Jignesh M. Patel, 4 7 /7/4 Jignesh M. Patel, 4 8

15 Roadmap Background and Introduction Primary Structure Sequence Matching Querying on Secondary Structure PiQA: : Integrated Query Algebra Summary /7/4 Jignesh M. Patel, 4 9 Summary Bioinformatics applications (urgently) need declarative and efficient query processing tools Current procedural methods reminiscent of pre-relational relational days Database researchers have a lot to contribute! Periscope is our effort in this direction PiQA: : algebraic framework for querying on primary and secondary structures Primary Structure: OASIS Declarative querying on secondary structure: Periscope/PS Current Status OASIS and Periscope/PS currently (beta-)deployed at UM Under development: PiQA powered PiQL queries on Periscope /7/4 Jignesh M. Patel, 4