Compression and Analysis of Biological Sequences. Why and How.

On work with Julie Bernal, Minh Duc Cao, Samira Jaeger, Chris Mears, David Powell, Tim Edgoose, Linda Stern*, Trevor Dix, and others at Monash U., .au., and *Melbourne U..

Will discuss these related problems:
Compression of DNA.

Compression of protein.

Uses of information content in sequence analysis.

Relating 2 or more (non-random) sequences.

-- Lloyd Allison
users.monash.edu/~lloyd/Seminars/2007-BioCompression/index.shtml

Abstract

Compression can be rather addictive. It has an obvious ``figure of merit'' -- the compressed size of some standard data-set, and it is natural to strive to beat your competitors on this measure. Biological sequences are hard to compress; the more compression the better but there is usually limited value in fighting over the third significant digit. Other properties of a compression method can be just as important or more important. This talk gives some reasons why it is challenging, interesting and useful to compress biological sequences. It also presents two simple models for compressing biological sequences (a possible sub-addiction in compression is to complicated models, but simple is often good); we get good results for DNA and protein.

Why Compress Biological Sequences?

Not to save time or space! (Not yet.)

For analysis.

Good model <=> good compression
  ( compression = modelling + coding )

Note, "masking-out" implies that low-information content = zero information. Not true!

Desirable features of a model:

(Good compression.)

Based on biological knowledge.

Simple, few parameters.

Can give per-symbol information content.

Efficient algorithm.

DNA

Alphabet = {A, C, G, T}.

Various kinds of repeat.

Usually approximate.

Reverse complementary repeats.

Protein

Amino acids.

|Alphabet| = 20.   (log220 = 4.3)

Very hard to compress.

1. Approximate repeats model (ARM) for DNA:

<--end copy--<  
base
part
   copy
editing
part
>--start copy-->  
(4+, 1-)
Note can replace base and editing sub-models.

Repeat Graph - all explanations under ARM.

2-D Probability Plot


Plasmodium falciparum, chromosome 2, subtelomeric regions.
Note,  P.f. is 80% AT.

1-D Information Content Sequence


P. falciparum, all of chromosome 2, (947,102 bp).
Note,  1-D => linear space.

2. A Browser


Cyanidioschyzon merolae, chromosome 1 (422,393 bp), smoothing window 1000,
[small] [med] [big]

Differencing


C. merolae chromosome 4 (513,790 bp), [med] [big], and . . .

. . . its difference from chromosome 4 given 18, smoothing window 1000,  I(c4) - I(c4 | c18),  [med] [big]

Two or more sequences


Plot for C. merolae 10,000 bp for chromosomes 1*, 4*, 5*, 18*, 6+ & 11+, smoothing window 100 (*elt P, +elt PH), [med] [big]

Using reversal


P. falciparumI(chr2) - rev( I( rev(complement(chr2))))

3. Experts model (XM)

A blend (weighted average) of:
 
  A global background expert, e.g., a low-order MM.
 
  A local context expert, e.g., a "windowed" MM.
 
  Forward copy experts (offset, prob. copy v. change).
 
  Reverse complement copy experts (ditto).
 
Good compression, simple, information per-symbol, fast.

XM on DNA, e.g.,

  BioC GenC DNAC DNAP CDNA ARM XM
HUMHBB 1.88 1.82 1.79 1.78 1.77 1.73 1.75
bits per base

XM on protein, e.g.,

  CP ProtComp LZ-CTW XM
SC,
Sacharomyces
Cerevisiae
4.15 3.94 3.95 3.89
bits per amino acid


(more results in DCC paper)

4. Low-information substrings and alignment

Low-information ~ repetitive ~ compressible.
 
Cause false +ve matches for standard software, so . . .
 
  (i) sometimes "masked-out" (discarded), and/or
 
  (ii) significance test: shuffle & re-align (PRSS program).

The M-alignment solution:

Build a sequence-model into the alignment algorithm.
 
  Fewer false +ves, and false -ves, than PRSS and Blast on
  real and artificial test data.
 
  Time-complexity unchanged (but worse constant).

. . .

m_align (
  seq_1, model_1,
  seq_2, model_2 )
  { ... };
 
Gives a natural significance test for given propulation model(s).
 
Changes rank-order of alignments cf. "uniform" model, as it should
 
(rather than est. significance of a bad alignment!).
 
model_i can be almost any model, e.g., ARM, or XM.

. . . ROC curves


blue: S-W raw score, green: PRSS p-value, red: M-align (MM 1), purple: M-align (blend mdl), NB. bottom right is good

Conclusions

Calculating information carefully, using it, not discarding it, gives accurate results, few false +ves or -ves.

Information content sequences are 1-D, linear-space. Many operations on them take linear-time. Feasible for millions of bases.

Also see

L. Allison, T. Edgoose, T. I. Dix. Compression of Strings with Approximate Repeats, Intell. Sys. in Molec. Biol., pp.8-16, 1998.
L. Stern, L. Allison, R. L. Coppel, T. I. Dix. Discovering patterns in Plasmodium falciparum genomic DNA, Molecular and Biochemical Parasitology, 118(2), pp.175-186, 2001.
D. R. Powell, L. Allison, T. I. Dix. Modelling alignment for non-random sequences, Springer Verlag, LNCS Vol.3339, pp.203-214, 2004.
T. I. Dix, D. R. Powell, L. Allison, J. Bernal, S. Jaeger, L. Stern. Comparative Analysis of Long DNA Sequences by Per Element Information Content Using Different Contexts, BMC Bioinformatics, 8(Suppl 2):S10, May 2007.
Minh Duc Cao, T. I. Dix, L. Allison, C. Mears. A Simple Statistical Algorithm for Biological Sequence Compression, DCC 2007, pp.43-52, March 2007.

And [Bioinformatics] [MML] [interests].


© Lloyd Allison Computer Science and Software Engineering, Monash University, Australia 3800.
Created with "vi (Linux & Solaris)",   charset=iso-8859-1