bio/util/contingency_table.rb - Statistical contingency table analysis for aligned sequences

Author

Trevor Wennblom <trevor@corevx.com>

Copyright

Copyright © 2005-2007 Midwinter Laboratories, LLC (midwinterlabs.com)

License

The Ruby License

Description¶ ↑

The Bio::ContingencyTable class provides basic statistical contingency table analysis for two positions within aligned sequences.

When ContingencyTable is instantiated the set of characters in the aligned sequences may be passed to it as an array. This is important since it uses these characters to create the table's rows and columns. If this array is not passed it will use it's default of an amino acid and nucleotide alphabet in lowercase along with the clustal spacer '-'.

To get data from the table the most used functions will be chi_square and contingency_coefficient:

ctable = Bio::ContingencyTable.new()
ctable['a']['t'] += 1
# .. put more values into the table
puts ctable.chi_square
puts ctable.contingency_coefficient  # between 0.0 and 1.0

The contingency_coefficient represents the degree of correlation of change between two sequence positions in a multiple-sequence alignment. 0.0 indicates no correlation, 1.0 is the maximum correlation.

Further Reading¶ ↑

• en.wikipedia.org/wiki/Contingency_table

• www.physics.csbsju.edu/stats/exact.details.html

• Numerical Recipes in C by Press, Flannery, Teukolsky, and Vetterling

Usage¶ ↑

What follows is an example of ContingencyTable in typical usage analyzing results from a clustal alignment.

require 'bio'

seqs = {}
max_length = 0
Bio::ClustalW::Report.new( IO.read('sample.aln') ).to_a.each do |entry|
  data = entry.data.strip
  seqs[entry.definition] = data.downcase
  max_length = data.size if max_length == 0
  raise "Aligned sequences must be the same length!" unless data.size == max_length
end

VERBOSE = true
puts "i\tj\tchi_square\tcontingency_coefficient" if VERBOSE
correlations = {}

0.upto(max_length - 1) do |i|
  (i+1).upto(max_length - 1) do |j|
    ctable = Bio::ContingencyTable.new()
    seqs.each_value { |seq| ctable.table[ seq[i].chr ][ seq[j].chr ] += 1 }

    chi_square = ctable.chi_square
    contingency_coefficient = ctable.contingency_coefficient
    puts [(i+1), (j+1), chi_square, contingency_coefficient].join("\t") if VERBOSE

    correlations["#{i+1},#{j+1}"] = contingency_coefficient
    correlations["#{j+1},#{i+1}"] = contingency_coefficient  # Both ways are accurate
  end
end

require 'yaml'
File.new('results.yml', 'a+') { |f| f.puts correlations.to_yaml }

Tutorial¶ ↑

ContingencyTable returns the statistical significance of change between two positions in an alignment. If you would like to see how every possible combination of positions in your alignment compares to one another you must set this up yourself. Hopefully the provided examples will help you get started without too much trouble.

def lite_example(sequences, max_length, characters)

  %w{i j chi_square contingency_coefficient}.each { |x| print x.ljust(12) }
  puts

  0.upto(max_length - 1) do |i|
    (i+1).upto(max_length - 1) do |j|
      ctable = Bio::ContingencyTable.new( characters )
      sequences.each do |seq|
        i_char = seq[i].chr
        j_char = seq[j].chr
        ctable.table[i_char][j_char] += 1
      end
      chi_square = ctable.chi_square
      contingency_coefficient = ctable.contingency_coefficient
      [(i+1), (j+1), chi_square, contingency_coefficient].each { |x| print x.to_s.ljust(12) }
      puts
    end
  end

end

allowed_letters = Array.new
allowed_letters = 'abcdefghijk'.split('')

seqs = Array.new
seqs << 'abcde'
seqs << 'abcde'
seqs << 'aacje'
seqs << 'aacae'

length_of_every_sequence = seqs[0].size  # 5 letters long

lite_example(seqs, length_of_every_sequence, allowed_letters)

Producing the following results:

i           j           chi_square  contingency_coefficient
1           2           0.0         0.0
1           3           0.0         0.0
1           4           0.0         0.0
1           5           0.0         0.0
2           3           0.0         0.0
2           4           4.0         0.707106781186548
2           5           0.0         0.0
3           4           0.0         0.0
3           5           0.0         0.0
4           5           0.0         0.0

The position i=2 and j=4 has a high contingency coefficient indicating that the changes at these positions are related. Note that i and j are arbitrary, this could be represented as i=4 and j=2 since they both refer to position two and position four in the alignment. Here are some more examples:

seqs = Array.new
seqs << 'abcde'
seqs << 'abcde'
seqs << 'aacje'
seqs << 'aacae'
seqs << 'akcfe'
seqs << 'akcfe'

length_of_every_sequence = seqs[0].size  # 5 letters long

lite_example(seqs, length_of_every_sequence, allowed_letters)

Results:

i           j           chi_square  contingency_coefficient
1           2           0.0         0.0
1           3           0.0         0.0
1           4           0.0         0.0
1           5           0.0         0.0
2           3           0.0         0.0
2           4           12.0        0.816496580927726
2           5           0.0         0.0
3           4           0.0         0.0
3           5           0.0         0.0
4           5           0.0         0.0

Here we can see that the strength of the correlation of change has increased when more data is added with correlated changes at the same positions.

seqs = Array.new
seqs << 'abcde'
seqs << 'abcde'
seqs << 'kacje'  # changed first letter
seqs << 'aacae'
seqs << 'akcfa'  # changed last letter
seqs << 'akcfe'

length_of_every_sequence = seqs[0].size  # 5 letters long

lite_example(seqs, length_of_every_sequence, allowed_letters)

Results:

i           j           chi_square  contingency_coefficient
1           2           2.4         0.534522483824849
1           3           0.0         0.0
1           4           6.0         0.707106781186548
1           5           0.24        0.196116135138184
2           3           0.0         0.0
2           4           12.0        0.816496580927726
2           5           2.4         0.534522483824849
3           4           0.0         0.0
3           5           0.0         0.0
4           5           2.4         0.534522483824849

With random changes it becomes more difficult to identify correlated changes, yet positions two and four still have the highest correlation as indicated by the contingency coefficient. The best way to improve the accuracy of your results, as is often the case with statistics, is to increase the sample size.

A Note on Efficiency¶ ↑

ContingencyTable is slow. It involves many calculations for even a seemingly small five-string data set. Even worse, it's very dependent on matrix traversal, and this is done with two dimensional hashes which dashes any hope of decent speed.

Finally, half of the matrix is redundant and positions could be summed with their companion position to reduce calculations. For example the positions (5,2) and (2,5) could both have their values added together and just stored in (2,5) while (5,2) could be an illegal position. Also, positions (1,1), (2,2), (3,3), etc. will never be used.

The purpose of this package is flexibility and education. The code is short and to the point in aims of achieving that purpose. If the BioRuby project moves towards C extensions in the future a professional caliber version will likely be created.